Domain Transfer based Data Augmentation for Neural Query Translation

Proceedings of the 28th International Conference on Computational Linguistics, pages 437–449 Barcelona, Spain (Online), December 8-13, 2020 437 Improving Abstractive Dialogue Summarization with Graph Structures and Topic Words Lulu Zhao Beijing University of Posts and Telecommunications, Beijing, China zhaoll@bupt.edu.cn Weiran Xu ∗ Beijing University of Posts and Telecommunications, Beijing, China xuweiran@bupt.edu.cn Jun Guo Beijing University of Posts and Telecommunications, Beijing, China guojun@bupt.edu.cn Abstract Recently, people have been beginning paying more attention to the abstractive dialogue summa- rization task. Since the information flows are exchanged between at least two interlocutors and key elements about a certain event are often spanned across multiple utterances, it is necessary for researchers to explore the inherent relations and structures of dialogue contents. However, the existing approaches often process the dialogue with sequence-based models, which are hard to capture long-distance inter-sentence relations. In this paper, we propose a Topic-word Guided Dialogue Graph Attention (TGDGA) network to model the dialogue as an interaction graph ac- cording to the topic word information. A masked graph self-attention mechanism is used to in- tegrate cross-sentence information flows and focus more on the related utterances, which makes it better to understand the dialogue. Moreover, the topic word features are introduced to assist the decoding process. We evaluate our model on the SAMSum Corpus and Automobile Master Corpus. The experimental results show that our method outperforms most of the baselines. 1 Introduction Due to the explosive growth of the textual information, text summarization, which is an important task in Natural Language Processing (NLP), has been widely studied for several years. It can be categorized into two types: extractive and abstractive. Extractive methods select sentences or phrases from the source text directly (Nallapati et al., 2017; Zhou et al., 2018; Zhang et al., 2018a; Wang et al., 2019), while abstractive methods, which are more similar to how humans summarize texts, attempt to understand the semantic information of source text and generate new expressions as the summary. Recently, neural network methods have led to encouraging results in the abstractive summarization of single-speaker documents like news, scientific publications, etc (Rush et al., 2015; Gehrmann et al., 2018; Xu et al., 2020). These approaches employ a sequence-to-sequence general framework where the documents are fed into an encoder network and another decoder network learns to decode the summary. With the popularity of phone calls, e-mails, and social network applications, people share information in more different ways, which are often in the form of dialogues. Different from news texts, dialogue is a dynamic information exchange flow, which is often informal, verbose and repetitive, sprinkled with false-starts, backchanneling, reconfirmations, hesitations, and speaker interruptions (Sacks et al., 1974). Besides, utterances are often turned from different interlocutors, which leads to the topic drifts, and lower information density. These problems need to be solved using natural language generation techniques with a high level of semantic understanding. Some early works benchmarked the abstractive dialogue summarization task using the AMI meeting corpus, which contains a wide range of annotations, including dialogue acts, topic descriptions, …

Content truncated to 500 words