In-Depth Seminar in the Software Department Explores the Use of Deep Reinforcement Learning in Query-Based Text Summarization

Duhaa Fadill Abbas
As part of the ongoing scientific activities held by the Department of Software at the College of Information Technology – University of Babylon, Ph.D. candidate Abeer Hussein Jabir Abbas delivered an in-depth seminar presenting her dissertation titled:
"Deep Reinforcement Learning and Topic Modeling Using BERT-Based Coherence for Extractive Query-Focused Summarization", under the supervision of Prof. Dr. Wafaa Mohammed Saeed, and in the presence of the department’s scientific committee.

The dissertation aims to develop an intelligent system capable of performing query-focused extractive summarization with enhanced precision and efficiency, leveraging advanced techniques in Natural Language Processing (NLP) and Machine Learning. The researcher proposes an innovative model that integrates Deep Reinforcement Learning, Topic Modeling, and a BERT-based Coherence Metric, to generate contextually coherent, non-redundant, and query-relevant summaries from multiple textual sources.

The proposed system follows a structured multi-phase approach. Initially, it utilizes Latent Dirichlet Allocation (LDA) combined with the K-Means clustering algorithm to semantically group documents based on their topic distributions. This step significantly narrows the search space by isolating the most topically relevant subset of documents in response to a user query.

Subsequently, the system trains an agent using a Deep Q-Network (DQN) to extract sentences that are not only highly relevant to the query but also novel (non-repetitive) and contextually coherent when assembled into a summary. This methodological design ensures that the final output meets user needs for focused, meaningful, and high-quality summarization.