The Semantic Similarity Measurement Of Sentence Pairs And Its Application

The semantic similarity of sentence pairs refers to the similarity between a sentence and another sentence in semantics.It plays an important role in many applications,such as the similarity matrix construction in abstraction text summarization,similar question retrieval in community question answering system,and the similar query retrieval in chatbot systems.The semantic similarity measurement methods can be classified into supervised model and unsupervised model by the labeled dataset.Unsupervised models mainly include edit distance,word vector method and BM25.Supervised models mainly include linear model,support vector machine model,convolutional neural network model,recurrent neural network model,and attention model.This paper proposes a sentence semantic similarity measurement model based on the siamese neural network and aligned word extraction.The model consists of three key components.The first key component is called the Sentence Related Information Representation.It is used to introduce the interrelated features between sentences and sentences,so that the neural network input contains not only the original sentence but also the correlation information between the sentences.The second key component is called the Siamese Recurrent Neural Network,The Recurrent Neural Network not only considers the current step input when processing the sequence information,but also considers the output of the previous step.It is suitable for processing sequence information.The siamese neural network is a network with two identical structures with parameter sharing and is suitable for extracting similar features between input data.This paper combines the siamese neural network with the recurrent neural network to extract similar features between sequence pairs.The last key component is the alignment word extraction.Alignment words defines that the words themselves have the same or similar semantics,and their contexts are also the same or similar.It is also used to extract similar features between sequence pairs.Finally,this model is applied to the International Semantic Evaluation dataset SemEval-2016 and got an excellent result.