Multilanguage chatbot with artificial intelligence for client support

University of Minho School of Engineering João Miguel Santos Sá Multilanguage chatbot with artificial intelligence for client support october 2023 University of Minho School of Engineering João Miguel Santos Sá Multilanguage chatbot with artificial intelligence for client support Master’s in Informatics Engineering Natural Language Processing and Artificial Intelligence Dissertation supervised by Paulo Jorge Freitas Oliveira Novais Dalila Alves Durães october 2023 Copyright and Terms of Use for Third Party Work This dissertation reports on academic work that can be used by third parties as long as the internationally accepted standards and good practices are respected concerning copyright and related rights. This work can thereafter be used under the terms established in the license below. Readers needing authorization conditions not provided for in the indicated licensing should contact the author through the RepositóriUM of the University of Minho. License granted to users of this work: CC BY-NC-SA https://creativecommons.org/licenses/by-nc-sa/4.0/ i Acknowledgements This work has been partially supported by the project “IBPS - Intelligent Batch Processing System” , with the reference POCI-01-0247-FEDER-069998, co-financed by the European Regional Development Fund (ERDF), through the Operational Programme for Competitiveness and Internationalization (COMPETE 2020), under the PORTUGAL 2020 Partnership Agreement. I would like to express my heartfelt gratitude to my supervisor for all of her assistance throughout the dissertation writing process, and for playing a critical role in the success of this study, from research and analysis of the state of the art to methodology selection, implementation, and completion. I want to offer my heartfelt appreciation to Alexandra for her focused instruction and amazing example. ii Statement of Integrity I hereby declare having conducted this academic work with integrity. I confirm that I have not used plagiarism or any form of undue use of information or falsification of results along the process leading to its elaboration. I further declare that I have fully acknowledged the Code of Ethical Conduct of the University of Minho. University of Minho, Braga, october 2023 João Miguel Santos Sá iii Resumo A inteligência artificial surgiu nos anos 50, com a criação de um espaço de estudo com o objetivo principal de desenvolver máquinas inteligentes. A sua evolução foi abrupta e hoje existem mesmo programas que, ao escreverem um conjunto de palavras, conseguem gerar imagens surpreendentes com base no que foi escrito em segundos, algo que um pintor famoso faria, mas que demoraria horas ou mesmo dias a concluir. Em vários setores empresariais, desde a educação aos cuidados de saúde, entre outros, tem havido um aumento notável do interesse e da utilização de chatbots. Esta tendência crescente abriu caminho para a implementação de chatbots em diversas áreas. Especificamente, estes chatbots servem como intermediários informativos em empresas como a Retail Consult , uma empresa de tecnologia especializada em desenvolvimento de software na área do retalho e na utilização de processamento em lote, que envolve o tratamento de grandes volumes de dados. Além disso, o sistema de chatbot foi concebido para reconhecer e compreender uma multiplicidade de línguas, incluindo o inglês, que será a língua principal do utilizador, bem como o português e o espanhol, assegurando uma comunicação na língua preferida do utilizador. Esta capacidade multilingue é particularmente crucial para a Retail Consult , dada a sua presença global com escritórios espalhados por vários países. Sem dúvida, esta versatilidade linguística acrescenta um valor imenso ao produto. Para conseguir este reconhecimento linguístico, foram utilizadas técnicas de inteligência artificial e de processamento da linguagem natural, que foram implementadas através de um quadro bem estruturado concebido para a criação de chatbots e de software similar, incluindo assistentes virtuais. O desenvolvimento de um serviço com estas características está preparado para aumentar significativamente a produtividade interna da organização. Por exemplo, um analista de sistemas pode perguntar diretamente ao chatbot o que pretende, poupando tempo que, de outra forma, teria sido gasto na procura e consulta de informações à base de dados. Quer se trate de determinar o número de trabalhos em execução ou de identificar os que apresentam falhas, este sistema procura simplificar as operações e melhorar a eficiência dentro da empresa. iv Com base no trabalho apresentado na introdução, onde discutimos os papéis cruciais do processamento em lote e da integração do chatbot no ambiente em específico, agora aprofundamos a dinâmica do setor de retalho nesta análise minuciosa. O objetivo está centrado na criação de um chatbot versátil capaz de comunicar com os utilizadores em várias línguas, fornecer respostas precisas e, acima de tudo, auxiliar na tradução de idiomas. Além disso, os resultados confirmam a competência do chatbot na identificação e tradução precisa de idiomas, com base em nossa análise avançada, onde investigamos a importância do processamento em lote e introduzimos a estrutura de IA conversacional RASA. Esses elementos, contribuem para uma experiência mais envolvente e satisfatória para o utilizador. Ou seja, este trabalho permitiu e ampliou o nosso conhecimento sobre a interação entre processamento em lote, tecnologia de chatbot e comunicação multilíngue no contexto do retalho. Palavras-Chave Processamento de linguagem natural (PNL), linguagem natural (LN), aprendizagem automática (AA), aprendizagem profunda (AP), inteligência artificial (IA), processamento em Lote (PL) v Abstract Artificial intelligence emerged in the 1950s with the establishment of a research field aimed at developing intelligent machines as its main objective. Its evolution was abrupt, and today there are even programs that, by writing a set of words, can generate astonishing images based on what has been written in seconds, something that a famous painter would do, but which would take hours or even days to complete. In various business sectors, from education to healthcare, among others, there has been a notable increase in interest in and use of chatbots. This growing trend has paved the way for the implementation of chatbots in various areas. Specifically, these chatbots serve as information intermediaries in companies such as Retail Consult , a technology company specializing in software development in the area of shredding and the use of batch processing, which involves handling large volumes of data. In addition, the chatbot system has been designed to recognize and understand a multitude of languages, including English, which will be the user’s main language, as well as Portuguese and Spanish, ensuring communication in the user’s preferred language. This multilingual capability is particularly crucial for Retail Consult , given its global presence with offices in several countries. Undoubtedly, this linguistic versatility adds immense value to the product. To achieve this linguistic recognition, artificial intelligence, and natural language processing techniques will be used, which will be implemented through a well-structured framework designed for the creation of chatbots and similar software, including virtual assistants. The development of a service with these characteristics is set to significantly increase the organization’s internal productivity. For example, a systems analyst can ask the chatbot directly what they want, saving time that would otherwise have been spent searching for and consulting information in the database. Whether it’s determining the number of jobs in progress or identifying those with faults, this system seeks to simplify operations and improve efficiency within the company. Building on the work presented in the introduction, where we discussed the crucial roles of batch processing and chatbot integration in the specific environment, we now delve into the dynamics of the retail sector in this in-depth analysis. Our goals are centered on creating a versatile chatbot capable of communicating with users in multiple languages, providing accurate responses, and, above all, assisting vi xiii Part I Introductory material 1 Chapter 1 Introduction 1.1 Contextualization and Motivation Retail sale is the sale of products or services to the final consumer, while wholesale is the vending of these to other companies/resellers, who will resell the products to the final consumer. In other words, in retail sales, we can sell several or a few products to several different buyers, while wholesale implies the sale of several products to a single buyer editorial de Conceito.de [2020]. Retail sales can be carried out in physical stores or online, while wholesale sales are usually carried out also by telephone, email, or orders on specialized e-commerce sites. Retail prices are generally higher than original wholesale prices because they include a profit for the reseller editorial de Conceito.de [2020]. The market where the retailer operates and all the business dynamics are important factors that, consequently, lead to several concerns such as the profitability factor, because like any company, retailers want to make profits to keep their business running, that is, they have to control their expenses and maximize their income. Another concern that retailers face is competition and this can come from companies in the same sector or another, as is the case with the online sale of products. The way to mitigate it would be for the retailer to find a way to stand out from the rest, suppliers are also part of this cake, as retailers must have a pleasant, healthy, and lasting relationship with their suppliers, as they depend a lot on these people to obtain their products. Another aspect to note is stock management, which must undoubtedly be managed efficiently to avoid having products that are expiring or not run out of products to sell, finally, and perhaps the most important thing is to have satisfied customers with the shopping experience and the product they purchase, to achieve this, the retailer needs to sell good quality products and offer excellent customer service Alves [2015]. Regarding the business, we imagine this scenario in which we are facing the process of selling a perishable good. Firstly, that has to be grown by a farmer, then the farmer if he wants to, can sell several units of his product, if so, the said goods are transported from the field to a central market and then a 2 certain amount of them is sold to a small local grocery store which, in turn, sells to the final consumer. In short, the grocery store functions similarly to the retailer, and the entire business dynamic is undeniably complex, as the grocery store does not only sell perishable goods and it is impossible to manage it without technological solutions, that is, to support retailers in the various aspects of their own business, it is necessary to resource to several computational applications, and this is exactly where Retail Consult comes in, offering this type of specialized telecommunications. Retail Consult is a multinational IT company, which provides strategic solutions such as implementation, development, training, and service support, to help its clients in their specific retail business with a strong focus on the solutions provided by Oracle Retail and related technology, these aforementioned solutions address a broad range of retail business needs, such as planning, omnichannel, supply chain management, merchandising, insights and science RetailConsult [2023]. All of these solutions serve to successfully transform the business from customers to users of the application suite that Oracle Retail provides. Retail Consult is one of Oracle’s leading software implementation partners, with an unrivaled track record of helping customers. When it comes to planning, with purchases becoming more and more complex, and customer expectations rising, having the right product is no longer enough. Hyper-personalized, omnichannel thinking is also essential to change, and this is the mindset that Oracle Retail delivers RetailConsult [2023]. Oracle’s omnichannel suite enables retailers to provide personalized cross-channel experiences for their customers by seamlessly integrating e-commerce, stores, customer relations, order management, and loss-prevention systems RetailConsult [2023]. The Oracle Retail Insights and Science suite combines artificial intelligence, machine learning, and decision science with data captured from Oracle Retail applications and third-party data. The unique property of these self-learning applications is that they detect trends, learn from results, and increase their accuracy the more they are used RetailConsult [2023]. Concerning merchandising, one of the main challenges in retail is efficiently and effectively providing the right amount of information available in a usable and consumable format. With the explosion of social media, omnichannel shopping, and the pressure to expand into new formats and countries, this has never been more important RetailConsult [2023]. Finally, with an effectively managed supply chain, retailers can reduce inventory and cut operating costs, while increasing sales. Oracle Retail supply chain management solutions empower retailers to plan and execute management strategies RetailConsult [2023]. Retailer businesses are done through the implementation of various applications strategically defined for a given purpose and the insertion of these numerous business areas. The integration of this application is only achieved through batch processing. 3 Batch processing is a crucial technique employed in handling extensive datasets, necessitating interaction with external systems and internal maintenance Oracle [2017]. It operates autonomously, sparing users from constant monitoring unless an error arises. In the event of an error within a specific job, an administrator is required to manually rectify and rerun the failed program. This entire process is orchestrated by scheduling software, making it fully automated. These automated tasks are commonly referred to as ”jobs.” Moreover, during batch processing, applications become temporarily inaccessible to users as they are locked out. However, this temporary inconvenience is balanced by the flexibility of setting batch processing preferences according to specific requirements. This means that administrators can easily identify each job’s unique identity, its start and end times, and gain insights into the duration of each task, among other details Oracle [2017]. This approach offers both advantages and disadvantages, which are worth exploring further. The batch execution circumstances encompass a diverse array of elements within the operational environment of the chatbot. This environment is characterized by a multifaceted composition that incorporates distributed systems, modular components, virtual machines, autonomous agents, central control nodes (referred to as masters), and a multitude of specific tasks, among other components. These aspects play a pivotal role in shaping the operational landscape of the chatbot, and they present numerous benefits and challenges when it comes to its integration and performance. It is also crucial to have significantly good batch operating and processing, as it is critical in the day-today dynamics of the retailer. However, the lot must be managed and monitored with great care and rigor, both on the side of those producing and providing the monitoring process service and the consumer. As a result, it is critical to build a solution, such as a chatbot, that mediators between the human component and batch processing, as well as the IT operations that are being run. In the future, examples of questions the chatbot could answer would be: ”Did that particular job have an error?” or ”What is the minimal execution time of the critical path for a specific day?” or ”When will the batch end?”. 4 1.2 Objectives In this section, we articulate the specific objectives of this dissertation. These objectives serve as the foundational framework guiding the research endeavor. They play a crucial role in delineating the direction and purpose of our study. By defining these objectives with precision, we establish a clear path for our research, enabling a focused exploration of chatbots and their applications. Using artificial intelligence technology and natural language processing, the purpose of this project is to develop a chatbot that can detect and recognize multiple languages and respond appropriately to questions in the same language. Customer support will improve and reach new heights with the development of a service that can assist with basic chores. This chatbot is software that can interact with customers via text messages, primarily in English. Furthermore, we have successfully extended its functionality to support multiple languages, with a particular emphasis on Portuguese and Spanish. This is significant for Retail Consult because it is a multinational company, and the incorporation of these languages into the chatbot has greatly benefited our business. Furthermore, we hope to emulate natural human dialogue to better align with everyday communication, bridging the gap between human and machine interaction. Besides, the product is to help the customer in the communication part, in case he has any doubts, the chatbot must be able to respond within a particular knowledge domain. Summarizing, we can specify what are the main objectives of the chatbot to organize our thought when we pass to the implementation phase. So the chatbot will be able to de Trabalho [2022]: • Recognize and answer in the language that is being used, such as Portuguese, Spanish, or English; • Switch language during conversation; • Engage users in the conversation using natural language processing, close to real assistants; • Answer users questions and offer information related to the batch processing, providing automatic answers; • Capable of making corrections, such as terminating a job as desired by the user. 5 1.3 Research Methodology The research methodology that we are using is the agile Scrum technique Sachdeva [2016] to create a multifunctional chatbot capable of answering queries and correcting commands in multiple languages. There were numerous steps to the development process Sachdeva [2016]. We began the project with a planning phase in which we defined the chatbot’s objectives, primary functionalities, and languages it should support. This stage was critical for creating a firm basis for future growth. During the chatbot’s development, the team used an iterative technique. Our primary focus was on developing capabilities that would allow the chatbot to answer assertively and clearly to users regardless of the language in which they interacted. This included enhancing the chatbot’s capacity to interpret user questions and offer correct and helpful responses. The chatbot’s quality maintenance did not involve rigorous testing processes. Instead, it relied solely on internal testing efforts, without incorporating usability studies to assess interface intuitiveness or command correction testing to improve the chatbot’s ability to handle complex requests, such as corrections to poorly written instructions. The progressive inclusion of support for many languages was a key component of the development. This enabled the chatbot to serve a diverse range of consumers in many languages, ensuring an effective multilingual experience. The Scrum approach was crucial in developing a flexible and responsive chatbot. This was critical to address the needs of consumers in a continually changing environment. Surprisingly, the lack of direct feedback from users throughout development had no detrimental impact on the Scrum methodology’s success in providing the anticipated multifunctional chatbot. 1.4 Dissertation Structure This dissertation digs into a investigation and the creation of a multilingual chatbot with multiple language support to aid in the demonstration of data generated by batch processing. The dissertation is broken into two main chapters: introduction and dissertation core. In the first main chapter, we provide context and motivation for this program in the Introduction section, where we discuss the objectives, the technique used, and the research methodology. Then in section Background and State-of-the-art, we provide background information and an assessment of the state of the art. We 6 also describe batch processing in the context of the existing business, examine technology and concept research, conduct a literature assessment, and identify the challenges and problems to be addressed. Finally, we presented some problems and challenge in the section Problem and Challenges. In the second main chapter it’s divide in four section: Contribution, Applications, Evaluations and Results, and Conclusions and Future Work, which highlights the contributions of chatbot development to science, practical applications, and implementation details, including features such as languages and automatic responses, among others. So, in the last section, Conclusions and Future Work, we summarize the key findings and recommend avenues for future research and application enhancement. 7 Chapter 2 Background and State of the Art 2.1 Enterprise Batch Processing In the technology area, batch processing is a method where we execute information suited to run programs that are called jobs. This approach is a low-cost solution for processing big volumes of data quickly. After starting the procedure, the computer will only halt if a mistake or irregularity is found and will alert the relevant employees or management. This efficient data processing approach stresses the significance of human-machine interaction, which is a key component in the greater landscape of technological systems. To delve deeper into the realm of batch processing and its implications, it is essential to establish three fundamental domains: the functional domain, the technological domain, and the human domain. These domains serve as the foundation for comprehending how batch processing systems operate and their impact on human-machine interaction. By dissecting these domains, we gain valuable insights into the intricate relationships between technology, functionality, and the human element, ultimately contributing to a holistic understanding of the role and significance of batch processing in modern computing environments. The functional domain is introduced into the batch execution process, and it is directly related to the functionalities that are conducted internally and how they are designed and constructed Spring [2023]. This execution process is made up of several modules, each of which operates in a different area of Retail Consult ’s business context. Each of these modules, which are made up of databases and applications, is assigned a specific agent, and the agent is assigned a specific task or set of tasks with certain dependencies. Finally, there is an agent with a higher hierarchical level than the others, known as the master, who determines the order of execution and job precedence Wikipedia [2023]. The technological domain corresponds to the environment where the batch is processed and, in turn, executed, as we mentioned earlier, an agent has a task and is associated with a module and each agent has a virtual machine assigned to it, which can be shared with another agent and they can run the same 8 application and also a specific port. As for the size of each machine, it was established according to the number of wires needed by customers Oracle [2017]. Moving on, the master agent gives tasks to the agents, and the execution is done using the FIFO method (first in first out), which means that the first agent to receive the job will be the first to perform it, automatically switching from passive to active mode. Furthermore, each agent must grant itself two crucial components: a maximum load and a weight per thread. If the entire weight of this agent’s threads is less than the maximum load, it is an agent to whom you may be assigned new tasks; otherwise, it is not Wikipedia [2023]. As previously said, batch processing requires a significant amount of time to finish since we are working with a big quantity of data. Unless there is an error that requires manual intervention, there is no human participation at this stage. Producing these processing mistakes can occasionally have a substantial negative impact on the business and the client waiting for service. As for the development, this can have an impact from the point of view of time and resource constraints, because if an error occurs during batch processing, the system will be stopped, which will suffer delays and aggravate the use of resources Wikipedia [2023]. Another critical aspect to emphasize is data integrity. Data damage or loss can result from batch processing mistakes, which can be damaging to the developer and system users. Furthermore, if the system experiences faults during batch processing, the user may have an unsatisfying experience. After all, they may be unable to obtain the information they seek. Finally, there is the consideration factor, because if errors in the batch cause serious problems in the system, they may negatively impact the developer’s and, in some cases, the system’s opinion. In conclusion, to minimize the impacts that errors cause in batch processing, the team or person who developed the system must have carefully designed and tested batch processing systems and implemented robust switches that handle and recover extremely well from mistakes. On the other hand, if there are errors during batch processing, the customer may also be affected. A potential impact would be the fact that the error itself prevents the necessary data from being processed outside the time requested by the customer, which would affect him, as he wouldn’t receive the expected results and could harm his commercial activity or cause other types of disorders. If the batch processing error leads to incomplete and inaccurate results, the customer will receive this data and, consequently, may lead to incorrect decisions in the business area. Finally, if there is a batch processing error for data loss, the customer may not have access to all the necessary information. 9 Figure 3: Definition of Chatbot Architecture Nacimiento-García et al. [2020]. 16 2.3 Literature Review In the previous topic, We spoke about the various virtual assistants and chatbots on the market, indicating those used the most by the public. Nevertheless, we are going to focus on explaining their specifications in detail. The requirements used in this study included whether the tools were open-source, whether they permitted the addition of new knowledge for personal or professional purposes, their features, benefits, and drawbacks, their ability to recognize both entities and various languages, and how they did so, their first interaction with the user, and how they implemented and developed this dialogue. Finally, to fully comprehend everything, we shall address a few crucial queries: 1. What are the common features? 2. What are the differences? 3. What groups do they insert? First of all, I would like to emphasize that research has been done on the following virtual assistants: Alexa, Google Home, the Facebook Messenger bot, RASA, and Deep Speech. Relatively to the first topic, what these assistants have in common are entity recognition and multilanguage recognition, and the best one is that they are all open-source, which means, they allow the public to use the source code to study, change, or modify the software without paying. Regarding multilanguage recognition, we have two important aspects to analyze, the way that they reach multilanguage and how the user speaks in another language to the assistant. Besides, they have these two things in common (entities and multilanguage recognition), but they have specific approaches to them. Yet to add, all of the virtual assistants will fail unless Apple allows a new knowledge base, permits us to train our models, and allows us to use them as we want and need. In this examination, we will give many virtual assistant comparison tables that address crucial characteristics and qualities to aid readers in selecting the best virtual assistant for their requirements. The ease of use of artificial intelligence has been linked to our daily encounters with technology via virtual assistants, which have evolved to become a vital part of our lives. However, it is crucial to note that not all tables may have the data needed throughout the search. This is a result of the field of artificial intelligence and virtual assistants constantly growing, as well as the accessibility and availability of data at the time of data gathering. 17 These tables are meant to give a thorough overview of each virtual assistant, enabling a comparison of its features and functionalities. With this knowledge, readers will be better equipped to select a virtual assistant that perfectly matches the demands of their projects or applications. To better understand the benefits and drawbacks of each virtual assistant, as well as the libraries utilized in their development, we give comparative tables of the virtual assistants examined. We have the ”Mozilla Deep Speech” voice recognition system compared in the table. It is Python-based open-source software created by Mozilla Firefox. The technology is renowned for text-to-audio conversion and for supporting entity and multilingual recognition. Good voice recognition performance and an emphasis on privacy are some benefits. The difficulty in interpreting spoken words, however, is a drawback. Through transfer learning, the system also enables the development of new knowledge bases NacimientoGarcía et al. [2020](Table 1). Table 1: Mozilla Deep Speech. Comparison Criteria Description Name Mozilla Deep Speech Nacimiento-García et al. [2020] Company / open source Mozilla Firefox / Yes relased under MPL (Mozilla Public License) Technology Python Features Multilanguage recognition Yes Entities recognition Yes Interaction virtual assistant? A stream of audio is inputted into DeepSpeech, which turns the stream of audio into a series of characters in the chosen alphabet. Application Field/Knowledge base General Settings (but this one is used to convert speech to text) Approach for multilanguage recognition You cannot use DeepSpeech because it does not detect the language (de/en/...) Your audio should be transcribed to see what has high confidence values. allow building new knowledge base Yes, you can remove specific layers from a pre-trained model and initialize new layers for your target data using DeepSpeech’s transfer-learning method, which also incorporates pre-trained data sets. But we can also include new sets) Advantages Excellent speech-to-text results, good privacy Disadvantage or limitations Voice words recognition difficult to understand 18 Comparison Criteria Description Libraries used They employ a model developed using machine learning methods and based on the Deep Speech research paper from Baidu. An overview of Amazon’s virtual assistant Alexa is provided in the Table 2. Alexa is a potent tool that offers recognition of over 50 languages and entities and is based on technologies similar to those shown. By using the word ”Alexa” to activate the assistant, Users can communicate with it naturally Hoy [2018]. Alexa’s multilingual capabilities are made possible by the self-supervised learning approach, which trains the models using massive monolingual datasets or unlabeled text. Entity resolution also guarantees that user requests are correctly interpreted Hoy [2018]. Alexa offers simplicity of use, online shopping, music playback, alarms, reminders, and more, highlighting benefits and features. It is crucial to note the system’s drawbacks, such as potential recognition failures, delayed answers, and disconnections. The chart also reveals that using Q&A and information from the company, a new knowledge base may be created Hoy [2018]. In conclusion, the table provides a clear and detailed review of the Alexa virtual assistant and its key characteristics (Table 2). Table 2: Virtual Assistant Alexa. Comparison Criteria Description Name Alexa Hoy [2018], Kepuska and Bohouta [2018] Company / open source Amazon / Yes Technology Javascript, java, Python, C#, Go, Ruby, or PowerShell Features Multilanguage recognition Yes (50+ languages) Entities recognition Yes Interaction virtual assistant? The words and phrases that users can use to direct Alexa to perform a task are defined by the voice interaction model. First interaction User says an awanke word - Alexa Application Field/Knowledge base General Settings Approach for multilanguage recognition Amazon used a self-supervised learning pretraining strategy, in which the models are trained on sizable, monolingual datasets or unlabeled texts, to give Alexa multilingual capabilities. 19 Comparison Criteria Description Approach for entities recognition The user’s request for a slot value is resolved by Alexa using entity resolution for a single, recognized entity. Approach for interaction Relatively to voice recognize Alexa uses ASR (automatic speech recognition) Allow building new knowledge base Yes (Q&A on your organization’s data) Advantages Is a tool that is simple to use, capable of online shopping, and equipped with nonstop music, timers, alarms, and reminders. MyAyan [2022] Disadvantage or limitations mishearing, slow response and bloatware, connection drops, stop responding MyAyan [2022] Libraries used Alexa Skills Kits provides a library of built-in intents In comparison to Google Home, a Virtual Assistant was created by Google that enables entity identification, voice interactions, support for many languages, a general knowledge base, and the capacity to build new intelligence clusters. While it has numerous benefits, like syncing and frequent updates, there are also drawbacks, such as the voice quality, which isn’t fully natural, and privacy issues PeopleBank [2018] (Table 3). Table 3: Google Assistant. Comparison Criteria Description Name Google Assistant/Google Home Hoy [2018], Kepuska and Bohouta [2018] Company / open source Google / Yes Technology Javascript, go, c++, java Features Multilanguage recognition Yes (44 languages) Entities recognition Yes Interaction virtual assistant? When you ask for assistance, Google Assistant can use information from your Google Account to provide you with what you need First interaction Say “Hey Google” and ask your question or give a command Application Field/Knowledge base General Settings Approach for multilanguage recognition We need to change the settings 20 Comparison Criteria Description Approach for entities recognition The entity type for each intent parameter specifies how data from an end-user expression should be extracted Allow building new knowledge base Yes (New Intelligence Clusters let you build devices that respond to whole-home context such as presence, for more helpful experiences) Advantages Synchronization, upgrade day by day, has many functionalities PeopleBank [2018] Disadvantage or limitations No natural voice, privacy concern PeopleBank [2018] Libraries used Google Assistant library for Python, supported by Raspberry Pi 3 A comparison of Cortana’s Virtual Assistant is shown in the table. To comprehend its features and functionalities, it highlights pertinent comparison criteria. The application for the Cortana virtual assistant focuses on general settings and supports the creation of additional knowledge bases. However, the settings for multilingual support must be changed. Several customer support functions, including meeting assistance and calendar management, are utilized by the assistant Microsoft [2022]. Cortana has a significant drawback in that it may be susceptible to malware infection, jeopardizing its security. Consequently, it enables a more unbiased study of its strengths and weaknesses for the objectives for which it is intended (Table 4). Table 4: Virtual Assistant Cortana. Comparison Criteria Description Name Cortana Hoy [2018], Kepuska and Bohouta [2018] Company / open source Microsoft / Yes Technology C++ Features Multilanguage recognition Yes (19 languages) Entities recognition Yes Interaction virtual assistant? EI - emotional intelligence 21 Comparison Criteria Description First interaction When we launch Cortana, the following message appears: Type a message for me in the text box or click the microphone icon to speak to me. Application Field/Knowledge base General Settings Approach for multilanguage recognition We need to change the settings. Approach for entities recognition NLP, DL, computer vision Approach for interaction Cortana’s exploration of emotional intelligence (EI) is based on computer vision and machine learning (ML) methods. Allow building new knowledge base In the most recent version of Cortana for Windows, users can search for documents and quickly compose emails. Advantages Various customer support functionalities (e.g. meetings, calendar assistance, etc.) Microsoft [2022] Disadvantage or limitations Cortana can be tricked into installing malware PandaSecurity [2022] Libraries used Cortana Skills Kit A comparison of virtual assistants is presented in the table, with a special emphasis on Apple’s Siri, its assistant. It includes crucial details about Siri, including its underlying technology, attributes, domain of use, and strategies for entity identification and language recognition Reuters [2022]. The popular and frequently used virtual assistant Siri was created by Apple and is optimized for iOS devices. With a Python technical foundation, Siri offers multilingual recognition, enabling users to communicate with the assistant in a variety of languages. Additionally, it contains entity identification capabilities that enable a deeper comprehension of user commands and inquiries. The assistant operates through voice interaction, with Siri’s servers handling command processing. Siri does have significant limits, such as privacy concerns and difficulties interacting with kids while being widely used and acclaimed for its usability and communication abilities. In summary, the table provides a summary of the key characteristics, benefits, and drawbacks of Siri, enabling a comparison between this personal assistant and other products on the market (Table 5). 22 Table 5: Virtual Assistant Siri. Comparison Criteria Description Name Siri Hoy [2018], Kepuska and Bohouta [2018] Company / open source Apple / Yes (has a version that is open source) Technology Python Features Multilanguage recognition Yes Entities recognition Yes Interaction virtual assistant? Siri servers receive and process your speech inputs. In every situation, Apple will get transcripts of your communications in order to fulfill your requests First interaction Siri speaks back to you when you ask her a question. After saying ”Hey Siri,” ask a question or make a request Application Field/Knowledge base General Settings: 98 percent of iPhone owners have at least once used this virtual assistant Approach for multilanguage recognition To give the computer a perfect depiction of the spoken text to learn, humans read passages in a variety of accents and dialects. They record a variety of noises as well. They then create a language model to forecast word sequences from there Approach for entities recognition Apple uses huge, datasets to provide Siri with an effective model of speech recognition which is then trained on varying datasets that are made up of voice samplings from lots of people, which allows Siri to recognize all sorts of accents, inflections, and pace of speech Allow building new knowledge base No (Siri Data and your requests are not used to build a marketing profile) Advantages Very easy to use, good in comunication Medium [2022] Disadvantage or limitations don’t work well with children, privacy issues, lack of function Medium [2022] The table provides a comparison of many features of the Facebook company’s ”Facebook Messenger bot,” which is housed inside the BlenderBot open-source project. Erlang is used for the chat component of this bot’s front, whereas Java and C++ are used elsewhere. The bot recognizes entities and offers recogni23 tion in several languages. It also enables the creation of a customized knowledge base. Greater consumer interaction and simplicity in scaling are two of its benefits. Despite its skills, it’s crucial to remember that this bot cannot completely replace human engagement in encounters Smutny and Schreiberova [2020]. The table’s content describes the methods for entity and entity-entity interaction, multilingual recognition, and messaging, as well as the libraries that are used, like the ”messenger-bot-library” (Table 6). Table 6: Facebook Messenger bot. Comparison Criteria Description Name Facebook Messenger bot Smutny and Schreiberova [2020] Company / open source Facebook / Yes (BlenderBot) Technology Php for frontend, erlang used for chat, java and c++ in other places Features Multilanguage recognition Yes Entities recognition Yes Interaction virtual assistant? Uses the same things as RASA First interaction A Messenger session is automatically started when someone clicks the ”Message” button on your Facebook page (or website), allowing them to type a query and start a conversation with your bot Application Field/Knowledge base Education Chatbots Approach for multilanguage recognition We need to change on the settings Approach for entities recognition As the parameter entities on yaml files Approach for interaction Use the same things of RASA Allow building new knowledge base Yes Advantages Higher Customer Engagement, easy scalability Martech [2022] Disadvantage or limitations It won’t replace human touch Martech [2022] Libraries used messenger-bot-library is a python library for facebook messenger bot The RASA framework, an open-source platform designed to produce chatbots and virtual assistants with advanced natural language understanding skills, is described in depth in the table. Python and NLU (Natural Language Understanding) are the foundations of RASA, which enables programmers to build interactive and customized conversational systems RASA [2023a]. The comparison criteria are first presented in the table, which is then followed by details regarding RASA. It gives a comprehensive summary of the most crucial elements, including the name of the company 24 (or whether it is open source), the technology employed, and the application domain of RASA, which includes programmers, conversational teams, and businesses in general. The table includes a list of some of the libraries used by RASA, including SpaCY and TensorFlow, to help readers better understand the framework. In summary, this table is a helpful resource for presenting, contrasting, and comprehending the key aspects of RASA, assisting teams and developers in selecting the best platform for their virtual assistant and chatbot projects (Table 7). Table 7: RASA framework. Comparison Criteria Description Name RASA Singh et al. [2019], Bocklisch et al. [2017] Company / open source RASA / Yes Technology python and NLU (Natural Language Understanding) Features Multilanguage recognition Yes (it is possible to implement a chatbot that allows multilingualism Entities recognition Yes Interaction virtual assistant? RASA uses intents and responses Application Field/Knowledge base Developers, conversational teams and enterprises Approach for multilanguage recognition We can train X models, one for each language, and then load X agents, selecting which agent should receive the message using a language detector Approach for entities recognition RASA uses spaCY for names, places, and organizations related to this issue. This system employs ducks for numerical capture, and if that isn’t enough, we can also use regex or neural approaches. Approach for interaction RASA communicates with the user through actions, as well as these units’ intentions (which are statements that the user may make to the assistant) and answers (which are groups of things that the assistant may make) Advantages Easy to integrate, interactive learning RASA [2023a] Disadvantage or limitations Anyone interested in using RASA to develop a virtual assistant should be familiar with chatbots and NLP (natural language processing) RASA [2023a] Libraries used SpaCY, TensorFlow, etc 25 Kuyven et al. [2018] Table 14: Chatbots in education: a systematic literature review. Comparison Criteria Description Title Chatbots in education: a systematic literature review Kuyven et al. [2018] Focus Natural language recognition Context Chatbots are a system that is increasingly found in the health area, and the content of the article focuses on education. However, the area where it appears the most is in the field of communication sciences. Data The six databases that gave rise to the publications extracted in the RSL, in descending order, were: Computers & Education(5 -31.3%); Revista Novas Tecnologias na Educação-RENOTE (4 -25.0%); IEEE Transactions on Learning Technologies (2 -12.5%); Brazilian Symposium of Informatics in Education-SBIE(2 - 12.5%); International Congress of Educational Informatics -TISE (2 -12.5%); Latin American Journal of Educational Technology -RELATEC-(1 -6.2%). Methods AIML (Artificial Intelligence Markup Language); PLN (Natural Language Processing, e.g., pattern matching, Hash tables for keywords). Evaluation Methods ML (Machine Learning); DL (Deep Learning); Issues and Limitations The need for a considerable knowledge base for a satisfactory conversation; the greater complexity and unpredictability of the dialog flow between the agent and the learner compared to a casual conversation Conclusion The machine learning and natural language processing parts show the best results, as does the use of deep learning, which is a type of machine learning, for a deeper search. Keywords Chatbots, conversational agents, artificial intelligence, and education A summary of the research paper ”A Classification Model for Named Entity Recognition” Silva [2020] can be found in the table below. The work focuses on employing word embeddings and vector representations to identify named items using contextual distribution and linguistic features. Reputable sources in the fields of educational technology and human-computer interaction provided the data for the evaluation. BERT (Bidirectional Encoder Representations from Transformers) and BiLSTM (Bidirectional LongTerm Memory), two well-liked neural network designs for named entity recognition, are the techniques employed in the study. As an example, the term ”Castelo Branco” can be both a place and a name, 32 making its classification difficult. The study’s conclusion emphasizes that results were mostly produced through neural networks and that further development and improvement of this area of study is possible (Table 15)Silva [2020]. Table 15: A classification model for the NER. Comparison Criteria Description Title A classification model for the Recognition of Named Entities Silva [2020] Focus Exploitation of traces based on the contextual distribution of entities through word embeddings and vector representations associated with linguistic traces Context The exploitation of traces based on the contextual distribution of entities through word embeddings and vector representations associated with linguistic traces Data Database used: ACM Transactions on Computer-Human Interaction (TOCHI); Congresso Internacional de Informática Educativa (TISE) ; Computers & Education; Creative Education; IEEE Revista Iberoamericana de Tecnologias del Aprendizaje, entre outras Methods The use of two neural network architectures such as BERT and LSTM (bidirectional neural networks) Evaluation Methods BERT ( Bidirectional Encoder Representations from Transformers) and BiLSTM (Bidirectional long short term memory) Issues and Limitations Difficulties in the classification of entities hampered the evaluation of the models (e.g. ”Castelo Branco” can be both a place and name) Conclusion Not talking directly about technologies to be used but about techniques. The results obtained were almost all obtained through the use of neural networks, and research with this type of technique is always open to improvement. Keywords Representation of a given language, neural networks, named entity recognition, embedded words This table provides an outline of the paper ”An Overview on Chatbots Technology” Adamopoulou and Moussiades [2020]. The study focuses on the context of virtual assistants as it examines the history, motivation, design, classification, and architecture of chatbots. Chatbots employ knowledge bases to respond to different user inquiries. The article outlines three main approaches for putting chatbots into practice: the rule-based model, which employs a set of predetermined rules to determine responses; the retrieval-based model, which is 33 more adaptable and uses APIs to retrieve data; and the generation-based model, which is more sophisticated, employs deep learning, and can deliver better results based on prior user messages. The article mentions RASA, an open-source platform, as well as closed platforms from Google and IBM that are made available for enterprise use. Finally, the paper covers key ideas in artificial intelligence, machine learning, and NLU (natural language understanding) to help readers comprehend the fundamentals of chatbot technology. Chatbot, chatbot architecture, artificial intelligence, machine learning, and NLU are all pertinent keywords (Table 16)Adamopoulou and Moussiades [2020]. Table 16: An Overview of Chatbot Technology. Comparison Criteria Description Title An Overview of Chatbot Technology Adamopoulou and Moussiades [2020] Focus History, motivation, design, classification, chatbot architecture Context Knowledge about virtual assistants Data Databases allow the chatbot to answer more types of questions from users (e.g. the knowledge base of the chatbot) Methods Rule-based model (most common; choose the system response based on a predefined set of rules such as lexical form recognition); retrieval-based model (less common, more flexible; query and analyze available resources using APIs); and generative-based model (chatbots closer to humans; use deep learning; harder to implement; better than the other two based on users’ old and current messages). Evaluation Methods RASA (open source), Google e IBM (closed plattforms, offered by enterprise) Conclusion Demonstrates information needed to understand the basic principles of a chatbot Keywords Chatbot, chatbot architecture, artificial intelligence, machine learning, NLU (natural language understanding) 34 Chapter 3 Problem and Challenges Developing a chatbot in a business context, especially in the area of batch processing, presents several complex challenges that require creative approaches and innovative solutions. In this chapter, we will look at some of the main problems and challenges faced during the course of this project, as well as the strategies adopted to overcome them. One of the most significant challenges encountered was ensuring that the chatbot was globally accessible in the business environment. This accessibility required support for several languages, considering the linguistic diversity that can be found in companies with international operations. The solution adopted to solve this problem was to create a robust and versatile translation system. This system combined Deep Translator, a flexible, free, and unlimited Python tool designed to efficiently translate between different languages using a variety of Organization [2023] translators, with Spacy’s core language identification models. Spacy, an open-source library for Natural Language Processing in Python, offers advanced features such as named entity recognition (NER), part-of-speech (POS) tagging, dependency parsing, word vectors, and much more SpaCy [2023]. This combination allowed the chatbot to be effective in different languages and to communicate effectively with a wide range of international audiences. Another critical challenge we faced was integrating RASA with external databases, which proved crucial in the context of batch processing. To ensure that the chatbot provided relevant and accurate responses to user requests, it was essential to access real-time information from databases such as IBPS and IPE. To achieve this integration, custom actions were defined in RASA, allowing the chatbot to access these databases efficiently and securely. In addition, SQL queries were implemented in Oracle Retail SQL Developer to extract the necessary data, ensuring that the chatbot had access to up-to-date and accurate information. This close integration with the external databases was crucial to the chatbot’s effectiveness, as it allowed it to provide real-time information and high-quality contextual responses to users. In summary, developing a chatbot in a business context with an emphasis on batch processing is a challenging task that involves overcoming obstacles related to global accessibility and effective integration 35 with external databases. The combination of advanced technologies such as Deep Translator and Spacy, along with the implementation of customized actions and SQL queries, made it possible to create an efficient and highly functional chatbot that met the specific needs of this business project. 36 Part II Core of the Dissertation 37 Chapter 4 Contribution This dissertation describes the creation and application of a cutting-edge chatbot that makes use of the RASA framework and aims to provide users who are fluent in English, Portuguese, and Spanish with a warm and welcoming interaction experience. The chatbot can access relevant data stored in a database and respond to queries contextually. It also stands out since it exhibits adaptive and continuous learning behaviour by being able to make wise corrections and skip particular steps during the interaction. The research presented in this dissertation makes a substantial contribution to artificial intelligence, especially in the creation of sophisticated and engaging chatbots. The following is an explanation of the major contributions: The design and implementation of a chatbot with a friendly and positive focus on user interaction is one of the key accomplishments of this dissertation. Advanced natural language processing (NLP) and a focus on sentiment analysis methods were used to accomplish this goal. The chatbot is careful to ask sometimes how the user is and can modify its responses to sound human depending on whether the user says he is good or sad, so he tries to understand the user’s emotions. This strategy tries to enhance the user’s interaction experience, boosting satisfaction and system receptivity. The proposal of an effective method for integrating the chatbot with a database containing pertinent data is another significant contribution of the disseration. The chatbot uses sophisticated data querying and structuring techniques to deliver precise and current responses to users’ inquiries. This capacity is especially useful in settings where timely, accurate information is required. The chatbot can assist a range of useful applications, from customer care to the distribution of knowledge, by optimizing this data access. One of the most creative contributions of this research is the use of intelligent corrections. The chatbot exhibits higher flexibility and adaptability during the engagement process by making intelligent modifications like skipping particular processes, depending on the user input, making it a more dependable and versatile solution. And finally, the chatbot’s support for the three major languages of English, Portuguese, and Spanish 38 makes a substantial contribution to cross-cultural communication and worldwide user service. Future research on artificial intelligence systems with multilingual assistance can use the strategy used to assure multilingualism’s effectiveness as a guide. The chatbot’s ability to operate in a variety of languages, which enables it to effectively connect with a wide range of consumers globally, significantly expands its reach and utility. In conclusion, the contributions of this dissertation represent advancements in the field of chatbot artificial intelligence. The research opens up new possibilities for the use of chatbots in various sectors, from customer service to education, by developing a highly functional and interactive system that can learn and adapt to users’ needs and operate in multiple languages, thus propelling the advancement of AI technology on a large scale. 39 Chapter 5 Applications 5.1 Description Implemation In this section, we will address aspects related to the implementation of the chatbot, present a broader and more generalized view of what was developed and for that, and address the following: • Configuration of the development environment; • Installation and configuration of RASA; • Creation of domains, intentions, entities, and the definition of conversation stories; • Designed and trained the chatbot model. Firstly, we chose Windows for our development work because we are familiar with the operating system made it easier to focus on creating the chatbot rather than having to resolve problems with the environment’s configuration. . The use of readily available and accessible tools and resources was made possible by the selection of Windows as the operating system. Regarding the installation and configuration of the components necessary for the development of this project, we can say that it was an obligation to insert all the mandatory dependencies for the use of TensorFlow. That is, TensorFlow is a dependency of RASA, and to be successfully installed on Windows 7 or a higher version, it was needed to install the C++ redistributable for the specific architecture of my machine, and for its application, it was necessary to restart it RASA [2023b]. After finishing the Tensorflow installation, Anaconda had to be set up. This decision is entirely up to the user, who may also choose to install the required parts directly on the computer. But since I think installing everything in an Anaconda environment will be simpler, I went with that option. For example, if you want to install a specific Python version, which is difficult if you do not use software like this or anything similar. Anaconda installation was nearly default, but to avoid difficulties, I opted to add it to 40 the environment variables because I didn’t have any version of Anaconda installed on this machine RASA [2023b]. Proceeding when the installation was finished, the next step was to open Anaconda’s command line and create a virtual environment, mainly to install the Python version more adequate to my needs, because Anaconda’s default environment, named as base, according to the installation process I did, came with version 3.9.13 and I didn’t want to use that version, I preferred to use an older version, for questions like code compatibility, specific dependencies, stability, among others. In other words, to create the new environment, I used the command conda create -n multilingua python=3.7 , where multilingua corresponds to the name I decided to give the environment, then I activated the environment with the command conda activate multilingua , then already inside multilingua I uninstalled and installed pip, to ensure that it came with the latest version developed and this was done through three commands that we will mention RASA [2023b]: • For uninstall pip: python -m pip uninstall pip ; • To ensure that the pip library can be installed again in the environment: python -m ensurepip ; • For install latest version of pip: python -m pip install -U pip The installation of RASA was done quite simply. It was necessary to run in the command line of Anaconda the command pip install rasa . However, by choice, I decided to install an older version, so the command was the following pip install rasa==2.3 . To verify that the installation was successful, I ran the command rasa -h RASA [2023b]. Afterwards, and already with the guarantee of having RASA in the system, I ran the command rasa init , which, in short, creates a project with example training data presented in the data folder, which are nlu, stories, rules files, and configuration files too RASA [2023b]. The following diagram shows the project structure created based on the command mentioned above. It is relevant to offer a more detailed explanation of some of the files mentioned to better understand the operation of RASA and deepen our understanding of these key concepts. This will also be addressed in the last two topics mentioned at the beginning of this section, which are the creation of domains, intentions, and entities, the definition of conversational stories, and the definition and training of the chatbot model. As we see in the project path, the models folder is empty, but we can generate a model, because we already have example training data, like natural language understanding data (NLU) and stories. To do that, we need to run the command rasa train . 41 is triggered when there is an insufficient knowledge base to answer the user’s inquiry, particularly if the question is unrelated to the case under consideration. Another important aspect to mention is that the illustrative example that we see below, it has information about a list of entities and slots, something that was not explained in the greeting example because, for that specific action, it was not necessary to use it. Briefly, the entities serve to capture a specific word from the message provided by the user. For example, and now speaking more intrinsically in the project, let’s imagine that the user wants to know what the functional area of a job is, he can send a message like ”What is the functional area of job x, where x represents the name of the job” What the chatbot will do in this situation is the following: it stores the word x and assigns its value to an entity called a job. This entity can store other values that are defined by me. Then the response to the user is sent informing the respective functional area. To conclude, I would like to point out and reinforce that another architecture diagram was not built, as this corresponds to the process of creating visual representations of the components of the software system, and in the following one components such as rectangles with specific RASA commands and files were added. After the analysis, it is expected that it will be easier to consolidate the knowledge of how everything is interconnected and processed (Figure 10). Figure 10: Explanatory diagram to aid writing (based on RASA [2023a]). 48 5.3 Chatbot Architecture In addition to using visual diagrams to help explain files and commands, the development of intelligent chatbot systems is an important aspect of modern communication systems. Advanced frameworks and components are used in these systems to enable interactive and efficient user-bot interactions. In the parts that follow, we will look at the high-level architecture of a chatbot system that has been meticulously developed to handle seamless chats, multilingual support, database integration, and integration with external modules. We hope to highlight the underlying components and their relationships by investigating their design, shedding light on the intricate workings of this intelligent communication system. The chatbot system’s high-level architecture is made up of various interconnected components that allow for effective communication between the user and the bot. At the heart of the architecture is a central box dubbed ”Chatbot,” which houses the main intelligence and logic that drives the bot’s activity. To offer a smooth conversational experience, the chatbot interacts with other components that were implemented with the help of RASA used to design and train the bot, which is one of the essential components coupled with the chatbot. RASA supports natural language comprehension and processing, allowing the bot to interpret human input and create suitable answers. A User Interface (UI) is linked to the chatbot to ease user interaction. The UI represents the chatbot visually, giving users an interface through which they may engage and get replies and is powered by RASA Botfront WebChat. The ”Multilingual” box includes the chatbot’s multilingual functionality and he can speak three languages, English, Portuguese and Spanish, being English the main one. The chatbot system also includes automated responses to frequently asked questions. These responses are saved in databases and can be retrieved by the bot as needed, with selections and corrections dependent on human input. Additionally, connections to the ”IBPS” and ”IPE” modules, as well as their associated databases, are also included in the architecture. The chatbot can communicate with these modules and get needed data and functionality thanks to these connections. Overall, this high-level design demonstrates the interconnection of numerous components, allowing the chatbot to interpret user inputs, provide suitable responses, support multiple languages, and effectively interface with external modules and databases (Figure 11). To conclude, I would like to mention that this high-level architecture of the project was defined with the help of the article Faria et al. [2022]. 49 Figure 11: Chatbot high level architecture. 50 5.4 Chatbot Incorporation I’ll start off by discussing how my chatbot was included into the business context by illustrating this architectural image before moving on to the technical context of the program (Figure 12). Due to its integration with Retail Consult and its incorporation into the Intelligence Batch Processing System (IBPS) module, the chatbot in question plays a crucial role as an internal tool of great value for the organization. In addition, a link is established to the Intelligence Processing Engine (IPE) module, which enables access to the same database. In terms of architecture, the chatbot is integrated into the application layer, which includes a module with functionalities such as task scheduling and forecasting. The main components are IBPS and IPE. In addition, Oracle Retail, machine learning, and artificial intelligence are all connected to the database, constituting the data layer. The chatbot acts as an internal assistant, providing information to the company’s employees. When requested, it can perform direct queries to the database, making specific ”selects”. This functionality is incorporated into the chatbot’s automatic responses, as highlighted in blue in the figure above. This ability to obtain information about the data quickly and accurately, without the need to manually query the data, is enabled by the integration of the chatbot into the organizational structure. To add, one of the main features of the chatbot is its ability to perform correction routines. For example, a user only needs to submit a request to skip a certain job. This functionality of skipping a job, offered by the chatbot, is an option present in the corrections, also highlighted in blue in the image. This functionality provides better administration and control of the system processes. The chatbot acts as a reliable assistant, providing quick and efficient responses to user requests, including, when necessary, the modification of specific routines, present in the patches. In short, the chatbot performs tasks and implements changes for the company’s employees, as well as providing relevant information. Its integration into Retail Consult ’s servers extends its usability and accessibility to all users, increasing efficiency and productivity in the work environment. 51 Figure 12: Incorporation of chatbot in the business context. 52 5.5 Multilingue Allowing the chatbot to converse in many languages was one of the key problems and needs in the creation of this dissertation. The usage of this in the context of a global firm that works in numerous countries, specifically Portugal, Germany, the United States, Brazil, Mexico, Chile, and China, was a major difficulty and demand in the production of this research. Although English is the chatbot’s native language and is frequently used in international settings, it is a common practice to speak in one’s mother tongue when conducting business. This is because cultural, historical, and geographic factors can generate considerable pronunciation and accent differences in English, depending on the country. Given the company’s global reach, it is imperative to comprehend how various regional tongues and dialects may influence the English language, resulting in a variety of accents and communication styles. Understanding that these variances represent rich cultural and linguistic variety rather than a language’s shortcomings is crucial for improving communication on a global scale. It was planned to create a chatbot that could understand and respond in several languages for these reasons. The system was split into two primary components to address this need: identifying the user’s language and providing a response in that language. When the chatbot appears, a welcome message with the following text is displayed: ”Welcome! Your virtual helper: I’m RC Bot. This strategy aims to give users a welcoming and approachable experience, regardless of the language they choose to communicate with the chatbot in (Figure 13). 53 Figure 13: Welcome message to the user. The user must tell the chatbot in English whatever language they want to speak for the chatbot to change, which is the sole constraint in the context of combining many languages. The chatbot cannot change the language if the request is not made in English. For instance, if you want to speak Portuguese or Spanish, just say ”speak Portuguese” or ”speak Spanish” in English to indicate your selection. The chatbot will only react in the chosen language moving forward until the user chooses to switch back to it. The future operation of this interaction will subsequently be visually demonstrated (Figure 14). Figure 14: Test witch Language - English to Portuguese. 54 Users can converse freely and naturally in their favourite language thanks to this tailored approach. A more inclusive and effective work environment is created by the chatbot since it encourages greater comprehension and interaction by speaking the same language as the user. When the user asked the chatbot to speak Portuguese, just as seen in the previous image, it answered by saying that it would, having previously done so. On the other hand, a conversation with the chatbot in Portuguese has not yet been demonstrated; it was only apparent that the language shift had been successfully implemented. The user and the chatbot can then be seen having a conversation in the following picture, considering the Portuguese language as a language and the progression of the conversation in the prior image (Figure 15). Figure 15: Test user-chatbot conversation in Portuguese. Additionally, I would want to draw attention to some additional routes that are connected to the multilingual area. For each of these paths, I will include an illustration to support the functionalities that will be explained. As we can see, the chatbot also speaks Spanish (Figure 16). 55 Figure 16: Test user-chatbot conversation in Spanish. If the user asks the chatbot to speak English and the language is already in English, it will inform them of this information and tell them which languages they can speak (Figure 17). Figure 17: Test user-chatbot conversation in English. Finally, the user has the option to switch languages as many as they like while speaking (Figure 18). 56 Figure 18: Test user-chatbot multiple languages. This system’s implementation required considerable consideration to support two-way communication with the user and allow for dynamic language change. The usage of a bespoke pipeline component—which the programmer can create from scratch or choose from preexisting options—enables this functionality. In this instance, it was decided to use Spacy, a library built into the RASA program that has various developed components, as was already discussed in the research section of this dissertation. The en_core_web_md is built with a CPU-optimized pipeline and is intended to handle English text quickly. The existing RASA NLU models can be improved by adding the en_core_web_md model to the RASA NLU configuration file. As previously indicated, these models can be altered to meet the unique requirements of the RASA [2023a] project. A portion of the RASA train command process will then be displayed, showing RASA attempting to load the model into rasa.nlu.utils.spacy_utils which is the component that gives others, in this case, my chatbot, access to the shared loaded SpaCy model. Next, we can see the model being added to the RASA rasa.nlu.components (Figure 19). Figure 19: Spacy model ’en_core_web_md’ loaded and added. It’s also necessary to note that this component is essential for automatically determining the user’s language. Spacy’s approach deciphers linguistic elements from the input text and establishes the domi57 5.6 Social Actions As we know, the chatbot can respond to specific questions by calling dynamic and non-dynamic queries and HTTP requests, making real-time corrections to the database, supporting workers, and providing individualized support by providing clear and unbiased answers. On the other hand, this section will address the interaction and communication activities of this chatbot, which were an important focus of this study. By establishing a consistent and friendly dialogue with users, they hope to provide a more pleasant and seamless experience to establish a deeper relationship with people, express gratitude and praise, and pose as a transparent bot. This method seeks to improve humanmachine interaction, making it more pleasant and efficient in a range of applications. The information that we used to improve the dialogue with the user was this table (Table 17) that contains the action name, example of sentences that the user writes to the chatbot and examples of sentences that the chatbot writes to the user: Table 17: Protocol Social Actions. Actions User examples phrases Chabot examples phrases action_utter_greet ”Hey”, ”Hello”, ”Hi” ”Hi! How are you?”, ”Hi there!”, ”Hey” action_utter_happy ”Perfect”, ”Great”,”Feeling like a king”, ”I’m good, thanks” ”Great, carry on.”, ”Great”, ”Great, how can I help you?” action_utter_unhappy ”My day was horrible”, ”I don’t feel very well”, ”I am sad” ”I’m sorry to hear that.”, How can I help you?”, ”I’m sorry to hear that you’re feeling sad.”, ”Sometimes it can help to share your thoughts and feelings with someone.” action_utter_goodbye ”Bye”, ”Goodbye”, ”Have a nice day” ”Bye!”, ”Goodbye!”, ”See you soon!”, ”See you later” action_utter_iambot ”What are you?”, ”Are you a bot?”, ”Are you a human?” ”I am a bot, powered by RASA.”, ”I’m a chatbot!”, ”Beep, boop, I’m a bot.” 64 action_ok_option ”Ok”, ”Okey”, ”Okay” ”Can I help you with something more?”, ”May I assist you with more?”, ”Do you need to know more information?”, ”If you need any further assistance, I’m here to help.” action_utter_thanks ”I appreciate that”, ”Thanks”, ”Thank you”, ”Good Help” ”You are welcome! I’m happy to be able to help. If you have any other questions, don’t hesitate to ask.” 5.7 User Interface We all know that as technology has advanced, there has been an increasing need for graphical interfaces since they offer a more effective and simple user experience, facilitating access to and perception of more complicated applications. For this reason, the features that were incorporated into the designed graphical interface will be examined and analyzed in this part. It is feasible to confirm that after the chatbot is closed, the message shows as though it were a notification as seen in the following image. (Figure 22). Figure 22: Chatbot close message. The chat interface displays the user’s messages with a grey backdrop to indicate the message time, while the chatbot’s messages have a white background. By pressing a button, users may go fullscreen, and it stays fullscreen even after closing and reopening. Click X to end fullscreen. The interface is titled ”RC BOT” and includes the subtitle ”Greet and Get Started.” (Figure 35) 65 Figure 23: Chatbot fullscreen mode. Assume that while the user is not writing any messages, an init payload with the message ”Type a message...” shows. In this image, we can see that the user is typing or has completed typing everything; here is the word ”good,” and that message has disappeared; what is now visible is the symbol to send the message. When you activate the bot, the message icon displays, as seen in the figure below. When we launch the chatbot, we notice an x that implies that tapping it would close the chatbot. The chatbot has an orange backdrop, while the user’s messages have a white background that complements the company’s color scheme. 66 5.8 Logic Actions Other activities implemented in this work that required more logic have not yet been fully discussed, such as joining both the IBPS and IPE database. Furthermore, I will now demonstrate all of the actions that have been developed in this bot, such as demonstrating which and how many jobs are active and not active, which jobs have pre and post-dependencies, which jobs are configured in batch processing, which is the functional area of a specific job, which is the critical path and the minimum execution time of the same for a given day, how long it will take to process a given job, provide information about the status of a job, and so on, for example, a task can be finished, running, submitted, or with an error, and the chatbot can advise you of the number of jobs in each stage, how long it will take to finish the batch if it is running, and ultimately, we can skip work from the database. The actions that deserve more attention are those with the ability to make dynamic queries, that is, queries that are made based on what the user is asking, and of this type of action, we have some of those that I mentioned above, starting with the one that provides information about the job’s status. When a user asks an inquiry about the status of jobs, the chatbot can perform a specific action called job status. It interprets the user’s language and the job status they wish to check using information stored in slots. 67 Algorithm 3 Determinação de Status de Trabalho 1: procedure JobStatus(dispatcher, tracker, domain) 2: status ←tracker.get_slot(‘status’) 3: if not status then 4: dispatcher.utter_message(‘Desculpe, não entendi...’) 5: return 6: end if 7: status ←status.strip().lower() 8: response ←‘Unknown status.’ 9: status_map ← {”submitted”:”T hesubmittedjobsare :”,”working”: ”Thejobsthatareworkingare :”,”finished”:”T hefinishedjobsare :”,”error”: ”Thejobsthatshowederrorsare :”} 10: if status in status_map then 11: response ←status_map[status] 12: end if 13: results, error ←some_query_function(status) 14: if not results then 15: dispatcher.utter_message(‘O lote não está em execução no momento’) 16: return 17: end if 18: if not error then 19: language ←tracker.get_slot(‘language’) 20: translated ←response 21: if language is ‘Portuguese’ then 22: translated ←GoogleTranslator(source=’en’, target=’pt’).translate(response) 23: else if language is ‘Spanish’ then 24: translated ←GoogleTranslator(source=’en’, target=’es’).translate(response) 25: end if 26: dispatcher.utter_message(translated) 27: else 28: dispatcher.utter_message(error) 29: end if 30: end procedure 68 The function indicated in the image is the one that will be explained in detail: The function begins by reading the tracker object’s values for the slot’s language and status. The slots are variables that hold vital information during the user conversation. The language column most likely contains the user’s language preference, and the status slot contains information on the status of the ”job” the user is inquiring about. As previously mentioned, this function handles the task of storing the information entered by the user. In such a scenario, if the user inquires about the status of a job that either doesn’t exist, is spelled incorrectly, or if they omit the job’s name altogether, the chatbot will respond by expressing that it didn’t quite understand the user’s request and will politely request them to provide a clearer phrasing. To ensure that comparisons with status keywords are case-insensitive (meaning it doesn’t care about letter case), the code strips away any extra spaces and converts the contents of the status slot to lowercase. This way, users can input the status in any format, and the chatbot will still understand it correctly. Once the status slot value has been processed, the function checks if it matches any of the known keywords, like ”submitted,” ”working,” ”finished,” ”error,” and so on. If there’s a match, it assigns a status code (S, W, F, or E) and crafts an appropriate response message in English. If the status slot value doesn’t match any of these known keywords, the chatbot will respond in English with ”Unknown status” to let the user know that the specified status is not recognized. The function then searches the database for job names that match the specified status. The results of this query are stored in the results variable, and the error variable helps determine whether any issues arose during the query. It’s worth noting that the question is constructed using words supplied by the user, which is what I wanted to highlight. In this scenario, we have an action that can run four different queries, making the query process dynamic. If the query returns no results (results == []), the chatbot will inform the user that no jobs are currently in progress. In case there’s an error during the database query (non-empty error), the chatbot will notify the user about the technical problem. 69 If the user specifies a language (like ”Portuguese” or ”Spanish”), the chatbot will use the Google Translator API to translate the English response into the chosen language. If no language is specified or if it’s not supported, the chatbot will reply in English. Using the tabulate library, the chatbot constructs the final response, comprising the translated (or English) response message and the names of the ”jobs” in a table format. This response is delivered to the user. To signify that the action has been done, the function returns an empty list []. Figure 24: Test job status. Another function that is unquestionably important to explain is the action that deals with determining the minimum batch execution time for a specific date because it praises the chatbot’s error handling, as well as its ability to help and assist the user if the user, for whatever reason, has difficulty typing what he wants. The verbal action is represented in the image; in addition to the code being discussed in depth, it will also be exhibited in execution. 70 Algorithm 4 Critical Path Execution Time 1: class CriticalPathExecutionTime(Action): 2: function name(self) return ”action_criticalpath_executiontime” 3: end function 4: procedure run(self, dispatcher, tracker, domain) 5: language ←tracker.get_slot(‘language’) 6: slot_value ←tracker.get_slot(‘date’) 7: if not slot_value or not validate_date_format(slot_value) then 8: dispatcher.utter_message(‘I’m sorry, I didn’t understand or please mention the date in the format...’) 9: return [] 10: end if 11: formatted_date ←convert_date_format(slot_value) 12: results, error ←run_query_IBPS( SQL_query_with_formatted_date ) 13: if error then 14: dispatcher.utter_message(error) 15: else 16: if results then 17: table ←tabulate(results, headers=…, tablefmt=‘fancy_grid’) 18: english_response ←‘For the critical path on the date...’ 19: translated ←GoogleTranslator(source=’en’, target=language).translate(english_response) 20: dispatcher.utter_message(translated) 21: else 22: dispatcher.utter_message(‘No results found for the specified date.’) 23: end if 24: end if 25: return [] 26: end procedure 71 Below we have the test against the function represented and explained above, and as we can see the chatbot interpreted the question correctly and answered by providing the time it takes to execute the minimum execution time Figure 25: Test critical path minimal execution time date. Finally, there is an action that allows the user to skip one or more jobs of his choice, with the condition that these jobs be in the submitted or error states; if the user requests to skip a job that is not in one of these two states, an informative message will be displayed. If a job is in one of these states, the user will see a message indicating that the skip to the job was successful. The developed code can be seen in the image below. 72 Algorithm 5 Skip Job Process 1: class skipJob(Action): 2: function name(self) 3: return ”action_skip_job” 4: end function 5: function run(self, dispatcher, tracker, domain) 6: language ←tracker.get_slot(”Language”) 7: jobName ←tracker.get_slot(”jobName”) 8: results, error ←run_query_IPE(SQL query on process instances) 9: if results is empty then 10: english_response ←”The mentioned process ID is neither submitted...” 11: else 12: if error exists then 13: dispatcher.utter_message(error) 14: return 15: end if 16: Execute healing command with parameters 17: english_response ←”The job {jobName} was skipped with success” 18: end if 19: if language == ”Portuguese” then 20: translated_responses ←Translate(english_response, to=’pt’) 21: else if language == ”Spanish” then 22: translated_responses ←Translate(english_response, to=’es’) 23: else 24: translated_responses ←english_response 25: end if 26: dispatcher.utter_message(translated_responses) 27: return 28: end function 73 Figure 35: Entity prediction confidence distribution. 80 Chapter 7 Conclusions and Future Work 7.1 Conclusions Before getting into the specifics of the chatbot’s creation and deployment, it is critical to address the constraints encountered both before and during these operations. These constraints were critical in influencing the project’s trajectory and outcomes. During the chatbot’s development, it was decided to display the values requested by the user in a list for ease of viewing. Let’s suppose the same thing, asking which jobs are active. In this case, some values of these identical jobs will be supplied, but it will also indicate the number of active jobs that exist in the database at the moment. As a result, the user will obtain more information about what he requested, eliminating the need for him to query how many positions are active in the future. To avoid overloading the user interface and create a more gratifying experience, the list was limited to 10 values. This decision was made to avoid presenting the user with a lengthy list of up to 300 values, which could make the interaction onerous and useless. It is critical to realize that this option can limit the viewing of all possible values. However, this decision was taken to achieve a compromise between delivering vital information to the user and preserving a clean and easy user interface. The Rasa chatbot includes a collection of functions that make it an effective and engaging virtual assistant for users. The capacity to detect and reply in different languages, switch languages during a conversation, engage users through natural language processing, answer queries, batch process associated information, and do simple activities such as automated answers and corrections are among the key characteristics. The chatbot was programmed to recognize and respond to users who spoke Portuguese, Spanish, or English. To do this, a Spacy model was utilized to recognize the language included in the user’s message. When the language is detected, the chatbot instructs Deep Translator to translate the text into the chatbot’s 81 primary language. This allows the chatbot to understand and answer effectively regardless of the user’s language. The chatbot’s capacity to switch languages throughout a discussion is an intriguing feature. If the user chooses to continue the conversation in a different language, the chatbot rapidly adapts to react in the new language. Users who are multilingual or who want to switch languages to better express themselves will have a more natural and pleasant experience as a result of this. The bot is intended to mimic the behavior of human assistants, resulting in an engaging conversational experience. To accomplish this, separate lexicons and vocabularies in Portuguese, English, and Spanish were created for each chatbot objective. These lexicons are a set of types of things that users can say to the chatbot during encounters. Natural language processing techniques are used by the system to understand and interpret users’ intentions based on these categories and to provide relevant responses. The conversational system can intelligently respond to user questions and provide batch processing details. The system has access to a large database including important information from the Intelligence Batch Processing System (IBPS) and the Intelligence Processing Engine (IPE) for this purpose. The chatbot employs queries to retrieve specific data from the database, delivering the needed information to the user quickly and accurately. The chatbot may execute simple tasks such as automatic responses and corrections in addition to answering queries and supplying information in bulk. For example, if a user wishes to skip a specific job, they can use the appropriate syntax to submit a request, and the chatbot will conduct the ”skip” action on the desired job. So, the Rasa-powered chatbot is a solution for multilingual interactions with users. It can adapt to different languages during a conversation, provide an engaging virtual assistant experience through natural language processing, answer users’ questions provide relevant information based on trustworthy databases, and perform simple tasks to streamline user interaction. Because of this mix of qualities, the chatbot is a useful tool for a variety of applications, such as customer service, technical assistance, and others. All of the objectives listed above were presented. However, another main aim of the dissertation was to improve the experience of users who engage with the generated product by acting as a liaison between batch processing and the individual or group of individuals who use it daily. I attempted to provide the chatbot’s users with a natural and intuitive interaction by combining the aforementioned elements. I expect that this technology will tremendously benefit Retail Consult and help streamline everyday activities and communication. 82 7.2 Prospect for future work This dissertation describes how to use the Rasa platform to create a chatbot that can communicate in Portuguese, English, and Spanish. However, in order to widen the scope and usage of the chatbot, we recommend the following subjects for further research and implementation: One of the most major improvements would be to add Mandarin, German, and French to the chatbot’s linguistic skills. In addition to serving Retail Consult ’s Chinese clients, adding German and French support would allow for a more direct and effective engagement with customers in German and French-speaking nations, increasing the chatbot’s reach and making it into a truly global tool. To improve the chatbot’s language capability, the lexicon and terminology employed should be expanded. By adding more terms and synonyms to the chatbot’s knowledge base, which is contained in the natural language understanding (NLU) file, it will be able to understand and respond to a broader range of expressions and questions asked, making the chatbot more powerful and reliable tool for user interactions. The chatbot currently offers a capability that allows the user to bypass specific tasks contained in the Intelligent Batch Processing System (IBPS) database. However, it would be fascinating to provide the option to start or stop certain jobs, which would be conducted by HTTP requests, as was the skip, allowing users to further control batch processing activities. To facilitate the scalability and deployment of the chatbot in diverse contexts, all produced code should be packaged in Docker containers. It is suggested that three containers be used: one to run the Rasa actions server and install the required dependencies, another to run the React application through npm, and a third to run Rasa with the API enabled. This strategy will result in a simpler and more efficient chatbot deployment, mostly by avoiding dependence issues on other machines. To improve the user experience, more actions that implement various inquiries and capabilities related to the chatbot environment can be added, offering more accurate and full replies to a wide range of questions and requests. The incorporation of pagination functionality is a tempting area for advancement. To avoid undue strain on the user interface, the chatbot currently displays a short list of ten values. A smart improvement would be to include a ”Show More” button, allowing consumers to access other options. Users can properly assess the values by wisely using paging, avoiding the unnecessary imposition of an exhaustive list all at once. This technique fosters a greater sense of flexibility and personalization, effectively catering to a wide range of user preferences and needs. Furthermore, if we want to improve the system, one of the aspects that we will take into account 83 will be the user’s expectations from the system and interactions and the quality of the chatbot in terms of appropriateness of language level, continuity of conversation, and success in task performance de Trabalho [2022]. By adopting these future works, the Rasa chatbot will become an even more valuable tool for Retail Consult and its users, improving the experience by increasing interaction efficiency and expanding its worldwide reach. Each of these enhancements will add to a more smart, adaptable, and valuable chatbot that will satisfy the demands of an ever-changing market. 84 Bibliography Eleni Adamopoulou and Lefteris Moussiades. An overview of chatbot technology. In IFIP International Conference on Artificial Intelligence Applications and Innovations , pages 373–383. Springer, 2020. Helena Isabel Oliveira Alves. Estudo e implementação de modelos de gestão de stocks em sistemas ERP para empresas retalhistas . PhD thesis, 2015. Mattia Atzeni and Maurizio Atzori. Askco: A multi-language and extensible smart virtual assistant. In 2019 IEEE Second International Conference on Artificial Intelligence and Knowledge Engineering (AIKE) , pages 111–112. IEEE, 2019. Nicolas Audebert, Catherine Herold, Kuider Slimani, and Cédric Vidal. Multimodal deep networks for text and image-based document classification. In Machine Learning and Knowledge Discovery in Databases: International Workshops of ECML PKDD 2019, Würzburg, Germany, September 16–20, 2019, Proceedings, Part I , pages 427–443. Springer, 2020. Tom Bocklisch, Joey Faulkner, Nick Pawlowski, and Alan Nichol. Rasa: Open source language understanding and dialogue management. arXiv preprint arXiv:1712.05181 , 2017. Giovanni Campagna, Rakesh Ramesh, Silei Xu, Michael Fischer, and Monica S Lam. Almond: The architecture of an open, crowdsourced, privacy-preserving, programmable virtual assistant. In Proceedings of the 26th International Conference on World Wide Web , pages 341–350, 2017. Davide Cucurnia, Nikolai Rozanov, Irene Sucameli, Augusto Ciuffoletti, and Maria Simi. Matilda-multiannotator multi-language interactivelight-weight dialogue annotator. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: System Demonstrations , pages 32–39, 2021. Plano de Trabalho. Calendarização e objetivos plano de trabalho. https://dcmei.di.uminho. pt/web#id=588&model=dissertation_admission.dissertation&view_type=form& cids=1&menu_id=100, 2022. 85 Equipe editorial de Conceito.de. Retalhista - o que é, conceito e definição. https://conceito.de/ retalhista, 2020. accessed 10.01.2023. Hugo Faria, Maria Araújo Barbosa, Bruno Veloso, Francisco S Marcondes, Celso Lima, Dalila Durães, and Paulo Novais. Edubot: A proof-of-concept for a high school motivational agent. In International Conference on Intelligent Data Engineering and Automated Learning , pages 223–232. Springer, 2022. Felipe Ramos Feitoza et al. Cecílio: um chatbot para automação do atendimento aos usuários em instituições federais de ensino superior. 2021. Matthew B Hoy. Alexa, siri, cortana, and more: an introduction to voice assistants. Medical reference services quarterly , 37(1):81–88, 2018. Marie-Claire Jenkins, Richard Churchill, Stephen Cox, and Dan Smith. Analysis of user interaction with service oriented chatbot systems. In International conference on human-computer interaction , pages 76–83. Springer, 2007. Veton Kepuska and Gamal Bohouta. Next-generation of virtual personal assistants (microsoft cortana, apple siri, amazon alexa and google home). In 2018 IEEE 8th annual computing and communication workshop and conference (CCWC) , pages 99–103. IEEE, 2018. Neiva Larisane Kuyven, Carlos André Antunes, Vinicius João de Barros Vanzin, João Luis Tavares da Silva, Aliane Loureiro Krassmann, and Liane Margarida Rockenbach Tarouco. Chatbots na educação: uma revisão sistemática da literatura. RENOTE , 16(1), 2018. Francisca Leitão Gonçalves do Vale Lima. Big data warehousing em tempo real: da recolha ao processamento de dados, 2017. Trang Nguyen Thi Mai and Shcherbakov Maxim. Enhancing rasa nlu model for vietnamese chatbot. International Journal of Open Information Technologies , 9(1):31–36, 2021. Martech. Pros and cons future facebook chatbots. https://martech.org/ pros-cons-future-facebook-chatbots/, 2022. accessed: 02.11.2022. Medium. Advantages and disadvantges of siri. https://macbook.medium.com/ the-advantages-and-disadvantages-of-siri-1cc83604a08f, 2022. accessed: 07.12.2022. 86 Microsoft. What can you do with cortana in windows. https://support.microsoft.com/en-us/ topic/what-can-you-do-with-cortana-in-windows-f57ef46f-716a-9940-a8fc-09b3433d05ea, 2022. accessed: 26.10.2022. MyAyan. Advantages and disadvantages of amazon alexa. https://www.myayan.com/ advantages-and-disadvantages-of-amazon-alexa, 2022. accessed: 25.10.2022. Eduardo Nacimiento-García, Carina S González-González, and Francisco L Gutiérrez-Vela. Free software virtual assistants for designing pervasive gaming experiences to promote active aging: State of the art. CEUR-WS. org , 2747, 2020. S Nithuna and CA Laseena. Review on implementation techniques of chatbot. In 2020 International Conference on Communication and Signal Processing (ICCSP) , pages 0157–0161. IEEE, 2020. Oracle. Batch schedule. https://docs.oracle.com/cd/E79623_01/rms/pdf/16021/html/ batch_schedule/merch-batch.htm, 2017. accessed: 17.12.2022. Python Organization. Deep translator from google translator. https://pypi.org/project/ deep-translator/, 2023. accessed: 15.03.2023. PandaSecurity. Cortana security flaw. https://www.pandasecurity.com/en/mediacenter/ mobile-news/cortana-security-flaw/, 2022. accessed: 25.10.2022. PeopleBank. The pros and cons of google home. https://www.peoplebank.com.au/blog/ 2018/04/the-pros-and-cons-of-google-home?source=google.com, 2018. accessed: 05.12.2022. Bhavika R Ranoliya, Nidhi Raghuwanshi, and Sanjay Singh. Chatbot for university related faqs. In 2017 International Conference on Advances in Computing, Communications and Informatics (ICACCI) , pages 1525–1530. IEEE, 2017. RASA. Conversational ai with rasa: Introduction to rasa. https://www.youtube.com/watch?v= Ap62n_YAVZ8&t=3s, 2021. accessed: 07.3.2023. RASA. Conversational ai with rasa: Introduction to rasa. https://rasa.com/, 2023a. accessed: 07.03.2023. RASA. Rasa installations procedures. https://www.youtube.com/watch?v=RVoFqxmG8p0, 2023b. accessed: 12.07.2023. 87 RetailConsult. Oracle retail consult. https://www.retail-consult.com, 2023. accessed: 16.01.2023. Reuters. Us apple siri. https://www.reuters.com/article/ us-apple-siri-idUSKBN16G0H3, 2022. accessed: 18.11.2022. Sakshi Sachdeva. Scrum methodology. Int. J. Eng. Comput. Sci , 5(16792):16792–16800, 2016. Andressa Vieira Silva. Um modelo de classificação para o Reconhecimento de Entidades Nomeadas . PhD thesis, Universidade de São Paulo, 2020. Abhishek Singh, Karthik Ramasubramanian, and Shrey Shivam. Introduction to microsoft bot, rasa, and google dialogflow. In Building an Enterprise Chatbot , pages 281–302. Springer, 2019. Pavel Smutny and Petra Schreiberova. Chatbots for learning: A review of educational chatbots for the facebook messenger. Computers & Education , 151:103862, 2020. Marcin Sowański and Artur Janicki. Leyzer: A dataset for multilingual virtual assistants. In International Conference on Text, Speech, and Dialogue , pages 477–486. Springer, 2020. SpaCy. Industrial-strength natural language processing. https://spacy.io/, 2023. accessed: 22.04.2023. Spring. Spring batch documentation. https://docs.spring.io/spring-batch/docs/ current/reference/html/index-single.html, 2023. accessed: 24.03.2023. Techtarget. Named entity recognition - ner. https://www.techtarget.com/whatis/ definition/named-entity-recognition-NER, 2023. accessed: 02.03.2023. U THIRUPALU and M SURESH. Virtual personal assistant using artificial intelligence. Journal of Engineering Sciences , 14(08), 2023. Wikipedia. Batch processing. https://en.wikipedia.org/wiki/Batch_processing, 2023. accessed: 13.11.2022. 88 Chapter 8 Listings 8.1 Attached 1 Listing 8.1: Classification intents report { "skipJob": { "precision": 1.0, "recall": 1.0, "f1-score": 1.0, "support": 18, "confused_with": {} }, "criticalPathDate": { "precision": 1.0, "recall": 1.0, "f1-score": 1.0, "support": 7, "confused_with": {} }, "jobstatus": { "precision": 1.0, "recall": 1.0, "f1-score": 1.0, "support": 15, "confused_with": {} }, "website": { "precision": 1.0, "recall": 1.0, "f1-score": 1.0, "support": 9, "confused_with": {} }, "mood_unhappy": { "precision": 1.0, "recall": 1.0, "f1-score": 1.0, "support": 21, 89