Extractive Summarization Method for Contact Center Dialogues based on Call Logs

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 500–508, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Extractive Summarization Method for Contact Center Dialogues based on Call Logs Akihiro Tamura1* , Kai Ishikawa2, Masahiro Saikou2, and Masaaki Tsuchida2 1Multilingual Translation Laboratory, MASTAR Project, National Institute of Information and Communications Technology, Kyoto, Japan 2 Information and Media Processing Laboratories, NEC Corporation, Nara, Japan akihiro.tamura@nict.go.jp, k-ishikawa@dq.jp.nec.com, m-saikou@ax.jp.nec.com , m-tsuchida@cq.jp.nec.com Abstract This paper proposes a novel extractive summarization method for speech dialo- gues between agents and customers in contact centers. The proposed method does not require any extra cost for apply- ing the method such as preparing rules or creating training data. Conventional me- thods such as the tf*idf method, which gives importance to characteristic words in an input text, can miss the essential points for contact center work. Our pro- posed method evaluates the importance of each utterance from the standpoint of call agents who report calls for managing or analyzing calls. Specifically, the pro- posed method includes information fre- quently reported by call agents in sum- maries using past call logs commonly recorded in the contact center. Evaluation using real data (call dialogues and call logs) shows that the proposed method can extract essential points in terms of contact center work and outperforms the conventional method. 1 Introduction In recent years, the role of contact centers has become more important in many companies. This is because each contact center is a main channel for individual customers to directly access a company and the voice of the customers can be used to improve the quality of the compa- ny's products and services. Many contact centers have two problems of human cost. The first one is that agents spend much time documenting logs. Logs are usually generated by agents based on their memories or handwritten memos after a call. Such logs are used for reporting, managing, and analyzing their calls. Figure 1 shows an example of a call dialo- gue and its corresponding call log (We simply refer to these as a dialogue and a log hereafter). The other cost is that managers also spend much time understanding the details of calls. This is because managers must listen to speech data or browse dialogue texts generated by au- tomatic speech recognition (ASR), which are lengthy and include many uninformative parts. Accordingly, automatic summarization of di- alogues is required as an effective solution to the above problems. For the first problem, the sum- mary can be an alternative to the logs or draft for agents. Byrd et al. (2008) have shown that auto- matic summarization helps to reduce the time for documenting calls. For the second one, the sum- mary can help to reduce the time spent listening to the speech data or browsing the dialogue texts. In previous works (Zechner, 1996; Edmund- son, 2004; Orasan et al., 2004; Murray et al., 2007; Higashinaka et al., 2010), the methods summarize an input text without considering the Call Dialogues …

Content truncated to 500 words