The Role of Underspecification in RC Attachment: Speed-accuracy tradeoff evidence

A number of studies have found evidence for the so-called ambiguity advantage, i.e., a speed-up in processing ambiguous sentences compared to their unambiguous counterparts. While a number of proposals regarding the mechanism underlying this phenomenon have been made, the empirical evidence so far is far from unequivocal. It is compatible with several theories, including strategic underspecification (Swets et al., 2008), race models (Van Gompel et al., 2000; Logacev and Vasishth, 2016), and a more recentcoactivation-based account (Dillon et al., 2019). While all three classes of theories make matching predictions for the average time to complete RC attachment in ambiguous compared to unambiguous sentences, their predictions diverge with regard to theminimum completion times. I used the speed-accuracy tradeoff procedure to test the predictions of all three classesof theories. According to a hierarchical Bayesian model, the speed-accuracy tradeoff functions (SATFs) for different RC attachment conditions (high, low or ambiguous) show an earlier departure from chance performance in the ambiguous condition than in either of the unambiguous conditions. The results further indicate increased asymptotic accuracy but no increase in processing rate in the ambiguous condition. Taken together, this pattern of results is compatible with the strategic underspecification model, and to a lesser degree with coactivation based accounts.

A software toolkit for modeling human sentence parsing: An approach using continuous-time, discrete-state stochastic dynamical systems

We present a new software toolkit for implementing a broad class oftheories of sentence processing. In this framework, processing a word ina sentence is viewed as a continuous-time random walk through a set ofdiscrete states that encode information about the emerging structure of thesentence so far. The state space includes one or more special absorbingstates, which, when reached, indicate the decision to move on to the nextword of the sentence. This setup allows us to ask how how long it takesto reach an absorbing state and what the probability of reaching this stateis. We summarize a number of important statistics that can be directlyrelated to human reading times and comprehension question performance.To illustrate the use of the toolkit, we model two types of garden paths,local coherence effects, and the ambiguity advantage using three qualitativelydifferent theories of sentence processing. While the modeler must still makedefensible theoretical and implementation choices, this framework representsan improvement over the descriptive, paper-pencil modeling that is thenorm in psycholinguistics by facilitating quantitative evaluations of modelperformance and laying the groundwork for Bayesian fitting of free parametersin a model. An open-source Python package is provided.

Absence of Evidence for Underspecification in Prenominal Relative Clause Attachment

Traxler et al. (1998) have found that relative clauses with ambiguous attachment are sometimes read faster than their unambiguous counterparts. Two broad classes of theories account for this phenomenon: Race-based models posit that ambiguous sentences are read faster due to a ‘race’ between several permissible analyses of the sentence. In contrast, the strategic underspecification account maintains that, under the right conditions, readers underspecify ambiguities in order to save time. We show the two accounts make qualitatively different predictions for structures with prenominal relative clauses, such as in Turkish. While the underspecification account predicts an ambiguity in Turkish, race-based accounts predict the absence of such an effect. We present data from two reading experiments in Turkish in which we find no evidence for an ambiguity advantage in the processing of ambiguous sentences with prenominal relative clauses and argue that this finding poses a major challenge for the strategic underspecification account.

Parsing preferences and individual differences in nonnative sentence processing: Evidence from eye movements

Using both offline and online measures, the present study investigates attachment resolution in relative clauses in English natives (L1) and nonnatives (L2). We test how relative clause resolution interacts with linguistic factors and participant-level individual differences. Previous L1 English studies have demonstrated a low attachment preference and also an “ambiguity advantage” suggesting that L1ers may not have as strong a low attachment preference as is sometimes claimed. We employ a similar design to examine this effect in L1 and L2 comprehension. Offline results indicate that both groups exhibit a low attachment preference, positively correlated with reading span scores and with proficiency in the L2 group. Online results also suggest a low attachment preference in both groups. However, our data show that individual differences influence online attachment resolution for both native and nonnatives; higher lexical processing efficiency correlates with quicker resolution of linguistic conflicts. We argue that the current findings suggest that attachment resolution during L1 and L2 processing share the same processing mechanisms and are modulated by similar individual differences.

Chapter 2. Eye-tracking research of reading mechanisms: processing of letters, words, sen- tences and texts

In Сhapter 2 we describe how verbal information is processed at different linguistic levels, from recognizing single letters to reading and comprehension of coherent texts. We present the results of several experimental studies on reading in Russian which has specific features like Cyrillic script, rich morphology and flexible word order. First, we show some features of Cyrillic letters recognition of different font types in the experiment with invisible boundary. Our results reveal that the font type affects the recognition of crowed letters (letters in Courier New were harder to identify than the ones in Georgia), while recognition efficiency of isolat- ed letters remains at the same level. Since crowded letters imitate real reading, we claim that Georgia is more readable font than Courier New. Second, we describe the lexical, syntactic and referential ambiguity processing emphasizing the role of semantic context. Thus, we show that the processing of ambiguous words does not depend on the type of their meaning (literal or non-literal) …, and the referential ambiguity advantage effect. Third, we compare the process- ing of literal and non-literal expressions in Russian. We try to tease apart different approaches to idioms as well as to give a better explanation of what units may be stored in the mental lex- icon and how syntactic processing may proceed. Finally, we demonstrate the influence of the text type, functional style and reading skills on text processing. We show that the text type is among the readability categories and it influences the effect of reading perspective: eye-track- ing parameters of reading a static text (descriptive sentences) and a dynamic text (sequence of events following swiftly on one another) differ a lot.

A new argument for co-active parses during language comprehension

One perennially important question for theories of sentence comprehension is whether the human sentence processing mechanism is parallel (i.e. it simultaneously represents multiple syntactic analyses of linguistic input) or serial (i.e. it constructs only a single analysis at a time). Despite its centrality, this question has proven difficult to address for both theoretical and methodological reasons (Gibson & Pearlmutter, 2000; Lewis, 2000). In the present study, we reassess this question from a novel perspective. We investigated the well-known ambiguity advantage effect (Traxler, Pickering & Clifton, 1998) in a speeded acceptability judgment task. We adopted a Signal Detection Theoretic approach to these data, with the goal of determining whether speeded judgment responses were conditioned on one or multiple syntactic analyses. To link these results to incremental parsing models, we developed formal models to quantitatively evaluate how serial and parallel parsing models should impact perceived sentence acceptability in our task. Our results suggest that speeded acceptability judgments are jointly conditioned on multiple parses of the input, a finding that is overall more consistent with parallel parsing models than serial models. Our study thus provides a new, psychophysical argument for co-active parses during language comprehension.

Together or apart: Learning of translation-ambiguous words

In a multiple-session training study, native English speakers learned foreign Dutch vocabulary items that mapped to English either in a one-to-one way (translation-unambiguous) or in a one-to-many way (translation-ambiguous), such that two Dutch words corresponded to a single English translation. Critically, these two translation-ambiguous Dutch words were taught on consecutive trials in the same session, or were presented separately, such that each word was taught in a separate session. Translation-ambiguous words were produced and recognized substantially less accurately than translation-unambiguous words on tests administered one and three weeks after training. An ambiguity advantage emerged, however, in a free-recall test. Interestingly, teaching both translations together led to superior performance over teaching them in separate sessions, in which case the translation learned first enjoyed a considerable advantage over that learned second. These findings underscore the importance of order of acquisition in second-language vocabulary learning, and have practical implications for language instruction.