# When Do Language-Grounded Explanations Help? A Graph-Bottleneck for Farm Monitoring Interpretable Sheep Facial Pain

arXiv: https://arxiv.org/abs/2609.20427 · HTML: https://mlsanity.com/paper/2609.20427/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 11 of its 26 references that are in the map.
- First posted on arXiv 17 Sep 2026.

## Builds on (most cited first, 11 of 11)

- [Learning Transferable Visual Models From Natural Language Supervision](https://mlsanity.com/paper/2103.00020/) (2021)
- [Semi-Supervised Classification with Graph Convolutional Networks](https://mlsanity.com/paper/1609.02907/) (2016)
- [Graph Attention Networks](https://mlsanity.com/paper/1710.10903/) (2017)
- [Concept Bottleneck Models](https://mlsanity.com/paper/2007.04612/) (2020)
- [Attention is not Explanation](https://mlsanity.com/paper/1902.10186/) (2019)
- [Learning Human-Object Interactions by Graph Parsing Neural Networks](https://mlsanity.com/paper/1808.07962/) (2018)
- [Do Concept Bottleneck Models Learn as Intended?](https://mlsanity.com/paper/2105.04289/) (2021)
- [Stability and Generalization of Graph Convolutional Neural Networks](https://mlsanity.com/paper/1905.01004/) (2019)
- [Human and Sheep Facial Landmarks Localisation by Triplet Interpolated Features](https://mlsanity.com/paper/1509.04954/) (2015)
- [Deeper Insights into Deep Graph Convolutional Networks: Stability and Generalization](https://mlsanity.com/paper/2410.08473/) (2024)
- [Sheep Facial Pain Assessment Under Weighted Graph Neural Networks](https://mlsanity.com/paper/2506.01468/) (2025)

## Abstract (arXiv)

Automated pain recognition from facial expression could make continuous welfare assessment practical in sheep, but adoption depends on trust: a stockperson cannot act on a score that arrives without justification. We ground a model in the Sheep Pain Facial Expression Scale (SPFES) by letting each detected facial region attend over text embeddings of the clinical descriptors and then test whether the resulting explanations mean anything. They do not. Ablating an entire descriptor changes the predicted logit by about 10-4, and the most-attended cue agrees with the predicted pain level in only 32.6% of regions, although the attention maps, the learned gate, and the generated text all proposed otherwise. We therefore remove the appearance bypass with a concept bottleneck whose classifier reads only SPFES concept scores, supervised by per-region state annotations that image-level pipelines discard. This costs 0.05--0.10 in Cohen's kappa but yields concepts that are demonstrably learned: minority pain-indicating states are recovered at 3.5--8.3× their base rates, and the ear and eye severity orderings emerge without severity supervision. Removing the supervision alone leaves kappa unchanged while concept accuracy falls to 0.109, showing that architectural necessity does not imply semantic validity. We also show that pooled concept accuracy is misleading under clinical imbalance and provide a cross-validated, protocol-matched benchmark of seven methods on this dataset.