Vidhisha Balachandran

Research Abstract:

Large language models have brought about a shift towards constructing large, general-purpose computational models of language, moving away from task-specific architectures. These models, trained on large unstructured data are opaque and challenging to control by design. Consequently, such data-driven models tend to overfit to spurious artifacts, perform poorly on underrepresented data, and fail in unpredictable ways. Thus, a paradigm shift towards developing trustworthy systems to ensure fairness, accountability, and robustness in their outcomes is essential. My research presents methods and solutions to improve the trustworthiness, transparency, and reliability of large-scale, data-driven language generation models, across various stages of the model pipeline. One axis of my research focuses on developing inherently interpretable models for various language processing tasks. I have designed feature attribution models for classification and keyphrase extraction and graph-based attention interpretations for dialogue generation and text summarization. In another related direction, I have also focused on efficient methods to evaluate and mitigate factual inconsistencies in language models. I have developed knowledge-integration measures for generalizable factual error detection and post-editing to correct a broad range of factual errors in abstractive text summarization. Overall my research highlights challenges in developing trustworthy language generation models and proposes solutions to improve their interpretability and factual reliability by design.

Bio:

Vidhisha Balachandran is a PhD candidate at the Language Technologies Institute at Carnegie Mellon University advised by Prof. Yulia Tsvetkov and a visiting student at the University of Washington. Her research focuses on interpretability and factuality in language models with the goal of developing reliable language systems. She has previously interned at Google Brain, Google AI, and at Allen Institute for AI (Ai2). She was selected as a Cadence Scholar in 2022 for her research and service.