연상작용과 샴 네트워크를 통한 문장 유사도 판별

Abstract

Modeling sentence similarity plays an important role in natural language processing tasks such as question answering and plagiarism detection. Measuring semantic relationship of two sentences is challenging because of the variability and ambiguity of linguistic expression. Previous studies on sentence similarity are focusing on the configuration of input data and classification model structure. However, we focus on the sentence understanding process of human. Human brain stimulates association effect when one tries to understand a sentence describing landscape or object. The association effect that transforms text into image makes human robust to expression changes and word order changes in a sentence. To implement the association effect, we propose a new sentence similarity model based on Siamese network and Text2image generative adversarial network (GAN). The role of Siamese network is to compute the similarity between two sentences with the shared network weights. Inside the Siamese network, two subnetworks are composed of Text2image GAN which transforms the text data into image data. Once the sentences are transformed into image, latent features are extracted through VGGNet. The sentence similarity is computed from the normalized distance between two feature vectors. To verify our proposed method, we modify the MSCOCO dataset and experimental results show that the proposed method outperforms the benchmarked models without association process.