Addressee and Response Selection for Multilingual Conversation

Proceedings of the 27th International Conference on Computational Linguistics, pages 3631–3644 Santa Fe, New Mexico, USA, August 20-26, 2018. 3631 Addressee and Response Selection for Multilingual Conversation Motoki Sato1∗ Hiroki Ouchi2,3 Yuta Tsuboi1∗ 1Preferred Networks, Inc., 2RIKEN Center for Advanced Intelligence Project 3Tohoku University sato@preferred.jp, hiroki.ouchi@riken.jp, tsuboi@preferred.jp Abstract Developing conversational systems that can converse in many languages is an interesting chal- lenge for natural language processing. In this paper, we introduce multilingual addressee and response selection. In this task, a conversational system predicts an appropriate addressee and response for an input message in multiple languages. A key to developing such multilingual re- sponding systems is how to utilize high-resource language data to compensate for low-resource language data. We present several knowledge transfer methods for conversational systems. To evaluate our methods, we create a new multilingual conversation dataset. Experiments on the dataset demonstrate the effectiveness of our methods. 1 Introduction Open-domain conversational systems, such as chatbots, are attracting a vast amount of interest and play their functional and entertainment roles in real-world applications. Recent conversational models are often built in an end-to-end fashion using neural networks, which require a large amount of training data (Vinyals and Le, 2015; Serban et al., 2016). However, it is challenging to collect enough data to build such models for many languages. Consequently, most work has targeted high-resource languages, such as English and Chinese (Shang et al., 2015; Serban et al., 2016). In this work, we aim to develop multilingual conversational systems that can return appropriate re- sponses in many languages. Specifically, we assume the two types of systems: (i) language-specific systems and (ii) language-invariant systems. A language-specific system consists of multiple conver- sational models, each of which returns responses in a corresponding language. By contrast, a language- invariant system consists of a single unified model, which returns responses in all target languages. A key to building these multilingual models is how to utilize high-resource language data to compensate for low-resource language data. We present several knowledge-transfer methods. To the best of our knowl- edge, this is the first work focusing on low-resource language enablement of conversational systems. One challenge when developing conversational systems is how to evaluate the system performance. For generation-based conversational systems, which generate each word for a response one by one, many studies adopt human judgments. However, it is costly and impractical to adopt this evaluation method for multilingual systems, especially for minor-language systems. Thus, as a first step, we develop retrieval-based conversational systems and evaluate the ability to select appropriate responses from a set of candidates. Fig. 1 shows the overview of our multilingual responding systems. This paper provides: (i) formal task definitions, (ii) several knowledge-transfer methods and (iii) a multilingual conversation dataset. First, we introduce and formalize the two task settings: single- language adaptation for language-specific systems and multi-language adaptation for language-invariant systems (Sec. 4). Second, we present several methods leveraging high-resource language data to com- pensate for low-resource language in the two settings (Sec. 5). Our basic method uses …

Content truncated to 500 words