Sentiment classification with syntactic relationship and attention for teaching evaluation texts

In this paper, teaching evaluation refers to the students' evaluation of teaching. To help complete the teaching evaluation work better, we construct the corpus of teaching evaluation texts and complete the sentiment classification on it. The corpus is collected from a university and processed, which includes 10,299 Chinese sentences. The annotators manually label texts according to the rules designed by educational experts. These texts are divided into three categories, which are positive, negative, and neutral. This paper proposes a sentiment classification method for teaching evaluation texts based on Attention and BiLSTM (Bi-directional Long Short-Term Memory) combined with Syntactic Relationships (BLASR). In this model, the syntactic relationships of sentences are fused into the BiLSTM for feature learning. The weights of different words in the sentences are calculated through an attention layer. The sentiment classification of the teaching evaluation texts is completed by a dense layer. The experimental results show that the classification accuracy of BLASR proposed in this paper on the dataset of teaching evaluation texts is 89.04%, which outperforms baselines. It can satisfy the needs of teaching evaluation in colleges.