Fine-grained semantic textual similarity measurement via a feature separation network

Semantic text similarity (STS), which measures the semantic similarity of sentences, is an important task in the field of NLP. It has a wide range of applications, such as machine translation (MT), semantic search, and summarization. In recent years, with the development of deep neural networks, the existing semantic similarity measurement has made great progress. In particular, pretraining models, such as BERT-based models, which have been good representations of sentence features, have set a new state-of-the-art on STS tasks. Although a large amount of corpus data are used in the pretraining stage, there is no fine-grained semantic analysis. We observe that many sentences, such as user reviews and the QA corpus, can be abstractly regarded as including two core parts: a) this sentence states a certain attribute; and b) this attribute is described by descriptive words. This feature is particularly prominent in the corpus of reviews. Motivated by the above observations, in this paper, we propose a feature separation network (FSN) model, which can further separate and extract attribute features and description features and then measure the semantic similarity according to the separated features. To better verify the effectiveness of our model, we propose an unsupervised approach to construct the semantic similarity dataset in the review domain. Experimental results demonstrate that our method outperforms the general semantic similarity measurement method.