From Sense to Action: A Word-Action Disambiguation Task in NLP

From Sense to Action: A Word-Action Disambiguation Task in NLP Shu-Kai Hsieh Graduate Institute of Linguistics National Taiwan University shukaihsieh@ntu.edu.tw Yu-Hsiang Tseng Graduate Institute of Linguistics National Taiwan University seantyh@gmail.com Chiung-Yu Chiang Graduate Institute of Linguistics National Taiwan University cychiang@ntu.edu.tw Richard Lian Graduate Institute of Networking and Multimedia National Taiwan University dclian@nlg.csie.ntu.edu.tw Yong-fu Liao Graduate Institute of Linguistics National Taiwan University mcku1115@gmail.com Mao-Chang Ku Graduate Institute of Linguistics National Taiwan University mcku1115@gmail.com Ching-Fang Shih Graduate Institute of Linguistics National Taiwan University r08142004@ntu.edu.tw Abstract Words are conventionalized symbols that pre- sent the function by which meaning is attached to form. The Word Sense Disambiguation, which has been taken as one of the core seman- tic processing tasks in the pipe-lined NLP ar- chitecture, aims to assign proper word sense to lemma form in varied contexts based on a word-sense inventory such as WordNet. How- ever, there are some theoretical assumptions unattested from a functional linguistic point of view. This paper proposes an alternative by in- troducing a novel task called word action dis- ambiguation task (WAD) concentrated on the observable pairs between words and actions. The accompanying dataset, which was manu- ally edited and compiled, is composed of 419 multiple-choice questions. We further verified the dataset through item evaluation with human rating data, and the semantic relations among the dataset were annotated automatically. A baseline performance with an accuracy of 38.64% was also provided with BERT models and 43.18% after incorporating paradigmatic knowledge with semantic graph. We expect the proposal of the WAD task and dataset would motivate computational models to incorporate more complex aspects of human language. 1 Introduction Due to its polysemous behavior, selecting the most appropriate sense for a word in a text has been one of the most important yet challenging NLP tasks over the years. Given a pre-defined sense inventory, computationally assigning each word in target texts with proper sense (thus Word Sense Disambigua- tion) is assumed to be crucial for MT, IR, QA, and other systems (Navigli, 2009). Although the sense inventory such as WordNet has been continuously maintained and implemented cross-linguistically, the issue regarding the extent to which the sense granularity (i.e., levels of semantic specificity) in the sense inventory would be sufficient for down- stream NLP tasks remains less explored. Three tacit and intertwined assumptions under- lying the conventional WSD task are (1) word senses can be operationalized as discrete and distin- guishable ones, (2) word senses (as included in the sense inventory) can be shared by the entire lan- guage community, and (3) WSD with the fine- grained sense specification can be successfully ap- plied to actual language data, and facilitate a wide range of downstream NLP tasks. However, the re- ported poor inter-annotator agreement (IAA) and low reliability of sense distinction/annotation in the task seem to falsify these assumptions and thus mo- tivate projects like OntoNotes (Hovy et al., 2006; Cinková et al., 2012). This paper aims to serve as a first attempt to pro- pose an alternative to the underlying …

Content truncated to 500 words