Open Domain Question Answering Using Early Fusion of Knowledge Bases and Text (1809.00782v1)

Abstract: Open Domain Question Answering (QA) is evolving from complex pipelined systems to end-to-end deep neural networks. Specialized neural models have been developed for extracting answers from either text alone or Knowledge Bases (KBs) alone. In this paper we look at a more practical setting, namely QA over the combination of a KB and entity-linked text, which is appropriate when an incomplete KB is available with a large text corpus. Building on recent advances in graph representation learning we propose a novel model, GRAFT-Net, for extracting answers from a question-specific subgraph containing text and KB entities and relations. We construct a suite of benchmark tasks for this problem, varying the difficulty of questions, the amount of training data, and KB completeness. We show that GRAFT-Net is competitive with the state-of-the-art when tested using either KBs or text alone, and vastly outperforms existing methods in the combined setting. Source code is available at https://github.com/OceanskySun/GraftNet .

Open Domain Question Answering Using Early Fusion of Knowledge Bases and Text

This paper examines the evolution of Open Domain Question Answering (QA) from traditional, complex pipelines to deep neural networks designed for end-to-end learning. The research focuses on the integration of two primary information sources: structured Knowledge Bases (KBs) and text documents, through a model named GRAFT-Net. This approach addresses scenarios where neither the KB nor the text alone is sufficient for comprehensive QA.

Key Contributions and Findings

1. Single-model Architecture: The authors propose GRAFT-Net, an early fusion model that combines text and KB data into a single, unified graph structure. The model uses graph representation learning to propagate information across this heterogeneous graph, thereby enabling the extraction of answer entities in a question-specific subgraph containing both text and KB elements.
2. Graph Construction and Learning: The GRAFT-Net leverages graph convolutional networks and includes innovations such as heterogeneous update rules tailored for diverse node types. Additionally, it implements directed propagation inspired by personalized PageRank, allowing efficient dissemination of embeddings along relevant paths in the graph.
3. Benchmark Tasks and Experimental Validation: To evaluate GRAFT-Net, the authors construct benchmark tasks that test the model's ability to handle varying degrees of question difficulty and KB completeness. Results indicate that GRAFT-Net is robust across these conditions, outperforming existing methods in scenarios that require early fusion of text and KBs.
4. Comparison with Existing Methods: The paper highlights that GRAFT-Net is competitive with state-of-the-art models in both specialized KB-only and text-only QA tasks, demonstrating its versatility and efficiency.
5. Implications and Future Directions: The research opens avenues for practical applications in AI systems where information must be extracted from multiple sources. It suggests potential improvements in real-world QA systems by combining diverse data types into a cohesive analysis method. Future work could explore the integration of open text spans as outputs and the enhancement of the subgraph retrieval process.

Implications for AI Development

The integration of structured KB and unstructured text sources into a single queryable framework represents a significant advancement in QA systems. It acknowledges the limitations of relying solely on either information modality and points towards a more integrated data-driven QA approach. Furthermore, the research underscores the importance of developing robust, flexible models capable of adapting to varying levels of KB completeness and question complexity, which is critical for the scalability of AI applications in dynamic and information-rich environments.

Numerical Results

In empirical tests, GRAFT-Net consistently showed strong performance improvements when compared to both late fusion methods and existing QA models, particularly in settings where the KB is incomplete, and the integration of information from text is necessary. These results underscore the model's capability to generalize across different task settings, further supporting the viability of early fusion strategies for multi-source QA applications.

In conclusion, this paper presents a comprehensive framework for QA through the early fusion of KBs and text, establishing a baseline for future advances in hybrid information source QA models. The research provides both a theoretical basis and practical algorithmic innovations, positioning itself as a robust tool for researchers and practitioners aiming to enhance QA systems with combined data inputs.