Cross-lingual Transfer of Correlations between Parts of Speech and Gaze Features

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 1330–1339, Osaka, Japan, December 11-17 2016. Cross-lingual Transfer of Correlations between Parts of Speech and Gaze Features Maria Barrett Centre for Language Technology University of Copenhagen Njalsgade 136 2300 Copenhagen S, Denmark barrett@hum.ku.dk Frank Keller School of Informatics University of Edinburgh 10 Crichton Street Edinburgh EH8 9AB, UK keller@inf.ed.ac.uk Anders Søgaard Dpt. of Computer Science University of Copenhagen Sigurdsgade 41 2200 Copenhagen N, Denmark soegaard@di.ku.dk Abstract Several recent studies have shown that eye movements during reading provide information about grammatical and syntactic processing, which can assist the induction of NLP models. All these studies have been limited to English, however. This study shows that gaze and part of speech (PoS) correlations largely transfer across English and French. This means that we can replicate previous studies on gaze-based PoS tagging for French, but also that we can use English gaze data to assist the induction of French NLP models. 1 Introduction The eye movements during normal, skilled reading are known to reflect the processing load associ- ated with reading. Recently, eye movement data has been integrated into natural language processing models for weakly supervised part-of-speech (PoS) induction (Barrett et al., 2016), sentence compres- sion (Klerke et al., 2016), supervised PoS tagging (Barrett and Søgaard, 2015a), and supervised pars- ing (Barrett and Søgaard, 2015b). Barrett et al. (2016) used eye movements from the English portion of a large eye tracking corpus, the Dundee corpus (Kennedy et al., 2003), for weakly supervised PoS induction for English, obtain- ing significant improvements over a baseline without gaze features. They used a second-order hidden Markov Model, which was type-constrained by Wiktionary dictionaries for their experiments. These re- sults suggest an approach to weakly supervised PoS induction using only a dictionary and eye movement data. Such an approach would be applicable for low-resource languages, for which it is difficult to find professional annotators. English French 0.0 0.2 0.4 0.6 0.8 1.0 X VERB PRT PRON NUM NOUN DET CONJ ADV ADP ADJ . Figure 1: Distribution of PoS in the English and French training sets. The present study further explores to which extent native readers’ processing of PoS gener- alizes across related languages. We use a sim- ilar model as Barrett et al. (2016), but perform cross-lingual experiments with both the French and the English portion of the Dundee Corpus. Contribution This is to the best of our knowledge the first study to explore how the eye movements of native readers that inform PoS models generalize from one language to another. We also introduce a new resource for studying the relation between grammatical class and eye movements in French: we provide PoS annotation for most of the French Dundee Corpus by aligning it with the morphosyntactic annotation of the French Treebank (Abeill´e et al., 2003). This work is licenced under a Creative Commons Attribution 4.0 International Licence. Licence details: http:// creativecommons.org/licenses/by/4.0/ 1330 ADJ ADP ADV CONJ DET NOUN NUM PRON PRT VERB X …

Content truncated to 500 words