Single Word Argument Projection: A Priming Study in Modern Eastern Armenian
by Sona Zohrabyan
A Master’s thesis submitted in partial fulfillment of the requirements for the degree of
Master of Science
(Clinical Linguistics)
at the Joint European Erasmus Mundus Master’s Programme in Clinical Linguistics (EMCL+)
UNIVERSITY OF GRONINGEN
August, 2020
Single Word Argument Projection: A Priming Study in Modern Eastern Armenian
Sona Zohrabyan
Under the supervision of Dr. Vânia de Aguiar at the University of Groningen and Dr. Byurakn Ishkhanyan at the Aarhus University
Abstract
Verb argument structure has been a topic of investigation in psycholinguistics for decades. It has been studied among both neurotypical individuals and individuals with language
disorders, mostly in word production experiments, but also in comprehension studies. These studies explained their results in the frameworks of weak or strong lexicalist theories of language production and comprehension. The strong lexicalist theories predict that single verbs would be processed faster by the human brain if they were primed by their argument nouns. The current study investigated this question by implementing a noun-verb priming paradigm with a lexical decision task in Modern Eastern Armenian. The nouns priming the experimental verbs were manipulated such that they would be: either semantically related or unrelated to the verb, either animate or inanimate, and either have an argument or non- argument status in relation to the verb. The analysis of the data from the lexical-decision task revealed that semantically related nouns had a priming effect on the verbs, as measured by the reaction times. However, contrary to the predictions drawn from strong lexicalist accounts of language production and comprehension, argument status or animacy manipulation of the nouns did not affect the processing times of the verbs due to higher than expected within and between participant variance.
Acknowledgments
The implementation of this thesis work would have been impossible without the support and expertise of my supervisors Dr. Vânia de Aguiar and Dr. Byurakn Ishkhanyan. I am grateful to Dr. de Aguiar for her guidance from the beginning of my thesis project and for providing me with professional advice and feedback which helped me to polish the theoretical aspects and writing of my work. I am thankful to Dr. Ishkhanyan for giving me valuable feedback on my statistical analysis and writing.
I would also like to thank each and every person who took part in my norming study and experiment. Without them my thesis would have been incomplete. I appreciate their help immensely. During the Coronavirus pandemic, a particularly difficult time for people
everywhere in the world, the participants believed in me and my project, and took the time to help me to move one step closer to my academic dream. I am grateful to Karen, Hasmik, Ferdinand and Mrs. Anahit for their great support in recruiting participants.
My family and friends, who have encouraged me and stood by me throughout EMCL, deserve the whole world and I thank them from the depths of my heart. Finally, I thank Robert for being my greatest sounding board. He has supported and believed in me more than I believed in myself.
TABLE OF CONTENTS
Abstract ... iii
Acknowledgments ... iv
List of Tables ... vii
List of Figures... viii
1 Introduction………..…………...1
1.1 Lexical Argument Structure………..……..2
1.1.1. Selectional Restrictions of Verbs for Their Arguments….………...5
1.2 Modern Eastern Armenian as a Testing Ground………..………...7
1.3 Representation of Lexical – Syntax in Word Production Models ……….8
1.4 Representation of Lexical – Syntax in Sentence Comprehension Models…………...11
1.5 Lexical Decision Tasks and Lexical Argument Structure in Psycholinguistic and Neurolinguistic Literature ... 14
1.6 Present Study ... 16
2 Method ... 19
2.1 Participants ... 19
2.2 Materials ... 20
2.3 Procedure ... 24
2.4 Power Calculations ... 26
2.5 Data Analysis ... 27
3 Results ... 30
3.1 Accuracy Analysis ... 31
3.2 Reaction Time Analysis ... 32
4 Discussion ... 37
4.1 Limitations and Future Directions ... 41
5 Conclusion………....43 6 Social Impact of the Current Study………..44 References………...45 Appendix A: Experimental Verbs and Nouns in Each Condition Used in the Lexical
Decision Task ………..…....52 Appendix B: Filler Verbs and Nouns Used in the Lexical Decision Task……….…....54 Appendix C: Nonword Targets and Real Word Prime Nouns Used in the Lexical
Decision Task………...56
List of Tables
1. Illustration of Predictions for RTs in Each Condition……….19
2. Structure of Conditions ………..………...…..20
3. Dunn Test Results for Semantic Relatedness between Conditions………...23
4. Descriptive Data of Accuracy……….………...31
5. Summary Results of Accuracy ………...………...32
6. Summary of Model 1………...………...32
7. Summary of Model 2………33
8. Summary of Model 3………34
9. Summary of Model 4 with Age as a Fixed Effect ……….…………..36
List of Figures
1. Experimental Design of Each Trial ………..…...25
2. Q-Q Plot of Raw RTs………..…………...………..27
3. Residual Plots of Model before Transformation……….………...28
4. Residual Plots of Model after Transformation…………..………..………...29
5. RTs Explained by Age of Participants………....………...35
6. RT Data of Individual Participants………...36
Single Word Argument Projection: A Priming Study in Modern Eastern Armenian In the last decades a number of research theories and principles have been spawned to unravel the mystery of language processing. Although still a subject of linguistic debate, language comprehension and production have been researched from various angles which, subsequently, has untangled a variety of knots in language processing. In both language comprehension and production, verbs and their argument structure have been well-researched among individuals with aphasia in comparison to neurotypical individuals (Bastiaanse et al., 2002; Dragoy & Bastiaanse, 2010; Kim & Thompson, 2000; Malyutina & Zelenkova, 2020;
Shapiro et al., 1993; Thompson, 2003). The investigation of verbs can enhance the approaches to design and carry out language treatments for individuals with aphasia. For example, in an aphasia intervention study the participant with chronic agrammatic aphasia improved sentence production after targeting the treatment of lexical argument structure of verbs (Whitworth et al., 2015).
Furthermore, studying verbs can reveal patterns of child language acquisition. For instance, investigating transitive and intransitive verbs in sentences showed evidence how young children learn lexical argument structure information from discourse (Naigles &
Maltempo, 2011). It is noteworthy that in the lexical argument structure of verbs, nominals (e.g. nouns, noun phrases) play an important part because the arguments of verbs are often realized by nominals (Haegeman, 1994).
From word recognition to sentence processing, lexical argument structure has played a pivotal role in investigations of language processing (Carlson & Tanenhaus, 1988).
Previously studied among individuals with and without brain disorders, lexical argument structure of verbs has been greatly researched in the context of language production paradigms (Barbieri et al., 2010; Thompson, 2003), as well as sentence comprehension paradigms (Shapiro & Levine, 1990). Outside of the sentential context, lexical argument
structure has been explored in limited studies (McRae et al., 2005; Thompson et al., 2007).
The objective of the current study was to explore lexical argument structure outside of sentential context, in a word-word priming paradigm, where nouns primed verbs.
Particularly, we set to explore nouns’ argument status projection onto verbs. The argument status of a noun signified whether a noun acted as an argument of a specific verb, or not.
In the introduction, section 1.1 presents relevant information about lexical argument structure, followed by section 1.1.1 addressing the selectional rules of features that verbs assign to their arguments. Section 1.2 presents information about the Modern Eastern Armenian language, as a testing ground, followed by sections 1.3 and 1.4 exploring representation of lexical-syntax in production and comprehension models. Furthermore, section 1.5 describing psycholinguistic and neurolinguistic evidence on argument structure is presented. Finally, in section 1.6 we turn to goals, hypothesis and predictions of the present study.
1.1 Lexical Argument Structure
Lexical argument structure is usually described as a word’s ability to be combined with other words and phrases through syntactic and semantic information. Lexical argument structure also shows what kind of arguments and how many of them a word can have. The combination of this information results in meaningful content. For instance, lexical argument structure of verbs, also referred to as verb argument structure, contains syntactic and semantic information about verbs (Trueswell & Kim, 1998; Müller & Wechsler, 2014).
Subcategorization frames store syntactic information (e.g. types of complements) associated with verbs, while thematic roles incorporate semantic information (e.g. Agent, Patient, Experiencer) about verbs and the arguments they take (Chomsky, 1969; Haegeman, 1994).
For example, the verb “kill” in example (1) below expresses an action that involves two obligatory participants (arguments). With respect to the semantic information it carries, the
verb “kill” assigns the thematic role of Agent to its subject “Maigret” and the thematic role of Patient to its direct object “Poirot” (Haegeman, 1994). Regarding the syntactic information carried by the verb “kill” in example (1), one can note that “kill” is a transitive verb taking the noun phrase (NP) “Poirot” as its complement.
(1) Maigret killed Poirot. (Haegeman, 1994, p. 49)
In previous literature, the importance of verbs’ role in sentence comprehension has been discussed thoroughly. Verbs are thought to be accountable for the grammatical structure of the language(Tomasello, 1992). This part of speech is thought to take the core responsibility of children’s development of grammatical proficiency (Tomasello, 1992). When it comes to sentence comprehension, the central role of verbs has been discussed within the framework of ambiguity resolution in sentences with direct objects or sentence complements, such as in examples (2) and (3) (Garnsey et al., 1997).
(2) The gossipy neighbor heard (that) the story had never actually been true. (Garnsey et al., 1997, p. 89)
(3) The gossipy neighbor heard (that) the house had never actually been sold. (Garnsey et al., 1997, p. 89)
In example (2) the NP “the story” can be interpreted in two ways, as a direct object of the verb “heard” or as a subject of the complement sentence “the story had never actually been true”. After conducting an eye-tracking study, Garnsey et al. (1997) concluded that it is not only the grammatical property of the verb (taking a direct object or a complement) that affects the interpretation of the sentence, but also the semantics of the NP following the verb.
The NP “the story” in example (2) is more likely to be encountered as a Theme for the verb
“heard” than the NP “the house” in example (3). Thus, it can be concluded that both syntactic and semantic information carried by the verb affect sentence processing. MacDonald,
Pearlmutter and Seidenberg’s (1994) constraint-based lexicalist theory of sentence processing
has been the central theory explaining ambiguity resolution studies such as the one mentioned above. This theory suggests that when a verb is encountered and recognized, its lexical argument structure information is activated (MacDonald et al., 1994). This is why there are multiple parsing options available during sentence processing.
Previous research has also addressed processing of lexical-semantic information, particularly, thematic role assignment based on experiential knowledge outside of a sentential context (McRae et al., 2005). The goal of the study by McRae and colleagues (2005) was to find out whether a single noun can carry information that allows the comprehender to
generate expectancy from nouns to verbs during a word-word priming paradigm. In McRae et al.’s (2005) priming study with a word-naming paradigm, nouns denoting Agents, Patients, Location, and Instruments primed their target verbs. Interestingly, they found that
comprehenders do generate information about the upcoming verbs based on the nouns’
thematic features and world-knowledge (e.g. Agent: assassin → killing; Patient: evidence → examined; Instrument: axe → chopping; and Location: casino → gambling). The findings of this study were explained by constraint-based probabilistic accounts (Trueswell et al., 1994), which assume that world knowledge of the comprehender leads to activation of nouns creating expectancies for the upcoming verbs. Similar to the study by McRae and colleagues (2005), the current study implemented a word-word priming paradigm, where nouns acted as primes for verbs. In contrast to McRae and colleagues’ (2005) study exploring expectancy generation from nouns based on event knowledge, the current study explored the argument status of nouns (i.e. whether the noun fulfils the thematic requirements to be an argument of a specific verb or not), inquiring whether the information about nouns’ argument status primed their verbs. In brief, our study pinpointed lexical-semantic restrictions rather than thematic information based on event knowledge the nouns carry to yield priming effects.
1.1.1 Selectional Restrictions of Verbs for their Arguments.
Syntactic categories of words, such as nouns or verbs, determine the environment in which these words can occur. This environment is also referred to as the words’ distribution.
Hence, in sentential context, words of one syntactic category cannot usually be exchanged with words of another syntactic category, as this will affect the grammaticality of the sentence (Haegeman, 1994). However, not all words of the same syntactic category have the ability to act interchangeably in a sentential context, as is the case in example (4).
(4) Jeeves will meet his castle at the employer. (Haegeman, 1994, p. 38)
Although not ungrammatical, the constituents of the above-mentioned sentence communicate semantic incompatibility. The disharmony in example (4) is due to the fact that the animacy of nouns has been disregarded. Animacy of nouns is a feature which defines whether a noun is living or not (Caplan et al., 1994). Furthermore, people’s general event knowledge dictates that the verb “meet” requires animate participants to accomplish the action (for an exhaustive explanation, see Haegeman, 1994).
Usually associated with predicates, there are semantic features that a verb requires the NPs, acting as verb’s arguments, to have (Caplan et al., 1994; Chomsky, 1969; Gayral et al., 2000; Grimshaw, 1990). These are the so-called selectional rules (Chomsky, 1969), that assign specific features of the verb to its arguments such as animacy and humanness, or the grammatical position of NPs (Caplan et al., 1994). In the case of the verb “ amuse”, the selectional rule of animacy requires a human argument to act as an Experiencer for it (Caplan et al., 1994, p. 551). In example (5), the animate noun “Anna” acts as an Experiencer.
Whereas in example (6) the NP “the book” cannot act as an Experiencer, as it does not fulfil the selectional rule of animacy of the verb “amuse”.
(5) The boy amused Anna.
(6) The boy amused the book.
Animacy of nouns is one of the most commonly investigated semantic features in the framework of selectional rules (Caplan et al., 1994; Gayral et al., 2000). Caplan and
colleagues’ (1994) investigation of grammatical positions and animacy with a behavioural paradigm yielded robust results indicating the crucial role of the subject’s and object’s animacy in sentence comprehension. Caplan et al. (1994) demonstrated that sentences containing animate subjects were comprehended faster and more accurately than sentences containing inanimate subjects. Besides, sentences with a verb selecting an animate subject were comprehended faster and more accurately than sentences with a verb selecting an animate object. Similar results were found in a study conducted by Weckerly and Kutas (1999), who investigated an extended question regarding animacy effects during
comprehension, particularly focusing on the time-course of this effect. Weckerly and Kutas (1999) juxtaposed relative clauses with an animate/inanimate subject, and main clauses with an animate/inanimate subject in object related sentences to find out whether animacy affects sentence comprehension (e.g. “The editor that the poetry depressed recognized the publisher of the struggling....” compared to “The poetry that the editor recognized depressed the publisher of the struggling….”) (p. 561). When a manipulation of animacy in main and relative clause nouns was created, the ERP patterns showed that animacy “was noted”
(Weckerly & Kutas, 1999, p. 565), which replicates the previous findings (Caplan et al., 1994). Weckerly and Kutas (1999) concluded that animacy was noted by participants, because two main clause nouns in two conditions matched on frequency and length and differing by animacy only, yielded different ERP patterns. Furthermore, they found that animacy came into effect quite early in sentence processing: at the moment when the noun is encountered. This finding was interpreted as a proof of early interaction between semantic and syntactic information.
1.2 Modern Eastern Armenian as a Testing Ground
Modern Eastern Armenian (Armenian), an independent branch of the Indo-European language family and a variety of Modern Armenian, was selected as a testing ground for the experiment in this study. Armenian is the standard language of the Republic of Armenia and Nagorno Karabakh. It is described as subject-object-verb (SOV) language with a somewhat free word order and a rich case marking of nouns (Tamrazyan, 1994). In Armenian, within the noun class, the humanness of nouns comes into play in the case system, particularly in marking of direct objects (Dum-Tragut, 2009). The dative case is used for human direct object nouns and the nominative case is used with non-human ones (Dum-Tragut, 2009). In Armenian the animate/inanimate distinction is based on humanness of the nouns. If the nouns denote a person, then they are considered animate. Conversely, the nouns not expressing a meaning of a person are considered to be inanimate (Dum-Tragut, 2009). The latter group includes animals and plants (Dum-Tragut, 2009). As reported by Dum-Tragut (2009), in informal speech the concept of humanness is being extended to animals. That is, when speaking of specific animals, they are treated as animate entities and are assigned the dative case (e.g. “Ես սիրում եմ իմ կատուներին” /jɛs sirum ɛm im katunɛɾin/ (I love my cats)).
Otherwise, when speaking of unspecified animals, they are assigned the accusative case (e.g.
“Ես սիրում եմ կատուներ” /jɛs sirum ɛm katunɛɾ/ (I love cats)).
The complexity of verbs is demonstrated in the variety of verbs including transitives, intransitives and inchoatives, as well as in the rich tense and aspect, and mood systems. As described by Dum-Tragut (2009), the differentiation of transitive and intransitive verbs can be achieved by morphological and analytic transformations. A transitive verb is not usually causativised morphologically, but rather with analytic means (Dum-Tragut, 2009). For instance, the causative form of the transitive verb “սպանել” /spɑnɛl/ (to kill) is formed by adding the verb “տալ” /tɑl/ (to give) to the first verb, resulting in the analytic causative form
“սպանել տալ” /spɑnɛl tɑl / (to make kill). Otherwise, the verb “սպանել” /spɑnɛl/ (to kill) would be causativized morphologically with the help of the bound morpheme “-եցն-” /ɛtshn/, resulting in the ungrammatical form “սպանեցնել” /spɑnɛtshnɛl/ (to make kill).
There are several reasons why Armenian was chosen as the research language for this experiment. Firstly, as an understudied language, the presence of various linguistic
phenomena (e.g associative and semantic relatedness effect, word frequency effect, word length effect) has not been documented in Armenian. Verb argument structure property effects, in particular, have not been explored in Armenian, a language with flexible syntactic structure and morphology. Many areas of research incorporate verb argument structure properties and their effects, including language acquisition studies, developmental disorder, and treatment studies. It is of crucial importance to study verb argument structure and its effects in Armenian to advance the research in areas such as aphasia and SLI studies, first and second language acquisition research and others.
Besides, various linguistic corpora are absent from the Armenian language which creates additional difficulties for researchers. Thus, with the help of the current research a limited corpus of words and nonwords fulfilling certain criteria would be available for future linguistic research in Armenian.
1.3 Representation of Lexical - Syntax in Word Production Models
A large body of knowledge pertaining to verb argument structure comes from studies investigating it in language production tasks and explaining the results in the framework of word-production models (Thompson, et al., 2010; Thompson, 2003; Thompson et al., 2007).
Furthermore, the claim that both language comprehension and production share the same structural reserves, especially when it comes to syntactic priming showed by Bock and colleagues (2007), allows us to explain verb argument structure within both production and
comprehension models. Hence, it is important to introduce key information about both types of theories of language processing.
Lexicalist word production models describe the processing of lexical-syntax through investigation of speech errors occurring during speech production (Dell, 1986), as well as through utilization of experimental designs with RTs (Levelt et al., 1999). These techniques made the development of theories and models about speech production and comprehension possible. Lexicalist theories of production are divided into weak and strong ones. The weak view poses that the retrieval of grammatical properties of a word is activated only if
necessary, while the strong view assumes that the retrieval of grammatical properties is always automatic and necessary (Goldrick et al., 2014).
Supporting the weak lexicalist view, Levelt and colleagues (1999) developed a serial two-step model, according to which word production starts when a concept activates the lexical-semantic level. The activation of this level then leads to activation of the lemma level.
These two levels remain activated simultaneously. The lemma level contains syntactic features of a lexical entry. At this level, grammatical choices, such as tense, gender, or other features, are made. After this information is activated, the model predicts activation of the phonological level where selection of word form is made. This model was designed to explain experimental results of unimpaired speakers. According to Levelt and colleagues (1999), word recognition from both spoken and written modalities activates its syntactic properties described as the “perceptual equivalent of the lemma” (p. 7). This assumption by Levelt and colleagues (1999) allows them to conclude that lemma level and all previous levels overlap for both production and comprehension. It is noteworthy, that in their model, Levelt and colleagues (1999) indicate that grammatical properties of a word are represented as empty lemma nodes (Nickels, 2001). Selection of the lemma node is mandatory for the selection of grammatical properties; however, selection of an activated grammatical property
is executed only if required (Roelofs et al., 1998). Therefore, in single word processing, syntactic properties are assumed not to be activated, in contrast to sentence processing (Goldrick et al., 2014).
Similar to the previous model, the interactive activation model (Dell et al., 1997) also presupposes the existence of a lexical-semantic level, which activates the lemma level
followed by a phoneme level. As the name interactive activation suggests, the model involves an interactive relationship between the levels, which differentiates it from Levelt et al.’s (1999) two-step model. This model also incorporates lemma nodes which include grammatical property information. An activated node at any level also activates any
connected nodes. This whole activation process leads to selection of the most activated node followed by phoneme selection at the phoneme level (for further details, see Biedermann et al., 2018; Nickels, 2001). This model predicts automatic and necessary activation of syntactic forms in single word production (Goldrick et al., 2014).
Among several dissimilarities between the two accounts described above, activation of lexical-syntactic information before the access of the phonological form is noteworthy. As described in previous paragraphs, regarding the access of lexical-syntactic information, Dell and colleagues’ (1997) account states that before the phonological form is accessed, this information is obligatorily selected. On the contrary, Levelt and colleagues (1999) argue that, although the lexical-syntactic information is activated, it is not necessarily selected before the access of the phonological form. The word necessarily is a key concept in this context, as it presupposes that, while in sentence processing grammatical information is activated, in single word processing the lemma level is activated but syntactic properties are only retrieved when required by the task at hand. Single word processing should not raise any need to activate syntactic properties, as these properties are characteristic to sentence construction (Chomsky
& Lightfoot, 2002).
It is reasonable to speculate that with lemma selection, the argument structure properties of verbs and nouns should be activated in both production and comprehension.
This speculation is based on the above-mentioned accounts, and particularly on Levelt and colleagues’ (1999) explicit remark of the overlap of all the levels of language processing down to the lemma level. The overlap assumes merging of production and comprehension processes. Furthermore, the speculation is based on Bock and colleagues’ (2007) claim of shared structural reserves between production and comprehension modalities. The mentioned reasons motivated us to base our hypothesis on both production and comprehension theories.
1.4 Representation of Lexical-Syntax in Sentence Comprehension Models Theories of sentence comprehension, including constraint-based accounts and the garden-path theory, also contribute to the knowledge of lexical argument structure processing (MacDonald et al., 1994; Frazier & Rayner, 1982). These theories aim to explain how
comprehenders resolve ambiguities in a sentence at different levels of language processing, such as at meaning or syntactic levels.
The constraint-based theories imply that sentence comprehension or ambiguity resolution is achieved by utilizing more than one so-called constraint (information source) as the sentence is being read or heard (Matsuki, 2013). According to MacDonald et al.’s (1994) constraint-based interactive activation theory, argument structure information is represented in a word’s lexical entry. Alongside argument structure information, a word’s lexical representation also involves information of properties that are phonological, orthographic, semantic, grammatical, frequency related, as well as syntactic (MacDonald et al., 1994).
Their view of syntactic ambiguity resolution, assumes a “winner-take-all” attitude (MacDonald et al., 1994, p. 686). During comprehension of the sentence “John cooked”
(MacDonald et a., 1994, p. 687), when the first word is encountered, all the information levels are activated at this point, including argument structure and thematic information.
When the second word is encountered, information about the second word’s lexical entry, including argument structure and thematic roles, is also activated. With all possible argument structure constraints, such as different thematic roles, the processing system satisfies those that correspond for both words in the sentence, thus leading to successful sentence
comprehension (i.e. ambiguity resolution). MacDonald and colleagues (1994) assume that syntactic ambiguity resolution may be supported by the production system, or other possible systems.
The garden-path model (Frazier & Rayner, 1982), proponing serial and modular theories of sentence comprehension, postulates that when coming across an ambiguity in a sentence, the human parser trails a single syntactic interpretation. This interpretation may or may not be correct, thus leading the parser to a garden path in case of an incorrect
interpretation. To correctly interpret the sentence, the human parser goes back and conducts a reanalysis, which makes the comprehension process longer. After reanalysis, only during the final interpretation of the sentence, other types of information, such as semantic or thematic properties, are utilized. The whole comprehension process is driven by phrase structure rules, which direct the initial analysis of the sentence. Thus, the initial analysis is based on syntactic information only, excluding even thematic information. The initial syntax-only activation is an inherently different point to MacDonald and colleagues’ (1994) constraint-based theory predicting parallel activation of different types of information.
Therefore, according to the constraint-based theory by MacDonald et al. (1994), when a word in a sentence is encountered, all its syntactic and lexical properties are activated. As per the garden-path theory (Frazier & Rayner, 1982), when a word in a sentence in
encountered, initially only its syntactic properties (based on phrasal rules), are activated.
Here, one can draw parallels between the interactive activation models of production and comprehension, and the two-stage models of production and comprehension based on the
types of information that is activated and used during an initial encounter with a word.
Similar to Dell et al.’s (1997) model predicting immediate activation of argument structure properties at the lemma level, MacDonald et al.’s (1994) theory predicts immediate activation of argument structure properties when a word is encountered. In the sentence “John cooked”
(MacDonald et al., 1994, p. 688), when encountering the word “John” its lexical
representation including the role of Agent, alongside other types of information, is already activated before moving on to the word “cooked”. When the word “cooked” is activated, similarly, its argument structure and other types of information are activated. As stated by MacDonald et al. (1994), during the processing of a lexical entry in a sentence, its activated properties will compete with each other similarly to competition during single lexical entry processing.
With respect to the two-stage models, Levelt et al.’s (1999) word production model predicts activation of argument structure information only when necessary, which is in a sentential context. Thus, when a single word is to be produced, its argument structure properties are not activated. Similarly, the garden-path model (Frazier & Rayner, 1982) of comprehension predicts that during the initial encounter of a word its argument structure information is not activated. The only initially activated information is syntactic, based on phrasal rules. The argument structure information is only activated after the first syntactic interpretation of the sentence is completed.
In a study investigating how event-specific world knowledge or thematic fit affects ambiguity resolution in a sentential context, McRae et al. (1998) explored the role of nouns as verbs’ argument, contrasting the two-stage and constraint-based models. They manipulated the thematic fit constraint in reduced relative sentences by varying the initial NP as an Agent or a Patient for the relevant verb (e.g. The waitress served by the trainee… vs. The customer served by the trainee…) (McRae et al., 1998, p. 307). They found that thematic fit
information in the initial NP shortened the reading times by 42 milliseconds (msec) at the Agents (e.g. The cop arrested…) and 38 msec at the Patients (The crook arrested…). This finding supports the constraint-based accounts (MacDonald et al., 1994), pinpointing that utilization of constraints to resolve ambiguities is a product of information access very early after encountering the words in the sentence. This information provides evidence against the garden-path model (Frazier & Rayner, 1982), which predicts late activation of information such as thematic fit. Furthermore, they provided evidence against the two-stage models by simulating reading time data as predicted by Frazier and Rayner’s (1982) garden-path model and comparing it to simulated data as predicted by the constraint-based models (e.g.
MacDonal et al., 1994). The simulated data predicted by the constraint-based models proved to be a significantly better predictor of the actual reading time data, as compared to the simulated data predicted by the garden-path model.
1.5 Lexical Decision Tasks and Lexical Argument Structure in Psycholinguistic and Neurolinguistic Literature
Verb processing and verb argument structure have been explored in studies implementing lexical decision paradigms, both among neurotypical individuals and individuals with brain damage (Shapiro et al., 1993; Thompson et al., 2010; Thompson et al., 2007). Shapiro and colleagues (1993), for instance, provided evidence that thematic properties of verbs are automatically and exhaustively activated in sentence comprehension when encountering a verb during a cross-modal lexical decision task, irrespective of the type of sentence (active, passive, cleft-object, cleft-subject). The experiment involved auditory sentence presentation and visual word/nonword lexical decision task during the unfolding of the sentence. This evidence came from experiments conducted among neurotypical participants and ones with Broca’s aphasia. In the case of the latter group, the activation was observed even without full comprehension of some types of experimental sentences (Shapiro et al., 1993).
Thompson and colleagues (2007, 2010) implemented a lexical decision paradigm in their studies investigating neural correlates of verb argument structure. It was found that in word comprehension one argument verbs yielded significantly longer RTs as compared to two- argument and three-argument verbs among neurotypical participants (Thompson et al., 2007).
This evidence was not consistent with their later study conducted among older individuals without brain disorders and age-matched individuals with aphasia (Thompson et al., 2010).
Nevertheless, in both studies brain activation in similar areas was observed when processing verb argument structure. Based on the evidence from the above-mentioned studies, one could conclude that argument structure properties play a role not only in sentence, but also in single word processing. The activation of argument structure properties is consistent with strong lexicalist theories assuming activation of lexical-syntactic properties of lexical entries at the lemma level.
Experiments not only with a lexical decision paradigm, but also with language production tasks yielded evidence about effects of argument structure on verb
processing. The argument structure complexity hypothesis (ASCH) (Thompson, 2003) assumes that production of verbs with a syntactically more complex (movement from d- structure to s-structure) or larger number of arguments is more challenging for
individuals with aphasia. This conclusion was drawn from results of narrative language production and single verb naming and comprehension tasks. The results from the verb naming (picture naming) and comprehension (word-to-picture matching) tasks, as single word processing tasks, are more relevant to the current study. According to these tasks, comprehension of the unaccusative (e.g. break) and the unergative verbs (e.g. jump) was unimpaired, as measured by accuracy rates. Whereas naming of the unaccusative verbs was significantly more impaired than naming of the unergative verbs. These results were attributed to the more complex argument structure of unaccusative verbs (having no
external argument, but only an internal one). According to Thompson (2003), this is due to the automatic access to argument structure properties at the lemma level,
demonstrated by the greater difficulty to select verbs with a more complex argument structure. These findings are in line with the strong lexicalist views of processing. We can speculate that measuring RTs alongside accuracy would have yielded evidence supporting the strong lexicalist views, as opposed to the null results in the
comprehension task. This is because RTs can act as a more sensitive measure for underlying mental processes than only accuracy rates.
1.6 Present Study
Historically, the psycholinguistics community has been divided between those who explore language production, and those investigating language comprehension. The dissociation between the subfields is striking in the studies of brain damage and language, where one can find examples of damaged production but intact comprehension of verb argument structure (Thompson, 2003); or studies of language acquisition where
comprehension and production seem to have different time-course (for a review, see Clark &
Hecht, 1983). Nevertheless, the link between production and comprehension has been acknowledged, at least in psycholinguistic models, such as the two-stage model by Levelt et al. (1999). From the description of production and comprehension models in the previous section, the overlap regarding the representation of lexical-syntactic and lexical-semantic knowledge was apparent. The two-stage models of both production and comprehension assume non-obligatory and delayed selection of lexical-syntactic information (Frazier &
Rayner, 1982; Levelt et al., 1999). Meanwhile the interactive activation models postulate immediate and obligatory selection of lexical-syntactic information (Dell et al., 1997;
MacDonald et al., 1994).
The present body of knowledge inspired us to pose two questions pertaining to the verb-argument structure and its projection from the arguments to verbs in light of two-stage and interactive activation models. First of all, does the human parser activate and project information such as argument status of nouns from nouns to verbs? Does animacy of nouns, as a selectional restriction and a constraint, play a role in priming verbs when manipulated alongside argument status? The current study employed a noun-verb priming paradigm to answer the aforementioned questions by testing Armenian-speaking individuals. A noun was labeled as having an argument status if it was more likely to be an argument of a certain verb, such as the noun “ուսուցիչ” /usutshitʃh/ (teacher) as a semantically related argument of the verb “կրթել” /kəɾthɛl / (to educate). As mentioned previously, a noun was labeled as not having an argument status if it was not likely to be an argument of a certain verb, such as the noun “մանկապարտեզ” /mɑnkɑpɑɾtɛz/ (kindergarten) as a semantically related noun but not an argument of the verb “կրթել” /kəɾthɛl / (to educate). A priming paradigm was implemented because word-word priming tasks are a conventional technique to study how related information is activated in the brain when hearing or reading a word (McRae et al., 2005). Word priming studies in word comprehension and production mainly focus on two types of prime-target relationships, namely semantic and associative priming (for review, see Hutchison, 2003). The focus of semantic relationships is the meaning relatedness of word pairs (Williams, 1994). On the other hand, associative relationship between prime and target is mainly concerned with word use, rather than meaning relatedness (Fischler, 1977).
Nevertheless, as reported by Perea and Rosa (2002), both semantic and associative priming experiments demonstrate similar priming effects. As mentioned previously, McRae and colleagues (2005) investigated noun-verb pairs to tap into semantic memory. Although McRae and colleagues (2005) did not deny the associative relationship between the word pairs, they attributed the priming effects to word event knowledge. According to McRae et al.
(2005), their choice of stimuli was based on people’s knowledge of events and appropriate types of arguments, which have a general associative relationship. The experimental stimuli in the current study did not exclude a general associative relationship between primes and targets in a similar way as in the case of the study by McRae et al. (2005). The experimental stimuli involved world knowledge information which is reflected in the argument status of the nouns.
Within the framework of interactive activation models, wehypothesized that due to immediate and obligatory activation of lexical-syntactic information, a stronger semantic priming effect would be elicited in the semantically related condition with argument status nouns, as compared to semantically related conditions with nouns not having an argument status. This effect would be shown by manipulating the argument status of nouns in the semantically related conditions, so that some semantically related conditions would include argument status nouns, while others would not. Furthermore, we postulated that animacy, as a constraint and a selectional restriction, would lead to additional priming effect based on the constraint-based account. It is noteworthy that McRae et al. (1998) singled out animacy as a special conceptually-based feature that affects ambiguity resolution. Thus, we predicted that the lexical decision task would yield the shortest RTs in milliseconds for verbs when primed with semantically related, argument status, animate (SRAA) nouns. In contrast, the longest RTs would be evinced when priming the verbs with semantically unrelated, non-argument status, inanimate (SUNI) nouns. Semantically related, non-argument status, inanimate (SRNI) nouns would prime verbs to be processed faster, as compared to priming with SUNI and semantically unrelated, non-argument status, animate (SUNA) nouns. Between SUNI and SUNA conditions, SUNA nouns would prime the verbs to be processed faster. Alternatively, if the grammatical information, particularly, argument structure properties, and the animacy constraint are not necessarily activated during word processing, as postulated by Levelt and
colleagues (1999) and as predicted by the garden-path model (Frazier & Rayner, 1982), SRAA and SRNI nouns would not yield significantly different verb priming effects.
Furthermore, SUNA and SUNI conditions would not yield significantly different RTs.
However, due to semantic priming, SRAA/SRNI nouns would prime the verbs to be
processed faster as compared to SUNA/SUNI nouns. The predictions are illustrated in Table 1.
Table 1
Illustration of Predictions for RTs in Each Condition
Pred. based on interactive activation models Alternative pred. based on two-stage models RT (SRAA)
<
RT (SRNI)
<
RT (SUNA)
<
RT (SUNI)
RT (SRAA) ≈ RT (SRNI)
<
RT (SUNA) ≈ RT (SUNI)
Note: Pred. = Predictions
Method 2.1 Participants
42 Armenian native speaking individuals participated in the priming experiment (mean age = 36.85 years, standard deviation (SD) = 10.48 years). Six participants dropped out due to incomplete experiments. One participant was excluded due to an extremely large number of incorrect answers both in the training session (90% incorrect) and the
experimental session (41.96% incorrect). From the remaining 35 participants, nine were excluded due to reporting non-corrected vision. The remaining 26 participants had normal or corrected-to-normal vision. The average education of the final participants was 17.48 years (SD = 2.38). The participants were recruited through social media and word-of-mouth. There was no monetary compensation for participation in the experiment. All the participants
provided their consent before participating in the experiment. The study was approved by the Research Ethics Committee of the Faculty of Arts at the University of Groningen.
2.2 Materials
The final experimental stimuli consisted of 28 transitive verbs in Armenian, as targets.
Each target verb was paired with a prime noun with properties such as semantic relatedness to the target verb, argument status and animacy. The structure of the experimental conditions is presented in Table 2. Verb transitivity in Armenian was checked by the means of
morphological causativisation. As mentioned previously, the verbs which cannot be causativised morphologically by means of the appropriate morphological suffix are considered to be strictly transitive (Dum-Tragut, 2009). The length of the selected verbs ranged between five and ten phonemes (mean = 5.96, SD = 1.2), with only one verb having ten phonemes (ձերբակալել /dzeɾphɑkɑlɛl/ (to arrest).
Table 2
Structure of Conditions Sem. rel.
of prime
Arg. stat.
of prime
Anim.
of prime
Cond. name Target
+ + + SRAA
ուսուցիչ /usutshitʃh/ (teacher)
VERBկրթել /kəɾt hɛl/ (educate)
+ - - SRNI
մանկապարտեզ /mɑnkɑpɑɾtɛz/
(kindergarten)
- - + SUNA
պինգվին /pingvin/ (penguin)
- - - SUNI
ամպրոպ /ɑmpɾɔp/ (thunderstorm)
Note: Sem. rel. = semantically related, Arg. stat = argument status, Anim. = animacy, Cond. = condition, + sign represents presence of appropriate property, - sign represents absence of appropriate property.
In the SUNA condition, names of animals were used. Due to the humanness constraint in Armenian, it was not ideal to incorporate animal names into our experiment.
However, when creating the stimuli, we were guided by the definition of animacy brought by Caplan and colleagues (1994) introduced above, according to which animals are living creatures; hence, they are considered to be animate. Besides, our choice of stimuli can be justified, because, as mentioned above, in Armenian informal speech animals are treated grammatically as animate. Furthermore, the fulfillment of the animacy constraint for the majority of the non-argument status nouns did not seem to be feasible simply because most of the experimental verbs assign thematic roles to humans but not to animals (e.g. խոհարար /χɔհɑɾɑɾ/ (cook) խաշել /χɑʃɛl/ (to boil) (Agent); թիթեռ /thithɛr/ butterfly/ խաշել /χɑʃɛl/ (to boil) (not likely to be Agent or Theme)).
The experimental materials were selected based on a norming study of semantic relatedness. The norming study was conducted in a similar fashion to studies investigating semantic and associative priming (Ferrand & New, 2004; Moldovan et al., 2015). The current norming study was adapted from the study by Moldovan and colleagues (2015). Nineteen Armenian native speaking individuals (mean age = 31.74, SD = 10.73) took part in the semantic rating study. During the norming study, the participants were asked to rate the degree of meaning relatedness between words in pairs according to a seven-point Likert scale. The score seven meant that the words were strongly related in meaning, and the score one meant that the words were unrelated in meaning. In previous studies, where priming with unrelated items was explored, the unrelated experimental stimuli were selected from either corpora or tasks yielding pairs of semantically unrelated words (Jouravlev & McRae, 2016).
Due to the absence of such corpora or experiments in Armenian, the semantically unrelated words were selected based on the event-knowledge of the researcher. Furthermore, the unavailability of resources in the Armenian language led to a decision to have cut off points
of three and four on the seven-point Likert scale for the meaning-relatedness study. The word pairs rated three points and below were considered as unrelated in meaning; meanwhile the pairs rated four points and above were considered to be related in meaning. Furthermore, due to the absence of word-pair occurrence frequency corpora in Armenian, we assumed that semantically related pairs would co-occur more, as compared to semantically unrelated ones.
After the norming study, due to unfavorable ratings, six verbs, alongside their noun pairs, were excluded from the initial stimuli list consisting of 34 target verbs, thus leaving 28 experimental verbs. Each verb was paired with seven nouns from each condition. Four lists with 28 experimental verbs paired with nouns of each condition were created. Each list also contained the same number of filler pairs, as well as, as many word-nonword items, as there were experimental and filler items combined (Sánchez-Casas et al., 2006). In total, the experimental stimuli comprised of 112 items per list. Latin square design was implemented during preparation of the stimuli for presentation. The nonwords were created by changing two letters of a real Armenian word in such a way that the nonword strings were
phonologically permissible in Armenian. The comprehenders were tested on one of the randomly chosen lists.
The verbs were in their infinitive form and the nouns were in their bare singular form.
Care was taken to ensure that it was possible for the target verbs to assign the role of Agent to the selected argument status animate nouns in the semantically related condition. To ensure that the nouns were always interpreted as the entities they were to refer to, nouns without additional adjectival meanings were selected as primes, and this was verified using the Explanatory Dictionary of the Modern Armenian Language (1969). There was one exception (the noun “ձիավոր” /dziɑvɔɾ/ (horseback rider)), which has both noun and adjectival meanings.
Semantic relatedness of the word pairs was controlled for, ensuring difference in semantically related and unrelated conditions, and no difference between semantically related conditions, and between semantically unrelated conditions, as measured by Dunn test. The results are presented in Table 3. The four experimental conditions had a similar average number of syllables (between 2.76 and 2.79),as measured by Kruskal-Wallis rank sum test (H(3) = 0.88, p = 0.83), and similar average number of phonemes (between 6.85 and 7.29), as measured by one-way ANOVA (F(3), p = 0.44) . The frequency of all items was measured by counting the number of hits on Google search and log-transforming them (Marusch et al., 2017), resulting in a similar average frequency number among the conditions (between 5.68 and 5.83), as measured by one-way ANOVA (F(3) = 1.52, p = 0.22).
Table 3
Dunn Test Results for Semantic Relatedness between Conditions
Comparison z-value Unadj. p-value Adj. p-value
SRAA & SRNI 0.88 0.38 0.76
SRAA & SUNA 7.20 0.00 0.00
SRNI & SUNA 6.31 0.00 0.00
SRAA & SUNI 6.61 0.00 0.00
SRNI & SUNI 5.73 0.00 0.00
SUNA & SUNI -0.58 0.56 0.56
Note: Unadj. = unadjusted, Adj. = adjusted.
All four conditions were also matched in phonological distance between the noun and the target verbs using the Levenshtein distance measure (between 6.18 and 6.75), as
measured by Kruskal-Wallis rank sum test (H(3) = 2.62, p = 0.45) (Sanders & Chin, 2009).
The Levenshtein distance was measured using an online distance calculator supporting the Armenian script (https://planetcalc.com/1721/)). As in the Armenian language some graphemes might express more than one phoneme and two graphemes might express less than two phonemes, the words were entered into the calculator according to their phonemic
representation, rather than regular written forms. The grapheme “ե” /jɛ/, expressing two phonemes was entered as “յե” /jɛ/ in the calculator to convey the appropriate phonemes. The grapheme “և” /jɛv/, expressing three (e.g. in ևս /jɛv/ (too)) or two (e.g. in բարև /barɛv/
(hello)) phonemes depending on the location in the word was entered as “յեվ” /jɛv/ or “եվ”
/ɛv/ in the calculator. The graphemes “ու” /u/ expressing one phoneme were entered in the calculator as “ւ”. Furthermore, when necessary, the grapheme “ը” /ə/ was entered to convey the schwa phoneme. For instance, the word “եղևնի” /jɛʁɛvni/ (spruce) was entered as յեղեվնի /jɛʁɛvni/, the word նվաճել /nəvatʃɛel/ (to conquer) was entered as նըվաճել /nəvatʃɛel/, and the word ուսուցիչ /usutshitʃh/ (teacher) was entered as ւսւցիչ /usutshitʃh/ . 2.3 Procedure
Participants received a participant number from the researcher and the link to the information leaflet, the consent form and the background information form, as well as a link to the online experiment. Before the experiment, the experimental procedure was explained to them via video/audio-chat or in writing, and some instructions were given regarding using a computer instead of a smartphone or a tablet, and sitting in a quiet room. Afterwards, participants were asked to read the consent form and the information leaflet and give their consent, as well as fill out the background questionnaire, before proceeding to the
experiment. Next, participants were instructed to start the experiment. First, written
instructions were presented on the computer screen. Participants were instructed to read two subsequent words presented on the computer screen silently. After reading each second word, participants had to make a lexical decision. They were instructed to press the “l” key on the keyboard if they thought that the second word was a real word (e.g. խմել /χəmɛl/ (to drink)), and to press the “a” key on the keyboard if they thought the second word was a non-word (e.g. “շնոդել” /ʃənɔdɛl/). The instructions were followed by a training session. During the training session participants read ten noun-verb pairs, which were not included in the
experimental session, and made a lexical decision. After the training session, participants were reminded of the experimental procedure and instructed to press the “space” key to start the experimental session. Each trial began with a fixation cross (+) presented centrally for 250 msec, followed by the prime for 200 msec. Immediately after the prime, the target was presented until the button was pressed by the participant. The inter-trial interval comprised 1500 msec (Figure 1).
Figure 1. Experimental design of each trial with examples from the experimental stimuli;
Trial starts with a fixation cross, and then appears the prime, followed by the target until button press.
A short stimulus onset asynchrony (SOA) paradigm for prime-target presentation was used to make sure that the participants did not have enough time to strategically generate expectancies for the target word (McRae et al., 2005). The button press after every target measured the response latency in seconds from the onset of the target. The stimuli presentation within the lists was randomized and each participant saw only one list. The experiment lasted approximately six to eight minutes. The experiment was constructed using PsychoPy v2020.1.3 (Peirce et al., 2019). The online platform Pavlovia (https://pavlovia.org/) was used to implement the online data collection.
+ 250 msec
Prime e.g. ուսուցիչ /usutshitʃh/ (teacher)
200 msec
Target e.g. կրթել /kəɾthɛl/ (educate)
Until press 1500 msec intertrial
2.4 Power Calculations
To make sure that the study was sufficiently powered, care was taken to conduct calculations of the statistical power with the given variables including the number of stimuli, the number of participants, as well as, argument status, animacy and semantic relatedness.
A Monte Carlo simulation in the R version 4.0.1 programming environment (R Development Core Team, 2015) was used for the power analysis in RStudio 1.1.456 (RStudio Team, 2016). The independent variables of semantic relatedness (related vs. unrelated), animacy (animate vs. inanimate), argument status (argument vs. non-argument), item number as a random effect (28 target verbs and 7 prime nouns in every condition for each participant) and the participant number as a random effect (30 participants) were used to simulate RT data with a within-participant standard deviation of 100 msec, a standard deviation for each verb as a random factor of 20 msec and a standard deviation for each participant as a random factor of 30 msec. The simulated data were fitted into a linear mixed model using the
“lmerTest” package (Kuznetsova, 2017) with the RT data as the response variable; sematic, animacy, argument status data as fixed effects and participant number as a random effect.
Verbs as a random effect were excluded from the model to avoid convergence problems. The Monte Carlo simulation was run 10000 times, with assumptions of baseline mean RT of 650 msec (McRae et al., 2005), the effect of semantic relatedness of 60 msec (Holderbaum & de Salles, 2011), effect of argument status of 40 msec and effect of animacy of 50 msec. The effect sizes of argument status and animacy were assumed to be similar to, but smaller than, semantic relatedness effect. These assumptions were sufficient to empower the experiment with a p < 0.05, leading to a power of 0.728 for argument status, 0.997 for semantic relatedness and 0.984 for animacy.
2.5 Data Analysis
The data analyses were conducted in R version 4.0.1 (R Development Core Team, 2015) programming environment in RStudio 1.1.456 (RStudio Team, 2016). The accuracy data were analyzed in terms of percentages of correct responses to filler and experimental items in total, each experimental condition separately, as well as, accuracy of answers by participants. Furthermore, generalized linear mixed-effects model (GLMM) from the “lme4”
package (Bates et al., 2015) was fit to the accuracy data to find out whether the response variable (accuracy) was affected by the independent variables of main interest (semantic relatedness and argument status).
CorrectPress ~ Sem. + Arg. + (1|participant)
The accuracy percentage, as well as, z-values, coefficients and standard errors of the GLMM are reported, where a z-value above 2 or below -2 reports a significant effect at the alpha level of 0.05. The RT data were analyzed after removing the inaccurate responses to the lexical decision task, as well as, all filler items, leaving 721 observations. The RTs were transformed from seconds to msec for convenience purposes. Initial visual observation of raw RTs with a normality quantile-quantile plot, shown in Figure 2, revealed an extreme outlier above 60000 msec, which was removed from the data, leaving 720 observations.
Figure 2. Q-Q plot of raw RTs
T-values, coefficients and standard errors of the linear mixed-effect model (LMM), from the “lme4” package (Bates et al., 2015), are reported, where a t-value above 2 or below -2 indicates a significant effect at the alpha level of 0.05. LMM was fit to the remaining data to check the LMM assumptions of linearity, homoscedasticity and normality. According to visual inspection, and tests of homoscedasticity (Levene’s test, F = 6.83, p = 0.009) and normality of residuals (Shapiro-Wilk test, p = 2.2e-16), none of the assumptions were met as presented in Figure 3.
A B
C D
Figure 3. Residual plots of model before transformation; Plot A shows the absence of linearity of residuals; Plot B shows the heteroscedasticity of the residuals; Plot C and Plot D show the not normally distributed residuals.
In order to avoid the extreme skewness of residuals, the raw RTs were subjected to Box-Cox transformation (Box & Cox, 1964), using the “MASS” package (Venables &
Ripley, 2002), as exemplified below.
𝑅𝑇 = (RT ^ lambda - 1)/lambda
The optimal lambda value of -0.7 was found by evaluating lambda values ranging between -5 to 5 in steps of 0.1 maximizing the log-likelihood of a normally distributed curve.
The transformed data met the assumptions of homoscedasticity of residuals measured by Levene’s test (F = 1.34, p = 0.25), however, it failed to meet the assumption of normality of residuals as measured by Shapiro-Wilk normality test (p = 5.484e-06). Nevertheless, visual inspection showed the large amount of improvement in normality, as presented in Figure 4.
A B
C D
Figure 4. Residual plots of model after transformation. Plot A shows the linearity of residuals; Plot B shows the homoscedasticity of the residuals; Plot C and Plot D show the normally distributed residuals.
This allowed us to fit the LMM to the RT data, as a response variable, to measure the effect of semantic relatedness, argument status and animacy, as fixed effects, and verb type and participant, as random effects on the response variable.
As the selected Armenian target verbs appeared to be longer in phonemes than the target English verbs in similar studies (Gomes et al., 1997), additional exploratory word- length correlation analysis was conducted. The purpose of the correlation analysis was to find out whether verbs’ length in phonemes was associated with the length of the response times.
As mentioned above, although most of the verbs had a similar number of phonemes (five to six phonemes), some verbs were eight phonemes long and one verb was ten phonemes long.
Nevertheless, the number of phonemes in verbs was not expected to be a contributing factor to the potential semantic relatedness, argument status and animacy effect sizes, as all the verbs were presented in every condition.
Results 3.1 Accuracy Analysis
In the dataset, the number of all items including fillers comprised 2912 observations, out of which 728 were experimental items. Among the participants, 97.73% (2846
observations) of all items received a correct answer. Besides, 97.3% (2152 observations) of all filler items received a correct answer. When the data were subsetted leaving only experimental items, it was revealed that the percentage of correct experimental items comprised 98.94% (721 observations), as presented in Table 4.
Next, the accuracy of the experimental items in each condition was measured. The results are as follows: in the SRAA condition 99.45% (181 observations) of all experimental items received accurate answers. In the SRNI condition the accuracy of response was 100%
(182 observations). Meanwhile, in the SUNA condition the accuracy rate decreased
comprising 98.35% (179 observations). In the SUNI condition the accuracy of experimental
conditions was measured to be 98.35% (179 observations), as presented in Table 4.
Subsequently, the accuracy rate of experimental conditions for each participant was measured which ranged between 92.86% and 100%.
Table 4
Descriptive Data of Accuracy
Total items 2912 Exp. items
728
Filler items 2184 Total correct items
2846 (97.73 %) Cor. exp. items
721 (99.04%)
Cor. filler items 2125 (97.3%) SRAA correct
181 (99.45%)
SRNI correct 182 (100 %)
SUNA correct 179 (98.35%)
SUNI correct 179 (98.35%) Note: exp. = experimental, cor. = correct.
Among all participants, 76.92% (20 individuals) answered the lexical decision task 100% accurately, while 19.23% (5 individuals) of participant had 96.43% accuracy and 3.84% (1 individual) of participants had 92.86% accuracy. Next, a generalized linear mixed model with a binomial link function was fit to the accuracy data. In the model, the correct response to the lexical decision task was the dependent variable. To avoid non-convergence of the GLMM due to a small number of incorrect responses, a simplified model with only the main effects of interest included as fixed effects (semantic relatedness and argument status) was fit with participant number included as a random effect.
CorrectPress ~ Sem. + Arg. + (1|participant)
The model did not yield any significant effects of accuracy, as predicted by semantic relatedness and argument status of the prime in the experimental conditions according to the z-values, as presented in Table 5.
Table 5
Summary Results of Accuracy
Fixed effects Estimate SE z-value
Intercept 4.37 0.79 5.54
Sem. rel. 21.60 190.98 0.11
Arg. stat. -20.49 190.98 -0.11
Note: SE = Standard error.
3.2 Reaction Time Analysis
To test the hypothesis, initially, a model was built including semantic relatedness, argument-status and animacy as fixed effects, and verb type and participant number as random effects.
RT ~ Sem. + Arg. + Anim. + (1|participant) + (1|verb)
The coefficient estimates revealed RTs explained by semantic relatedness and argument status, as compared to the intercept. The results of Model 1 are presented in Table 6.
Table 6
Summary of Model 1
Fixed effects Estimate SE t-value
Intercept 1.4180 0.0004 3451.36
Sem. rel. -0.0003 0.0002 -1.57
Arg. stat. -0.0003 0.0003 -0.84
Animacy 0.0001 0.0002 0.25
Note: SE = Standard error.
Next, a reduced model was built by removing animacy from Model 1 as it had the lowest absolute t-value, leaving only two fixed effects, namely semantic relatedness and argument status. The random effects remained unchanged.
RT ~ Sem. + Arg. + (1|participant) + (1|verb)
The results of the model are presented in Table 7.
Table 7
Summary of Model 2
Fixed effects Estimate SE t-value
Intercept 1.4181 0.0004 3576.62
Sem. rel. -0.0004 0.0002 -1.96
Arg. stat. -0.0002 0.0002 -0.93
Note: SE = Standard error.
Model 1 was then compared to Model 2 with a likelihood ratio test, which revealed that animacy did not improve the model (χ2 (1) = 0.065, p = 0.79). Afterwards, a third model was fit by removing argument status and leaving only semantic relatedness, as a fixed effect, while keeping the initial random effects.
RT ~ Sem. + (1|participant) + (1|verb)
The results of Model 3 are presented in Table 8. Model 3 was compared to Model 1 with a likelihood ratio test, revealing that the two effects of interest, namely argument status and animacy, did not improve the model significantly (χ2 (2)= 0.9414, p= 0.62).
Table 8
Summary of Model 3
Fixed effects Estimate SE t-value
Intercept 1.4181 0.000 3575.53
Sem. rel. -0.0005 0.0002 -3.05
Note: SE = Standard error.
Lastly, an intercept only model was built which excluded all the fixed effects. This model was then compared to Model 3.
RT ~ 1 + (1|participant) + (1|verb)
It was revealed that semantic relatedness alone improved the model significantly (χ2 (1) = 9.27, p= 0.002). This finding revealed the statistical significant effect of semantic relatedness on RTs with Model 3 being the most suitable model to explain the data, as reported in Table 8. To find the effect size of semantic relatedness on RTs in milliseconds, the Box-Cox transformed RT data were reversed revealing the slope with a value of 66.61 msec and the intercept with a value of 1114.19 msec, as presented in Figure 5. Overall, no evidence supporting our hypothesis pertaining to effects of argument status or animacy was found.
To find out whether the interaction of semantic relatedness and animacy affected the RTs significantly, another model was fit including interactions between semantic relatedness and animacy.
RT~ Sem. * Anim. + (1|participant) + (1|verb)
This model was compared to a model excluding argument status.
RT ~ Sem. + Anim. + (1|participant) + (1|verb)
The likelihood ratio comparison did not reveal a significant effect of interaction between semantic relatedness and animacy on RTs (χ2 (1) = 0.7074, p= 0.40).
Figure 5. RTs of the statistically significant semantic relatedness effect plotted from the Model 3, as compared to argument status, and animacy RTs (statistically not significant) plotted from the Model 1; error bars represent the standard errors of the mean effects;
Semantic_rel. = semantic relatedness.
As an exploratory analysis, age was added to Model 1 as a fixed effect to find out the largeness of age effect on the response variable in our heterogeneous population.
RT ~ Sem. + Arg. + Anim. + Age + (1|participant) + (1|verb)
The results did not reveal significant effect of age on RTs, as presented in Table 9.
Table 9
Summary of Model 4 with Age as a Fixed Effect
Fixed effects Estimate SE t-value
Intercept 1.4148 0.0013 1073.02
Sem. rel. -0.0003 0.0002 -1.5692
Arg. stat. -0.0003 0.0003 -0.8403
Animacy 0.0000 0.0002 0.2555
Age 0.0000 0.0000 0.9854
Note: SE = Standard error.
To make sure that age as a fixed effect did not affect the outcomes of Model 3, step() function was implemented to Model 4, which revealed Model 3 to be the optimal model even when age was present as a fixed effect. Besides, when the mean response latencies were grouped by age and plotted, no visual evidence was found that age influenced average RTs as presented in Figure 6.
Figure 6. RTs explained by age of participants.
Next, in each condition RTs of individual participants were averaged and presented visually in Figure 7. As displayed in Figure 7, although not statistically significant, visually the mean RT data of four participants in SRAA condition were the shortest as compared to
