Sentence Classification with Transfer Network

Abstract

Sentence classification is a fundamental task in natural language processing. In this paper, clickbait detection is taken as an example to study the sentence classification with a transferring network. Clickbait are news headlines that exaggerate the facts or hide partial facts headlines. Statistics show that clickbaits are prevalent among all languages. However, previous research on clickbait detection mainly focus on English, exploiting lexical or syntactical features that are not shared by other languages. On the other hand, it would be both time-consuming and labor-intensive to annotate a clickbait dataset by humans. Recently, neural language model that represent each word by a real-valued, dense vector show that words with similar meanings across languages are close to each other in the vector space. Inspired by this, transfer learning is proposed to be applied to transfer the model on clickbait detection from a source language to other languages with very few annotations. This paper trains the source model on English corpus and transfers it to corpus in Chinese. Experimental results show that transfer learning model in this paper can achieve similar performance on the target language using less annotation, showing the effectiveness and robustness of this model.