Discourse-Aware Neural Extractive Text Summarization

In the era of information overload, effective text summarization plays a pivotal role in distilling vast amounts of content into concise, digestible insights. Traditional approaches to extractive summarization often fall short in capturing the nuanced discourse structure of text, leading to summaries that lack coherence and context. DiscoBERT emerges as a pioneering solution, integrating discourse-aware methodologies to bridge this gap and revolutionize the landscape of text summarization.


Enter DiscoBERT, a groundbreaking project poised to revolutionize how we distill vast volumes of text into succinct, coherent summaries. Designed to navigate the intricate landscape of discourse structures, DiscoBERT offers a beacon of clarity amidst the sea of information overload. By integrating Enhanced Discourse Units (EDUs) segmentation and Rhetorical Structure Theory (RST) parsing, DiscoBERT unveils the hidden patterns and relationships within text, transforming it into digestible insights. With its innovative approach and cutting-edge methodology, DiscoBERT represents a beacon of hope for those seeking to unravel the complexities of textual information. Join us on a journey through the world of DiscoBERT, where every word counts, and clarity reigns supreme

What We Did

In our quest to enhance extractive text summarization, we embarked on a journey to develop DiscoBERT. Leveraging insights from discourse analysis and neural network architectures, we devised a novel approach that incorporates Enhanced Discourse Units (EDUs) segmentation and Rhetorical Structure Theory (RST) parsing. By integrating these techniques, DiscoBERT not only identifies relevant content but also captures the inherent relationships and rhetorical patterns within the text.

Technologies Used

Challenges Faced During Model Training

  1. Navigating the Discourse Maze: One of the foremost challenges we encountered in training DiscoBERT was navigating the intricate maze of discourse structures within text. Like explorers in a dense forest, we delved deep into the tangled web of relationships between sentences, striving to extract meaningful insights amidst the complexity.
  2. Taming the Hyper-parameter Hydra: Hyper-parameter tuning resembled a battle against the mythical Hydra, with each adjustment spawning new challenges. Balancing parameters such as batch size, use of discourse graphs, and trigram blocking required strategic finesse akin to facing a multi-headed beast.
  3. The EDU Conundrum: Deciding whether to use Enhanced Discourse Units (EDUs) as selection units presented a conundrum akin to choosing the right path in a labyrinth. We grappled with the trade-offs between using EDUs or sentences, each option offering its own set of challenges and rewards.
  4. Coreference Chaos: Coreference resolution emerged as a realm of chaos, akin to untangling a web of interconnected threads. Resolving references and constructing Coreference Graphs demanded meticulous attention to detail amidst the tangled linguistic landscape.
  5. The RST Puzzle: Parsing Rhetorical Structure Theory (RST) trees felt like solving an intricate puzzle, where each piece represented a rhetorical relationship waiting to be deciphered. Achieving accurate RST parsing required patience and perseverance akin to solving a cryptic enigma.
  6. Model Performance Odyssey: Navigating the vast expanse of model performance metrics felt like embarking on an epic odyssey, with each evaluation resembling a new adventure. Optimizing ROUGE F-1 scores demanded a journey through uncharted territory, with twists and turns at every step.
  7. The Graph Expedition: Constructing discourse graphs resembled embarking on an expedition into uncharted territory, with each node and edge representing a discovery waiting to be made. Encoding relationships within Coreference and RST Graphs required strategic mapping akin to charting a course through unknown waters.
  8. The Preprocessing Expedition: Preprocessing datasets and preparing model inputs felt like embarking on an expedition through a dense jungle, with each step revealing new challenges. Tokenization, sentence splitting, and coreference resolution demanded meticulous attention akin to blazing a trail through dense foliage.
  9. The Training Marathon: Training DiscoBERT resembled running a marathon through a rugged terrain, with each epoch representing a grueling stretch of the journey. Fine-tuning model parameters and architecture demanded endurance akin to pushing through fatigue to reach the finish line.
  10. The Performance Balancing Act: Achieving balance between summary length and content coverage resembled walking a tightrope suspended between two extremes. Selecting the optimal number of units for summarization demanded precision akin to maintaining equilibrium amidst shifting winds.

Key Features

DiscoBERT boasts several key features that distinguish it from traditional summarization models:

  1. Discourse-Aware Methodology: By incorporating EDU segmentation and RST parsing, DiscoBERT captures the intricate discourse structure of text, resulting in summaries that are both informative and coherent.
  2. Graph-based Representation: The model constructs discourse graphs, including Coreference Graphs and RST Graphs, to encode relationships between entities and rhetorical structures, enhancing summary quality.
  3. Flexible Training Framework: Built on AllenNLP, DiscoBERT offers a customizable training framework with adaptable hyper-parameters, facilitating seamless model customization and evaluation.

Featured Images


Our system performed really well in tests. It beat other methods in making money from trading stocks. We also made sure our system was safe by closing deals if we started losing too much money.


In conclusion, DiscoBERT represents a significant advancement in extractive text summarization, offering a sophisticated solution that transcends traditional approaches. By leveraging discourse-aware methodologies and graph-based representations, DiscoBERT not only captures the essence of the input text but also delivers summaries that are coherent and contextually rich. With its promising results and innovative features, DiscoBERT holds immense potential for transforming the landscape of information retrieval and document understanding.