Fine-Grained Sentence Functions for Short-Text Conversation

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3984–3993 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 3984 Fine-Grained Sentence Functions for Short-Text Conversation Wei Bi1, Jun Gao1,2, Xiaojiang Liu1, Shuming Shi1 1Tencent AI Lab, Shenzhen, China 2School of Computer Science and Technology, Soochow University, Suzhou, China {victoriabi, jamgao, kieranliu, shumingshi}@tencent.com Abstract Sentence function is an important linguistic feature referring to a user’s purpose in utter- ing a specific sentence. The use of sentence function has shown promising results to im- prove the performance of conversation mod- els. However, there is no large conversa- tion dataset annotated with sentence functions. In this work, we collect a new Short-Text Conversation dataset with manually annotated SEntence FUNctions (STC-Sefun). Classifi- cation models are trained on this dataset to (i) recognize the sentence function of new data in a large corpus of short-text conver- sations; (ii) estimate a proper sentence func- tion of the response given a test query. We later train conversation models conditioned on the sentence functions, including information retrieval-based and neural generative models. Experimental results demonstrate that the use of sentence functions can help improve the quality of the returned responses. 1 Introduction The ability to model and detect the purpose of a user is essential when we build a dialogue system or chatbot that can have coherent conversations with humans. Existing research has analyzed var- ious factors indicating the conversational purpose such as emotions (Prendinger and Ishizuka, 2005; Zhou et al., 2018; Shi and Yu, 2018), topics (Xing et al., 2017; Wang et al., 2017), dialogue acts (Lis- combe et al., 2005; Higashinaka et al., 2014; Zhao et al., 2017) and so on. This work describes an ef- fort to understand conversations, especially short- text conversations (Shang et al., 2015), in terms of sentence function. Sentence function is an impor- tant linguistic feature referring to a user’s purpose in uttering a specific sentence (Rozakis, 2003; Ke et al., 2018). There are four major sentence func- tions: Interrogative, Declarative, Imperative and Exclamatory (Rozakis, 2003). Sentences with dif- ferent sentence functions generally have different structures of the entire text including word orders, syntactic patterns and other aspects (Akmajian, 1984; Yule, 2016). Some work has investigated the use of sentence function in conversation models. For example, Li et al. (2016) propose to output interrogative and imperative responses to avoid stalemates. Ke et al. (2018) incorporate a given sentence function as a controllable variable into the conditional varia- tional autoencoder (CVAE), which can encourage the model to generate a response compatible with the given sentence function. Considering the importance of sentence func- tion in conversation modeling, it is surprised to find that no large conversation dataset has been annotated with sentence functions. In Ke et al. (2018), they only labeled a small dataset with 2,000 query-response pairs. Sentence function classifiers are trained and tested on this dataset and the best model only achieves an accuracy of 78%, which is unsatisfactory to serve as an …

Content truncated to 500 words