An NLP-Based Approach for Automated Task Identification in Unstructured Chat Conversations

Abstract

The task of identifying relevant actions from short workplace messages is difficult because of the lack of contextual information, use of informal language, and code-mixing.   Although transformer models have demonstrated remarkable success in minimal-context settings, their performance in such settings has not been adequately explored. This paper presents the application of binary task identification in a single-utterance classification paradigm, where messages are assessed independently without any conversational context. Three different methods are employed for the task: a traditional lexical approach using TF-IDF with logistic regression, a multilingual transformer model with XLM-RoBERTa-base, and an optimized transformer model. Empirical evaluation on 1,190 real-world workplace messages with class imbalance and Hinglish code-mixing demonstrates that lexical approaches are robust baselines, and optimized transformer models perform best. The results show that increased model complexity does not necessarily translate to improved performance and the need for empirical analysis in designing task identification systems for minimal-context workplace communication