摘要： 部分文本中隐喻本体与喻体位置相距较远,导致模型学习文本语境信息的难度增大,以及所提取的特征中重要信息不明显。为此,提出一种基于全局与局部特征加权融合的隐喻识别模型。首先,设计了局部特征提取模块(LFEM),通过对不同范围以及更大感受野下文本局部特征的关注来达到学习词语周围不同距离语境信息的目的;其次,使用双向长短时记忆(BiLSTM)与多头注意力构成全局特征提取模块(GFEM),学习宏观句子级语义信息;最后,设计了特征加权融合模块(FWFM),对提取得到的2种特征进行自适应动态融合,以较少的噪声获得鲁棒性更强且重要信息更为集中的特征。实验结果表明,相比RoBERTa+Transformer+GCN模型,所提模型在VUA ALLPOS、TOEFL ALLPOS以及CCL 3个数据集上的F1值分别提升了1.1、1.2和3.2百分点,所提模型具有更高的隐喻识别精度。

关键词: 隐喻识别, 全局特征, 局部特征, 特征加权, 注意力机制, 双向长短时记忆

Abstract: This study presents a metaphor recognition model based on the weighted integration of global and local features to address the problem of the location of a metaphor body and the metaphor body being far apart in some texts. This issue makes it difficult for a model to learn the contextual information of texts, and consequently the important information of extracted features remaining unclear. First, a Local Feature Extraction Module (LFEM) is designed to learn contextual information at different distances around words by focusing on local features across different ranges and larger receptive fields. Second, Bidirectional Long Short-Term Memory (BiLSTM) and multi-head attention are used to construct a Global Feature Extraction Module (GFEM) to learn macrosentence-level global features. Finally, a Feature Weighted Fusion Module (FWFM) is designed to perform adaptive dynamic fusion of the two extracted features and obtain more robust features with less noise and more concentrated important information. Experimental results show that compared to the RoBERTa+Transformer+GCN model, the F1 value of the proposed model increases by 1.1, 1.2, and 3.2 percentage points on the VUA ALLPOS, TOEFL ALLPOS, and CCL datasets, respectively. The proposed model has higher metaphor recognition accuracy.

Key words: metaphor recognition, global feature, local feature, feature weighting, attention mechanism, Bidirectional Long Short-Term Memory (BiLSTM)

中图分类号: 

• TP391.1