CN-HIT-MI.T at SemEval-2019 Task 6: Offensive Language Identification Based on BiLSTM with Double Attention

Proceedings of the 13th International Workshop on Semantic Evaluation (SemEval-2019), pages 564–570 Minneapolis, Minnesota, USA, June 6–7, 2019. ©2019 Association for Computational Linguistics 564 CN-HIT-MI.T at SemEval-2019 Task 6: Offensive Language Identification Based on BiLSTM with Double Attention Yaojie zhang, Bing Xu, Tiejun Zhao Laboratory of Machine Intelligence and Translation, Harbin Institute of Technology School of Computer Science and Technology, Harbin Institute of Technology, Harbin, China yjzhang@hit-mtlab.net, hitxb@hit.edu.cn, tjzhao@hit.edu.cn Abstract Offensive language has become pervasive in social media. In Offensive Language Iden- tification tasks, it may be difficult to pre- dict accurately only according to the surface words. So we try to dig deeper semantic in- formation of text. This paper presents use an attention-based two layers bidirectional long- short memory neural network (BiLSTM) for semantic feature extraction. Additionally, a residual connection mechanism is used to syn- thesize two different deep features, and an emoji attention mechanism is used to extract semantic information of emojis in text. We participated in three sub-tasks of SemEval 2019 Task 6 as CN-HIT-MI.T team. Our macro-averaged F1-score in sub-task A is 0.768, ranking 28/103. We got 0.638 in sub- task B, ranking 30/75. In sub-task C, we got 0.549, ranking 22/65. We also tried some other methods of not submitting results. 1 Introduction Recognition of Offensive information has research and application value in many aspects. With the popularity of social media, people’s comments on social media has become an important part of pub- lic opinion. Although freedom of speech is ad- vocated, there are still some unacceptable words. The study of offensive language has only recently arisen. With the deepening of the research, we need to consider the different sub-tasks of its de- composition. In OffensEval tasks (Zampieri et al., 2019b), of- fensive content was divided into three sub-tasks taking the type and target of offenses into account. Sub-task A is offensive language identification. We should identify a short text sentence as offen- sive or non-offensive. Sub-task B is automatic cat- egorization of offense types. We need to classify a sentence as having an attack target or not, if this is an offensive sentence. Sub-task C is offense target identification. Its purpose is to identify the target of an attack sentence with an attack target. The target is individual, group or other. User’s comments on Twitter are usually clut- tered. In order to clean up the data, a series of pre- processing for the data is necessary. Then, pre- trained word vectors are helpful to extract seman- tic features from deep learning model. For clas- sification model, bidirectional long-short memory neural network can catch the contextual informa- tion in text, in order to get offensive semantics from text. A residual connection cascade the first layer’s output and the second layer’s output can get features of text at different levels. Attention mechanism is used for the final output. Besides, referring to ZHANG et al. (2018), emojis in a sen- tence have a significant impact on sentiment. We assume them may affect the offensive semantics of …

Content truncated to 500 words