基于RoBERTa-CNN-BiLSTM-CRF的高中数学知识实体识别

Entity Recognition of High School Mathematics Knowledge Based on RoBERTa-CNN-BiLSTM-CRF

摘要: 命名实体识别是自然语言处理中的一个重要研究步骤，也是自然语言中许多下游任务的前置任务。传统的命名实体识别方法通常采用简单线性或非线性模型进行识别，实体识别的准确率不高。随着深度学习的引入，能够处理更为复杂的非线性问题，使用神经网络模型来提高实体识别的准确率。本文提出一种基于RoBERTa-CNN-BiLSTM-CRF模型，用于高中数学知识实体的识别任务。首先利用RoBERTa模型中的双向Transformer编码层对数据的特征进行提取与分析生成相应的词向量，然后利用卷积神经网络(CNN)中的卷积层和池化层实现降维操作，提取句子中的局部特征，最后构建适合高中数学知识点实体识别的LSTM-CRF模型进行训练和处理。经过实验表明，该模型具有较高的准确性。准确率、召回率和F1分别达到94.32%、94.58%和94.45%。

Abstract: Named entity recognition is an important research step in natural language processing, and it is also the pre-research of many downstream tasks in natural language. The traditional method usually adopts a simple linear or nonlinear model for entity recognition, and its accuracy is not high. With the introduction of deep learning, it can deal with more complex nonlinear problems, and use neural network models to improve the accuracy of entity recognition. In this paper, a RoBERTa-CNN-BiLSTM-CRF model is proposed for recognition of mathematical knowledge entities in high school. Firstly, the bidirectional Transformer coding layer in RoBERTa model is used to extract and analyze data features and generate corresponding word vectors. Then, the convolution layer and pooling layer in a convolutional neural network (CNN) are used to achieve dimensionality reduction and extract local features in sentences. Finally, the LSTM-CRF model suitable for entity recognition of mathematics knowledge points in high school is constructed for training and processing. Experiments show that the model has high accuracy. Precision, recall and F1 reached 94.32%, 94.58% and 94.45%, respectively.