Improving the Representation of Sentences with Reinforcement Learning and AMR graph

Sentence Embedding is a technique that represents the meaning of sentences in vector form, playing a crucial role in various natural language processing tasks such as question-answering, sentiment analysis, and information retrieval. Therefore, understanding the meaning and structure of sentences is essential. We propose a novel approach to improve the performance of Sentence Embedding by utilizing Abstract Meaning Representation(AMR) parsing and reinforcement learning. We generate Sentence Embeddings using AMRBART, a type of AMR parser, and evaluate them in Question Answering (QA) tasks. In this process, we measure the similarity between the AMR graphs of two sentences using the Weighted Walks with Lookahead score and employ the Deep Deterministic Policy Gradient algorithm, a reinforcement learning algorithm, to enhance this score. By integrating AMR syntactic analysis and reinforcement learning into the Sentence Embedding generation process, we enable a more accurate understanding of natural language sentences.