Distinguishing between different syntactic roles of identical words in normal reading: an ERP study Lakretz Yair (yair.lakretz@post.tau.ac.il) Tel-Aviv University 69978, Tel-Aviv, Israel Marjanovic Katarina (kmarjano@sissa.it) SISSA, Via Bonomea 265 34136 Trieste, Italy Gu YuQiao (yuqiaogu@sissa.it ) SISSA, Via Bonomea 265 34136 Trieste, Italy Treves Alessandro (ale@sissa.it ) SISSA, Via Bonomea 265 34136 Trieste, Italy Abstract In this study, we present a novel task for event-related po- tentials (ERP) studies which enables distinguishing between Separating between semantic and syntactic aspects of language semantic and syntactic aspects of language processing in nor- processing in the brain is a difficult task. In an attempt to dis- tinguish between the two, many studies so far have measured mal reading. In the experimental design, participants are pre- responses to semantic or syntactic violations in reading com- sented with target words which differ by their syntactic role prehension tasks. However, this methodology may be inac- but have similar semantic content, and are otherwise (e.g. curate in describing semantic and syntactic processing during normal reading. In this study, we use a novel task, measuring orthographically) identical. For this, we make use of noun- responses to identical target words as they assume different plus-noun constructions in English, in which the first noun syntactic roles. All sentences presented in the task are syntac- preserves its meaning while changing its position on the syn- tically correct sentences without lexical-semantic anomalies. We present results from a behavioral experiment, testing the tactic tree, moving from the position of the specifier of a head validity of the experimental design, and results from a pilot in a noun phrase (NP), to the head of the NP in a simple sen- ERP study, measuring brain responses to the difference in the tence without such construction. For example, compare be- syntactic role of the target words. We conclude that the pro- posed design is valid and may be used to shed light on seman- tween the word ’family’ in ’It’s a family discount’ to the same tic and syntactic processing during language comprehension, word in ’It’s a family from Sweden’. in normal reading. We conducted two experiments, a behavioral and an ERP Keywords: Language comprehension; syntactic violation; experiment. The behavioral experiment was designed to test ERPs; normal reading task; noun-plus-noun constructions. the validity of the novel task described below, and it is also used to select pertinent stimuli for the task. We then con- Introduction ducted an ERP pilot study using the task and the stimuli Semantic and syntactic aspects of language processing are as- which were selected according to the behavioral experiment. sociated with characteristic electrophysiological responses to We present here qualitative results from this pilot study. language stimuli. For example, many studies have shown that the N400 component systematically correlates with lexical- Related literature semantic aspects of language processing, and is concluded to ERP signatures of semantic and syntactic aspects of reflect lexical-integration processing, e.g. (Kutas & Hillyard, 1980, 1984). Other studies have shown that the syntactic as- language processing pect of language processing correlates with the P600 compo- In order to study the first language (L1) syntactic aspect of nent, e.g. (Friederici, Pfeifer, & Hahne, 1993) - See related language processing, many ERP studies adopted violation literature section. paradigms where non-grammatical sentences are compared Many of these studies use reading comprehension tasks, in- with correct sentences, which are otherwise similar to the volving stimuli of sentences which contain either a semantic violation stimuli. These studies assume that when all other or syntactic violation. The type of violation is used to reveal linguistic variables are held constant, the brain response to the syntactic or semantic process in question. However, this the target stimulus, compared to the control stimulus, re- methodology offers no insight into how these types of pro- flects processes which are related to the grammatical rule in cessing are distinguished in normal reading, that is, without question. The major ERP signatures reported in L1 sentence violating the syntactic rules or semantic expectations. processing are (for a recent review see (Caffarra, Molinaro, 611 Davidson, & Carreiras, 2015)): P600 for both types of violations (Ojima, Nakata, & Kakigi, 2005), as compared to natives. However, participants with The early left anterior negativity (ELAN) The ELAN high proficiency showed similar response to that observed component peaks at around 200 ms, with left-anterior distri- with L1 controls. The authors concluded that, at high-enough bution, in response to violations of an obligatory phrase struc- L2 proficiency levels, an L1-like brain response can be ob- ture. It is ascribed to automatic early syntactic parsing pro- served, reflecting early automatic parsing processes followed cesses, during which an initial phrase structure is built - see, by late processes of re-analysis and repair. e.g. (Hahne & Friederici, 1999; Friederici, 2002; Friederici Noun-plus-noun constructions in English & Weissenborn, 2007; Steinhauer & Drury, 2012). Noun-plus-noun (NNs) constructions are composite nominals The left anterior negativity (LAN) The LAN component in which both the head and the attributive dependendat(s) peaks at around 400 ms, with left-anterior distribution, in are nouns (Garnier, 2011), e.g. family discount, bus driver. response to morphosyntactic violations such as grammati- These types of NNs are a common type of constructions in cal agreement violations, tense-marking violations and case- the English language, however their role in its grammar is marking violations. It is ascribed to difficulties in integrating nevertheless still an ongoing debate among linguists, as some morphosyntactic information within a sentence structure with classify them as a phrase, originating in the syntax (Garnier, the final goal of thematic role assignment, or mismatch detec- 2011), while the others claim them as compounds, originat- tion during linking processes of agreement computation - see, ing in the lexicon (Giegerich, 2004). Another group of stud- e.g. (Molinaro, Vespignani, Zamparelli, & Job, 2011; Moli- ies claims they can belong to both categories (Payne & Hud- naro, Barber, & Carreiras, 2011; Friederici, 2002; Barber & dleston, 2002). Furthermore, (Giegerich, 2004) also distin- Carreiras, 2005). guishes between the fore-stressed and end-stressed NNs, as- signing them into two different origins. Avoiding this debate, and in order to keep a homogeneous set of stimuli in the ex- N400 The N400 component peaks at around 400ms, with periment, we therefore chose NNs with a fore-stress only. centro-posterior distribution, in response to lexical-semantic anomalies. It is ascribed to difficulties in processing lexical- Methods semantic information - see, e.g. (Kutas & Federmeier, 2011; Federmeier, 2007; Hagoort, 2003; Traxler & Gernsbacher, Participants 2011; Kutas & Federmeier, 2000). The participants of this study are native Italian speakers, stu- dents in Scuola Internazionale Superiore di Studi Avanzati P600 The P600 component peaks at around 600 ms, with (SISSA, Italy), with high proficiency of English. All the posterior distribution, in response to various violations of participants are between 18 and 40 years of age and right- syntactic and morphosyntactic features, thematic-rule struc- handed. In the behavioral experiment, 9 participants have ture violations, temporary ambiguities, semantic anomalies, taken part. In the ERP experiment, 5 participants have par- and long-distance dependencies - see, e.g. (Friederici et al., ticipated who did not take part in the behavioral experiment. 1993; Molinaro, Vespignani, et al., 2011; Molinaro, Barber, Stimuli & Carreiras, 2011; Carreiras, Salillas, & Barber, 2004). It is ascribed to processes of syntactic reanalysis and repair, and The stimuli consist of 52 syntactically and semantically cor- to late integration processes which are not syntactic specific - rect sets of English sentences. Each of the 52 sets consists of see, e.g. (Friederici, 2002; Molinaro, Vespignani, et al., 2011; a short quiz question, a beginning of an answer (including the Molinaro, Barber, & Carreiras, 2011; van de Meerendonk, target word), and 4 possible completions. There are 52 differ- Kolk, Vissers, & Chwilla, 2010; Brouwer, Fitz, & Hoeks, ent quiz questions, 26 different beginnings of answers (each 2012). of these repeats twice, in each of the two conditions), and 26 * 4 different completion options. First language (L1) - second language (L2) similarity Design and procedure In this study, we tested native Italian speakers, with high pro- We present a novel task for ERP studies which enables to dis- ficiency in English, on a task in English. The brain responses tinguish between semantic and syntactic aspects of language of these participants were recorded in an ERP design. Previ- processing in normal reading. To do so, participants are pre- ous studies have found that participants with high proficiency sented with target words which differ by their syntactic role in L2 have similar brain responses compared to L1 speakers. but have similar semantic content, and are otherwise (e.g. or- Several ERP experiments were conducted by Rossi, Gugler, thographically) identical. We make use of noun-plus-noun Friederici, & Hahne (2006) on L2 speakers presenting sen- constructions in English to contrast between, e.g. the word tences with morphosyntactic and phrase structure violations. ’family’ in Results show that low-proficiency L2 speakers did not show (S1) It’s a family discount. (noun-plus-noun condition) a LAN effect for morphosyntactic violations with a delayed (S2) It’s a family from Sweden. (single noun condition) 612 However, such a comparison is only possible if the syn- The list of quiz questions contained 51 quizzes from the tactic expectation of the participant in (S1) is the desired one noun-plus-noun condition, and 46 quizzes from the simple- while reading the target-word. For example, such a compari- sentence condition. All quizzes were presented to the partic- son would fail if the participant comprehends ’family’ in (S1) ipants in a random order. as a pre-head only after having completed and reanalysed the sentence. We therefore manipulate the syntactic expectation of the participant by preceding the sentence with a quiz ques- tion the answer to which requires the desired syntactic role only. Continuing with the above example, we precede (S1) with the following quiz question: (Q): It will get you a cheaper entrance to the pool. What is it? The quiz is then followed by a beginning of an answer: (A) ’It’s a family ’. Note that whether the participant knows the correct an- swer to the above quiz question is unimportant. Even without knowing the correct answer to (Q), we hypothesised that one would expect the answer (A) to the type of question in (Q) to Figure 1: The design of the behavioral experiment. end with a noun, thus reading ’family’ as an adjective. We assume that this kind of expectations are also enhanced after the practice block. Note also, that data analysis focuses on the ERP experiment Brain responses of another group of par- time during which the participant reads the target word (e.g. ticipants, none of whom participated in the behavioral exper- ’family’), before she is asked to complete the sentence. We iment, were recorded with 128-channel EEG. Stimuli were therefore regard it as normal reading, and are not concerned presented to the participants on a computer screen, in a simi- with other processes that may follow. lar manner to that in the behavioral experiment. Before start- Importantly, while creating syntactic expectation, the pre- ing the experiment, participants execute a practice section ceding quiz question must not have created semantic expec- containing 10 quiz questions which are different from the tation to the target word. Therefore, all words in the quiz ones later presented in the experiment. The experiment con- questions were made sure not to be semantically related to the sists of 5 blocks, each containing 52 quiz questions, to which target word in their answers (as can be assessed with Latent four possible answers are given (see section Stimuli). The list Semantic Analysis). For example, no word in (Q) semanti- of sentences within each block is presented to participants in cally primes the target-word in (A). An additional benefit to a random manner, which is different for each block. this manipulation is that it enhances the engagement of the The quiz question is presented on the screen until the par- participants in the task, by challenging them with quiz ques- ticipants decide to continue by pressing a key. After having tions. pressed a key, the following three screens are presented: (1) ”It’s a”, (2) target word (e.g. ’family’), (3) dashed line ( ) In order to test our hypothesis that the quiz questions in- (ISI=300ms). The last screen of the dashed line is then fol- duce the correct syntactic expectation, we ran a behavioral lowed by an option screen, containing four possible answer experiment that assesses the syntactic expectation of the par- completions, randomly ordered in each trial and block (Fig- ticipant when reading the target word. This experiment is ure 2). The five blocks are separated by breaks, during which described below. According to the results of the behavioral the participants remain in their position in front of the screen. experiment, we chose the quizzes which best manipulated the The participants determine the length of the breaks by them- syntactic expectations of the participants. These quizzes were selves. then used in the ERP experiment, described in the following section. EEG recording Behavioral experiment In order to test the syntactic expec- The EEG was continuously recorded using the ActiveTwo tation of the quiz question, we ran a behavioral experiment in BioSemi EEG system (BioSemi V.O.F., Amsterdam, Nether- which participants are asked to complete target sentences af- lands) with 128 channels covering the entire scalp. EEG sig- ter reading the quiz. The design of this task is shown in Figure nals were sampled at 512 Hz with band-pass filters set at 1. Participants are presented with 5 subsequent screens: (1) 0.1 − 100 Hz. A quiz question, (2) ”It’s a” (beginning of the answer), (3) fixation cross, (4) target word (e.g. ’family’), (5) Textbox. Data analysis participants are asked to read the quiz question and then to Acquired data is analysed using EEGLAB, open source MAT- complete the sentence in a textbox, using the keyboard, af- LAB (The Matworks, Natick, MA) toolbox for EEG process- ter having read the beginning of the sentence. Responses and ing. Data is first high-pass filtered at 1 Hz and low-pass fil- reaction times are recorded during the experiment. tered at 40Hz, and re-referenced. Next, follows an extrac- 613 We found that the manipulations for the noun-plus-noun construction achieve higher average accuracy (average accu- racy = 0.80 ± 0.15), in comparison to the single nouns (aver- age accuracy = 0.54 ± 0.25). For the ERP experiment, we then chose for each of the two conditions 26 quiz questions with the highest accuracy score (NNs - 0.86 ± 0.09, single nouns - 0.72 ± 0.17). These quiz questions are then used in the ERP experiment. ERP experiment Since the experiment is still ongoing, we present ERP results from our pilot study (Figure 4). Figure 2: The design of the ERP experiment. tion of the epochs of the two conditions. In both epoch sets, the answer of the participants are divided into three groups: the correct answer (e.g. ’discount’ in (Q)); the semi-correct answer (e.g. ’heritage’ in (Q)), which is syntactically cor- rect, suggesting that the participant comprehended the tar- get word in the desired syntactic role; and wrong answers (e.g. ’from Sweden’ and ’from Mongolia’ in (Q)). The an- swers are counterbalanced over conditions such that, for ex- ample, ’from Sweden’ is the correct answer, ’from Mogolia’ the semi-correct, and ’heritage’ and ’discount’ are the wrong answers in the second condition. All the wrong answers are omitted and only the correct and semi-correct answers are in- cluded in the analysis. The data is pre-processed and artifacts are omitted using independent component analysis (ICA). Results Figure 4: Results of the ERP experiment for an arbitrarily chosen participant: (A) The ERP signatures for syntactic and Behavioral experiment semantic processing, revealing different responses between We tested whether quiz questions can induce the desired syn- the two conditions (B) brain response during processing of tactic expectation when participants read the target sentence. single nouns (left) and noun-plus-noun constructions (right). Figure 3 presents the results for this experiment, showing for each quiz the mean accuracy, calculated over all participants. Summary and discussion In this study, we present a novel paradigm to disentangle the syntactic aspect of language processing from its semantic as- pect. The paradigm avoids the use of syntactic violations, and focuses on normal reading of correct sentences. We make use of noun-plus-noun constructions in English to contrast between two conditions which differ by their syntactic aspect only. In order to manipulate the syntactic expectation of the participant before reading the target word, we use quiz ques- tions, such that the answer to the questions allows only one possible syntactic role to the target word. We tested the method of manipulating the syntactic expec- tations of the participants in a behavioral experiment. Results support the validity of the proposed experimental design, and are also used to choose quiz questions with highest score of Figure 3: Results of the behavioral experiment (A) Correct manipulation. Results suggest that the syntactic manipula- completion rates for each quiz question, for the condition of tions are more effective for the noun-plus-noun conditions, noun (N) (B) Correct completion rates for each quiz, for the but are reasonably effective for both conditions for the se- condition of noun-plus-noun construction (NNs). lected group of stimuli. 614 Following the behavioral experiment, we ran an ERP pilot Hahne, A., & Friederici, A. D. (1999). Electrophysiological study using the novel paradigm and selected quiz questions. evidence for two steps in syntactic analysis: Early auto- Albeit only qualitative, pilot results are showing promising matic and late controlled processes. Journal of Cognitive disclosure of different ERP signatures for the two conditions, Neuroscience, 11(2), 194–205. opening a new window into syntactic processing during lan- Kutas, M., & Federmeier, K. D. (2000). Electrophysiology guage comprehension in normal reading. We believe that it reveals semantic memory use in language comprehension. may provide a way to relate observable signals in the human Trends in cognitive sciences, 4(12), 463–470. brain to hypothesised mechanisms in models of latching dy- Kutas, M., & Federmeier, K. D. (2011). Thirty years and namics (Russo, Namboodiri, Treves, & Kropff, 2008; Russo counting: Finding meaning in the n400 component of the & Treves, 2012), in particular at the transition between words event related brain potential (erp). Annual review of psy- (Pirmoradian & Treves, 2013). chology, 62, 621. Kutas, M., & Hillyard, S. A. (1980). Reading senseless sen- References tences: Brain potentials reflect semantic incongruity. Sci- ence, 207(4427), 203–205. Barber, H., & Carreiras, M. (2005). Grammatical gender and Kutas, M., & Hillyard, S. A. (1984). Brain potentials during number agreement in spanish: An erp comparison. Journal reading reflect word expectancy and semantic association. of Cognitive Neuroscience, 17(1), 137–153. Nature(307), 161–3. Brouwer, H., Fitz, H., & Hoeks, J. (2012). Getting real about Molinaro, N., Barber, H. A., & Carreiras, M. (2011). Gram- semantic illusions: rethinking the functional role of the matical agreement processing in reading: Erp findings and p600 in language comprehension. Brain Research, 1446, future directions. cortex, 47(8), 908–930. 127–143. Molinaro, N., Vespignani, F., Zamparelli, R., & Job, R. Caffarra, S., Molinaro, N., Davidson, D., & Carreiras, M. (2011). Why brother and sister are not just siblings: Repair (2015). Second language syntactic processing revealed processes in agreement computation. Journal of Memory through event-related potentials: an empirical review. Neu- and Language, 64(3), 211–232. roscience & Biobehavioral Reviews. Ojima, S., Nakata, H., & Kakigi, R. (2005). An erp study of Carreiras, M., Salillas, E., & Barber, H. (2004). Event- second language learning after childhood: Effects of profi- related potentials elicited during parsing of ambiguous rel- ciency. Cognitive Neuroscience, Journal of , 17(8), 1212– ative clauses in spanish. Cognitive Brain Research, 20(1), 1228. 98–105. Payne, J., & Huddleston, R. D. (2002). Nouns and noun phrases. Federmeier, K. D. (2007). Thinking ahead: The role and roots Pirmoradian, S., & Treves, A. (2013). Encoding words into of prediction in language comprehension. Psychophysiol- a potts attractor network. In Proceedings of the thirteenth ogy, 44(4), 491–505. neural computation and psychology workshop (ncpw13) on Friederici, A. D. (2002). Towards a neural basis of auditory computational models of cognitive processes, world scien- sentence processing. Trends in cognitive sciences, 6(2), tific press, singapore (pp. 29–42). 78–84. Russo, E., Namboodiri, V. M., Treves, A., & Kropff, E. Friederici, A. D., Pfeifer, E., & Hahne, A. (1993). Event- (2008). Free association transitions in models of cortical related brain potentials during natural speech processing: latching dynamics. New Journal of Physics, 10(1), 015008. Effects of semantic, morphological and syntactic viola- Russo, E., & Treves, A. (2012). Cortical free-association tions. Cognitive brain research, 1(3), 183–192. dynamics: Distinct phases of a latching network. Physical Friederici, A. D., & Weissenborn, J. (2007). Mapping sen- Review E, 85(5), 051920. tence form onto meaning: The syntax–semantic interface. Steinhauer, K., & Drury, J. E. (2012). On the early left- Brain research, 1146, 50–58. anterior negativity (elan) in syntax studies. Brain and lan- Garnier, M. (2011, septembre). Correcting erroneous N+N guage, 120(2), 135–162. structures in the production of French users of English as Traxler, M., & Gernsbacher, M. A. (2011). Handbook of part of the CorrecTools project (regular paper). In EU- psycholinguistics. Academic Press. ROCALL, Nottinghma, 31/08/2011-03/09/2011 (pp. 157– van de Meerendonk, N., Kolk, H. H., Vissers, C. T. W., & 165). http://www.eurocall-languages.org: Eurocall Associ- Chwilla, D. J. (2010). Monitoring in language perception: ation. mild and strong conflicts elicit different erp patterns. Jour- Giegerich, H. J. (2004). Compound or phrase? english noun- nal of Cognitive Neuroscience, 22(1), 67–82. plus-noun constructions and the stress criterion. English Language and Linguistics, 8(01), 1–24. Hagoort, P. (2003). Interplay between syntax and semantics during sentence comprehension: Erp effects of combining syntactic and semantic violations. Cognitive Neuroscience, Journal of , 15(6), 883–899. 615