Self-Rationalization in the Wild: A Large Scale Out-of-Distribution Evaluation on NLI-related tasks

Jing Yang1, Max Glockner2, Anderson Rocha1, Iryna Gurevych2
1 Artificial Intelligence Lab (Recod.ai), Institute of Computing, University of Campinas, Brazil
2 UKP Lab, Department of Computer Science, Technical University of Darmstadt, Germany

Transactions of the Association for Computational Linguistics (TACL), 2025
Paper Code arXiv

Abstract

Free-text explanations are expressive and easy to understand, but many datasets lack annotated explanation data, making it challenging to train models for explainable predictions. To address this, we investigate how to use existing explanation datasets for self-rationalization and evaluate models’ out-of-distribution (OOD) performance. We fine-tune T5-Large and OLMo-7B models and assess the impact of fine-tuning data quality, the number of fine-tuning samples, and few-shot selection methods. The models are evaluated on 19 diverse OOD datasets across three tasks: natural language inference (NLI), fact-checking, and hallucination detection in abstractive summarization. For the generated explanation evaluation, we conduct a human study on 13 selected models and study its correlation with the Acceptability score (T5-11B) and three other LLM-based reference-free metrics. Human evaluation shows that the Acceptability score correlates most strongly with human judgments, demonstrating its effectiveness in evaluating free-text explanations. Our findings reveal: 1) few annotated examples effectively adapt models for OOD explanation generation; 2) compared to sample selection strategies, fine-tuning data source has a larger impact on OOD performance; and 3) models with higher label prediction accuracy tend to produce better explanations, as reflected by higher Acceptability scores.

Out-of-distribution evaluation on 19 datasets

The table presents the 19 datasets we included in our evaluation, from three tasks: natural language inference, fact-checking and hallucination detection of abstractive summarization.

How much data is needed for generalization?

Is it important to always have more data to fine-tune your model? Not really. In our study, we fine-tuned two open-source models on two data sources, and found that with only 384 samples, the model already reached similar performance as fine-tuning on 50,000 samples.
What about data selection? Unfortunately we found that different sample selection methods have much smaller effects compared to data size and source.

F1‑score of fine‑tuned models versus training‑set size
Out‑of‑distribution performance of fine‑tuned models on 19 datasets covering NLI, fact‑checking and hallucination detection. Colors indicate different data selection methods.

Acceptability score (explanation quality) is positively related to label prediction performance

Does explanation improve when models have better performance? Yes, our study shows models that perform better on label prediction also generate better explanations.

Relationship between explanation quality and label prediction performance of models under different fine-tuning factors, with the x-axis showing the Acceptability score, and the y-axis the macro F1 score (scores are averaged over all datasets).

Video Presentation

Poster

BibTeX

@article{yang2025self,
        title={Self-Rationalization in the Wild: A Large-scale Out-of-Distribution Evaluation on NLI-related tasks},
        author={Yang, Jing and Glockner, Max and Rocha, Anderson and Gurevych, Iryna},
        journal={Transactions of the Association for Computational Linguistics},
        volume={13},
        pages={314--342},
        year={2025},
        publisher={MIT Press}
      }