A Computational Study on Word Meanings and Their Distributed Representations via Polymodal Embedding

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 214–223, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP A Computational Study on Word Meanings and Their Distributed Representations via Polymodal Embedding Joohee Park∗ Korea Advanced Institute of Science and Technology james.joohee.park@navercorp.com Sung-hyon Myaeng Korea Advanced Institute of Science and Technology myaeng@kaist.ac.kr Abstract A distributed representation has become a popular approach to capturing a word meaning. Besides its success and practical value, however, questions arise about the relationships between a true word meaning and its distributed representation. In this paper, we examine such a relationship via polymodal embedding approach inspired by the theory that humans tend to use di- verse sources in developing a word mean- ing. The result suggests that the existing embeddings lack in capturing certain as- pects of word meanings which can be sig- nificantly improved by the polymodal ap- proach. Also, we show distinct character- istics of different types of words (e.g. con- creteness) via computational studies. Fi- nally, we show our proposed embedding method outperforms the baselines in the word similarity measure tasks and the hy- pernym prediction tasks. 1 Introduction Word representations based on the distributional hypothesis of Harris (1954) have become a domi- nant approach including word2vec (Mikolov et al., 2013) and GloVe (Pennington et al., 2014), which show remarkable performances in a wide spec- trum of natural language processing. However, a question arises about a relationship between a true word meaning and its distributed representa- tion. While the context-driven word representa- tions seem to be able to capture word-to-word re- lations, for example, men is to women as king is to queen, it still remains unclear what aspects of ∗ Currently at Search Solutions Inc., Seongnam, 13561, Korea word meaning they capture and miss. For exam- ple, a word, coffee, can be understood from mul- tiple perspectives. It may be associated with a ce- ramic cup filled with dark brown liquid from the perceptual perspective or an emotion such as hap- piness or tranquility. It may provoke other related concepts like bagel or awakening. We raise the question of how well the current distributed repre- sentation captures such aspects of word meanings. In order to help answering this question, we pro- pose a polymodal word representation based on the theory that humans tend to use diverse sources in developing a word meaning. In particular, we construct six modules for polymodality including linear context, syntactic context, visual percep- tion, cognition, emotion, and sentiments based on the human cognitive model proposed by Maruish and Moses (2013). They are combined to build a single word representation. We conduct a series of experiments to examine the relationships between word meanings and their distributed representations and compare the re- sults with other representations such as word2vec, GloVe, and meta-embedding (Yin and Sch¨utze, 2015). We attempt to understand how well the model capture the diverse aspects of word mean- ings via two experiments: the property norms analysis and the sentiment polarity analysis. The …

Content truncated to 500 words