A Deep Neural Network based Approach for Entity Extraction in Code-Mixed Indian Social Media Text

A Deep Neural Network based Approach for Entity Extraction in Code-Mixed Indian Social Media Text Deepak Gupta, Asif Ekbal, Pushpak Bhattacharyya Indian Institute of Technology Patna, India {deepak.pcs16, asif, pb}@iitp.ac.in Abstract The rise in accessibility of web to the mass has led to a spurt in the use of social media making it convenient and powerful way to express and exchange information in their own language(s). India, being enormously diversified country have more than 168 millions users on social media. This diversity is also reflected in their scripts where a majority of users often switch between their native languages to be more expressive. These linguistic variations make automatic entity extraction both a necessary and a challenging problem. In this paper, we report our work for entity extraction in a code-mixed environment. Our proposed approach is based on the popular deep neural network based Gated Recurrent Unit (GRU) archirecture that automatically discovers the higher level features from the text. We do not make use of any handcrafted features or rules, and therefore our proposed model is quite generic in nature. Our experiments on two benchmark datasets of English-Hindi and English-Tamil language pairs show the F-scores of 66.04% and 53.85%, respectively. Keywords: Code-Mixing, Gated Recurrent Unit (GRU), Entity Extraction, Indian Language 1. Introduction The phenomenal growth in user-generated contents on pop- ular social media platforms like Twitter and Facebook has established a new perspective in Natural Language Pro- cessing (NLP). It incurs several challenges ranging from the incorrect spellings to the highly creative words (“f9” for “fine” ), word lengthening (“gooooood” for “good”), phonetic typing, abbreviation (“IDK” for “I don’t know”) and so on. In recent times there has been a rising interest in building several applications on social media contents such as mood identification, opinion mining, etc. How- ever, a majority of the existing research is restricted to- wards resource-rich languages such as English, some of the European languages and few Asian languages. The grow- ing contents of underprivileged languages in the Web has necessitated investigating automated techniques that could build solutions involving these. Code-mixing refers to the mixing of more than one language that often makes the task more complex for building automated techniques. Non- native English speakers do not always use unicode encod- ing scheme to write in social media. Instead, they often use the most frequently used transliterated forms with the English words or phrases. These multilingual speakers of- ten mix multiple languages in addition to anglicisms to be more expressive. There are 22 official languages mentioned in the Indian constitution. The report also reveals that 30 languages are spoken by more than one million native speakers, and 122 are spoken by more than 10,000 people. Language diversity and dialect changes instigate frequent code-mixing in Indian languages. Hence, Indians are multi- lingual by adaptation and necessity, and frequently change and mix languages while writing in social media platforms. These pose additional difficulties in building automated tools for social media analytics. Named Entity Recognition (NER) is a primary task …

Content truncated to 500 words