2021 2nd International Conference on Computers, Information Processing and Advanced Education | 2021
Siamese-ERNIE: Topic Relevance Check for English Writing
Abstract
Topic relevance is an important criterion for English writing assessment. It has always been hard work for teachers to evaluate the writing assignments from students. The development of deep learning and natural language processing can help by checking the semantic relevance of article title and content. CNN and RNN have shown good performance in semantic text similarity. However, they usually require large corpora and huge computational power for training. In the case of English writing topic relevance checking, it is hard to collect an exhaustive corpus due to the topic variety. BERT has established an outstanding performance in different natural language processing tasks based on transfer learning. ERNIE set a new level for natural language tasks, since it uses entity information based on BERT. Siamese-ERNIE is a modification of the pretrained ERNIE network that uses siamese network structures to derive semantically meaningful sentence embeddings. We use Siamese-ERNIE to train (fine-tuned) the title and content semantic relevance on our corpus based on a pretrained-model. As the article title and content embedding can be calculated by our Siamese-ERNIE model, we can then measure the topic relevance by computing the semantic cosine distance of its embedding.