[2405.00722] LLMs for Generating and Evaluating Counterfactuals: A Comprehensive Study

View a PDF of the paper titled LLMs for Producing and Evaluating Counterfactuals: A Complete Research, by Van Bach Nguyen and three different authors

View PDF
HTML (experimental)

Summary:As NLP fashions change into extra complicated, understanding their selections turns into extra essential. Counterfactuals (CFs), the place minimal modifications to inputs flip a mannequin’s prediction, supply a solution to clarify these fashions. Whereas Giant Language Fashions (LLMs) have proven outstanding efficiency in NLP duties, their efficacy in producing high-quality CFs stays unsure. This work fills this hole by investigating how nicely LLMs generate CFs for 2 NLU duties. We conduct a complete comparability of a number of widespread LLMs, and consider their CFs, assessing each intrinsic metrics, and the affect of those CFs on knowledge augmentation. Furthermore, we analyze variations between human and LLM-generated CFs, offering insights for future analysis instructions. Our outcomes present that LLMs generate fluent CFs, however battle to maintain the induced modifications minimal. Producing CFs for Sentiment Evaluation (SA) is much less difficult than NLI the place LLMs present weaknesses in producing CFs that flip the unique label. This additionally displays on the info augmentation efficiency, the place we observe a big hole between augmenting with human and LLMs CFs. Moreover, we consider LLMs’ potential to evaluate CFs in a mislabelled knowledge setting, and present that they’ve a robust bias in direction of agreeing with the supplied labels. GPT4 is extra strong in opposition to this bias and its scores correlate nicely with automated metrics. Our findings reveal a number of limitations and level to potential future work instructions.