Abstract
Reading and writing research papers is a fundamental skill for researchers, yet it remains a significant hurdle for newcomers. We propose the Deep ReAder-Writer (DRAW) network, an intelligent agent designed to bridge the gap between information extraction and text generation. DRAW consists of a Reader that extracts and enriches knowledge graphs (KGs) from input text, a Writer that generates novel paragraphs from these graphs, and a Reviewer that provides multi-faceted feedback via reinforcement learning. Our experiments on the AGENDA and M-AGENDA datasets demonstrate that DRAW significantly outperforms existing state-of-the-art methods in generating fluent, factually aligned research abstracts.
Introduction
The exponential growth of scientific literature presents a daunting challenge for researchers. With thousands of papers published daily, the ability to not only summarize individual works but also synthesize "potential knowledge"—hidden connections across multiple sources—is crucial. Traditional AI models often specialize in either reading (Information Extraction) or writing (Data-to-Text). However, a truly intelligent research assistant must do both simultaneously, understanding structured knowledge to generate coherent, novel scientific text.
The DRAW Framework
The DRAW network mimics the human research process through a tripartite architecture: the Reader, the Writer, and the Reviewer.
1. The Reader (Text-to-Graph)
The Reader module extracts entities and relations from source abstracts to construct a Knowledge Graph. Beyond simple extraction, it employs a Graph Neural Network (GNN) for link prediction. By analyzing multi-hop neighborhoods, the Reader can predict latent relationships between entities that were not explicitly stated in the source text, effectively "discovering" new research hypotheses.
2. The Writer (Graph-to-Text)
The Writer utilizes a graph encoder to capture the global context and local topology of the enriched KG. A Transformer-based decoder then translates these structured representations into natural language. This approach ensures that the generated text is grounded in the discovered knowledge rather than being a mere statistical hallucination.
3. The Reviewer (Feedback Rewards)
To overcome the limitations of standard "teacher-forcing" training, the Reviewer module provides three reinforcement learning signals:
- Quality Reward: Measures basic fluency using metrics like BLEU.
- Adversarial Reward: A discriminator trained to distinguish between human and machine-written text, pushing the Writer toward more natural phrasing.
- Alignment Reward: Ensures the generated paragraph accurately reflects the entities and relations present in the KG.
Experimental Results
We evaluated DRAW on the AGENDA dataset (40,000 AI abstracts) and a newly introduced M-AGENDA dataset designed for multi-source synthesis.
As shown in the results above, DRAW (Writer-Reviewer) achieves significant improvements over baselines like GraphWriter and CGE-LW. Notably, the CIDEr score—which measures the similarity to human-written consensus—jumped to 45.21, a substantial increase over the previous state-of-the-art.
In human evaluations, DRAW demonstrated impressive realism. In a Turing-style test, human participants identified DRAW-generated abstracts as human-written 48% of the time. For comparison, participants only correctly identified actual human-written text 68% of the time, suggesting that the gap between machine-generated and human-authored scientific summaries is narrowing.
Conclusion
The DRAW network represents a significant step toward autonomous research assistants. By using Knowledge Graphs as a structured "thought space" and a multi-objective Reviewer for refinement, DRAW can synthesize information across papers to generate high-quality scientific text. Future work will focus on expanding the model's scope from single paragraphs to full-length research papers and enabling the discovery of entirely new scientific concepts.
Code and supplementary materials are available at https://github.com/menggehe/DRAW.
Summary updated