A Gold Anaphora Annotation Layer on an Eye Movement Corpus

A Gold Anaphora Annotation Layer on an Eye Movement Corpus Olga Seminck, Pascal Amsili Laboratoire de Linguistique Formelle (Universit´e Paris Diderot & CNRS) olga.seminck@cri-paris.org, pascal.amsili@linguist.univ-paris-diderot.fr Abstract Anaphora resolution is a complex process in which multiple linguistic factors play a role, and this is witnessed by a large psycholinguistic literature. This literature is based on experiments with hand-constructed items, which have the advantage to filter influences outside the scope of the study, but, as a downside, make the experimental data artificial. Our goal is to provide a first resource allowing to study human anaphora resolution on natural data. We annotated anaphorical pronouns in the Dundee Corpus: a corpus of ∼ 50k words coming from newspaper articles read by humans of whom all eye movements were recorded. We identified all anaphoric pronouns — in opposition to non-referential, cataphoric and deictic uses — and identified the closest antecedent for each of them. Both the identification of the anaphoricity and the antecedents of the pronouns showed a high inter-annotator agreement. We used our resource to model reading time of pronouns to study simultaneously various factors of influence on anaphora resolution. Whereas the influence of the anaphoric relation on the reading time of the pronoun is subtle, psycholinguistic findings from settings using experimental items were confirmed. In this way our resource provides a new means to study anaphora. Keywords: Dundee Corpus, Personal Pronouns, Mixed Effects Modeling, Eye-Tracking 1. Introduction How the human mind interprets pronouns has been a sub- ject studied in psycholinguistics for a long time now. Many linguistic features have been shown to be of influence. They were discovered using experimental setup, for exam- ple by using ambiguous sentences and asking participants to choose an interpretation: (1) Fred loves Pete, because he is always smiling. In (1), people could be asked who the pronoun he refers to, and most would answer Pete, whereas in fact both Fred and Pete are possible answers. The preference for Pete depends mostly in this case on the verb loves, which belongs to the family of implicit causality verbs (Garvey and Caramazza, 1974). When reading (1), people automatically wonder what is so lovely about Pete. By manipulating different features in experiments (e.g. syntactic role, semantic role, information structure, implicit causality), many factors of influence were discovered. The effect of some factors is quite robust and was confirmed by studies using various experimental methods. However, we are not aware of any study on these factors that did not used hand-constructed experimental items. This is a reason why we wanted to test whether factors of influence on anaphora resolution would also show in other settings, like natural text reading. Find- ing an influence of these factors would strengthen the proof of their robustness, because —unlike in an experimental setting— unwanted properties cannot be filtered out. As natural data we used the English part of the Dundee Cor- pus: a resource of eye movement data from ten participants reading newspaper articles. We annotated all anaphoric personal pronouns in this corpus by …

Content truncated to 500 words