Context-aware Information Extraction from Multi-thread Business Conversations

Abstract

This paper primarily focuses on developing an end-to-end solution which can process multi-threaded conversations and perform information extraction (IE) specific to a domain and intended business task. The challenges of IE in a conversation are a) context understanding, which consists of two elements: topic and sense of expression and b) establishing context flow. Since the target is free-flow dialogue, understanding the change in contexts is crucial. In this research, we attempt to build a solution that can infer and connect these contexts and reflect the same in the extracted information, taking care of things like negotiations. The proposed approach has three main steps; The first step is domain-dependent which performs topic classification at the sentence level. The second step is domain-independent, and it categorizes sentences into different semantic classes, to understand the conversation flow and parse it into multiple conversation threads. In the final step, we carry out morphological parsing to extract the target value, utilizing the predicted sentence class labels along with the conversation flow. A buyer-seller chat conversation is taken as the sample domain and the target IE is towards information for purchase order generation.