Entity Recognition at First Sight: Improving NER with Eye Movement Information

Proceedings of NAACL-HLT 2019, pages 1–10 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 1 Entity Recognition at First Sight: Improving NER with Eye Movement Information Nora Hollenstein ETH Zurich noraho@ethz.ch Ce Zhang ETH Zurich ce.zhang@inf.ethz.ch Abstract Previous research shows that eye-tracking data contains information about the lexical and syn- tactic properties of text, which can be used to improve natural language processing models. In this work, we leverage eye movement fea- tures from three corpora with recorded gaze information to augment a state-of-the-art neu- ral model for named entity recognition (NER) with gaze embeddings. These corpora were manually annotated with named entity labels. Moreover, we show how gaze features, gen- eralized on word type level, eliminate the need for recorded eye-tracking data at test time. The gaze-augmented models for NER using token- level and type-level features outperform the baselines. We present the benefits of eye- tracking features by evaluating the NER mod- els on both individual datasets as well as in cross-domain settings. 1 Introduction The field of natural language processing includes studies of tasks of different granularity and depths of semantics: from lower level tasks such as tokenization and part-of-speech tagging up to higher level tasks of information extraction such as named entity recognition, relation extraction, and semantic role labeling (Collobert et al., 2011). As NLP systems become increasingly prevalent in society, how to take advantage of information passively collected from human readers, e.g. eye movement signals, is becoming more interesting to researchers. Previous research in this area has shown promising results: Eye-tracking data has been used to improve tasks such as part-of-speech tagging (Barrett et al., 2016), sentiment analysis (Mishra et al., 2017), prediction of multiword ex- pressions (Rohanian et al., 2017), and word em- bedding evaluation (Søgaard, 2016). However, most of these studies focus on either relatively lower-level tasks (e.g. part-of-speech tagging and multiword expressions) or relatively global properties in the text (e.g. sentiment analy- sis). In this paper, we test a hypothesis on a differ- ent level: Can eye movement signals also help im- prove higher-level semantic tasks such as extract- ing information from text? The answer to this question is not obvious. On one hand, the quality improvement attributed to eye movement signals on lower-level tasks implies that such signals do contain linguistic information. On the other hand, it is not clear whether these sig- nals can also provide significant improvement for tasks dealing with higher-level semantics. More- over, even if eye movement patterns contain sig- nals related to higher-level tasks, as implied by a recent psycholinguistic study (Tokunaga et al., 2017), noisy as these signals are, it is not straight- forward whether they would help, if not hurt, the quality of the models. In this paper, we provide the first study of the impact of gaze features to automatic named entity recognition from text. We test the hypothesis that eye-tracking data is beneficial for entity recogni- tion in a state-of-the-art neural named entity tag- ger augmented with embedding …

Content truncated to 500 words