Activity Modeling in Email

Proceedings of NAACL-HLT 2016, pages 1452–1462, San Diego, California, June 12-17, 2016. c⃝2016 Association for Computational Linguistics Activity Modeling in Email Ashequl Qadir School of Computing University of Utah Salt Lake City, UT 84112, USA asheq@cs.utah.edu Michael Gamon Microsoft Research One Microsoft Way Redmond, WA 98052, USA mgamon@microsoft.com Patrick Pantel Microsoft Research One Microsoft Way Redmond, WA 98052, USA ppantel@microsoft.com Ahmed Hassan Awadallah Microsoft Research One Microsoft Way Redmond, WA 98052, USA ahmed.awadallah@microsoft.com Abstract We introduce a latent activity model for work- place emails, positing that communication at work is purposeful and organized by activities. We pose the problem as probabilistic infer- ence in graphical models that jointly capture the interplay between latent activities and the email contexts they govern, such as the recipi- ents, subject and body. The model parameters are learned using maximum likelihood estima- tion with an expectation maximization algo- rithm. We present three variants of the model that incorporate the recipients, co-occurrence of the recipients, and email body and subject. We demonstrate the model’s effectiveness in an email recipient recommendation task and show that it outperforms a state-of-the-art gen- erative model. Additionally, we show that the activity model can be used to identify email senders who engage in similar activities, re- sulting in further improvements in recipient recommendation. 1 Introduction Activities are a prominent characteristic of a work- place, typically governed by people’s job roles and work responsibilities. Examples of workplace activ- ities can include organizing a conference, purchas- ing equipment, managing candidate interviews, etc. Activities can be viewed as a collaborative work practice involving a set of people each playing a different role in the activity (Dredze et al., 2006). Although emails are an integral part of workplace communication, current email clients offer little sup- port for the activity oriented use of email (Khous- sainov and Kushmerick, 2005). Discussions can get split across long email threads and communications about all activities can get intermixed, making activ- ity management difficult (Balakrishnan et al., 2010). In this work, we model activities as latent proba- bility distributions personalized to the email sender. We present three variants of the activity model, in- corporating: (1) email recipients, (2) email recip- ient pairs which account for co-occurrence of the email recipients, and (3) email body and subject to- kens along with email recipient pairs. Additionally, we experiment with lexical (bag of words), syntac- tic (nouns and verb phrases), and semantic (things of interest in an email) representations of the body and subject tokens of an email. The parameters of the generative model are learned using an expectation maximization (EM) algorithm. For evaluation, we formulate a real world task set- ting for email recipient recommendation, where we assume that all but the last recipient of an email has been entered by the sender, and we test the effective- ness of our activity model in recommending the last recipient. Such a system has practical applications, such as reminding an email sender about a poten- tially forgotten recipient or recommending the next recipient as …

Content truncated to 500 words