Parameter estimation of Japanese predicate argument structure analysis model using eye gaze information

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 2861–2869, Osaka, Japan, December 11-17 2016. Parameter estimation of Japanese predicate argument structure analysis model using eye gaze information Ryosuke Maki Hitoshi Nishikawa Department of Computer Science Tokyo Institute of Technology {maki.r.aa@m, {hitoshi, take}@c}.titech.ac.jp Takenobu Tokunaga Abstract In this paper, we propose utilising eye gaze information for estimating parameters of a Japanese predicate argument structure (PAS) analysis model. We employ not only linguistic information in the text, but also the information of annotator eye gaze during their annotation process. We hypothesise that annotator’s frequent looks at certain candidates imply their plausibility of being the argument of the predicate. Based on this hypothesis, we consider annotator eye gaze for estimating the model parameters of the PAS analysis. The evaluation experiment showed that introducing eye gaze information increased the accuracy of the PAS analysis by 0.05 compared with the conventional methods. 1 Introduction In recent years, there have been many attempts of annotating corpora with various kinds of informa- tion, as supervised machine learning (ML) techniques had been a significant device for natural language processing (NLP) (Pustejovsky and Stubbs, 2012). The annotated corpus is used as training data for constructing a task model, considering the annotated information as expected output of the model. In the current framework, however, only annotated information in the corpus is used for training. In this paper, we propose utilising the information of annotator behaviour during the annotation process as well as resulting annotated information for training the task model (Tokunaga et al., 2013). We take the predicate argument structure (PAS) analysis of Japanese texts as the target task in the present work. Predicate argument relations are usually marked by case particles denoting grammatical cases in Japanese, therefore identifying dependencies marked by the major obligatory cases, ga (nomina- tive), wo (accusative) and ni (dative) is the main task. However, since ellipses are ubiquitous in Japanese texts, arguments might be identified beyond the sentence including the target predicates (inter-sentence arguments) as well as within the sentence (intra-sentence arguments). This feature is different from the PAS analysis in English in which arguments can be found in the same sentence of the target predicate in most cases. Another complication is the case alternation caused by certain types of auxiliary verbs such as causative and passivisation verbs. In such cases, the original case should be recovered in the PAS analysis. In this respect, the Japanese PAS analysis shares the similarity with semantic role labelling in English (Gildea and Jurafsky, 2002). Furthermore, treating “event nouns” (Komachi et al., 2007) as predicates, we identify their arguments as well as the arguments of verbs and adjectives. Currently, several PAS annotated Japanese corpora such as NAIST Text Corpus (NTC) (Iida et al., 2007b) and BCCWJ-DepParaPAS (Ueda et al., 2015; Maekawa et al., 2014) are available. We use the latter in this study. In order to improve the PAS analysis performance, we propose to utilise the information of annota- tor behaviour, particularly eye …

Content truncated to 500 words