Enhancing Social Robots' Understanding: Predicate Grounding with Large Language Models

Enhancing Social Robots’ Understanding: Predicate Grounding with Large Language Models Giulia Berettieri1, Anna Allegra Bixio1, Lucrezia Grassi1, Carmine Tommaso Recchiuto1, and Antonio Sgorbissa1 Abstract—This work investigates whether Large Language Models (LLMs) can support predicate grounding in Child–Robot Interaction (CRI). We introduce a synthetic dataset of 1,800 dialogues covering explicit and implicit mental states, derived from the Applied Behavior Analysis (ABA) framework. The dataset was validated by domain experts and used to evaluate GPT-4o’s ability to classify children’s mental states from dialogue contexts. Results highlight systematic differences between explicit and implicit classifications and reveal the impact of interaction length on accuracy. These findings establish a foundation for applying LLM-based inference to real child–educator dialogues, advancing the development of socially aware robotic systems. Index Terms—Child–Robot Interaction, Predicate Grounding, Large Language Models, Theory of Mind I. INTRODUCTION Socially Assistive Robots (SARs) operating in dynamic environments are expected to continuously interpret and respond to diverse situations. To interact effectively with humans, these robots require cognitive profiling and adaptive capabilities. Such skills entail the recognition of a wide range of mental states (beliefs, desires, intentions, imagination, emotions), a capability known as Theory of Mind (ToM) [1]. In addition, robots should be able to adapt their behavior based on the interpretation and prediction of these mental states [2]. One approach to addressing this challenge is the development of software architectures that enable verbal interaction, as proposed in our previous work [3]. This approach draws inspiration from Applied Behavior Analysis (ABA), particularly the principles of functional analysis. ABA is an evidence-based framework designed to improve socially significant behaviors by identifying the environmental variables responsible for behavioral change [4]. A central objective is to determine the functions or purposes of behaviors considered “challenging” and to apply appropriate strategies for managing them. This involves assessing stimuli occurring before the behavior (antecedents) and after the behavior (consequences). ABA techniques are already widely implemented in educational settings due to their proven effectiveness in behavior management [5], [6]. Integrating these strategies into social robotics has the potential to greatly enhance the quality of child–robot interactions (CRI) in learning environments. 1All authors are with the RICE Lab, DIBRIS Department, University of Genoa, Via All’Opera Pia 13, 16145 Genoa, Italy. Corresponding author: [email protected] This work has been supported by the Italian Ministry of University and Research (MUR) under the PNC (Piano Nazionale Complementare) “Fit4MedRob – Fit for Medical Robotics” (PNC0000007), Mission 3. This paper investigates whether informed assumptions about mental states can be derived through dialogue by leveraging Large Language Models (LLMs). The main contributions are: •The creation of a dataset comprising 1,800 synthetic child–robot interactions across different scenarios. •An evaluation of GPT-4o’s ability to accurately classify mental states based on dialogues. II. METHOD A. Predicates and Dataset The mental states (hereafter referred to as “predicates”) considered in this study are derived from the ABA framework introduced above. We focused on states likely to emerge during playful activities, particularly those related to two behavioral functions: •Gain a Tangible: e.g., when the educator asks the child to stop an activity because time is running out, but the child wants to continue. •Escape: e.g., when the child wishes to stop the ongoing activity for some reason. Table I lists each predicate chosen and explains how we interpreted them in the analysis. Predicate Description hard The user finds the task too challenging, difficult, or frustrating. easy The user finds the task not sufficiently challenging, engaging, or motivating. bored The user perceives the task as repetitive, or the task is not interesting or engaging enough. tired The user feels physically or mentally tired or both. hungry The user would like to eat something, or they are thinking about food. succeed The user exhibits a competitive attitude while performing the task. fussy The user has a perfectionist attitude about everything. curious The user is interested in the task outcome or in exploring different things or activities. uncomfortable The user feels uncomfortable. The discomfort could be caused by the environment, the presence of the educator, or an inner feeling of pain or unease. TABLE I: Child’s mental states and their descriptions. Based on these predicates, we generated four different datasets of implicit and explicit CRIs. Here, explicit indicates that the predicate is explicitly cited, while implicit refers to 2025 I-RIM Conference October 17-19, Rome, Italy ISBN: 9788894580570 10.5281/zenodo.17629740 113 interactions where the predicate is not directly mentioned but rather implied. The datasets are as follows: •D1: 50 explicit interactions per predicate generated leveraging GPT-4o model. •D2: 50 implicit interactions per predicate derived from D1 leveraging GPT-4o model. •D3: Italian translation of D2 leveraging GPT-4o model. •D4: 50 implicit interactions per predicate generated using the DeepSeek online tool. Each interaction consists of eight sentences in a child–robot dialogue (four per speaker), covering various play scenarios. To simulate realistic chatbot settings, we extracted sliding windows of 2, 4, or 6 utterances from each dialogue. This approach both evaluates contextual inference from partial conversations and increases the dataset size from 1,800 to 18,000 samples. The complete dataset is available here. B. Expert Evaluation To ensure dataset reliability and quality, two experts – one in education and one in social robotics – manually reviewed all interactions in D1 and D2. A structured questionnaire, implemented via Google Forms, guided the evaluation. The process followed a qualitative approach to intercoder reliability, consistent with the guidelines in [7]. The evaluation proceeded in two stages: (1) the experts independently completed the questionnaire for each interaction, and (2) they compared classifications and resolved discrepancies through discussion. As a result, each dialogue received a label indicating which predicate it contained. C. Classification Classification accuracy is computed separately for each dataset. For each predicate, we considered 50 interactions representing the target state (i.e., the interactions created for that predicate) along with an equal number of counterexamples in which the predicate was absent. The selection of counterexamples varied across datasets: •D1 and D2: randomly selected from dialogues without the target predicate. •D3: as a translation of D2, it inherits the same labels and counterexample selection strategy. •D4: selected using a correlation matrix (computed from D1 and D2) to avoid predicates with high co-occurrence. Accuracy was computed for each predicate four times, once for each possible utterance window. Each sample was independently submitted to the OpenAI GPT-4o model via API for classification. The binary classification prompt was structured as follows: "You are provided with an interaction about {self.context}. Do you think that {self.gt}? Answer only with ’yes’ or ’no’." Because of their opposite nature, hard and easy were treated with a ternary classification scheme rather than binary. The prompt was structured as follows: "You are provided with an interaction about {self.context}. Among these mental states, which one better suits the interaction: {self.gt[0]}, {self.gt[1]}. Answer only with ’hard’, ’easy’, or ’None’." Here, self.context refers to the dialogue context, and self.gt specifies the predicate(s). III. RESULTS Building on the expert evaluation, which provided manual labels for the explicit and implicit English dialogues in two phases, we first computed the correlation matrix among predicates to analyze their co-occurrence patterns. The primary evaluation metric is classification accuracy, derived from confusion matrices and calculated for each predicate across different utterance windows and datasets. For brevity, not all the results are reported. The mean accuracy across the 9 predicates ranges from 0.90 to 0.93 for D1, depending on window length; 0.79 to 0.86 for D2; 0.77 to 0.86 for D3; and 0.86 to 0.93 for D4. IV. CONCLUSION This paper presented a methodology for grounding children’s mental states in CRI through LLMs. We introduced a synthetic dataset of 18,000 samples, validated its quality through expert review, and evaluated GPT-4o’s ability to classify explicit and implicit mental states. Although detailed results could not be reported here, our findings indicate clear trends regarding interaction length and the distinction between explicit and implicit states. We observed consistent accuracy across different LLMs used to generate samples, as well as unexpectedly strong performance on Italian translations. These results provide a foundation for extending the approach to real-world child–educator dialogues, paving the way for more adaptive and socially aware robotic systems. REFERENCES [1] D. Premack and G. Woodruff, “Does the chimpanzee have a theory of mind?” Behavioral and brain sciences, vol. 1, no. 4, pp. 515–526, 1978. [2] S. Rossi, F. Ferland, and A. Tapus, “User profiling and behavioral adaptation for HRI: A survey,” Pattern Recognition Letters, vol. 99, pp. 3–12, 2017. [3] G. Berettieri, A. A. Bixio, L. Grassi, C. T. Recchiuto, and A. Sgorbissa, “A novel approach for behavior management and real-time adaptation during child-robot interaction,” in Proceedings of the Workshop on Advanced AI Methods and Interfaces for Human-Centered Assistive and Rehabilitation Robotics (a Fit4MedRob event) - AIxIA 2024, ser. CEUR Workshop Proceedings, vol. Vol-3932, 2024, cEUR-WS, ISSN 1613-0073. [Online]. Available: https://ceur-ws.org/Vol-3932/paper7.pdf [4] J. O. Cooper, T. E. Heron, W. L. Heward, et al.,Applied behavior analysis. Pearson/Merrill-Prentice Hall Upper Saddle River, NJ, 2007. [5] B. S. Parsonson, “Evidence-based classroom behaviour management strategies.” Kairaranga, vol. 13, no. 1, pp. 16–23, 2012. [6] K. B. Geiger, J. E. Carr, and L. A. LeBlanc, “Function-based treatments for escape-maintained problem behavior: A treatment-selection model for practicing behavior analysts,” Behavior Analysis in Practice, vol. 3, pp. 22–32, 2010. [7] N. Cofie, H. Braund, and N. Dalgarno, “Eight ways to get a grip on intercoder reliability using qualitative-based measures,” Canadian medical education journal, vol. 13, no. 2, pp. 73–76, 2022. 114