R Raghav

Raghav R, Adarsh Vemali, Rajdeep Mukherjee · cmu

Structured Sentiment Analysis (SSA) deals with extracting opinion tuples in a text, where each tuple (h, e, t, p) consists of h, the holder, who expresses a sentiment polarity p towards a target t through a sentiment expression e. While prior works explore graph-based or sequence labeling-based approaches for the task, we in this paper present a novel unified generative method to solve SSA, a SemEval2022 shared task. We leverage a BART-based encoder-decoder architecture and suitably modify it to generate, given a sentence, a sequence of opinion tuples. Each generated tuple consists of seven integers respectively representing the indices corresponding to the start and end positions of the holder, target, and expression spans, followed by the sentiment polarity class associated between the target and the sentiment expression. We perform rigorous experiments for both Monolingual and Cross-lingual subtasks, and achieve competitive Sentiment F1 scores on the leaderboard in both settings.

Peter Carragher, Abhinand Jha, R Raghav et al.

Large Language Models (LLMs) demonstrate remarkable capabilities in question answering (QA), but metrics for assessing their reliance on memorization versus retrieval remain underdeveloped. Moreover, while finetuned models are state-of-the-art on closed-domain tasks, general-purpose models like GPT-4o exhibit strong zero-shot performance. This raises questions about the trade-offs between memorization, generalization, and retrieval. In this work, we analyze the extent to which multimodal retrieval-augmented VLMs memorize training data compared to baseline VLMs. Using the WebQA benchmark, we contrast finetuned models with baseline VLMs on multihop retrieval and question answering, examining the impact of finetuning on data memorization. To quantify memorization in end-to-end retrieval and QA systems, we propose several proxy metrics by investigating instances where QA succeeds despite retrieval failing. In line with existing work, we find that finetuned models rely more heavily on memorization than retrieval-augmented VLMs, and achieve higher accuracy as a result (72% vs 52% on WebQA test set). Finally, we present the first empirical comparison of the parametric effect between text and visual modalities. Here, we find that image-based questions have parametric response rates that are consistently 15-25% higher than for text-based questions in the WebQA dataset. As such, our measures pose a challenge for future work, both to account for differences in model memorization across different modalities and more generally to reconcile memorization and generalization in joint Retrieval-QA tasks.

Peter Carragher, Nikitha Rao, Abhinand Jha et al.

Vision language models (VLM) demonstrate sophisticated multimodal reasoning yet are prone to hallucination when confronted with knowledge conflicts, impeding their deployment in information-sensitive contexts. While existing research addresses robustness in unimodal models, the multimodal domain lacks systematic investigation of cross-modal knowledge conflicts. This research introduces \segsub, a framework for applying targeted image perturbations to investigate VLM resilience against knowledge conflicts. Our analysis reveals distinct vulnerability patterns: while VLMs are robust to parametric conflicts (20% adherence rates), they exhibit significant weaknesses in identifying counterfactual conditions (<30% accuracy) and resolving source conflicts (<1% accuracy). Correlations between contextual richness and hallucination rate (r = -0.368, p = 0.003) reveal the kinds of images that are likely to cause hallucinations. Through targeted fine-tuning on our benchmark dataset, we demonstrate improvements in VLM knowledge conflict detection, establishing a foundation for developing hallucination-resilient multimodal systems in information-sensitive environments.