Automatic Action Extraction for Short Text Conversation Using Unsupervised Learning

Abstract

Collaboration tools are important for workplace communication. The amount of conversation data produced in workplaces are increasing rapidly, placing a burden on workers. There is a necessity to analyze large amounts of data automatically to extract actionable information. Multiple studies were conducted on action extraction to identify actions such as promises and requests. Most of these studies used supervised learning methods. The key problem discussed in this paper are (i) the automatic extraction of action types from short text conversations in collaboration tools such as Twitter and Slack, and (ii) leveraging large amounts of data using unsupervised learning. Data labelling is an important issue when dealing with large datasets for training and extending the corresponding algorithms across different actions and domains. In this paper, we propose an unsupervised learning approach using a combination of relation extraction techniques and word embedding to leverage large amounts of data. The action2vec model is created to identify specific actions from short text of conversation data. We have evaluated our unsupervised method against supervised learning and the results are comparable. The action type extractor is integrated with Slack to provide assistance for action type extraction. Thus, the contributions of this paper include an unsupervised learning method to utilize large amounts of data, an automatic extraction of action types from short text and the integration of our approach with state of the art collaboration tools.