Ramit Sawhney - Invited Talks

Keynote; NLDB 2022; Valencia, Spain

Natural Language Processing for Industrial Financial Predictive Analysis and Stock Trading

Abstract: In the financial industry, risk modelling, trading strategy design, and profit generation heavily rely on accurately predicting stock movements. Stock movements are influenced by varied factors beyond the conventionally studied historical prices, such as social media and correlations among stocks. The rising ubiquity of online content and knowledge mandates an exploration of models that factor in such multimodal signals for accurate stock forecasting. In this talk, I introduce a set of modern AI and NLP-centric methods and techniques using alternate sources of data - social media text, financial disclosures, documents, and multimodal data such as audio from financial earnings calls for building financial models in the industry to trade stocks and cryptocurrency. I then delve into the architecture of these models - covering multimodal, sequential, and graph neural networks - and analyse them across a diverse spectrum of metrics through an industry lens - quantitative performance, profitability, qualitative analysis, computational complexity, gender bias, and improvements over conventional financial analysis methods.

Recording: https://www.youtube.com/watch?v=qCeRi9ITVlQ

Invited Talk; Alan Turing Institute, UK

Title: Modeling User Context for Mental Health Assessment on Social Media

Abstract: Social Media's rising ubiquity has enabled greater global accessibility that presents a complex ecosystem of user behaviour and interaction. Recent psychological studies show users find it easier to exhibit thoughts related to their mental health, such as suicide ideation on social media. Understanding the buildup and behaviour of users online is a crucial step in building safer spaces, online moderation, and suicide prevention. While there exist methods that analyze user-generated text in isolation, it is rare that a user's online behaviour is independent of the rich social ecosystem they interact with. A user's actions are often linked with their historical activity and their connections and interactions with other users on social media. Contextualizing a user's behaviour and posts on social media through their social network (graph) and history (sequence) presents a challenging but important problem. In this talk, we'll cover how we augment NLP methods with time-aware sequence models and hyperbolic graph neural networks to contextualize user behaviour, inspired by sociological theories of user activity. We'll navigate user context from multiple lenses - quantitative, qualitative & interpretability, ethical, privacy preservation and data minimization, and practical applicability in human-in-the-loop systems. We'll conclude by discussing the generalizability of such contextual modelling for a variety of user-centric NLP tasks - hate speech detection, suicide risk assessment, and even financial sentiment modelling!

Invited Talk; Data Science Seminar Program; University of Marburg, Germany

HYPMIX - Method for Interpolative Data Augmentation in Hyperbolic Space on Text, Speech, and Vision

Invited Talk; Universitat Hamburg, Germany

Modeling User Context on Social Media

Invited Talk; Zalando Berlin

Modeling User Context on Social Media