The Dots Have Their Values: Exploiting the Node-Edge Connections in Graph-based Neural Models for Document-level Relation Extraction

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 431–440 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 431 Dialogue Generation on Infrequent Sentence Functions via Structured Meta-Learning Yifan Gao1, Piji Li2, Wei Bi2, Xiaojiang Liu2, Michael R. Lyu1, and Irwin King1 1 Department of Computer Science and Engineering, The Chinese University of Hong Kong, Shatin, N.T., Hong Kong 2 Tencent AI Lab, Shenzhen, China 1{yfgao,king,lyu}@cse.cuhk.edu.hk 2{pijili,victoriabi,kieranliu}@tencent.com Abstract Sentence function is an important linguistic feature indicating the communicative purpose in uttering a sentence. Incorporating sen- tence functions into conversations has shown improvements in the quality of generated re- sponses. However, the number of utterances for different types of fine-grained sentence functions is extremely imbalanced. Besides a small number of high-resource sentence func- tions, a large portion of sentence functions is infrequent. Consequently, dialogue generation conditioned on these infrequent sentence func- tions suffers from data deficiency. In this pa- per, we investigate a structured meta-learning (SML) approach for dialogue generation on in- frequent sentence functions. We treat dialogue generation conditioned on different sentence functions as separate tasks, and apply model- agnostic meta-learning to high-resource sen- tence functions data. Furthermore, SML en- hances meta-learning effectiveness by promot- ing knowledge customization among different sentence functions but simultaneously preserv- ing knowledge generalization for similar sen- tence functions. Experimental results demon- strate that SML not only improves the informa- tiveness and relevance of generated responses, but also can generate responses consistent with the target sentence functions. 1 Introduction Humans express intentions in conversations through sentence functions, such as interrogation for acquiring further information, declaration for making statements, and imperative for making re- quests and instructions. For machines to interact with humans, it is therefore essential to enable them to make use of sentence functions for dialogue gen- eration. Sentence function is an important linguis- tic feature indicating the communicative purpose of a sentence in a conversation. There are four Query 今天心情不错 I am in a good mood today Sentence Function Positive Declarative Response 1 什么让你这么开心？ What makes you so happy? Sentence Function 1 Wh-style Interrogative Response 2 我今天心情很不好 I feel bad today Sentence Function 2 Negative Declarative Figure 1: Query-response pairs with fine-grained sen- tence functions. Responses under different sentence functions are completely different in global structures. major sentence functions: Declarative, Interroga- tive, Exclamatory and Imperative (Rozakis, 2003). Each major sentence function can be further de- composed into fine-grained ones according to dif- ferent purposes indicated in conversations. For example, Interrogative is divided into Wh-style In- terrogative, Yes-no Interrogative and other types. These fine-grained sentence functions have great influences on the structures of utterances in conver- sations including word orders, syntactic patterns, and other aspects (Akmajian, 1984; Yule, 2016). Figure 1 presents how sentence functions influence the responses. Given the same query expressed in Positive Declarative, the responses expressed in Wh-style Interrogative and in Negative Declarative are completely different. Although the use of sentence functions improves the overall quality of generated responses (Ke et al., 2018), it suffers from the data imbalance …

Content truncated to 500 words