Urdu Sentiment Corpus (v1.0): Linguistic Exploration and Visualization of Labeled Dataset for Urdu Sentiment Analysis

摘要

The significance of the labeled dataset is not obscure from artificial intelligence practitioners. We have seen much phenomenal work, in natural language processing, for many languages (like English, Chinese, and Arabic, etc.), due to the reason for the availability of substantial data. For the Urdu language, despite the third largest spoken language in the world, very little research work is shown; hence, it is adjudged as a ‘morphologically rich’ but ‘resource-poor’ language. Further, the researchers working on Urdu natural language processing are in a quandary due to the lack of availability of labeled/annotated datasets. This paper shares the data, “Urdu Sentiment Corpus” (USC), and insights therein, of Urdu tweets for the sentiment analysis and polarity detection. The dataset is consisting of tweets, such that it casts a political shadow and presents a competitive environment between two separate political parties versus the government of Pakistan. Overall, the dataset is comprising over 17, 185 tokens with 52% records as positive, and 48 % records as negative. This paper shares the visual insights (from document-level to word-level) into the textual similarities, manifold-learning, etc. In addition to it, this paper also presents a Part-of-Speech wise analysis and an unpretentious technique for the extraction of sentiment lexicons from the corpus.