Can Symbol Grounding Improve Low-Level NLP? Word Segmentation as a Case Study

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 2298–2303, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. Can Symbol Grounding Improve Low-Level NLP? Word Segmentation as a Case Study Hirotaka Kameko†, Shinsuke Mori‡, and Yoshimasa Tsuruoka† †Graduate School of Engineering, The University of Tokyo Hongo, Bunkyo-ku, Tokyo, Japan {kameko, tsuruoka}@logos.t.u-tokyo.ac.jp ‡Academic Center for Computing and Media Studies, Kyoto University Yoshida Honmachi, Sakyo-ku, Kyoto, Japan forest@i.kyoto-u.ac.jp Abstract We propose a novel framework for im- proving a word segmenter using informa- tion acquired from symbol grounding. We generate a term dictionary in three steps: generating a pseudo-stochastically seg- mented corpus, building a symbol ground- ing model to enumerate word candidates, and filtering them according to the ground- ing scores. We applied our method to game records of Japanese chess with com- mentaries. The experimental results show that the accuracy of a word segmenter can be improved by incorporating the gener- ated dictionary. 1 Introduction Today we can easily obtain a large amount of text associated with multi-modal information, and there is a growing interest in the use of non- textual information in the natural language pro- cessing (NLP) community. Many of these studies aim to output natural language sentences from a nonlinguistic modality, such as image (Farhadi et al., 2010; Yang et al., 2011; Rohrbach et al., 2013). Kiros et al. (2014) showed that multi-modal infor- mation improves the performance of a language model. Inspired by these studies, we explore a method for improving the performance of a low-level NLP task using multi-modal information. In this work, we focus on the task of word segmentation (WS) in Japanese. WS is often performed as the first processing step for languages without clear word boundaries, and it is as important as part-of-speech (POS) tagging in English. We assume that a large set of pairs of non-textual data and sentences describing them is available as the information source. In our experiments, the pairs consist of game states in Shogi (Japanese chess) and textual comments on them, which were made by Shogi experts. We enumerate substrings (character se- quences) in the sentences and match them with Shogi states by a neural network model. The ra- tionale here is that substrings which match with non-language data well tend to be real words. Our method consists of three steps (see Figure 1). First, we segment commentary sentences for a game state in various ways to produce word can- didates. Then, we match them with game states of a Shogi playing program. Finally, we compile the symbol grounding results at all states and incorpo- rate them to an automatic WS. To the best of our knowledge, this is the first result reporting a per- formance improvement in an NLP task by symbol grounding. 2 Stochastically Segmented Corpus Before symbol grounding, we need to segment the text into words that include probable candi- date words. For this purpose, we use a stochasti- cally segmented corpus (SSC) (Mori and Takuma, 2004). Then we propose …

Content truncated to 500 words