Papers by Dhanya Sridhar
Causal Inference in Natural Language Processing: Estimation, Prediction, Interpretation and Beyond (2022.tacl-1)
Copied to clipboard
Amir Feder, Katherine A. Keith, Emaad Manzoor, Reid Pryzant, Dhanya Sridhar, Zach Wood-Doughty, Jacob Eisenstein, Justin Grimmer, Roi Reichart, Margaret E. Roberts, Brandon M. Stewart, Victor Veitch, Diyi Yang
| Challenge: | causality has not had the same importance in natural language processing, says aaron e. smith . he says research on causality in NLP remains scattered across domains without unified definitions . |
| Approach: | They propose to consolidate research on causality in NLP across academic areas . they explore potential uses of causal inference to improve robustness, fairness, interpretability . |
| Outcome: | The proposed method is a unified overview of causal inference for the NLP community. |
Heterogeneous Supervised Topic Models (2022.tacl-1)
Copied to clipboard
| Challenge: | Researchers in the social sciences are interested in the relationship between text and an outcome of interest. |
| Approach: | They develop a probabilistic approach to text analysis and prediction using a joint model of text and outcomes to find heterogeneous patterns. |
| Outcome: | The proposed model outperforms other methods on eight datasets and consistently outperformed other models. |
Causal Effects of Linguistic Properties (2021.naacl-main)
Copied to clipboard
| Challenge: | Social scientists have long been interested in the causal effects of language, studying questions like: How should political candidates describe their personal history to appeal to voters? |
| Approach: | They propose an algorithm for estimating causal effects of linguistic properties that leverages distant supervision and a pre-trained language model to adjust for the text. |
| Outcome: | The proposed method outperforms other methods when estimating the effect of Amazon review sentiment on semi-simulated sales figures. |
From Isolation to Entanglement: When Do Interpretability Methods Identify and Disentangle Known Concepts? (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods to evaluate features disentangle concepts from activations of neural networks are limited by their quality . current methods for concept identification and steering are sparse autoencoders, but they are not reliable. |
| Approach: | They propose to evaluate how well featurization methods disentangle one concept from another . they use sentiment, domain, voice, and tense to steer these features . |
| Outcome: | The proposed evaluations show that featurization methods are insufficient to establish steering selectivity . the results suggest that steering a feature affects many concepts despite a near absence of interaction effects. |