Coreference Assignment and Anaphora Resolution in Arabic

Abstract

This dissertation examines the role of locality, complexity and features in resolving syntactic dependencies of the type adjectives, relative clauses and pronominals in Arabic. While several theories have argued for a universal paradigm in which speakers preferentially opt to attach new incoming linguistic input to the lowest phrase currently being processed, this view has been challenged by a number of cross-linguistic observations that show conflicting evidence. Similarly, in the context of relative clause complexity, it has been argued that a universal complexity hierarchy is in effect whereby subject structures are easier to process than object ones. The predictions of this hierarchy are tested and discussed. Finally, another hierarchy that is presumed to be operative crosslinguistically concerns the organization of nominal phi-features of number and gender whereby number information occupies a more privileged position than gender. The predictions of this hierarchy are also evaluated. An additional goal of this dissertation is to explore how the language parser behaves in Arabic as a first language (L1) and a second language (L2). Chapter 1 gives some preliminaries to situate the present studies. Chapter 2 presents the results of two experiments looking at the attachment preferences in two types of modifiers: adjective (AP) and relative clause (RC). Using a self-paced reading paradigm, the first experiment evaluates the relative strength of structural relations compared to semantic relations in order to determine the default strategy used by the language parser in attaching AP modifiers. The results show a mirror image effect where the exact opposite pattern of results was obtained for the two language groups. The L1 group showed a lower attachment preference across two postnominal modifiers (i.e. adjectives and relative clauses), whereas the L2 group demonstrated a higher attachment in both modifiers. The effects were modulated by structural cues (i.e. hierarchical dominance) in the case of L1 and semantic cues (i.e. animacy) in the case of L2. This asymmetry in attachment preferences is explained in terms a parsing strategy that is guided by thematic roles in L1 (i.e. preference for subject position), and by locality and semantic relations in the L2 (i.e. preference for animate and first position). A second experiment in this chapter gauges the attachment preferences of L1 and L2 Arabic in a relative clause context. The results replicated the same findings of the AP modifier by the animate condition, suggesting a strong bias toward low attachment for L1 and high attachment for L2 regardless of the modifying constituent type and length. In chapter 3, the processing of relative clauses is taken a step further to examine the influence of complexity in establishing syntactic dependencies. Complexity being defined in structural terms whereby a marked order (i.e. Object) is more complex than an unmarked one (i.e. Subject), as well as in linear terms in which more distant elements (i.e. center-embedded) are more complex than local ones (i.e. right-branching). Therefore, two types of complexity dimensions are investigated: filler-gap dependency in subject and object relatives and subject-verb dependency in center-embedded and right-branching conditions. The results show that the parser is sensitive to both the extraction site of the RC and its position within the sentence. This study replicates a well-established effect in the literature by showing that object relatives are costlier to process. However, while the results of L1 and L2 Arabic converge with respect to the filler-gap dependency, they diverge when RCs are in a right-branching position. The results are interpreted in light of feature similarity-based interference that is both structural and linear. The effect of center-embedding in object relative clauses is best captured through models such as Cue-Based Retrieval (Van Dyke & McElree, 2006) and Dependency Locality Theory (Gibson, 1998) in which shared features (e.g. subjecthood) incur a greater processing cost. By contrast, the effect of relative clause position is better understood in terms of an interruption account (Miller & Chomsky, 1963) for the L1 and by the Information Flow Hypothesis (Gibson et al., 2005) for the L2. Right-branching structures were found to improve the complexity of relative clauses in the L1, which fits with the linguistic intuition that center-embedded clauses are interruptive to the matrix clause and comprehension. With respect to the L2 group, which showed a surprisingly greater difficulty with the right-branching structure, the results are consistent with the Information Flow Hypothesis in which right-branching clauses lead to a surprisal effect due to the unexpected appearance of old information after new one. Building on the results of experiment 3 which established the complexity of object relatives by both groups, experiment 4 attempts to investigate the role of phi-features of gender and number in mitigating the complexity effect observed in object relatives. In particular, it asks whether a feature-hierarchy principle is in effect in mitigating the complexity observed in experiment 3. The results of this experiment demonstrate that not all features have the same effect in reducing the cost associated with object RC. In particular, the experiment shows that, in the L1 group, changing the feature specification of the embedded noun by adding a feature (e.g. number), not already included in the featural set of the distractor noun, showed the most facilitation. The L2 group by contrast benefited more from a gender feature mismatch. The results are interpreted in terms of a grammatical approach that considers (dis)similarity of the interfering noun to create disjointed feature sets between the head noun and the interfering noun. In the last two experiments of this dissertation, I investigate the role of gender and number in resolving anaphoric constructions with and without an overt pronoun. Experiment 5 attempts to establish some common grounds about the resolution preferences of Arabic speakers when a null pronoun (a phonologically empty element) is used and when an overt pronoun (and overtly realized element) is used. That is, the attachment preferences of L1 and L2 were gauged by pronoun type. In this experiment, only one feature (i.e. gender) was tested to establish attachment preferences independently of other features. The results show a clear division of labor between null and overt pronoun. A null pronoun expects to find an antecedent in the subject position while an overt pronoun prefers one in a lower position. The results were significant by the L1 but not the L2, even though the L2 showed a similar pattern but with no interaction. The last experiment, asks the same basic question that was asked in the context of experiment 4, which is the extent to which the gender and number features are able to reduce the complexity of the structure. In particular, it asks whether these features interact with the pronoun assignment strategy when the null pronoun is grammatically disambiguated towards a dispreferred position (i.e. object). The results show no evidence of any interaction with antecedent biases by the L1 but some gender (i.e. feminine) effects by the L2. The main contribution of the last study is to show that the nominal phi-features of gender and number carry different weights in (re)directing the language parser to the intended referent of the sentence (e.g. Carminati, 2005). Another contribution of the third, final, study is the finding that overt and null subjects in Arabic have intrinsically different interpretation biases. This adds to a growing body of literature demonstrating that overt and empty categories are processed differently (Huang, 2000). Taken together, these three studies suggest that strength of the cues triggering the computation of long-distance dependencies in Arabic vary both within the language itself, and between the language learning groups. It is shown that, although both language groups share similarities, L1 parsing strategy is driven primarily by phrase-structure information, while L2 is primarily guided by semantic and discourse information.