HeLP: The Hebrew Lexicon project

Vol.:(0123456789) Behavior Research Methods (2024) 56:8761–8783 https://doi.org/10.3758/s13428-024-02502-4 ORIGINAL MANUSCRIPT HeLP: The Hebrew Lexicon project RoniStein1· RamFrost1,2· NoamSiegelman1 Accepted: 17 July 2024 / Published online: 9 September 2024 © The Author(s) 2024 Abstract Lexicon projects (LPs) are large-scale data resources in different languages that present behavioral results from visual word recognition tasks. Analyses using LP data in multiple languages provide evidence regarding cross-linguistic differences as well as similarities in visual word recognition. Here we present the first LP in a Semitic language—the Hebrew Lexicon Project (HeLP). HeLP assembled lexical decision (LD) responses to 10,000 Hebrew words and nonwords, and naming responses to a subset of 5000 Hebrew words. We used the large-scale HeLP data to estimate the impact of general predictors (lexicality, frequency, word length, orthographic neighborhood density), and Hebrew-specific predictors (Semitic structure, presence of clitics, phonological entropy) of visual word recognition performance. Our results revealed the typical effects of lexicality and frequency obtained in many languages, but more complex impact of word length and neighborhood density. Considering Hebrew-specific characteristics, HeLP data revealed better recognition of words with a Semitic structure than words that do not conform to it, and a drop in performance for words comprising clitics. These effects varied, however, across LD and naming tasks. Lastly, a significant inhibitory effect of phonological ambiguity was found in both naming and LD. The implications of these findings for understanding reading in a Semitic language are discussed. Keyword Visual word recognition; Mega studies; Reading; Cross-linguistic differences Introduction The ability to rapidly identify printed sequences of letters as individual words and automatically access their phonological and semantic representations has intrigued scientists for decades, and is still the focus of extensive research in cognitive science. Tasks that measure participants reaction times (RTs) and recognition accuracy of printed letter sequences have provided important insights regarding the computations underlying visual word recognition. Perhaps the most common experimental paradigm in such studies is the lexical decision (LD) task, where participants are presented with letter strings, one at a time, and are required to provide fast responses as to whether or not they represent existing words. Another common task is the naming task, in which participants are required to pronounce, as fast and as accurately as possible, a visually presented word. RT and accuracy data using the LD and naming tasks were taken to reveal the underlying computations in recognizing printed words presented in isolation, retrieving their phonological structure, and accessing their semantic representation. Across many studies, they demonstrated highly replicable effects which today are the landmarks of visual word recognition: As a non-exhaustive list, words are recognized faster than nonwords (i.e., the lexicality effect, e.g., Forster & Chambers, 1973; Monsell etal., 1989); frequent words are recognized faster than infrequent words (i.e., the word frequency effect, e.g., Broadbent, 1967; Brysbaert etal., 2017); shorter words are processed faster than longer words (i.e., the word length effect, e.g., Fredriksen & Kroll, 1976; Hudson & Bergman, 1985); and the number and frequency of orthographic neighbors affect decision times (i.e., the neighborhood density effect; e.g., Andrews, 1992; Grainger etal., 1989). Lexicon projects indifferent languages In the first decades of visual word recognition studies, researchers typically employed experimental designs in which verbal stimuli were selected to represent factors of * Noam Siegelman [email protected] 1 Department ofPsychology, The Hebrew University ofJerusalem, Mount Scopus Campus, 9190501Jerusalem, Israel 2 BCBL, Basque Center ofCognition, Brain andLanguage, SanSebastian, Spain 8762 Behavior Research Methods (2024) 56:8761–8783 interest (e.g., imageability, concreteness, morphological complexity, homography, homophony), measuring performance for these stimuli. Generally, these studies used a relatively small set of stimuli, often not representative of the variety found across the full lexicon, thus limiting their external validity. One prominent example is the focus on monosyllabic words in English in visual word recognition studies, despite the fact that in many languages they represent a small portion of the words in the language (e.g., less than 15% of words, Ferrand etal., 2010). An elegant solution to these limitations was the English Lexicon Project (ELP, Balota etal., 2007), which presented an open large-scale data resource that included over 40,000 words along with their respective behavioral data in both the LD and the naming tasks. The ELP made it possible, for the first time, to test a range of hypotheses regarding word recognition computations without the need to construct a targeted experiment with its inevitable limitations. Instead, researchers could generate hypotheses regarding the impact of any variable, and simply extract the behavioral data for all relevant stimuli from the database (a process sometimes referred to as “virtual experiments”, see, e.g., Kuperman, 2015). The ELP has been used extensively since its publication in order to explore the underlying computations of word recognition in English, and to date has been cited over 3000 times. It has been useful in validating and testing the impact of a range of psycholinguistic factors on reading, such as words’ semantic transparency (Kim etal., 2018), word length (e.g., New etal., 2006), imageability (e.g., Dymarska etal., 2023); orthographic–phonological regularities (e.g., Chee etal., 2020; Siegelman etal., 2020), and orthographic–semantic consistency (Siegelman etal., 2022). Importantly, the ELP has also inspired the creation of parallel lexicon projects (LPs) in other languages, to provide a critical cross-linguistic perspective in reading research. These LPs include British English, to be distinguished from American English (Keuleers etal., 2012); Dutch (Keuleers etal., 2010); French (Ferrand etal., 2010); Spanish (Aguasvivas etal., 2018); Persian (Nemati etal., 2022); Malay (Yap etal., 2010); German (Schreuter & Schroeder, 2017); Portuguese (Soares etal., 2019); and Chinese (Tse etal., 2017). Together, the wide scale of cross-linguistic data offered by the different LPs have provided important insights into how basic word recognition processes vary across languages, and how the properties of a given writing system shape the cognitive computations involved in reading one language compared to another. From this perspective, the specific characteristics of a writing system can be taken as an “experimental manipulation” to examine their impact on reading performance (see Frost, 2012, for discussion). But note that whereas all LPs have the same aim, methodologically there is substantial variability in how they were constructed. For example, while all available LPs provide LD data, not all include naming data. Further, although all LPs do provide a broader coverage of a language’s words than typical experiments, there is still substantial variability in the number of words they employed (e.g., from 1800 words and nonwords in Persian, Nemati etal., 2022; to 40,481 words and nonwords in English, Balota etal., 2007). LPs also differ in the number of participants providing responses to each word in the project (e.g., from 25 responses per word in the French LP, Ferrand etal., 2010, to 300 in the Spanish LP, Aguasvivas etal., 2018), and in other design characteristics such as the word–nonword ratio in the LD task. Table1 summarizes the methodological properties of existing LPs, as well as the characteristics of the languages studied. In spite of the substantial methodological variability, large-scale LP data have nonetheless largely replicated multiple well-established visual word recognition effects across languages, demonstrating substantial similarities in computations. For example, LPs have consistently revealed frequency and lexicality effects, as well as orthographic neighborhood density effects. Importantly, however, the multilingual comparisons that LPs inspired have also demonstrated significant differences in reading behavior across languages, providing illuminating insights into how the properties of a language and its writing system shape the computations involved in processing printed information. For example, LD data from the Persian LP showed no word length effect, in contrast to other LPs. This has led authors to assume that the specific properties of the Persian orthography—specifically, the under-specification of vowels in the printed language or the strong correlation between word length and orthographic neighborhood size—are the reason for the absence of the word length effect (Nemati etal., 2022). In stark contrast to Persian, the Malay LP revealed that the length of words in Malay is the strongest predictor of word recognition time. This again was tied to the structure of the writing system: the Malay language has a shallow orthography, transparent morphology, and a simple syllabic structure, which presumably lead Malay readers to rely primarily on serial conversion of letters to phonemes during word recognition. Indeed, further analysis of the Malay LP revealed great differences between Malay and English in the contribution of predictors such as frequency, orthographic neighbors, and word length to word recognition performance (Yap etal., 2010). The cross-linguistic variations revealed by large-scale data highlight how the basic properties of a writing system impact word recognition behavior across languages, leading to a deep understanding of the universal principles involved in orthographic processing across the world’s languages. 8763Behavior Research Methods (2024) 56:8761–8783 Table 1 A summary of large-scale Lexicon Project in different languages, and the properties of those languages Linguistic characteristics Lexicon project details Lexicon Project Topological family (branch) Script (script type) Morphological typology Orthographic transparency Number of words LD/naming task Average observations per word Comments Chinese, Mandarin (simplified script) (Tsang etal., 2018) Sino-Tibetan (Sinitic) Chinese (logographic) Analytic Opaque 12,578 LD 42 See also Sze etal., 2014 Chinese, Cantonese (traditional script) (Tse etal., 2017) Sino-Tibetan (Sinitic) Chinese (logographic) Analytic Opaque 25,286 LD 33 All stimuli were twocharacter compound words Dutch (Keuleers etal., 2010) Indo-European (West Germanic) Latin (alphabetic) Synthetic, fusional Moderate 14,089 LD 39 Monosyllabic and disyllabic words English (American) (Balota etal., 2007) Indo-European (West Germanic) Latin (alphabetic) Moderately analytic Opaque 40,481 LD & naming LD34, Naming25 English (British) (Keuleers etal., 2012) Indo-European (West Germanic) Latin (alphabetic) Moderately analytic Opaque 14,365 LD 39 Monosyllabic and disyllabic words French (Ferrand etal., 2010) Indo-European (Romance) Latin (alphabetic) Moderately analytic Moderate 38,840 LD 25 German (Schröter & Schroeder, 2017) Indo-European (West Germanic) Latin (alphabetic) Synthetic, fusional Moderate 1152 LD & naming LD18, Naming20 Developmental project: Participants from 1st graders to adults Malay (Yap etal., 2010) Austronesian (Malayo-Polynesian) Latin (alphabetic) Synthetic, agglutinative Transparent 9592 LD & naming 44 Participants got feedback if they were wrong Persian (Nemati etal., 2022) Indo-European (Western Iranian) Arabic Synthetic, fusional Abjad 1800 LD Words 3–8 letters long European Portuguese (Soares etal., 2019) Indo-European (Romance) Latin (alphabetic) Synthetic, fusional Opaque 1920 LD & naming 55 Spanish (Aguasvivas etal., 2018) Indo-European (Romance) Latin (alphabetic) Synthetic, fusional Transparent 44,853 LD 333 Word to nonwords ratio of 7:3 8764 Behavior Research Methods (2024) 56:8761–8783 Since languages naturally differ in their scripts and how their writing systems represent sound and meaning, a good theory of reading should be able to explicate how this variation impacts the processing of printed words. When a given language diverges from others in terms of readers’ performance, important evidence is furnished regarding universal principles of reading (see Frost, 2012, for discussion). Such research enterprise, however, requires large-scale data from many diverse writing systems. The main goal of the present mega-study is to contribute to this important research effort by providing, for the first time, systematic data from a Semitic language—Hebrew—a language that has often been shown to produce contrastive results relative to European languages. The Hebrew language: Animportant test case Hebrew is a Semitic language, as are Arabic, Amharic, and Maltese. Many words in Semitic languages are root-derived, so that their base is a root morpheme, usually consisting of three consonants, which conveys the core meaning of the word. Semitic words are constructed by intertwining root morphemes with word pattern morphemes—abstract phonological structures, consisting of vowels or of vowels and consonants, in which there are “open slots” for the root’s consonants to fit into. In general, word patterns provide vague morphosyntactic information. For example, in Hebrew, the root K.Š.R., which conveys the general notion of “tying”, and the word pattern /ti–o-et/, which is mostly used to denote feminine nouns, form the word /tikšoret/, meaning “communication”. Embedding the root K.Š.R. in the word pattern /-i-u-/ produces the word /kišur/, meaning “link”, etc. The root consonants can be dispersed within the word in many possible positions, and there is little a priori information regarding their location. Word patterns have a well-defined internal structure. Their onset comprises a restricted number of consonants (mainly /h/, /m/, /t/, /n/, /l/), and the order and identity of subsequent consonants and vowels is rigid. Because there are no a priori constraints regarding the location of root consonants in the word, the main clue regarding their identity is the well-defined phonological structure of the word pattern that allows the root consonants to stand out (Deutsch etal., 1998, 2021; Lador-Weizman & Deutsch, 2022). Overall, there are about 3000 roots in Hebrew, about 100 nominal word patterns, and seven verbal patterns. Another major characteristic of the Hebrew writing system is its extreme phonological under-specification. Hebrew print consists of 22 letters, of which five have a finite letter form. The letters represent mostly consonantal information, and most vowel information is not conveyed in print (see Shimron, 2006; Ravid, 2011, for review). Two letters—one for both /o/ and /u/, and one for /i/—may convey the vowel information; however, in certain contexts, these letters also convey the consonants /v/ and /y/, respectively. This results in heavy phonological decoding demands, since a substantial part of the phonological information is missing. The missing vowels lead to an extensive homography, with many printed letter strings having multiple pronunciations and meanings (e.g., the printed word “ספר”, “SFR”, most commonly read as /sefer/, has seven possible pronunciations, each with a different meaning, depending on different vowel configurations). But since the structure of spoken words is highly constrained by the relatively small number of Semitic word patterns, readers can converge on a given word quite easily during text reading, since the printed form typically determines the appropriate word pattern with relatively high reliability, and once a word pattern has been recognized, the full vowel information is available to the reader, even if it is not specified by the printed form (Deutsch etal., 2021; Frost, 2006). Hence, Hebrew print provides a perfect example of optimization of information, where substantial morphological (and therefore semantic) information is provided along with sufficient phonological cues using minimal orthographic symbols (see Frost, 2012).1 From the perspective of orthographic depth (Frost etal., 1987; Katz & frost, 1992; Schmalz etal., 2015), Hebrew is considered to have a very deep orthography, given multiple features related to an underrepresentation of phonology in print. The first is the missing vowel information discussed above. In contrast to the depth of the English writing system, which results mainly from phonological inconsistency of vowel letters (e.g., EA is pronounced differently in DEAR, HEAD, and STEAK), in Hebrew, the vowel information is generally not inconsistent but missing. This poses a challenge in measuring and defining the phonological uncertainty of Hebrew words (Frost, 1994, 1995). A second source of phonological uncertainty in the Hebrew orthography is a feedforward and feed-backward inconsistency of a few consonantal letters, where some can be mapped to different phonemes, and some phonemes can be represented by different letters. For example, the Hebrew letter “כ” can be pronounced as /x/ or /k/, ״ב״ can be pronounced as /v/ or /b/, and ‘פ’ can be pronounced as /f/, or /p/. Conversely, /t/ can be written with the letters "ת"or "ט", /s/ can be written as “ס” and “ש”, /k/ can be written as ‘כ’’ and “ק”, and /x/ can be written as “כ” and “ח”. Lastly, another important feature of Hebrew is that morpho-syntactic information that is conveyed in European 1 Hebrew also has a diacritical variant of the writing system, where vowels are depicted by points (appearing mostly under printed letters). These vowel marks are typically taught in the first grade, assisting teachers in developing decoding skills during reading acquisition, but starting from the end of the second grade, printed and written Hebrew does not normally include diacritical marks (see Share & Bar-On, 2018, for a detailed discussion). Apart from poetry and religious texts, adult reading material is always un-pointed. 8765Behavior Research Methods (2024) 56:8761–8783 languages by function words (e.g., “the”, “from”, “to”, “and”) is conveyed in Hebrew as single letters which are attached to the word (i.e., clitics; e.g., “the” = > ה, “from “ = > מ, “to “ = > ל, “and” = > ו). For example, the fourword English sequence “and from the house” is printed in Hebrew as one word, where the three letters conveying and/ from/the are attached to the base word “house” (“״ומהבית, printed as VMHBYT, read as /vemehabayit/, see Ravid, 2011). Clitics are abundant in Hebrew printed input, although their impact on visual word recognition is currently largely unknown. This is due to the tendency of previous studies in the word recognition literature (in Hebrew as in other languages) to focus on the processing of base words, rather than on more naturalistic stimuli which better reflect the distribution of printed words in a language. Visual word recognition inHebrew: Aseries ofdivergent findings Given the unique characteristics of the Hebrew writing system, it has been the focus of extensive research investigating how these impact reading. In fact, Hebrew has been used as a case study for investigating a variety of relevant domains, such as the predictors of eye movements during reading (e.g., Deutsch etal., 2003; Velan etal., 2013), the study of the developmental trajectory of literacy (e.g., Share, 1999; Share & Bar-on, 2018), the impact of different scripts on reading in a second language (e.g., Mor & Prior, 2020, 2021), and how the orthographic structure of Hebrew is reflected in various forms of dyslexia (e.g., Friedmann & Lukov, 2008; Friedmann & Rahamim, 2007). Given the focus on the current study, however, our review centers on how the properties of the Hebrew writing system impact visual word recognition processes. Generally, and in line with earlier studies in English and other languages described above, previous work has tended to focus on how specific features of the Hebrew orthography (e.g., homography, phonological ambiguity, morphological structure) lead to divergent patterns of visual word processing. Then, the impact (or lack thereof) of the studied feature was discussed within a broader framework, towards understanding the universal principles that drive word recognition processes across languages. Underlying this research agenda is the theoretical supposition that if language A (e.g., English) shows a given pattern of behavior, and language B (e.g., Hebrew) does not, this points to a higher-order principle that simultaneously explains both phenomena. Within this research enterprise, substantial work has focused on Hebrew’s extreme phonological under-specification and homography. Some studies suggested that lexical decisions in Hebrew are made prior to phonological disambiguation (Bentin & Frost, 1987), and reflect the computation of a phonological impoverished code (see Frost, 1998, for discussion). Using the naming task, Frost etal. (1987) showed that frequency and semantic priming effects are stronger in Hebrew than in languages with shallower orthographies, resembling the effects revealed in LD. These results were taken to indicate that readers of languages with deep orthographies such as Hebrew rely more heavily on lexical information than readers of languages with shallow orthographies (e.g., Finnish, Spanish, German, Dutch), in order to compute phonology from print. This early finding is in line with the claim that readers of different writing systems rely on different informational cues in light of their writing system’s structure (e.g., Hirshorn & Harris, 2022; Lallier & Carreiras, 2018; Rau etal., 2015; Seidenberg, 2011; Seymour etal., 2003). However, all visual word recognition studies in Hebrew involved only a few dozen words that were selected as stimuli in each of the experiments. Moreover, most of these studies focused on nouns, typically disyllabic, without clitics or inflections—a partial set of stimuli which do not represent the variety of words in the Hebrew language. In the same vein, extensive work has focused on how the morphological structure of Hebrew words affects visual word recognition (e.g., Deutsch etal., 1998; Feldman etal., 1995; Frost etal., 1997, 2000). Overall, these studies suggested that the root consonants are the core target of word recognition, and that lexical organization in Hebrew follows morphological principles. For example, Frost and colleagues (2005) showed that in contrast to English, French, or Spanish, full orthographic overlap between primes and targets in Hebrew results in very weak masked orthographic priming, interpreted as evidence that the lexical architecture of Hebrew probably does not align, store, or connect words by virtue of their full sequence of letters. Indeed, considering the overall body of research using masked priming in Semitic languages, reliable facilitation is consistently obtained whenever primes consist of the root letters, irrespective of what the other letters are (e.g., Frost etal., 1997, 2000; Velan etal., 2005; Perea etal., 2010, but see Perea etal., 2014, for significant form priming effects in Arabic). Another important finding concerns letter position flexibility. In Indo-European languages, disrupting the order of the letters within a word has little impact on readers’ ability to recognize and read it correctly (e.g., Duñabeitia etal., 2007; Perea & Carreiras, 2006a, 2006b, 2008; Perea & Lupker, 2003, 2004; Schoonbaert & Grainger, 2004). In contrast, Hebrew readers reveal extreme letter position rigidity, and reading is significantly impaired when words involve transposed letters (Velan & Frost, 2007, 2009, 2011; and see Friedmann & Gvion, 2001, 2005, for letter position dyslexia of Hebrew readers). This cross-linguistic difference in letter position flexibility again reflects the morphological structure of Hebrew: since many Hebrew roots share a subset of letters but differ in their order (e.g., Z.M.R., “to 8766 Behavior Research Methods (2024) 56:8761–8783 sing”; R.M.Z., “to hint”; Z.R.M., “to flow”), letter position in Hebrew must be rigid rather than flexible in order to access the correct root (see Lerner etal., 2014, for computational evidence, and the recent PONG model, Snell, in press). However, to complicate things ever further, not all Hebrew words have a Semitic structure. Many words from various origins (e.g., Greek, Persian, English) have permeated Hebrew throughout history, and are not root-derived (and see a similar state of affairs for Maltese, e.g., Geary & Ussishkin, 2018). Indeed, these words have been shown to be processed differently from Semitic Hebrew words (Bitan etal., 2020; Haddad etal., 2018; Velan & Frost, 2011; Velan etal., 2013). However, the distributional properties of such non-Semitic Hebrew words are not yet known, and will be examined in this work. The current study: The Hebrew Lexicon Project Since the Hebrew writing system represents a stark contrast to other alphabetic orthographies, and because word recognition experiments in Hebrew have revealed divergent findings with important cross-linguistic implications, contrasting Hebrew reading performance with other languages promises to provide important insights for reading research. Hence, a database of behavioral data on Hebrew words has far-reaching implications. Here we present an open data source on Hebrew visual word recognition, the Hebrew Lexicon Project (HeLP), which is the first to examine printed word recognition in a Semitic language on a large scale. The project reports data from two tasks: LD and naming. It assembles LD responses to 10,000 words and nonwords, with 5000 of the words also having additional associated naming data. Importantly, HeLP employs an ecologically valid set of Hebrew words sampled from a natural Hebrew printed corpus (see Methods), so that all types of words are included, including words with clitics, prefixes, and suffixes, Semitic and non-Semitic, inflected and derived—a variety that reflects the words Hebrew readers encounter in their daily lives. In line with previous mega-studies and open-science studies more broadly, HeLP is meant to enable researchers to tackle a large number of questions regarding word recognition in Hebrew and its similarities and differences to other languages. The goal of this first paper, of course, is not to cover all such potential explorations. Rather, in the current paper, we demonstrate the utility of the HeLP data by addressing a series of foundational questions related, on the one hand, to the structure of the Hebrew writing system, and on the other to the predictors of word recognition in that language. In particular, as detailed below, our analyses examine the prevalence and impact of phonological uncertainty and homography in LD and naming tasks, the distributions of word lengths and neighborhood densities and their impact in a root-based orthography, and the distribution and behavioral consequences of Semitic and non-Semitic structure, as well as the number of clitics. Together, these analyses result in mapping, on a large scale, the behavioral impact of general predictors thought to reflect word recognition across languages (lexicality, frequency, word length, orthographic neighborhood density), but also predictors that are relevant specifically to reading in Hebrew (Semitic structure, presence of clitics, and phonological ambiguity). The structure of the paper is as follows: First, we present results confirming the reliability of the HeLP data, using split-half estimates (at both the item and participant level), meant to ensure that the collected data are of sufficient quality for subsequent analyses. We then introduce descriptive statistics for the data, and present the basic effects revealed in the LD and the naming tasks pertaining to both general and Hebrew-specific predictors, as reviewed above. Finally, the implications of the results are discussed. Methods Lexical decision task Stimuli The project assembled LD data for 10,000 Hebrew words and 10,000 nonwords. Words were sampled from the Hebrew portion of the Subs2vec corpus (van Paridon & Thompson, 2021) which has 170 million tokens. First, a list of the 50,000 most frequent Hebrew words was extracted from the corpus, and all proper names and misspelled words were manually removed from that list. We then randomly sampled 2500 words from the 5000 most frequent words in the list, and 7500 words from the remainder of the frequency range of the filtered list, which together made the 10,000 targets for the LD task. Nonwords were generated by shuffling letters of all words from the 50,000-word filtered list. Words with four or fewer letters had all their letters reshuffled. Words with five or more letters had their beginning, middle, or end shuffled randomly (first, middle, or last letters of the word). The number of letters shuffled ranged from 3 to n − 1. Following this procedure, 10,000 nonwords were selected randomly and then inspected manually to ensure they truly had no meaning in Hebrew. The shuffling of different numbers of letters at different locations within words avoided manual decisions that may introduce bias, and was meant to provide a variety of items which could be mined to study the determinants of ease or difficulty in rejecting letter strings as potential Hebrew words. As such, our “shuffling” approach resulted, for example, in items that have a clearer expected pronunciation (e.g., ביגלרת, טגל), along with others that are more phonologically ambiguous (e.g., רמלש); in nonwords with a pseudo-morphological Semitic structure (e.g., האחבטה, 8767Behavior Research Methods (2024) 56:8761–8783 להגהר); and in nonwords that vary in bigram frequency. In this first paper we do not analyze predictors of responses to nonwords, but the nonword data are made fully available with the rest of the dataset for future research. Twenty sublists, each comprising 500 words and 500 nonwords (1000 targets overall), were created to serve as stimuli for each experimental session in the LD task. To ensure that each sublist included words from the full frequency range, the first word was assigned to the first sublist, the second word to the second sublist, etc., and the 21st word was assigned again to the first sublist and so on. Results from the 20 sublists were analyzed to examine the general effects such as lexicality (see Supplementary Materials S1). Since Hebrew-specific predictors required manual coding, only words from the 10 sublists that were also employed in the naming task were used in analyses considering these predictors (see details on naming task stimuli, below). Nonwords were randomly assigned to the sublists, and Welch t-tests ensured that the words and the nonwords in every sublist did not differ significantly in terms of length (t(19,970) = − 0.4, p = 0.68; see Fig.1A). In contrast, words and nonwords did differ in their mean orthographic Levenshtein distance 20 (OLD20), a measure of orthographic neighborhood density defined as the mean Levenshtein distance of the 20 closest orthographic neighbors of an item (Yarkoni etal., 2008), with a mean OLD20 of 1.65 for words and 2.17 for nonwords (t(18,909) = − 74.5, p < 0.001, see Fig.1B).This is expected because words share more orthographic resemblance to other words, whereas nonwords (resulting from shuffling letters) have less resemblance to existing words. Participants A total of 273 participants completed the experimental sessions. Data for nine participants were excluded from the analysis due to technical difficulties and poor performance (mean accuracy lower than 75%, or more than 12% of responses under 300ms). Overall, the results for 264 participants (201 female) had valid LD data. Due to manual coding of stimuli from only a partial set of overlapping lists that were employed in both the LD and naming tasks, our central models which include Hebrew-specific predictors were conducted on data from 178 participants. Participants were students at the Hebrew University of Jerusalem and were recruited using the Psychology Department’s participant recruitment platform. The average age of the participants was 24.14years (SD = 3.53years). All participants declared that they had no attention or reading disabilities, that their first language was Hebrew, and that they had normal or corrected-to-normal vision. Participants received credit or payment for their participation after each experimental session. Each participant could take part in as many experimental sessions as they wished, up to 20 (see below), but they were required to take at least a 15-min break between sessions, and could not take part in more than two sessions a day. Procedure In the LD task, all experimental sessions were performed online from home (using laptop or desktop computers only, i.e., not via smartphones or tablets). The LD task was built using PsychoPy (Peirce etal., 2019), version 2021.2.3, and hosted on Pavlovia. After signing a consent form and confirming eligibility criteria, participants were instructed that in each trial they would be presented with a letter string on the computer screen, to which they had to respond as rapidly and as accurately as possible as to whether it formed an existing Hebrew word (pressing the “L” key) or not (pressing the “S” key). The session began with 10 practice trials consisting of five words and five nonwords, followed by the experimental stimuli. Each stimulus remained on the screen until a response was recorded, with a timeout of two Fig. 1 Properties of words and nonwords in the LD task. A Distribution of length (in letters). B Distribution of OLD20 8768 Behavior Research Methods (2024) 56:8761–8783 seconds, and a blank screen for 500ms between stimuli. Targets were presented in the middle of the screen, in white font on a dark gray background, with their size set to 10% of a participant’s screen (the absolute dimensions varied given the online nature of the task). There were three breaks during the session, after 250, 500, and 750 trials. Each experimental session typically lasted between 20 and 30min. Naming task Stimuli From the 20 sublists (i.e., 10,000 words) that were used in the LD task, we sampled 10 sublists, with a total of 5000 words to serve as stimuli in the naming task, maintaining the same frequency distribution as in the full set of 10,000 words. The 5000 words were redivided into six sublists for the naming experiment, four of them with 800 words and two with 900 words. Each of the six naming sublists included words from the full frequency range as described above (25% from the 5000 most frequent words, and 75% from the entire range of frequencies following the 5000 most frequent words). Participants A total of additional 151 participants (101 female) completed the naming experimental sessions, using the same recruitment procedure. The mean age of participants was 24.6years (SD = 4.4years). As in the LD task, participants could take part in as many experimental sessions as they wished (up to six, the number of sublists). Procedure In contrast to the web-based LD task, all experimental sessions in the naming task were performed in the laboratory. The naming task was built using the NeuroBehavioral Systems software, Presentation, version 23.010.27.21. Participants sat in front of a computer screen in a quiet experiment room, wearing a headset. They were told that Hebrew words would appear on the screen, one at a time, and they should read every word aloud as fast and as accurately as they could. Before each word, a fixation cross appeared in the middle of the screen for 800ms. The word disappeared from the screen once a participant initiated a voice key with a spoken response or after a timeout of 1.5s. The experiment started with 10 practice trials, to ensure that the voice key operated correctly (experimenters adjusted the threshold if needed). There were two breaks during the experiment, after 250 and 500 words in the 800-word sessions, and after 300 and 600 words in the 900-word sessions. Words were presented in the middle of the screen, in 55-pt. white font on a dark gray background, taking about 10% of the vertical dimension of the screen. RTs were measured from the appearance of a word on the screen to the activation of the voice key. Similar to the LD task, participants who wished to participate in multiple sessions had to take a break of at least 15min between sessions, and could not participate in more than two sessions in one day. Each experimental session typically lasted around 30min. Responses in the naming task were recorded and were later coded by a team of five trained research assistants. Each response was coded both for its accuracy (correct/ incorrect; and in rare cases, “unclear”—see below) and for the validity of the RT data (i.e., whether the first recorded auditory signal, which triggered the voice key, should be used in RT analysis). A response could be coded as correct/incorrect while still not having valid associated RT: coding as “invalid RT” was automatically assumed in cases where the responses were faster than 200ms, as well as in additional cases where the coder noticed another (invalid) response that triggered the voice key (e.g., in cases where there was an initial sound such as a cough or murmur before the participant read the target). In such cases, we used the correctness data but not the RT data in the analyses below. A total of 94.5% of responses were coded as having a valid RT associated with them, whereas 99.4% of responses had a correct/incorrect coding (i.e., only 0.6% of responses had “unclear” as the coding for correctness; in analyses below, we treat these as NA in all models). To ensure the reliability of the coding of naming responses, we further randomly sampled six sessions from six participants, with 5000 naming trials in total (four participants had 800 trials; two participants had 900 trials). The re-coding of these sessions was done by another (blinded) research assistant, who used the same coding scheme as the original coders. We found an inter-rater reliability estimate of Cohen’s 𝜅=0.62 , a value representing “substantial” agreement between raters (Landis & Koch, 1977), with estimates in the data from the six randomly sampled participants separately ranging from 𝜅=0.52 to 𝜅=0.85 . These estimates suggest that the reliability of the naming response coding was substantial overall and at a minimum moderate in individual participants. Predictors ofvisual word recognition performance: Word‑level variables Figure2 summarizes the different psycholinguistic predictors made available for the different sets of stimuli with the current release of the HeLP data. In what follows, we provide more information about these predictors. 8769Behavior Research Methods (2024) 56:8761–8783 General visual word recognition predictors The models below use as predictors the following general variables, which are known to impact word recognition across many languages: (a) lexicality, whether a target is a word or a nonword (in the LD task); (b) word frequency (string frequency of the surface form, log-transformed), based on the Subs2vec corpus (van Paridon & Thompson, 2021); (c) word length (in number of letters); and (d) OLD20, computed using the filtered list of 50,000 words from the Subs2vec corpus, examining for each item the number of substitution, insertion, or deletion operations required to turn that item into any of the other 50,000 words in the list. Then, OLD20 was defined as the mean number of alterations in the 20 words that required the minimal number of alterations (Yarkoni etal., 2008). Hebrew‑specific predictors In addition to these general predictors, our models consider predictors potentially relevant specifically to word recognition in Hebrew (and other Semitic languages), given previous research and considering the properties of the writing system. As detailed below, obtaining these measures required excessive manual coding of responses. Hence, we focused on the 5000 words that were used as stimuli in both the naming and the LD tasks (rather than on the full set of 10,000 words in the LD task). Semitic structure. As reviewed in the Introduction, an important feature of Hebrew is that while many words have a Semitic structure and comprise root and word pattern combinations, there are also many words without Semitic structure (see Velan & Frost, 2011), which have been assimilated into Hebrew throughout history. The Semitic tagging of words was manually performed such that each word was tagged into one of four categories: “clearly Semitic”, “clearly non-Semitic”, “undetermined”, or “other”. Words were classified as “clearly Semitic” if they had an unequivocal root which was productive and used in different phonological patterns, creating a variety of words with distinct meanings. For example, the word "כתבתי" (KTBTI, meaning I wrote) is constructed from the root letters “כ,ת,ב” (K.T.B.), and appears in many Hebrew patterns to create distinct words (e.g., "מכתב", MKTB, /mixtav/—a letter; "מכתיב", MKTIB, /maxtiv/— dictates; "כתב", KTB, /katav/—he wrote; see, e.g., Frost etal., 1997). Words were tagged as clearly non-Semitic if they could not be decomposed into a productive root and a word pattern (e.g., the word "לימון", LIMWN, /limon/, meaning a lemon, which was assimilated into Hebrew and is not derived from a productive root or pattern). The “undetermined” category was used for words that could be read or analyzed in more than one way (e.g., "אטום", ATWM, can be read as / atom/, meaning an atom, a non-Semitic word, or as /atum/ meaning “sealed”, derived from the root A.T.M.), or words that could not be unequivocally classified as having a Semitic structure. In the “other” category there were prepositions, adverbs, and pronouns, which do not follow either a Semitic or non-Semitic structure. Our tagging revealed that 75% of the 5000 words were clearly Semitic, 18.9% were defined as clearly non-Semitic, and 6.1% were undetermined/other (Fig.3A). Only words with a clear Semitic or non-Semitic structure were included in the models below. Fig. 2 Information about included HeLP stimuli and their available lexical properties 8776 Behavior Research Methods (2024) 56:8761–8783 Participants again showed the expected frequency effect (Z = 15.67, p < 0.001). However, in contrast to the LD results, they made fewer naming errors in shorter words (Z = − 2.75, p = 0.006; although, when removing OLD20 from the model, the length effect flipped, see Supplementary Materials S2). There was a significant OLD20 effect, indicating that participants were more accurate when a word had fewer orthographic neighbors (Z = 7.26, p < 0.001). In terms of interactions (Fig.9), there was a significant interaction between word length and OLD20 (Z = − 6.32, p < 0.001): participants made more errors naming short words with many orthographic neighbors, and made more errors naming longer words with few orthographic neighbors. There was also a significant interaction between log frequency and word length (Z = − 3.02, p = 0.003), with length effects revealed only for words in the mid-frequency range and above (but not in low-frequency words). Fig. 7 Visual depiction of significant Interactions in the accuracy model, LD data. A Interaction between word length and word log frequency. B Interaction between OLD20 and word log frequency. C Interaction between OLD20 and word length Fig. 8 Visual depiction of effects of interest in the RT model, naming data. A Interaction between OLD20 and word length (the only significant interaction in this model). B Estimated mean log-transformed RT for words as a function of the number of clitics 8777Behavior Research Methods (2024) 56:8761–8783 As for Hebrew-specific predictors, there was no difference in naming accuracy between Semitic and non-Semitic words (Z = 0.63, p = 0.53). Participants made more errors reading words with one (Z = − 4.75, p < 0.001), two (Z = − 2.87, p = 0.004), and three clitics (Z = − 3.29, p < 0.001) than words without clitics. Also, as expected, participants made more errors reading words with higher pronunciation entropy (Z = − 3.71, p < 0.001). Discussion How the basic characteristics of writing systems impact visual word recognition behavior has been the focus of extensive research (see, e.g., Frost, 2012, for review and discussion). In this vein, studies conducted in different languages, particularly those including large-scale LPs, have provided important insights regarding the different computations readers employ during reading in their writing system, highlighting high-order principles of word recognition. From this perspective, evidence from Hebrew has continuously shaped theories and models of reading, given the unique characteristics of its writing system. However, word recognition studies in Hebrew to date have employed small, targeted experiments, covering only a limited part of the language’s lexicon. To address this gap, we present here for the first time a large-scale dataset of reading behavior in a Semitic language, Hebrew, comprising LD responses to 10,000 words and nonwords, and naming responses to 5000 words. In this first paper, we then utilize the data from the Hebrew LP (HeLP) to examine the contribution of general predictors (lexicality, frequency, length, and orthographic neighborhood), and Hebrew-specific predictors (Semitic structure, clitic letters, and extent of phonological ambiguity), to visual word recognition performance. As we discuss below in detail, our findings offer important insights regarding the computations involved in the processing of printed words in a writing system such as Hebrew, suggesting a set of universal computations involved in print processing. General predictors ofword recognition Unsurprisingly, the benchmark effects of frequency and lexicality emerged in Hebrew as in any LP, confirming that these principles of lexical search are similar across writing systems. Of theoretical interest, therefore, are findings in which Hebrew seems to diverge from the well-researched European languages. A first finding of interest is the effect (or lack thereof) of word length, mainly in LD (the parallel effects in naming are somewhat weaker). When interpreting the HeLP findings for length, an important factor to consider is the collinearity between word length and OLD20 (r = 0.79). That shorter words have more orthographic neighbors is typical of many writing systems, as revealed in other LPs (e.g., a correlation of 0.77 in French, Ferrand etal., 2010; a correlation of 0.79 in European Portuguese, Soares etal., 2019). But note that in the vast majority of LPs, length effects were significant even when this high correlation was partialed out in the analyses. For example, in the Malay LP, the several length measures that were used showed correlations ranging from 0.47 to 0.77 with the OLD20 scores, and yet, independently, word length was the strongest predictor of performance in both LD and naming tasks (Yap etal., 2010). In light of those previous results, the consistent lack of word length effect for words in Hebrew stands out, and models that also included OLD20 often revealed a reversed length effect, where longer words incurred faster recognition time and greater accuracy. Compared with other existing LPs, these results seem to align Fig. 9 Visual depiction of significant Interactions in the accuracy model, naming data. A Interaction between word length and word log frequency. B Interaction between OLD20 and word length 8778 Behavior Research Methods (2024) 56:8761–8783 with findings in the Persian LP (Nemati etal., 2022). A possible common factor to Hebrew and Persian is that in both languages, most vowels are omitted from orthographic representation of words. Since longer words on average include more vowel letters, they incur less phonological ambiguity, pointing to a lexical candidate more rapidly and resulting in faster recognition.4 More broadly, our findings resonate with previous claims that word length effects are stronger in more transparent writing systems (e.g., Cuetos & Suarez-Coalla, 2009; Ellis & Hooper, 2001; and see Weiss etal., 2015, for related evidence from pointed vs. un-pointed Hebrew). Our results go a step further to suggest that in Hebrew, word length effects are sometimes reversed, arguably due to lower levels of phonological ambiguity in longer words. A second point of interest is the different impact of orthographic neighbors in the LD versus the naming task. While the typical facilitatory effect of orthographic neighborhood was observed in LD, our findings show either that orthographic neighborhood was not predictive of naming RT and accuracy, or that having many orthographic neighbors of a word resulted in an inhibitory effect. This contrasts with studies showing that the presence of many orthographic neighbors has a facilitatory effect on naming English words (e.g., Yarkoni etal., 2008). Importantly, there are documented cross-linguistic differences in orthographic neighborhood effects in naming. For example, similar to our present finding in Hebrew, Chang etal. (2016) reported an inhibitory effect of neighborhood size in Chinese, and demonstrated through computational modeling that the division of labor between phonological and semantic pathways in deep orthographies is the key to accounting for the inhibitory effect of neighborhood size (and see Peereman & Content, 1997, for differences between English and French). The difference between LD and naming, then, reflects substantial differences in computations in a deep orthography like Hebrew. While the presence of many orthographic neighbors could contribute to a fast decision of whether a letter string is a Hebrew word or not, naming requires one to identify, select, and pronounce specific lexical candidates that often differ in vowel configurations. Hence, for Hebrew, the presence of many competitors seems to slow response times, rather than accelerate them. Hebrew‑specific effects Given the unique properties of Hebrew, understanding the role of the Hebrew-specific predictors provides important insights with regard to visual word recognition in that language. Hence, in this section we review findings pertaining to the three Hebrew-specific predictors tested: Semitic structure, presence of clitics, and phonological ambiguity. With regard to Semitic structure, even though the Semitic tagging we employed was conservative (defining a word as “clearly Semitic” only if its root letters were clearly productive), 75% of the 5000 words were tagged as such. Most of our statistical models showed that participants’ performance improved when presented with words having a Semitic structure. The present findings of HeLP indicate then that Hebrew readers become attuned to the statistical properties of Hebrew with its non-concatenated morphology, and are more efficient in processing Hebrew words when they conform to the highly prevalent Semitic form of intertwined root and word pattern morphemes. We assume that through statistical learning, readers become increasingly efficient in detecting the root letters within printed Semitic words, enabling the fast decomposition of printed words into their constituent morphemes (see, e.g., Feldman etal., 1995; Velan etal., 2013). This provides readers with the missing vowel information, and leads to fast lexical access when words are organized by morphological rather than simple orthographic principles (see Frost etal., 2005, for discussion). Regarding the effects of clitics, our findings indicate that the number of clitic letters impacted performance, with responses to words with clitics generally being less accurate and incurring slower RTs. This finding suggests that when words appear in isolation without disambiguating context, clitic letters add further complexity to the process of decomposing the printed word into its morphemic constituents. An interesting deviation from this pattern, however, was found in the naming task: participants read aloud words with clitics faster (although still with more errors). Given the high prevalence of clitics and the relative systematicity of their pronunciation (e.g., “and” in most cases is pronounced as \ ve\, “the” as \ha\, etc.), and in light of the time constraints in the naming task, we cannot rule out the possibility that participants initiated the pronunciation of the initial clitics before fully recognizing the word (and the higher error rate for words with clitics indeed supports this explanation). As naming latencies reflect the time course of the initial utterance, our overall speeded responses in the naming task could reflect participants’ high confidence in initiating a vocal response to the initial clitic letter. The third important characteristic of Hebrew is its phonological under-specification. While previous research simply counted the number of empty vowel slots to assess phonological uncertainty (e.g., Frost, 1995), in the present 4 Some readers may wonder why the correlation between word length and phonological entropy, while negative, is quite small, r = − 0.16 (Table3 above). In this context, it is important to keep in mind that our measure of phonological entropy reflects the number of meaningful pronunciations a word has, given the heterophonic aspect of Hebrew. As such, it does not fully capture a word’s degree of phonological ambiguity (i.e., uncertainty or inconsistency): It does not reflect the difficulty in generating unique pronunciations given the number of missing vowels in the orthographic sequence. Hence two words can have a unique lexical pronunciation (and hence a similar score in the measure of phonological entropy), but one with few missing vowels and one with many. 8779Behavior Research Methods (2024) 56:8761–8783 work we captured the level of phonological uncertainty when pronouncing a word by mathematically quantifying phonological entropy across responses in the full sample of participants (see also De Simone etal., 2021). Our results show that in all models, higher pronunciation entropy hindered performance, for both naming and LD. Whereas the impact of phonological entropy in naming is anything but surprising, the parallel finding for LD is striking. The orthographical depth hypothesis (Frost etal., 1987) has argued that readers of shallow orthographies rely strongly on phonological cues in visual word recognition, while readers of deep orthographies such as English or Hebrew rely more on orthographic or semantic cues. It was assumed that in deep orthographies, the phonological information of a word is mediated by the internal lexicon (Frost etal., 1987; Katz & Feldman, 1981). Our present data suggests that phonological information is computed not only when pronouncing a word, but also when simply identifying it in the LD task. This finding accords with the claim that early and fast phonological computations in visual word recognition characterize the reading process in any orthography whether shallow or deep (for discussion see Frost, 1998; Rastle & Brysbaert, 2006). Indeed, Rueckl etal. (2015) have shown that readers of different languages with different orthographic depths demonstrate similar neuronal processing of printed words, including in brain areas associated with phonological processing. Our results support this line of research, showing early phonological processing in Hebrew, which is considered a highly deep orthography. That said, we should caution that our measure of phonological entropy also potentially reflects uncertainty in the mapping between print and meaning; this is because, in Hebrew, multiple pronunciations of the same word form also often have multiple distinct meanings, and therefore more homographic word forms also typically carry more ambiguity in the orthographic–semantic mapping. Future work should carefully disentangle the effects of different types of ambiguity by providing and validating wordlevel measures of print–speech and print–meaning regularities in Hebrew. We expect the HeLP data to be crucial in the validation of such measures (for parallel work in English using the ELP data, see, e.g., Chee etal., 2020; Marelli & Amenta, 2018; Siegelman etal., 2020, 2022). What stands out? What isuniversal? As outlined in our Introduction, our theoretical approach to LPs is that they go far beyond the descriptive statistics of yet another writing system. Divergent findings in LPs point to higher-order computational principles that explicate the difference in results in one language relative to another. Here we argue that to account for the range of findings revealed in HeLP, visual word recognition should be considered as a process of uncertainty reduction with respect to the identity of lexical candidates and their phonological structure, as represented by the printed forms. This universal principle accounts for cross-linguistic differences by weighting the set of constraints that drive uncertainty reduction in a given writing system. The main problem in reading Hebrew is in converging on an unequivocal lexical and phonological solution to a range of parsing and decoding alternatives. While reading in context significantly reduces uncertainty, leading in most cases to a single solution, words in isolation incur significant uncertainty. Uncertainty in Semitic languages concerns competing parsing possibilities (e.g., whether the initial letter is a clitic letter, a word pattern letter, or a root letter, which is critical for identifying the correct lemma) and also competing phonological representations. This perspective accounts, for example, for the faster responses to longer words, since often they incur lower entropy than shorter words. It offers a possible explanation for the faster responses to words with Semitic structure, since these words typically contain several cues for correct morphological decomposition (and see BarOn etal., 2017, 2019, 2021, for discussions of uncertainty reduction in Hebrew). Considering visual word recognition as a process of uncertainty reduction also outlines the range of dimensions to consider when comparing performance across languages, and generates predictions regarding cross-linguistic differences in reading. It shifts the scope of analysis from unidimensional factors such as print–speech transparency, or morphological complexity, to regard performance in visual word recognition in terms of constraint satisfaction, where multidimensional constraints interact to determine the outcome of processing. Future directions Our present findings offer compelling evidence of how the unique morphological and phonological properties of Hebrew play an important role during visual word recognition. However, our present analyses are but a first step which involves coarse-grained quantification of words’ properties. Following in the steps of previous LPs, the HeLP project adheres to the principles of open science, making all data available for secondary analyses, which we hope will facilitate future investigation into the exact predictors of visual word recognition in Hebrew. As mentioned briefly above, one important avenue for future research is the development of more subtle and precise quantification of orthographic–phonological regularities in Hebrew. In the current work, we only used a measure of pronunciation entropy, which was calculated given the actual pronunciations that participants uttered in the naming task. Although this measure has ecological validity as it includes all pronunciations that were expressed in practice by our 8780 Behavior Research Methods (2024) 56:8761–8783 sample of participants, it is limited in two important ways. First, there are other phonological expressions for the words we employed that were not included in the measure’s calculation. For example, the printed word "כמורה" (KMWRH) has five different phonological forms that bear meaning in Hebrew (/kmura/, /kemore/, /kemora/, /kamore/, /kamora/), but only three of them were produced by our participants. Perhaps more importantly, considering only actual pronunciations does not capture the full extent to which different graphemes predict phonemes in the language. We leave it for future work to develop precise corpus-based metrics of the links between orthography and phonology in Hebrew. This work will most likely involve adapting measures developed in English and other European languages (e.g., Chee etal., 2020; Siegelman etal., 2020) to capture the unique properties of Hebrew (e.g., the fact that in Hebrew, in contrast to English, substantial irregularities also exist in the mapping of consonant letters into phonemes, rather than mostly vowel letters). Note that producing precise measures of orthographic–phonological entropy is but one step in assessing uncertainty in an Abjad writing system such as Hebrew. Other avenues would include the impact of phonological Levenshtein distance, as well as measures of orthographic–semantic regularities (e.g., Marelli & Amenta, 2018; Siegelman etal., 2022). Multiple other avenues of analyses using the HeLP data can replicate studies using the ELP data in a writing system with a divergent structure, including those in psycholinguistic ratings such as word concreteness, (Brysbaert etal., 2014), age of acquisition (Kuperman etal., 2012), and body–object interaction (Pexman etal., 2019), to name a few. Future analyses should also consider the possible interactions that may exist between Hebrew-specific predictors and general psycholinguistic properties (e.g., the fact that Semitic structure or extent of phonological ambiguity may impact processing differently across word lengths). Merging all these dimensions together would enable the alignment of writing systems for cross-linguistic comparisons, = while simultaneously considering the possible interactions of their orthographic, phonological, and semantic properties. Supplementary information The online version contains supplementary material available at https:// doi. org/ 10. 3758/ s1342802402502-4. Acknowledgements Work in this paper was supported by the following funding sources: the European Research Council (ERC) Advanced Grant, project 692502-L2STAT, under the Horizon 2020 research and innovation program (awarded to RF); the Israel Science Foundation (ISF) Grant, project 705/20 (awarded to RF); the Israel Science Foundation (ISF) Grant, project 1034/23 (awarded to NS); and an Azrieli Early Career Faculty Fellowship (awarded to NS). We thank Gollan Ankori, Shaked Sukiennik, Einav Avraham, Noam Davidov, and Ellah Richter for their work on manually coding the Hebrew-specific predictors and help in data collection. Funding Open access funding provided by Hebrew University of Jerusalem. Work in this paper was supported by the following funding sources: the European Research Council (ERC) Advanced Grant, project 692502-L2STAT, under the Horizon 2020 research and innovation program (awarded to RF); the Israel Science Foundation (ISF) Grant, project 705/20 (awarded to RF); the Israel Science Foundation (ISF) Grant, project 1034/23 (awarded to NS); and an Azrieli Early Career Faculty Fellowship (awarded to NS). Data availability The full HeLP data are available via the Open Science Framework (OSF) website for secondary data analyses, along with the full lists of items used in the studies: https:// osf. io/ nxq8g/. Code availability The project’s OSF repository also includes the code used for analyses reported in this paper. Declarations Ethics approval This study was performed in line with the principles of the Declaration of Helsinki. Approval was granted by the Institutional Review Board of the Hebrew University of Jerusalem (Approval No.: 17032022, granted March 17th, 2022). Consent to participate Informed consent was obtained from all individual participants included in the study by signing a physical or virtual consent form. Consent for publication Only de-identified data are available in the project’s repository. Participants agreed to having their de-identified data shared as part of a publication. Conflicts of interest The authors have no relevant financial or nonfinancial interests to disclose. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. References Aguasvivas, J. A., Carreiras, M., Brysbaert, M., Mandera, P., Keuleers, E., & Duñabeitia, J. A. (2018). SPALEX: A Spanish lexical decision database from a massive online data collection. Frontiers in Psychology, 9, 2156. Andrews, S. (1992). Frequency and neighborhood effects on lexical access: Lexical similarity or orthographic redundancy? Journal of Experimental Psychology: Learning, Memory, and Cognition, 18(2), 234. Balota, D. A., Yap, M. J., Hutchison, K. A., Cortese, M. J., Kessler, B., Loftis, B., ... & Treiman, R. (2007). The English lexicon project. Behavior Research Methods, 39, 445–459. Bar-On, A., Dattner, E., & Braun-Peretz, O. (2019). Resolving homography: The role of post-homograph context in reading aloud 8781Behavior Research Methods (2024) 56:8761–8783 ambiguous sentences in Hebrew. Applied Psycholinguistics, 40(6), 1405–1420. Bar-On, A., Oron, T., & Peleg, O. (2021). Semantic and syntactic constraints in resolving homography: A developmental study in Hebrew. Reading and Writing, 34, 2103–2126. Bar-On, A., Dattner, E., & Ravid, D. (2017). Context effects on heterophonic-homography resolution in learning to read Hebrew. Reading and Writing, 30, 463–487. Bentin, S., & Frost, R. (1987). Processing lexical ambiguity and visual word recognition in a deep orthography. Memory & Cognition, 15(1), 13–23. Bitan, T., Weiss, Y., Katzir, T., & Truzman, T. (2020). Morphological decomposition compensates for imperfections in phonological decoding. Neural evidence from typical and dyslexic readers of an opaque orthography. Cortex, 130, 172–191. Broadbent, D. E. (1967). Word-frequency effect and response bias. Psychological Review, 74(1), 1. Brysbaert, M., Lagrou, E., & Stevens, M. (2017). Visual word recognition in a second language: A test of the lexical entrenchment hypothesis with lexical decision times.Bilingualism: Language and Cognition,20(3), 530–548. Brysbaert, M., Warriner, A. B., & Kuperman, V. (2014). Concreteness ratings for 40 thousand generally known English word lemmas. Behavior Research Methods, 46, 904–911. Chang, Y. N., Welbourne, S., & Lee, C. Y. (2016). Exploring orthographic neighborhood size effects in a computational model of Chinese character naming. Cognitive Psychology, 91, 1–23. Chee, Q. W., Chow, K. J., Yap, M. J., & Goh, W. D. (2020). Consistency norms for 37,677 English words.Behavior Research Methods,52(6), 2535–2555.s Cuetos, F., & Suárez-Coalla, P. (2009). From grapheme to word in reading acquisition in Spanish. Applied Psycholinguistics, 30(4), 583–601. De Simone, E., Beyersmann, E., Mulatti, C., Mirault, J., & Schmalz, X. (2021). Order among chaos: Cross-linguistic differences and developmental trajectories in pseudoword reading aloud using pronunciation Entropy. PLoS ONE, 16(5), e0251629. Deutsch, A., Frost, R., & Forster, K. I. (1998). Verbs and nouns are organized and accessed differently in the mental lexicon: Evidence from Hebrew. Journal of Experimental Psychology: Learning, Memory, and Cognition, 24(5), 1238. Deutsch, A., Frost, R., Pelleg, S., Pollatsek, A., & Rayner, K. (2003). Early morphological effects in reading: Evidence from parafoveal preview benefit in Hebrew. Psychonomic Bulletin & Review, 10(2), 415–422. Deutsch, A., Velan, H., Merzbach, Y., & Michaly, T. (2021). The dependence of root extraction in a non-concatenated morphology on the word-specific orthographic context. Journal of Memory and Language, 116, 104182. Duñabeitia, J. A., Perea, M., & Carreiras, M. (2007). Do transposedletter similarity effects occur at a morpheme level? Evidence for Morpho-Orthographic Decomposition. Cognition, 105(3), 691–703. Dymarska, A., Connell, L., & Banks, B. (2023). Weaker than you might imagine: Determining imageability effects on word recognition. Journal of Memory and Language, 129, 104398. Ellis, N. C., & Hooper, A. M. (2001). Why learning to read is easier in Welsh than in English: Orthographic transparency effects evinced with frequency-matched tests. Applied Psycholinguistics, 22(4), 571–599. Feldman, L. B., Frost, R., & Pnini, T. (1995). Decomposing words into their constituent morphemes: Evidence from English and Hebrew. Journal of Experimental Psychology: Learning, Memory, and Cognition, 21(4), 947. Ferrand, L., New, B., Brysbaert, M., Keuleers, E., Bonin, P., Méot, A., Augustinova, M., & Pallier, C. (2010). The French lexicon project: Lexical decision data for 38,840 French words and 38,840 pseudo words. Behavior Research Methods, 42(2), 488–496. Fredriksen, J. R., & Kroll, J. F. (1976). Spelling and sound: Approaches to the internal lexicon. Journal of Experimental Psychology: Human Perception and Performance, 2, 361–379. Friedmann, N., & Gvion, A. (2001). Letter position dyslexia. Cognitive Neuropsychology, 18(8), 673–696. Friedmann, N., & Gvion, A. (2005). Letter form as a constraint for errors in neglect dyslexia and letter position dyslexia. Behavioural Neurology, 16(2–3), 145–158. Friedmann, N., & Lukov, L. (2008). Developmental surface dyslexias. Cortex, 44(9), 1146–1160. Friedmann, N., & Rahamim, E. (2007). Developmental letter position dyslexia. Journal of Neuropsychology, 1(2), 201–236. Frost, R. (2006). Becoming literate in Hebrew: The grain size hypothesis and Semitic orthographic systems. Developmental Science, 9(5), 439. Frost, R. (1995). Phonological computation and missing vowels: Mapping lexical involvement in reading. Journal of Experimental Psychology: Learning, Memory, and Cognition, 21(2), 398. Frost, R. (1994). Prelexical and postlexical strategies in reading: Evidence from a deep and a shallow orthography. Journal of Experimental Psychology: Learning, Memory, and Cognition, 20(1), 116. Frost, R. (1998). Toward a strong phonological theory of visual word recognition: True issues and false trails. Psychological Bulletin, 123(1), 71–99. Frost, R. (2012). Towards a universal model of reading. Behavioral and Brain Sciences, 35(5), 263–279. Frost, R., Deutsch, A., Gilboa, O., Tannenbaum, M., & MarslenWilson, W. (2000). Morphological priming: Dissociation of phonological, semantic, and morphological factors. Memory & Cognition, 28(8), 1277–1288. Frost, R., Forster, K. I., & Deutsch, A. (1997). What can we learn from the morphology of Hebrew? A masked-priming investigation of morphological representation. Journal of Experimental Psychology: Learning Memory and Cognition, 23(4), 829–856. Frost, R., Kugler, T., Deutsch, A., & Forster, K. I. (2005). Orthographic Structure VersusMorphological Structure: Principles of Lexical Organization in a Given Language. Journal of Experimental Psychology: Learning, Memory, and Cognition, 31(6), 1293–1326. Frost, R., Katz, L., & Bentin, S. (1987). Strategies for Visual Word Recognition and Orthographical Depth: A Multilingual Comparison. Journal of Experimental Psychology: Human Perception and Performance, 13(1), 104–115. Forster, K. I., & Chambers, S. M. (1973). Lexical access and naming time. Journal of Verbal Learning and Verbal Behavior, 12(6), 627–635. Geary, J. A., & Ussishkin, A. (2018). Root-letter priming in Maltese visual word recognition. The Mental Lexicon, 13(1), 1–25. Grainger, J., Kevin O’regan, J., Jacobs, A. M., & Segui, J. (1989). On the role of competing word units in visual word recognition: The neighborhood frequency effect. Perception & Psychophysics, 45(3), 189–195. Haddad, L., Weiss, Y., Katzir, T., & Bitan, T. (2018). Orthographic transparency enhances morphological segmentation in children reading Hebrew words. Frontiers in Psychology, 8, 2369. Hirshorn, E. A., & Harris, L. N. (2022). Culture is not destiny, for reading: Highlighting variable routes to literacy within writing systems. Annals of the New York Academy of Sciences, 1513(1), 31–47. Hudson, P. T., & Bergman, M. W. (1985). Lexical knowledge in word recognition: Word length and word frequency in naming and 8782 Behavior Research Methods (2024) 56:8761–8783 lexical decision tasks. Journal of Memory and Language, 24(1), 46–58. Katz, L., & Feldman, L. B. (1981). Linguistic coding in word recognition: Comparisons between a deep and a shallow orthography. In A. Lesgold & C. Perfetti (Eds.), Interactive Processes in Reading (pp. 85–99). Lawrence Erlbaum Associates. Katz, L., & Frost, R. (1992). The reading process is different for different orthographies: The orthographic depth hypothesis. Advances in Psychology, 94, 67–84. Keuleers, E., Diependaele, K., & Brysbaert, M. (2010). Practice effects in large-scale visual word recognition studies: A lexical decision study on 14,000 Dutch mono-and disyllabic words and nonwords. Frontiers in Psychology, 1, 174. Keuleers, E., Lacey, P., Rastle, K., & Brysbaert, M. (2012). The British Lexicon Project: Lexical decision data for 28,730 monosyllabic and disyllabic English words. Behavior Research Methods, 44, 287–304. Kim, S. Y., Yap, M. J., & Goh, W. D. (2018). The role of semantic transparency in visual word recognition of compound words: A megastudy approach. Behavior Research Methods, 51, 2722–2732. Kuperman, V. (2015). Virtual experiments in megastudies: A case study of language and emotion. Quarterly Journal of Experimental Psychology, 68(8), 1693–1710. Kuperman, V., Stadthagen-Gonzalez, H., & Brysbaert, M. (2012). Age-of-acquisition ratings for 30,000 English words. Behavior Research Methods, 44, 978–990. Kuznetsova, A., Brockhoff, P. B., & Christensen, R. H. B. (2017). lmerTest Package: Tests in Linear Mixed Effects Models. Journal of Statistical Software, 82(13), 1–26. Lador-Weizman, Y., & Deutsch, A. (2022). The contribution of consonants and vowels to auditory word recognition is shaped by language-specific properties: Evidence from Hebrew. Journal of Experimental Psychology: Human Perception and Performance, 48(5), 401. Lallier, M., & Carreiras, M. (2018). Cross-linguistic transfer in bilinguals reading in two alphabetic orthographies: The grain size accommodation hypothesis. Psychonomic Bulletin & Review, 25, 386–401. Landis, J. R., & Koch, G. G. (1977). The measurement of observer agreement for categorical data. Biometrics, 33(1), 159–174. Lerner, I., Armstrong, B. C., & Frost, R. (2014). What can we learn from learning models about sensitivity to letter-order in visual word recognition? Journal of Memory and Language, 77, 40–58. Marelli, M., & Amenta, S. (2018). A database of orthographysemantics consistency (OSC) estimates for 15,017 English words. Behavior Research Methods, 50(4), 1482–1495. Monsell, S., Doyle, M. C., & Haggard, P. N. (1989). Effects of frequency on visual word recognition tasks: Where are they? Journal of Experimental Psychology. General, 118(1), 43–71. Mor, B., & Prior, A. (2020). Individual differences in L2 frequency effects in different script bilinguals. International Journal of Bilingualism, 24(4), 672–690. Mor, B., & Prior, A. (2021). Frequency and predictability effects in first and second language of different script bilinguals. Journal of Experimental Psychology: Learning, Memory, and Cognition, 48(9), 1363. Nemati, F., Westbury, C., Hollis, G., & Haghbin, H. (2022). The Persian Lexicon Project: Minimized orthographic neighbourhood effects in a dense language. Journal of Psycholinguistic Research, 51(5), 957–979. New, B., Ferrand, L., Pallier, C., & Brysbaert, M. (2006). Reexamining the word length effect in visual word recognition: New evidence from the English Lexicon Project. Psychonomic Bulletin & Review, 13, 45–52. Peereman, R., & Content, A. (1997). Orthographic and phonological neighborhoods in naming: Not all neighbors are equally influential in orthographic space. Journal of Memory and Language, 37(3), 382–410. Perea, M., Abu Mallouh, R., & Carreiras, M. (2014). Are root letters compulsory for lexical access in Semitic languages? The case of masked form-priming in Arabic. Cognition, 132, 491–500. Perea, M., & Carreiras, M. (2006a). Do transposed-letter effects occur across lexeme boundaries? Psychonomic Bulletin & Review, 13(3), 418–422. Perea, M., & Carreiras, M. (2006b). Do transposed-letter similarity effects occur at a prelexical phonological level? Quarterly Journal of Experimental Psychology, 59(9), 1600–1613. Perea, M., & Carreiras, M. (2008). Do orthotactics and phonology constrain the transposed-letter effect? Language and Cognitive Processes, 23(1), 69–92. Perea, M., Mallouh, R. A., & Carreiras, M. (2010). The search for an input-coding scheme: Transposed-letter priming in Arabic. Psychonomic Bulletin & Review, 17(3), 375–380. Perea, M., & Lupker, S. J. (2004). Can CANISO activate CASINO? Transposed-letter similarity effects with nonadjacent letter positions. Journal of Memory and Language, 51(2), 231–246. Perea, M., & Lupker, S. J. (2003). Does jugde activate COURT? Transposed-letter similarity effects in masked associative priming. Memory & Cognition, 31(6), 829–841. Perea, M., Lupker, S. J., & Kinoshita, S. (2003). Transposed-letter confusability effects in masked form priming.Masked priming: State of the art, 97–120. Peirce, J., Gray, J. R., Simpson, S., MacAskill, M., Höchenberger, R., Sogo, H., ... & Lindeløv, J. K. (2019). PsychoPy2: Experiments in behavior made easy. Behavior Research Methods, 51, 195–203. Pexman, P. M., Muraki, E., Sidhu, D. M., Siakaluk, P. D., & Yap, M. J. (2019). Quantifying sensorimotor experience: Body–object interaction ratings for more than 9,000 English words. Behavior Research Methods, 51, 453–466. Rastle, K., & Brysbaert, M. (2006) Masked phonological priming effects in English: Are they real? Do they matter? Cognitive Psychology, 53(2), 97–145. Rau, A. K., Moll, K., Snowling, M. J., & Landerl, K. (2015). Effects of orthographic consistency on eye movement behavior: German and English children and adults process the same words differently. Journal of Experimental Child Psychology, 130, 92–105. Ravid, D. D. (2011).Spelling morphology: The psycholinguistics of Hebrew spelling(Vol. 3). Springer Science & Business Media. Rueckl, J. G., Paz-Alonso, P. M., Molfese, P. J., Kuo, W. J., Bick, A., Frost, S. J., ... & Frost, R. (2015). Universal brain signature of proficient reading: Evidence from four contrasting languages. Proceedings of the National Academy of Sciences, 112(50), 15510–15515. Schmalz, X., Marinus, E., Coltheart, M., & Castles, A. (2015). Getting to the bottom of orthographic depth. Psychonomic Bulletin & Review, 22, 1614–1629. Schröter, P., & Schroeder, S. (2017). The Developmental Lexicon Project: A behavioral database to investigate visual word recognition across the lifespan. Behavior Research Methods, 49, 2183–2203. Schoonbaert, S., & Grainger, J. (2004). Letter position coding in printed word perception: Effects of repeated and transposed letters. Language and Cognitive Processes, 19(3), 333–367. Seymour, P. H., Aro, M., & Erskine, J. M. (2003). Foundation literacy acquisition in European orthographies. British Journal of Psychology, 94, 143–174. Share, D. L. (1999). Phonological recoding and orthographic learning: A direct test of the self-teaching hypothesis. Journal of Experimental Child Psychology, 72(2), 95–129. 8783Behavior Research Methods (2024) 56:8761–8783 Share, D. L., & Bar-On, A. (2018). Learning to read a Semitic abjad: The triplex model of Hebrew reading development. Journal of Learning Disabilities, 51(5), 444–453. Shimron, J. (2006).Reading Hebrew: The language and the psychology of reading it. Routledge. Seidenberg, M. S. (2011). Reading in different writing systems: One architecture, multiple solutions. In P. McCardle, B. Miller, J. R. Lee, & O. J. L. Tzeng (Eds.), Dyslexia across languages: Orthography and the brain–gene–behavior link (pp. 146–168). Paul H Brookes Publishing. Shimron, J., & Sivan, T. (1994). Reading proficiency and orthography: Evidence from Hebrew and English. Language Learning, 44, 5–27. Siegelman, N., Kearns, D. M., & Rueckl, J. G. (2020). Using information-theoretic measures to characterize the structure of the writing system: The case of orthographic-phonological regularities in English. Behavior Research Methods, 52, 1292–1312. Siegelman, N., Rueckl, J. G., Lo, J. C. M., Kearns, D. M., Morris, R. D., & Compton, D. L. (2022). Quantifying the regularities between orthography and semantics and their impact on groupand individual-level behavior. Journal of Experimental Psychology: Learning, Memory, and Cognition, 48(6), 839. Snell, J. (in press). PONG: A computational model of visual word recognition through bi-hemispheric activation, Psychological Review. Soares, A. P., Lages, A., Silva, A., Comesaña, M., Sousa, I., Pinheiro, A. P., & Perea, M. (2019). Psycholinguistic variables in visual word recognition and pronunciation of European Portuguese words: A mega-study approach. Language, Cognition and Neuroscience, 34(6), 689–719. Sze, W. P., Rickard Liow, S. J., & Yap, M. J. (2014). The Chinese Lexicon Project: A repository of lexical decision behavioral responses for 2,500 Chinese characters. Behavior Research Methods, 46, 263–273. Tsang, Y. K., Huang, J., Lui, M., Xue, M., Chan, Y. W. F., Wang, S., & Chen, H. C. (2018). MELD-SCH: A megastudy of lexical decision in simplified Chinese. Behavior Research Methods, 50, 1763–1777. Tse, C. S., Yap, M. J., Chan, Y. L., Sze, W. P., Shaoul, C., & Lin, D. (2017). The Chinese Lexicon Project: A megastudy of lexical decision performance for 25,000+ traditional Chinese two-character compound words. Behavior Research Methods, 49, 1503–1519. Van Paridon, J., & Thompson, B. (2021). subs2vec: Word embeddings from subtitles in 55 languages. Behavior Research Methods, 53, 629–655. Velan, H., & Frost, R. (2007). Cambridge University versus Hebrew University: The impact of letter transposition on reading English and Hebrew. Psychonomic Bulletin & Review 2007 14:5, 14(5), 913–918. Velan, H., & Frost, R. (2011). Words with and without internal structure: What determines the nature of orthographic and morphological processing? Cognition, 118(2), 141–156. Velan, H., Deutsch, A., & Frost, R. (2013). The flexibility of letterposition flexibility: Evidence from eye movements in reading Hebrew. Journal of Experimental Psychology: Human Perception and Performance, 39(4), 1143. Velan, H., Frost, R., Deutsch, A., & Plaut, D. C. (2005). The processing of root morphemes in Hebrew: Contrasting localist and distributed accounts. Language and Cognitive Processes, 20(1–2), 169–206. Velan, H., & Frost, R. (2009). transposition effects are not universal: The impact of transposing letters in Hebrew. Journal of Memory and Language, 61(3), 285–302. Voeten, C. C. (2019). Using ‘buildmer’ to automatically find & compare maximal (mixed) models. R Package Version, 1(6), 1–7. Weiss, Y., Katzir, T., & Bitan, T. (2015). The effects of orthographic transparency and familiarity on reading Hebrew words in adults with and without dyslexia. Annals of Dyslexia, 65, 84–102. Yap, M. J., Rickard Liow, S. J., Jalil, S. B., & Faizal, S. S. B. (2010). The Malay Lexicon Project: A database of lexical statistics for 9,592 words. Behavior Research Methods, 42(4), 992–1003. Yarkoni, T., Balota, D., & Yap, M. (2008). Moving beyond Coltheart’s N: A new measure of orthographic similarity. Psychonomic Bulletin and Review, 15(5), 971–979. Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.