摘要：

针对谓语中心词识别模型中存在缺失跨度长度信息和多尺度跨度关联信息等问题, 提出一种融合多尺度跨度特征的汉语谓语中心词识别模型。首先, 使用ChineseBERT预训练语言模型和双向长短期记忆(BiLSTM)网络提取文本中包含上下文信息的字符向量序列; 其次, 利用线性神经网络对字符向量进行初步识别, 形成跨度遮蔽矩阵; 然后, 将字符向量序列二维化表示为跨度信息矩阵, 使用多尺度卷积神经网络(MSCNN)对跨度信息矩阵进行运算, 提取跨度的多尺度关联信息; 最后, 采用特征嵌入神经网络嵌入跨度的长度信息, 丰富跨度的特征向量以识别谓语中心词。实验结果表明, 该模型能够有效融合跨度的多尺度关联信息和长度信息, 提升谓语中心词识别的性能, 相比于同类模型中性能最优的谓语中心词识别模型的F1值提升了0.43个百分点。

关键词: 谓语中心词识别, 多尺度卷积, ChineseBERT预训练语言模型, 跨度长度信息, 多尺度跨度关联信息

Abstract:

To address the issues of missing span length and multiscale span correlation information in predicate center word recognition models, this study proposes a Chinese predicate center recognition model fused with multiscale span features. First, a Chinese Bidirectional Encoder Representations from Transformers (ChineseBERT) pre-trained language model and a Bidirectional Long Short-Term Memory (BiLSTM) network extract character vector sequences containing contextual information from the text. Second, a linear neural network performs the initial recognition of character vectors, forming a span-masking matrix. The character vector sequence is then represented in a two-dimensional format as a span information matrix, and a Multiscale Convolutional Neural Network(MSCNN) processes the span information matrix and extracts multiscale correlation information from the spans. Finally, a feature-embedding neural network embeds the length information of the spans, enriching the feature vectors of the spans for predicate head recognition. The experimental results demonstrate that this model can effectively integrate multiscale correlation and span length information, thereby enhancing the performance of predicate head recognition. Compared to the best-performing existing predicate center word recognition model, the proposed model achieves an improvement of 0.43 percentage points in F1 score.

Key words: predicate center word recognition, multiscale convolution, ChineseBERT pre-trained language model, span length information, multiscale span correlation information