Constructing Interpretive Spatio-Temporal Features for Multi-Turn Responses Selection

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 44–50 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 44 Constructing Interpretive Spatio-Temporal Features for Multi-Turn Responses Selection Junyu Lu†, Chenbin Zhang†, Zeying Xie, Guang Ling, Chao Zhou, Zenglin Xu† † SMILE Lab, University of Electronic Science and Technology of China, Sichuan, China {cs.junyu, aleczhang13, swpdtz, zacharyling}@gmail.com, tom.chaozhou@foxmail.com, zenglin@gmail.com Abstract Response selection plays an important role in fully automated dialogue systems. Given the dialogue context, the goal of response se- lection is to identify the best-matched next- utterance (i.e., response) from multiple can- didates. Despite the efforts of many previ- ous useful models, this task remains challeng- ing due to the huge semantic gap and also the large size of candidate set. To address these issues, we propose a Spatio-Temporal Matching network (STM) for response selec- tion. In detail, soft alignment is first used to obtain the local relevance between the con- text and the response. And then, we con- struct spatio-temporal features by aggregating attention images in time dimension and make use of 3D convolution and pooling operations to extract matching information. Evaluation on two large-scale multi-turn response selec- tion tasks has demonstrated that our proposed model significantly outperforms the state-of- the-art model. Particularly, visualization anal- ysis shows that the spatio-temporal features enables matching information in segment pairs and time sequences, and have good inter- pretability for multi-turn text matching. 1 Introduction Fully automated dialogue systems (Litman and Silliman, 2004; Banchs and Li, 2012; Lowe et al., 2017; Zhou et al., 2018) are becoming increas- ingly important area in natural language process- ing. An important research topic in dialogue sys- tems is response selection, as illustrated in Figure 1, which aims to select an optimal response from a pre-defined pool of potential responses (Kum- merfeld et al., 2018). Practical methods to re- sponse selection are usually retrieval-based, that focus on matching the semantic similarity between the response and utterances in the dialogue his- tory (Shang et al., 2015; Zhang et al., 2018). Recently, convolutional operation, as a useful attempt to explore local correlation, has been in- Figure 1: Examples of the Ubuntu dataset provided by NOESIS 1. Text segments with the same color symbols across context and response can be seen as matched pairs. vestigated to extract the matching features from the attention grid (Wu et al., 2017; Zhou et al., 2018). Unfortunately, these methods usually do not perform well when there are many candidate responses. In fact, in multi-turn dialogues, the next sen- tence is generally based on what was presented be- fore and tends to match a recent local context. This is because the topic in a conversation may change over time, and the effective matching between the dialogue may only appear in a local time period. This phenomena generally appear in video pro- cessing (Hara et al., 2018; Tran et al., 2014), im- age caption (Chen et al., 2017) and action recog- nition (Girdhar and Ramanan, 2017). Therefore, …

Content truncated to 500 words