Deep Reinforcement Learning Based Chat Bot Using Semantic Parsing Method

Semantic parsing plays a key role in chatbots as it enables the bot to understand the user inputs and process accurately for coherent interactions. To improve text coherence of chatbots, the proposed research focused on using semantic parsing method integrated with Deep Reinforcement Learning (DRL). Initially, the Task Oriented Parsing (TOP) dataset is considered and pre-processed to extract the grammar, convert the ground truth tree and identify the label types. To predict the masked tokens, structured focused semantic parsing is used enhancing the model's ability. The DRL framework with encoder-decoder model uses a Sequence-2-Sequence (Seq2Seq) approach with attention mechanism to effectively capture the semantic correlation between input output sequences. The reward functions then evaluate the weighted sum of the factors like semantic content, coherence, and fluency to balance with hyperparameters. A cumulative learning strategy simulated the dialogues over multiple turns while reducing expected future reward with defined loss function. Experimental results showed better accuracy in semantic parsing with 98.91% and BiLingual Evaluation Understudy (BLEU) score of 3.85 when compared to the results of conventional methods like Deep Learning based Reinforcement Learning (DL-RL) and Transformer based Seq2Seq DRL.