JFCKB: Japanese Feature Change Knowledge Base

JFCKB: Japanese Feature Change Knowledge Base Tetsuaki Nakamura†, Daisuke Kawahara†‡ †Kyoto University, ‡JST PRESTO Yoshida-Honmachi, Sakyo-ku, Kyoto, Japan tnakamura@nlp.ist.i.kyoto-u.ac.jp, dk@i.kyoto-u.ac.jp Abstract Commonsense knowledge plays an essential role in our language activities. Although many projects have aimed to develop language resources for commonsense knowledge, there is little work focusing on connotational meanings. This is because constructing com- monsense knowledge including connotational meanings is challenging. In this paper, we present a Japanese knowledge base where arguments in event sentences are associated with various feature changes caused by the events. For example, “my child” in “my wife hits my child” is associated with some feature changes, such as increase in pain, increase in anger, increase in disgust, and decrease in joy. We constructed this knowledge base through crowdsourcing tasks by gathering feature changes of arguments in event sentences. After the construction of the knowledge base, we conducted an experiment in anaphora resolution using the knowledge base. We regarded anaphora resolution as an antecedent candidate ranking task and used Ranking SVM as the solver. Experimental results demonstrated the usefulness of our feature change knowledge base. Keywords: emotion, commonsense knowledge, knowledge base, crowdsourcing, anaphora resolution 1. Introduction Commonsense knowledge plays an essential role in our language activities. Such knowledge also plays an important role for computers to understand texts. In the area of language resource development, most studies on commonsense knowledge focus on denota- tional meanings such as predicate-argument structures (Baker et al., 1998; Palmer et al., 2005). Meanwhile, there are few studies focusing on connotational meanings (Nakamura and Kawahara, 2016; Rashkin et al., 2016). Even in such studies, abstract features such as polarities are used. However, concrete knowledge (that is, fine- grained knowledge) is better for computers than abstract knowledge because events actually cause various concrete feature changes to participants in the events or those who know of the events. There are two approaches to acquiring commonsense knowledge. One is the automatic acquisition approach; the other is the manual acquisition approach. The au- tomatic acquisition approach uses machine learning tech- niques and pattern matching methods. Although this ap- proach is useful when the amount of data to be acquired is extremely large, the quality of acquired knowledge might be low. Moreover, the existence of reporting bias in texts (Gordon et al., 2010; Gordon and Durme, 2013) shows that infrequent events tend to appear more in texts than quotid- ian events. The manual approach is useful for gathering subjective information (such as emotional information) and quotidian commonsense knowledge, which are difficult to acquire automatically. This approach may use a fully man- ual technique (at a very early stage of artificial intelligence studies) or collective intelligence (e.g., crowdsourcing and games with a purpose). In this study, we aim to solve some problems in previ- ous work on connotational meanings by constructing a large-scale knowledge base of concrete feature changes of arguments in event sentences with a controlled gran- ularity. For example, “my child” in “my wife hits my child” is associated with some feature changes such …

Content truncated to 500 words