Towards Making a Dependency Parser See

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 1500–1506, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 1500 Towards Making a Dependency Parser See Michalina Strzyz David Vilares Carlos G´omez-Rodr´ıguez Universidade da Coru˜na, CITIC FASTPARSE Lab, LyS Research Group, Departamento de Computaci´on Campus de Elvi˜na, s/n, 15071 A Coru˜na, Spain {michalina.strzyz,david.vilares,carlos.gomez}@udc.es Abstract We explore whether it is possible to lever- age eye-tracking data in an RNN dependency parser (for English) when such information is only available during training – i.e. no aggre- gated or token-level gaze features are used at inference time. To do so, we train a multi- task learning model that parses sentences as sequence labeling and leverages gaze features as auxiliary tasks. Our method also learns to train from disjoint datasets, i.e. it can be used to test whether already collected gaze fea- tures are useful to improve the performance on new non-gazed annotated treebanks. Ac- curacy gains are modest but positive, showing the feasibility of the approach. It can serve as a first step towards architectures that can bet- ter leverage eye-tracking data or other comple- mentary information available only for train- ing sentences, possibly leading to improve- ments in syntactic parsing. 1 Introduction Eye trackers and gaze features collected from them have been recently applied to natural lan- guage processing (NLP) tasks, such as part-of- speech tagging (Duffy et al., 1988; Nilsson and Nivre, 2009; Barrett and Søgaard, 2015a), named- entity recognition (Tokunaga et al., 2017) or read- ability (Gonz´alez-Gardu˜no and Søgaard, 2018). Eye-movement data has been also used for pars- ing. For example, Barrett and Søgaard (2015b) rank discriminative features to predict syntactic categories (e.g. subject vs. object) and use them to improve a transition-based parser, trained on a structured perceptron with discrete features (Collins, 2002; Zhang and Nivre, 2011). However, the experiments were carried out on a parallel toy treebank and the performance was relatively low. Lopopolo et al. (2019) follow the inverse path, and use dependency parsing features to predict eye- regression during training, i.e. cases where the reader goes back to a word of the sentence. In this context, how to retrieve and leverage eye-tracking data has become an active area of re- search in different NLP fields. Previous studies (Barrett and Søgaard, 2015b; Barrett et al., 2016) suggest that real-time eye-tracking data can be col- lected at inference time, so that token-level gaze features are used during training but also at test time. However, even if in the near future every user has an eye tracker on top of their screen – a scenario which is far from guaranteed, and raises privacy concerns (Liebling and Preibusch, 2014) – many running NLP applications that process data from various Internet sources will not expect to have any human being reading massive amounts of data. Other studies (Barrett and Søgaard, 2015a; Hollenstein and Zhang, 2019) instead derive gaze features from the training set: in particular, …

Content truncated to 500 words