Information Bottleneck Inspired Method For Chat Text Segmentation

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 194–203, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP Information Bottleneck Inspired Method For Chat Text Segmentation S Vishal1,⋆, Mohit Yadav2,⋆,$, Lovekesh Vig1 and Gautam Shroff1 1TCS Research New Delhi, India 2University of Massachusetts, Amherst s.vishal3@tcs.com, ymohit@cs.umass.edu, lovekesh.vig@tcs.com, gautam.shroff@tcs.com Abstract We present a novel technique for segment- ing chat conversations using the infor- mation bottleneck method (Tishby et al., 2000), augmented with sequential conti- nuity constraints. Furthermore, we utilize critical non-textual clues such as time be- tween two consecutive posts and people mentions within the posts. To ascertain the effectiveness of the proposed method, we have collected data from public Slack conversations and Fresco, a proprietary platform deployed inside our organiza- tion. Experiments demonstrate that the proposed method yields an absolute (rel- ative) improvement of as high as 3.23% (11.25%). To facilitate future research, we are releasing manual annotations for seg- mentation on public Slack conversations. 1 Introduction The prolific upsurge in the amount of chat con- versations has notably influenced the way peo- ple wield languages for conversations. Moreover, conversation platforms have now become preva- lent for both personal and professional usage. For instance, in a large enterprise scenario, project managers can utilize these platforms for various tasks such as decision auditing and dynamic re- sponsibility allocation (Joty et al., 2013). Logs of such conversations offer potentially valuable information for various other applications such as automatic assessment of possible collaborative work among people (Rebedea et al., 2011). ⋆ indicates that both authors contributed equally. $ indi- cates that the author was at TCS Research New-Delhi during the course of this work. It is thus vital to invent effective segmentation methods that can seperate discussions into small granules of independent conversational snippets. By ’independent’, we meant a segment should as much as possible be self-contained and discussing the same topic, such that a segment can be sug- gested if any similar conversation occurs again. As an outcome of this, various short text simi- larity methods can be employed directly. Seg- mentation can also potentially act as an empower- ing preprocessing step for various down-streaming tasks such as automatic summarization (Dias et al., 2007), text generation (Barzilay and Lee, 2004), information extraction (Allan, 2012), and conver- sation visualization (Liu et al., 2012). It is worth noting that chat segmentation presents a number of gruelling challenges such as, the informal na- ture of the text, the frequently short length of the posts and a significant proportion of irrelevant in- terspersed text (Schmidt and Stone). Research in text segmentation has a long his- tory going back to the earliest attempt of Koz- ima (1993). Since then many methods, includ- ing but not limited to, Texttiling (Hearst, 1997), Choi’s segmentation (Choi, 2000), representation learning based on semantic embeddings (Alemi and Ginsparg, 2015), and topic models (Du et al., 2015a) have been presented. Albeit, very little re- search effort has been proposed for segmenting in- formal chat text. For instance, …

Content truncated to 500 words