Information extraction from digital social trace data with applications to social media and scholarly communication data

Abstract

Shubhanshu is a Machine Learning Researcher at Twitter working on the Content Understanding Research team. He finished his Ph.D. at the iSchool, University of Illinois at Urbana-Champaign, where he worked as a research assistant with Dr. Jana Diesner and Dr. Vetle Torvik on projects funded by NIH, NSF, KISTI, and Army Research Lab. Shubhanshu's research work is at the intersection of machine learning, information extraction, social network analysis, and visualizations. His work won then Best Student Paper Award at ASIS&T SIGMET workshop. Shubhanshu has published multiple open source projects as well as open datasets related to his research. More information about Shubhanshu can be found at: https://shubhanshu.com. In his thesis, Shubhanshu worked on unifying the concept of information extraction from social media data and scholarly data under the framework of Digital Social Trace Data (DSTD). This framework allow us to answer questions like conceptual novelty, expertise in scientific communities as well as influential users in social media corpus. He further proposed improved information extraction techniques for social media text by using active, semi-supervised, and multi-task machine learning along with the inclusion of meta-data. Additionally, his work resulted in the development of a visualization framework for DSTD along with multiple open source tools and datasets. For more details visit https://shubhanshu.com/phd_thesis/.

Exploring multi-task multi-lingual learning of transformer models for hate speech and offensive speech identification in social media

“…For this paper, we extend some of the techniques that we have used in TRAC 2020 in Mishra et al (2020b) as well as Mishra (2019Mishra ( , 2020a, and apply them to the HASOC data-set Mandl et al (2019). Furthermore, we extend the work that we did as part of the HASOC 2019 shared task by experimenting with multi-lingual training, back-translation based data-augmentation, and multi-task learning to tackle the data sparsity issue of the HASOC 2019 data-set.…”

“…In order to automate hate speech detection the Natural Language Processing (NLP) community has made significant progress which has been accelerated by organization of numerous shared tasks aimed at identifying hate speech (Mandl et al, 2019;Kumar et al, 2020Kumar et al, , 2018. Furthermore, there has been a proliferation of new methods for automated hate speech detection in social media text (Salminen et al, 2018;Mishra et al, 2020b;Mishra, 2020a;Waseem et al, 2017;Struß et al, 2019;Mandl et al, 2019;Mondal et al, 2017). However, working with social media text is difficult (Eisenstein, 2013;Mishra and Diesner, 2016;Mishra et al, 2014;Mishra and Diesner, 2019;Mishra, 2019Mishra, , 2020b, as people use combinations of different languages, spellings and words that one may never find in any dictionary.…”

“…Furthermore, there has been a proliferation of new methods for automated hate speech detection in social media text (Salminen et al, 2018;Mishra et al, 2020b;Mishra, 2020a;Waseem et al, 2017;Struß et al, 2019;Mandl et al, 2019;Mondal et al, 2017). However, working with social media text is difficult (Eisenstein, 2013;Mishra and Diesner, 2016;Mishra et al, 2014;Mishra and Diesner, 2019;Mishra, 2019Mishra, , 2020b, as people use combinations of different languages, spellings and words that one may never find in any dictionary. A common pattern across many hate speech identification tasks Mandl et al (2019); Kumar et al (2020); Waseem et al (2017); Zampieri et al (2019); Basile et al (2019); Struß et al (2019) is the identification of various aspects of hate speech, e.g., in HASOC 2019 (Mandl et al, 2019), the organizers divided the task into three sub-tasks, which focused on identifying the presence of hate speech; classification of hate speech into offensive, profane, and hateful; and identifying if the hate speech is targeted towards an entity.…”

“…Our aim is to be able to perform as good if not better than the best model for each task language combination. Our approach is inspired from the promising results of multi-task learning on some of our recent works (Mishra, 2019(Mishra, , 2020b. Additionally, while building a unified model which can perform well on all tasks is challenging, an important benefit of these models is their computational efficiency, achieved by reduced compute and maintenance costs, which can allow the system to trade-off slight accuracy for efficiency.…”

“…Our framework encompasses a variety of modelling architectures which can either train on all tasks, all languages, or a combination of both. We extend the our prior work in ; Mishra et al (2020b); Mishra (2020bMishra ( ,a, 2019 to develop efficient models for hate speech identification and benchmark them against the HASOC 2019 corpus, which consists of social media posts in three languages, namely, English, Hindi, and German. We open source our implementation to allow its usage by the wider research community.…”