Disentangling Online Chats with DAG-structured LSTMs

Proceedings of the 10th Conference on Lexical and Computational Semantics, pages 152–159 August 5–6, 2021, Bangkok, Thailand (online) ©2021 Association for Computational Linguistics 152 Disentangling Online Chats with DAG-Structured LSTMs Duccio Pappadopulo∗1, Lisa Bauer∗2, Marco Farina1, Ozan ˙Irsoy1, Mohit Bansal2 1Bloomberg 2UNC Chapel Hill {dpappadopulo, mfarina19, oirsoy}@bloomberg.net {lbauer6, mbansal}@cs.unc.edu Abstract Many modern messaging systems allow fast and synchronous textual communication among many users. The resulting sequence of messages hides a more complicated structure in which independent sub-conversations are in- terwoven with one another. This poses a chal- lenge for any task aiming to understand the content of the chat logs or gather information from them. The ability to disentangle these conversations is then tantamount to the success of many downstream tasks such as summariza- tion and question answering. Structured in- formation accompanying the text such as user turn, user mentions, timestamps, is used as a cue by the participants themselves who need to follow the conversation and has been shown to be important for disentanglement. DAG- LSTMs, a generalization of Tree-LSTMs that can handle directed acyclic dependencies, are a natural way to incorporate such information and its non-sequential nature. In this paper, we apply DAG-LSTMs to the conversation disen- tanglement task. We perform our experiments on the Ubuntu IRC dataset. We show that the novel model we propose achieves state of the art status on the task of recovering reply-to re- lations and it is competitive on other disentan- glement metrics. 1 Introduction Online chat and text messaging systems like Face- book Messenger, Slack, WeChat, WhatsApp, are common tools used by people to communicate in groups and in real time. In these venues multiple independent conversations often occur simultane- ously with their individual utterances interspersed. It is reasonable to assume the existence of an underlying thread structure partitioning the full conversation into disjoint sets of utterances, which ideally represent independent sub-conversations. ∗ Equal contribution [12:19] <tapia> and some dist-upgrade [...] [12:19] <microhaxo> lol thats what im using [12:19] <bob2> tapia: you’re using hoary [...] [12:19] <microhaxo> and it wont connect meh [12:19] <tapia> bob2: no, breezy χi φi τj ψj = [φj; τj] sij Figure 1: Excerpt from the IRC dataset (left) and our reply-to classifier architecture (right). Blue dots repre- sent a unidirectional DAG-LSTM unit processing the states coming from the children of the current node. Red dots represent the GRU units performing thread encoding. At this point in time, we are computing the score (log-odds) of fifth utterance replying to the third. The task of identifying these sub-units, disentan- glement, is a prerequisite for further downstream tasks among which question answering, summa- rization, and topic modeling (Traum et al., 2004; Shen et al., 2006; Adams and Martell, 2008; El- sner and Charniak, 2010). Additional structure can generally be found in these logs, as a particular utterance could be a response or a continuation of a previous one. Such reply-to relationships implic- itly define threads as the connected components of the resulting graph topology, and can then be used …

Content truncated to 500 words