Emotion Driven Prediction of Conversational Thread Depth

Abstract

While chatbots have become enhanced in human-computer interaction, they hardly provide quality and interesting interactions. This study emphasizes on conversational thread depth prediction, which quantifies the ability of a conversational thread to stay productive and valuable with a text-only emotion-based flow. Conversational thread depth is determined using the emotional features taken from textual data, then used to enhance chatbot responses for their appropriateness to the context or sensitivity to the emotions being expressed. The model developed under this study makes use of feature extraction through the use of Keras and Word2Vec, and implemented using Long Short-Term Memory (LSTM) with attention mechanism. This enables the model to detect more important affective features of text, thus making the model better suited to predicting the depth of conversation. Given experimental findings, it is noted that the text-only scenario is rather efficient, with training accuracy equal to 80% and validation accuracy increased to 76%. This means that the model can generalize well given the limitations of text-based nature of the data but some of the issues such as the problem of over fitting and accommodating for variety in patterns of human conversational exchanges. Chatbot engagement and automated customer service are the real-world uses in this study. Although this study is limited to the text-only mode, the current basis suggested for subsequent additions and improvements. Future studies will use audio-visual information to extend the range of rankings and broaden the emotions and context levels.