Lin: Unsupervised Extraction of Tasks from Textual Communication

Proceedings of the 28th International Conference on Computational Linguistics, pages 1815–1819 Barcelona, Spain (Online), December 8-13, 2020 1815 Li˙n: Unsupervised Extraction of Tasks from Textual Communication Parth Diwanji, Hui Guo, Anup K. Kalia∗, Munindar P. Singh North Carolina State University, Raleigh, NC ∗IBM Thomas J. Watson Research Center, Yorktown Heights, NY {pdiwanj, hguo5}@ncsu.edu, anup.kalia@ibm.com, mpsingh@ncsu.edu Abstract Commitments and requests are a hallmark of collaborative communication, especially in orga- nizational settings. Identifying specific tasks being committed to or requests from emails and chat messages can enable important downstream tasks, such as producing to-do lists, reminders, and calendar entries. State-of-the-art approaches for task identification rely on large annotated datasets, which are not always available, especially for domain-specific tasks. Accordingly, we propose Li˙n, an unsupervised approach of identifying tasks that leverages dependency parsing and VerbNet. Our evaluations show that Li˙n yields comparable or more accurate results than supervised models on domains with large training sets, and maintains its excellent performance on unseen domains. 1 Introduction In organizational settings where team members interact with each other, tasks are constantly exchanged through communications. For example, a team leader may request a team member to accomplish a task via email. Team members may chat with each other to make commitments about how tasks are assigned within a team. Efficient team collaborations, including creating to-do lists and meetings, presume that the tasks are clear. How can natural language processing (NLP) support such practices? We define a task as a verb phrase that specifies a single action to be carried out. One or more tasks could arise in a message from emails or chats, from a sender to a receiver. We consider the root verb of such a verb phrase as its main verb. Therefore, we assume identifying main verbs is essential to identifying tasks. Consider the following sentences indicating commitments or requests:(1) I will send the QA later today and (2) please reschedule the meeting to next week. Each sentence contains only one verb, which can easily be identified as a main verb. However, in some cases, main verbs could be difficult to identify. For example, consider the following sentences containing multiple verbs, and thus, the main verb of a task (in bold) is not obvious (other verbs are underlined): (1) let me know what I need to do to be ready; (2) go ahead and start working on this; (3) before you arrange a meeting, we should think about a few things.; and (4) I would like to meet to discuss or appeal to Greg. Several contributions exist on identifying tasks from emails and chats. Bennett and Carbonell [2005] provide supervised machine learning classifiers for detecting whether a sentence includes a task. Lam- pert et al. [2010] provide a binary classifier to detect requests. Kalia et al. [2013] propose binary classi- fiers to detect the operational forms of business tasks such as commitments of different types. Wang et al. [2019] categorize commitments into three types—Request Information, Schedule Meeting, and Promise Action—and train a deep learning model …

Content truncated to 500 words