Is a Concept Label Worth More Than a Class Label in a Low-Resource Language?

Interpretable-by-design text classifiers do not predict a label directly. They first score a small set of human-readable concepts, such as "insults a named person" or "invokes a religious blessing," and then predict the label from those scores alone. Every published version of these models has been built and tested on English. People working on African and other low-resource languages, meanwhile, get the most out of scarce data by adapting black box encoders to the target language. This proposal asks whether the concept layer can carry its own weight in these languages, as a way to spend scarce annotation better rather than as an interpretability add-on.

The hypothesis is simple. The step from concepts to label ("blessing + praise of a person → positive") is largely shared across languages. Only the step from text to concepts needs language-specific data. If so, a native speaker's time is better spent marking a few concepts in a few hundred tweets than labeling more tweets, especially for languages the model has never seen. The same concept layer also lets native speakers correct the model at test time. We test this on AfriSenti and AfriHate under strictly matched annotator time, against the strongest Africa-adapted black box.

Core Claim

A concept bottleneck model is a classifier forced to route its prediction through $k$ named concept scores. We train one on top of an Africa-adapted multilingual encoder, with a concept-to-label head shared across languages. We claim it beats an identically sized black box fine-tuned on the same encoder in two settings. The first is few-shot adaptation to a new language when both receive equal native-speaker annotation time. The second is zero-shot transfer to unseen languages. We further claim that its concept scores stay faithful, meaning that changing a concept actually moves the prediction, and that test-time correction by native speakers yields measurable accuracy gains.

What would be new is the matched-budget comparison of concept versus label supervision in low-resource languages, and the use of a shared concept-to-label head as the unit of cross-lingual transfer. Existing work on concept models does neither.

Relation to Prior Work

The motivation comes from two places. Concept Bottleneck Models introduced the concept-then-label architecture and reported that on knee X-rays a concept model trained on 25% of the data matched a black box trained on all of it. That data-efficiency claim has never been tested in language, let alone across languages. Explainability and Interpretability of Multilingual Large Language Models: A Survey finds that only 8% of interpretability publications concern low-resource languages, and that English-designed explanation tools are routinely applied to them unchanged. The same survey reports that multilingual models internally route non-English text through English-like intermediate representations. That finding is the reason to expect a shared concept space to exist at all.

The method draws on three text concept models, all English-only. Interpretable-by-Design Text Classification with Iteratively Generated Concept Bottleneck has GPT-4 propose about 30 concepts and score them at inference. Adaptive Concept Discovery for Interpretable Few-Shot Text Classification (StructCBM) refines concepts with an LLM loop and classifies by embedding similarity, reaching 85.45% on 10-shot AGNews. Concept Bottleneck Large Language Models scores concepts with sentence-embedding similarity and documents information leaking around the bottleneck.

We borrow their LLM-assisted concept proposal step. We reject LLM scoring at inference, because LLM quality degrades on African languages (AfriqueLLM). We instead train lightweight concept detectors on the adapted encoder, supervised by native-speaker concept annotations. Concept Matching for Low-Resource Classification (PARCUS) is the closest low-data relative: prototypes plus human-highlighted rationale tokens give 40.3 F1 against 23.4 for fine-tuned BERT on 60 SPOUSE examples. But it is English, token-level, and has no named global bottleneck.

Several papers constrain the design. Are Concept Bottleneck Models Effective as Decision-Support Systems? found that concept explanations help humans only when concepts are objectively grounded and users can manipulate them. Subjective concepts eroded trust. We therefore filter concepts by inter-annotator agreement, and we measure intervention rather than display concepts passively. The original CBM paper reports a joint concept model at 19.9% error versus 17.5% for a black box on CUB. That is a warning that in-distribution accuracy may cost something, and our design must report this honestly. Findings of the WASSA 2024 EXALT Shared Task frames cross-lingual explanation as trigger-word highlighting and covers only European languages. Our concept layer offers an alternative that does not depend on word boundaries, which matter in agglutinative languages, i.e. languages that build long words by stacking prefixes and suffixes.

The backbone comes from the low-resource side. AfroXLMR-Social adapts XLM-R to African social media and reports F1 gains of 1% to 30% over its base model, and it is our encoder for every system. Intent Identification and Entity Extraction for Healthcare Queries in Indic Languages shows that translating into English is viable but loses accuracy when vocabularies differ culturally. That result motivates our translate-then-explain baseline.

Research Program

The first step builds the concept inventory per task. For AfriSenti sentiment and AfriHate hate speech, an LLM proposes candidate concepts from English and high-resource-language training examples, following the proposal step of the iterative-bottleneck work. Native speakers of at least four AfriSenti languages then add culture-specific concepts the LLM misses, for example particular blessing formulas, proverbs, or ethnic slur categories. Each concept is piloted on 50 tweets per language by two annotators. Concepts below a fixed agreement threshold are dropped, which follows the decision-support study's finding that subjective concepts backfire. We expect roughly 15 to 30 concepts per task.

The second step collects annotations while timing every annotator. For each language, native speakers annotate a few hundred tweets with concept presence, and separately annotate additional tweets with labels only. Recording annotator-minutes per item lets every later comparison be made at equal human cost rather than equal item count.

The model has three parts. The encoder $E$ is AfroXLMR-Social. Concept detectors $g_j$ are small heads on $E$, one per concept $j$. A sparse linear head $f$ maps the concept vector to the label and is shared across all languages. Training uses the joint objective of the original CBM paper,

$$\mathcal{L} = \mathcal{L}_{task}(y, f(g(E(x)))) + \lambda \sum_{j=1}^{k} \mathcal{L}_{concept}(c_j, g_j(E(x))),$$

where $y$ is the class label, $c_j$ the annotated presence of concept $j$, and $\lambda$ trades label fit against concept fit. To limit leakage, concept scores pass through a sigmoid and are trained hard, and $f$ is refit on detector outputs (the "sequential" regime) for all faithfulness tests.

For a new language there are two regimes. In the few-shot regime we adapt only the detectors $g$ using that language's concept annotations, with $f$ frozen. In the zero-shot regime we adapt nothing.

Baselines

The naive reference point is the AfroXLMR-Social encoder fine-tuned as a standard classifier on whatever target-language labels are available, with zero in the zero-shot case. Every system's gain is measured against this floor.

The strongest existing competitor is AfroXLMR-Social with task-adaptive pretraining (TAPT), meaning continued masked-language pretraining on unlabeled task text before fine-tuning. Its authors report that TAPT improved baseline F1 by 0.55% to 15.11%. For absolute anchors on AfriSenti, SemEval-2023 Task 12 reports these scores:

Setting Best team Weighted F1
Monolingual track NLNDE 71.31
Multilingual track NLNDE 75.06
Zero-shot to Oromo and Tigrinya UCAS-IIE-NLP 58.15

We retrain this baseline with the same encoder, the same unlabeled TAPT data, the same compute, and the same hyperparameter search budget as our model. We then give it labels whose total annotator time equals what our model spent on concept annotations. Our model beats it if it achieves higher macro and weighted F1 on held-out target-language test sets at matched annotator time. In the zero-shot track, beating it means exceeding both the retrained TAPT black box and the reported 58.15 on Oromo and Tigrinya. If the black box wins at matched budget on in-distribution languages but loses on unseen ones, we report the claim as transfer-only.

The key ablation keeps the identical architecture, a $k$-dimensional sigmoid layer feeding a sparse shared head, but drops the concept loss ($\lambda = 0$). It isolates whether gains come from concept supervision or merely from a narrow shared bottleneck. A second ablation compares concept inventories: the LLM-proposed set alone versus the set with native-speaker additions. Finally, we include an LLM-scored concept bottleneck, in which GPT-class models rate concept presence directly in the target language as the iterative-bottleneck work does. This tests whether trained detectors are needed at all.

Why not just…?

Why not just fine-tune AfroXLMR-Social on more labels? This is the real threat. The original CBM results show a small in-distribution accuracy cost for concept models. AfroXLMR-Social is already strong, and labels are cheaper per item than concept annotations. We do not argue around this; the matched annotator-time curve settles it. If label-only training wins at every budget and in zero-shot transfer, the data-efficiency claim is false. The remaining contribution would then be faithfulness and intervention, priced at whatever F1 gap we measure.

Why not just translate into English and use an existing English concept model? The Indic healthcare study reports noticeable drops when local vocabularies diverge from English, and machine translation for Oromo or Tigrinya is far weaker than for Hindi. Translation also breaks any link between an explanation and the original words. We run this pipeline explicitly: translate the test set, then apply an English bottleneck trained on the same concepts. We compare both accuracy and how often native speakers judge the resulting concept scores correct for the original tweet.

Why not just ask an LLM to label and explain in the target language? The multilingual survey reports that free-text explanations often do not reflect the model's actual computation. A free-text explanation also offers no handle to correct. A wrong cited reason cannot be toggled to change the output, whereas a concept score can. The LLM-scored bottleneck baseline gives this alternative its best structured form. If it matches our trained detectors on Oromo and Tigrinya, then trained detectors are unnecessary and we will say so.

Why not just highlight trigger words, as in EXALT? Highlighting assumes the word is the unit of meaning, which fails when a single isiZulu or Oromo word bundles several morphemes. Highlighting also cannot express abstract concepts such as "call to violence." We measure this with a deletion test applied to both methods. For trigger words, we remove the highlighted tokens; for concepts, we zero the concept scores. In each case we record how much the prediction changes, which is a standard faithfulness proxy.

Proposed Experiments

The first decisive check is the matched-budget curve. On at least four AfriSenti languages and the AfriHate counterparts, we plot macro F1 against annotator-minutes for the concept model, the label-only TAPT black box, and the unsupervised-bottleneck ablation.

The second is unseen-language transfer. The concept head is trained on the AfriSenti training languages. We evaluate it on Oromo and Tigrinya with zero target annotations, and again with 50 to 200 concept-annotated tweets, comparing against the reported 58.15 and the retrained black box.

The third is faithfulness. For each concept, we flip its detector output and measure the change in prediction, comparing against the ablation. A model that ignores its concepts fails here regardless of its accuracy.

The fourth is intervention. Native speakers correct concept scores on test items the model gets wrong, with a budget of two concepts per item, mirroring the original CBM intervention protocol. We measure how much accuracy is recovered, compared with the same users viewing label-only output, as in the decision-support study design.

The idea is confirmed if the concept model wins or ties at matched budget and beats the black box on unseen languages, while also passing the flip test and showing intervention gains. It fails if the unsupervised-bottleneck ablation matches it, because that would mean the concepts are decorative.