Automatic conversion of sentence-end expressions for utterance characterization of dialogue systems

Automatic conversion of sentence-end expressions for utterance characterization of dialogue systems Chiaki Miyazaki Toru Hirano Ryuichiro Higashinaka Toshiro Makino Yoshihiro Matsuo NTT Media Intelligence Laboratories 1-1 Hikarinooka, Yokosuka, Kanagawa, Japan {miyazaki.chiaki, hirano.tohru, higashinaka.ryuichiro, makino.toshiro, matsuo.yoshihiro}@lab.ntt.co.jp Abstract Building characters for dialogue agents is im- portant in making the agents more friendly and human-like. To build such characters, utterances suitable for the designated charac- ters are usually manually prepared. However, it is expensive to do this for a large num- ber of utterances for various types of char- acters. We propose a method for automati- cally converting system utterances into those that are characteristic of designated personal attributes, such as gender, age and area of res- idence, to characterize agents. In particular, we focus on converting sentence-end expres- sions, which are considered to greatly affect personal attributes in Japanese. Conversion is done by (i) automatically collecting con- version candidates from various utterances on the Web (e.g., Twitter postings), and (ii) us- ing syntactic and semantic filters to suppress the generation of ill-formed utterances. Ex- perimental results show that our method can convert approximately 95% of utterances into those that are grammatically and semantically acceptable and approximately 90% of utter- ances into those that are perceived to be ac- ceptable for designated personal attributes. 1 Introduction Dialogue agents, which can carry out various tasks according to user demand, have been gaining in pop- ularity due to their convenience and potential in ca- sual conversations with humans. To make the agents more attractive as conversation partners, characteri- zation is important since it makes the agents more friendly and human-like. Characterization here means adding particular personal characteristics to agent utterances; for example, adding the character- istics of a particular person (Mizukami et al., 2015), Big Five personalities (Mairesse and Walker, 2007), or personal attributes such as gender, age and area of residence (which is closely related to dialects). To characterize agents, utterances suitable for the desig- nated characteristics are usually manually prepared. However, it is expensive to do this for a large num- ber of utterances. To reduce this cost, we propose a method for au- tomatically converting utterances into those that are suitable for various characters. In particular, the method automatically modifies ‘how to say it’ (i.e., linguistic expressions) without changing ‘what to say’ (i.e., contents of the utterances). Conversion is done by (i) collecting conversion candidates from various utterances on the Web (e.g., Twitter post- ings), which are annotated with their authors’ per- sonal attributes (this paper deals especially with gen- der, age, and area of residence), and (ii) using syn- tactic and semantic filters to suppress the generation of ill-formed utterances. The rest of the paper is organized as follows. Section 2 introduces studies related to characteriza- tion, Section 3 discusses the features of Japanese sentence-end expressions, Section 4 presents our method for converting sentence-end expressions, Section 5 shows our experimental results, and Sec- tion 6 concludes the paper and refers to future work. 2 Related work Studies related to characterization of …

Content truncated to 500 words