Feature Representation for Effective Action-Item Detection

This paper reports on a new task: automated action-item detection, in order to flag emails that require responses, and to highlight the specific passage(s) indicating the request(S) for action.

Abstract

E-mail users face an ever-growing challenge in managing their inboxes due to the growing centrality of email in the workplace for task assignment, action requests, and other roles beyond information dissemination. Whereas Information Retrieval and Machine Learning techniques are gaining initial acceptance in spam filtering and automated folder assignment, this paper reports on a new task: automated action-item detection, in order to flag emails that require responses, and to highlight the specific passage(s) indicating the request(s) for action. Unlike standard topic-driven text classification, action-item detection requires inferring the sender’s intent, and as such responds less well to pure bag-of-words classification. However, using enriched feature sets, such as n-grams (up to n=4) with chi-squared feature selection, and contextual cues for action-item location improve performance by up to 10 % over unigrams, using in both cases state of the art classifiers such as SVMs with automated model selection via embedded cross-validation.