Abstract
Toxic Language in social media is a newly emerging virtual disorder of human society. Detecting toxic language is an NLP task that requires a Dataset of utterances [1]. For the Bangla language, very few datasets have been developed on toxicity or similar concepts [2]. A dataset has been developed using user-generated content from Facebook and that will cover the demographic and thematic distribution of Bangla toxic language generated on the web. Therefore, 2207590 comments have been collected, annotated, and thus extract about 1959 unique bigrams as utterances, which were considered as base-entry of a toxic language dataset. The core derivatives of the dataset are bigram-based wordlists, which are annotated inductively and divided into 08 thematic classes that give some ideas on toxicity variations found in the Bengali community. These thematic classes cover political hate speech [3] and misogynist bullies dominantly. However, these thematic labels will serve as classifiers in the text classification process through machine learning. In addition to the thematic classification labels, this dataset includes some additional features such as imprecise meanings in English, IPA transliteration, real occurrences in the source pages, spelling standards, and degree of toxicity. As this is a dataset of utterance, it has de-identified and anonymous entries and no difficulties for public disclosure. Therefore, we consider this dataset as Toxic lexicon (Toxlex) as an exhaustive wordlist that is essentially a curated value-added and analyzed dataset which can be used as classifier material to detect toxicity in social media.