Matching Document Pairs using Multi-Feature Semantic Fusion Based on Knowledge Graph

Yibo Chen, Zuping Zhang, Xin Huang, Xing Xiang, Zhiqiang He, Yunsheng Chen, Qihui HuResearch 2021 · DOI

Abstract

Discriminating the homology and heterogeneity of two documents in information retrieval is a very important and difficult step. Existing methods mainly focus on word-based document duplicate checking or sentence pairs matching except manual verification which need a lot of human resource cost. The word-based document duplicate checking can not judge the similarity of two documents from the semantic level and the matching sentence pair methods can not effectively mine the semantic information from a long text which is frequent retrieval results. A concept-based Multi-Feature Semantic Fusion Model (MFSFM) is proposed. It employs multi-feature enhanced semantics to construct a concept map for represent the document, and employs a multi-convolution mixed residual CNN module to introduce local attention mechanism for improve the sensitivity of conceptual boundary information. To improve the feasibility of the proposed MFSFM based on concept maps, two multi-feature document data sets are set up. Each of them consists of about 500 actual scientific and technological project feasibility reports. Experimental results based on the actual datasets show that the proposed MFSFM converges quickly while expanding the latest methods of natural language matching at the accuracy rate.

Introduction

In modern information retrieval, determining whether two documents are homologous (representing the same project or topic) or heterogeneous is a significant challenge. This is particularly true for long-form technical content where manual verification is prohibitively expensive. Traditional word-based methods like TF-IDF are often "semantic-blind," while deep-learning sentence-matching models struggle to scale to documents containing tens of thousands of words.

The MFSFM Framework

To address these limitations, this paper introduces the Multi-Feature Semantic Fusion Model (MFSFM). The core innovation lies in representing long documents as "Concept Graphs" rather than flat sequences of text. This approach allows the model to capture the logical and structural relationships between technical concepts.

Concept Graph Construction

Instead of processing every word linearly, the document is transformed into a weighted graph. Vertices represent "concepts"—clusters of related keywords—and edges represent the interaction strength between them. Sentences from the original text are then mapped to their most relevant concept vertex, preserving semantic context within a manageable topological structure.

Mapping text sentences to a concept graph structure

Multi-Feature Encoding and GCN Fusion

The model utilizes several layers to process these graphs:

  1. Contextual Multi-Feature Embedding (CMFE): Enhances standard word embeddings with Part-of-Speech (POS) tags and positional information.
  2. MCMR-CNN: A local attention mechanism that identifies conceptual boundaries within the text.
  3. Graph Convolutional Network (GCN) Fusion: Aggregates local matching vectors across the document's topology, allowing the model to judge similarity based on the overall structure of the arguments.
The transformation and aggregation process using GCN layers

Datasets and Evaluation

The researchers evaluated MFSFM using two new benchmarks derived from real-world scientific and technological project feasibility reports:

  • CNSR (Chinese Feasibility Study Same Project): 4,678 pairs labeled for project identity.
  • CNSI (Chinese Feasibility Study Same Subject): 2,464 pairs labeled for general topic similarity.

These documents are exceptionally long, averaging over 9,000 words and reaching up to 32,000 words, far exceeding the context window of standard Transformer models like BERT.

Key Results and Insights

  • Significant Accuracy Gains: MFSFM improved classification accuracy by 13.17% on the CNSR dataset and 19.82% on the CNSI dataset compared to previous state-of-the-art methods.
  • Outperforming BERT: While BERT fine-tuning achieved 80-85% accuracy, MFSFM reached over 88% by effectively bypassing the truncation issues inherent in standard attention mechanisms.
  • Efficiency: By using community detection to group keywords into concepts, the researchers reduced graph complexity, slashing training time by over 50% without sacrificing accuracy.

Conclusion

The MFSFM demonstrates that for long-form technical documents, structure is as vital as individual word semantics. By fusing multi-feature embeddings with graph-based representations, the model provides a robust and efficient solution for identifying document homology in complex industrial datasets.

Summary updated

Abstract

Abstract Discriminating the homology and heterogeneity of two documents in information retrieval is very important and difficult step. Existing methods mainly focus on word-based document duplicate checking or sentence pairs matching except manual verification which need a lot of human resource cost. The word-based document duplicate checking can not judge the similarity of two documents from the semantic level and the matching sentence pair methods can not effectively mine the semantic information from a long text which is frequent retrieval results. A concept-based Multi-Feature Semantic Fusion Model (MFSFM) is proposed. It employs multi-feature enhanced semantics to construct a concept map for represent the document, and employs a multi-convolution mixed residual CNN module to introduce local attention mechanism for improve the sensitivity of conceptual boundary information. To improve the feasibility of the proposed MFSFM based on concept maps, two multi-feature document data sets are set up. Each of them consists of about 500 actual scientific and technological project feasibility reports. Experimental results based on the actual datasets show that the proposed MFSFM converges quickly while expanding the latest methods of natural language matching at the accuracy rate.
BibTeX
@inproceedings{chen2021matching,
  author = {Yibo Chen and Zuping Zhang and Xin Huang and Xing Xiang and Zhiqiang He and Yunsheng Chen and Qihui Hu},
  title = {{Matching Document Pairs using Multi-Feature Semantic Fusion Based on Knowledge Graph}},
  booktitle = {Research},
  year = {2021},
  doi = {10.21203/rs.3.rs-907473/v1},
}
Figure