Primitive-based Learning Model for Intent Detection in Human-AI Agent Negotiation

Abstract

Intent detection task is one of the core problems in the natural language understanding (NLU) as it enables the conversational agents to comprehend the human language for meaningful interaction. But human-AI agent communication is not fully rational due to the dynamic behaviour of the human. Several research works have been carried out to ascertain intent detection methods for social media (Twitter, Facebook) chat, email conversation, and product reviews etc. However, we cannot deploy existing intent detection methods in a real-time human-AI agent negotiation scenario. In this article, we present the first primitive-based learning model, which analyses human (buyer) dynamic behavioral patterns at content level, and automatically predict their purchase or non-purchase intent during a bilateral business negotiation. Thereby, by evaluating and keeping track of two key human behaviors “affective” & “disappointment”, an AI agent can therefore, better comprehend the human intent and engage them in the negotiation, and achieve more agreements (deals). We employ negotiation primitives by setting them to pertinent (human-human) communication units for enhancing the cognition and decision-making capability of AI-agents and evaluate its performance on various machine learning models including BERT. With 93% accuracy, the proposed model itself has proven to be a more efficient method for classifying e-negotiation behavioral patterns.