[month] [year]

Harika Abburi – Discriminatory content in social web

Harika Abburi received  her doctorate in Computer Science and Engineering (CSE). Her research work was supervised by Prof. Vasudeva Varma. Here’s a summary of her research work on Identifying and analyzing the discriminatory content in social web:

Social media has enabled individuals to easily communicate with others while freely sharing their ideas and thoughts by providing anonymity. Due to the substantial growth of user-generated web content in general and social media networks in particular, the amount of discriminatory content is also steadily increasing, leading to significant toxicity in abusive behaviours and online discourse. One type of such harmful content is hate speech, which directly attacks or promotes hate towards an individual member or a group based on their actual or perceived aspects of identity, such as ethnicity, religion, and discrimination. Developing models for this type of data can be difficult due to the fact that the data available on online media may be inadequate, unstructured, and highly imbalanced. In this thesis, we focus on discrimination, especially the form of discrimination perpetrated based on sex.
Sexism, a form of oppression based on one’s sex, manifests itself in numerous ways and causes enormous suffering. During the last years, the role of women within online platforms has gained attention, unfortunately, because of the growing hatred and abuse against them. In view of the growing number of experiences of sexism reported online (accounts of sexism) and the increase in sexist posts in online conversations, automatically categorizing both sorts of posts can assist the fight against sexism, as it can facilitate effective analyses by gender studies researchers and government officials involved in policy making. In this thesis, we investigate the detection and categorization of both types of sexist posts. The existing work on sexism classification has certain limitations in terms of the categories of sexism used and/or whether they can co-occur.
To the best of our knowledge, we work with considerably more categories of sexism than any prior work through our 23-class problem formulation. We present the first semi-supervised work for the multi-label classification of accounts describing any type(s) of sexism. We devise self-training based techniques tailor-made for the multi-label nature of the problem to utilize unlabeled samples for augmenting the labeled set. We identify high textual diversity with respect to the existing labeled set as a desirable quality for candidate unlabeled instances and develop methods for incorporating it into our approach. We also explore ways of infusing class imbalance alleviation for multi-label classification into our semi-supervised learning, independently and in conjunction with the method involving diversity. In addition to data augmentation methods, we develop a neural model which combines biLSTM and attention with
a domain-adapted BERT model in an end-to-end trainable manner. Further, we formulate a multi-level training approach in which models are sequentially trained using categories of sexism of different levels of granularity. Moreover, we devise a loss function that exploits any label confidence scores associated with the data.
We also propose a knowledge-based cascaded multi-task framework for fine-grained multi-label sexism classification that allows for shared learning across multiple tasks through common layers/weights and a combined loss function. We leverage several supporting tasks, including homogeneous and heterogeneous auxiliary tasks. Homogeneous tasks are set up without incurring any manual labeling cost and heterogeneous tasks are set up which have a high correlation with the accounts of sexism. Unlabeled accounts of sexism are utilized through unsupervised learning to help construct our multi-task setup. In addition, we incorporate a knowledge module within the framework to infuse domain-specific external knowledge features into the learning process. Further, we investigate transfer learning that employs weakly labeled accounts of sexism and transfers the learning to the multi-label sexism classification. We also devise objective functions that exploit label correlations in the training data explicitly.
In order to evaluate the versatility of our approaches, we experiment with adapting them for the toxic comment classification dataset. Multiple proposed multi-task methods outperform the state-of-the-art multi-label sexism classification and toxic comment classification across five standard metrics.
In addition to analysing the textual accounts describing sexism suffered or observed in a multi-label multi-class setup, we also analyze how sexism is expressed in online conversations, especially in the social networks, wherein the text perpetrates sexism. Moreover, we also aimed to address the problem of detecting sexism rather than only the categorization. Specifically, we explore the problem of detecting whether a tweet/gab post is sexist or not. We further discriminate the detected sexist post into one of the fine-grained sexism categories in a single-label multi-class manner since labeled data (text perpetrates sexism) is not available wherein categories of sexism can occur together. We propose a neural model for sexism detection and classification that can combine representations obtained using the RoBERTa
model and linguistic features such as Empath, Hurtlex, and Perspective API by involving recurrent components. We also leverage the unlabeled sexism data to infuse the domain-specific transformer model into our framework. Our proposed framework also features a knowledge module comprised of emoticon and hashtag representations to infuse the external knowledge-specific features into the learning process. Several proposed methods outperform various baselines across several standard metrics