Computing semantic similarity of texts by utilizing dependency graph.

The problem of Semantic Textual Similarity (STS) is a significant issue in Natural Language Processing (NLP). STS recognizes and measures semantic relations between two texts. Since the ability to determine the degree of the semantic relationship between sentence pairs is an integral part of machines that understand and infer natural language, we intend to improve the performance of the neural network systems computing the degree of the semantic relation. We propose a graph-U-Net model that operates on a dependency graph and is placed on top of a transformer. Our proposed model indicates the importance of the words in the sentence by assigning the words to several levels while a score as a degree of importance is computed for each level. These scores are used as a weighted average to produce the final result. The importance of the words is new information that our proposed model extracts from the STS and Paraphrase Identification (PI) datasets. We examine the effect of the proposed model on the performance of some transformers in computing semantic relation scores. We use STS2017 and MRPC datasets to evaluate our proposed model. Experimental evaluations show that compared to the transformers, our proposed model obtains a higher value of Pearson and Spearman correlation coefficients and also generates valuable representations for each input so that they improve the Pearson and Spearman values of the systems computing the degree of semantic equivalence between two texts. [ABSTRACT FROM AUTHOR]