基于依存结构学习的中学数学术语鲁棒抽取

Summary

The paper addresses automatic term extraction from Chinese middle-school mathematics texts. Earlier extractors ignored sentence syntax, and those that did use it suffered from segmentation and parsing mistakes. The authors propose DSL-ATRE, which treats extraction as character-level sequence labelling with the tags B-T, M-T, E-T, S-T and O.

The model works in four steps:

• Encoding: BERT produces character embeddings, and a BiLSTM captures context.
• Graph building: Stanford CoreNLP segments and parses each sentence, and the word-level dependency links become undirected character-to-character edges.
• Denoising: ATT-GCN, a graph network adapted from A-GCN, uses dependency-label attention to reweight edges and dampen unreliable arcs.
• Fusion and decoding: multi-head attention combines the structural and contextual views, and a CRF decodes the tags.

Experiments use a self-built corpus of math textbook and teaching sentences whose annotation was refined for this work. DSL-ATRE beats the BERT-BiLSTM-CRF baseline by 2.21 points in precision and 1.22 in F1. It also handles hard cases better, such as ambiguous single-character terms, nested terms and implicit double-predicate sentences. This shows that syntax can help term extraction even when the parser output is noisy.

Key points

• Final scores on the middle-school math term dataset are P = 83.82% and F1 = 82.51%, gains of 2.21 and 1.22 points over the baseline.
• The corpus has 10,934 sentences from new 人教版 (PEP) textbooks, exam syllabi and lesson plans. Labels were refined using Baidu Baike, textbooks and the 数学大辞典 dictionary, and the authors also built a separate set of complex-context examples.
• Annotation quality was checked against a third-party annotator using C-rate, Kendall and Fleiss Kappa. After refinement these rose by 2.48%, 11.44% and 25.72%, reaching κ = 0.9960.
• ATT-GCN computes edge weights by combining dependency-label embeddings with node states through dot-product attention. These learned weights replace the binary adjacency matrix in a syntax-fused GCN layer.
• The authors describe this as the first term-extraction work to use dependency structure while explicitly mitigating segmentation and parse errors.