--- license: cc-by-4.0 language: - es task_categories: - text-classification task_ids: - natural-language-inference pretty_name: ESNLIR — human majority labels vs discourse-connector labels (1,971 pairs) size_categories: - 1K **An Analysis of the Performance of Large Language Models in Spanish NLI Datasets with Causal Relationships** > Nicolás Pérez, Johan R. Portela, Ruben Manrique — Universidad de los Andes, Bogotá (IBERAMIA 2026, to appear) Code: [`Pacolas/NLI-via-LLM`](https://github.com/Pacolas/NLI-via-LLM) · Collection: [**ESNLIR-LLM**](https://huggingface.co/collections/Flaglab/esnlir-llm-6a74a6b9ce149918f48ebba5) ## Why this exists ESNLIR, like SciNLI and MSciNLI, builds NLI pairs by *distant supervision*: take two adjacent sentences joined by a discourse connector, delete the connector, and let it stand as the gold label. The method scales to 400k pairs without annotators. It also assumes the connector is recoverable from the sentences alone. This file measures how often that assumption holds. It is the evidence behind the paper's **Cohen's κ = 0.32** — and the part that matters most, the 999 pairs where the two label sources diverge, exists nowhere else. [`Flaglab/esnlir-human-validated`](https://huggingface.co/datasets/Flaglab/esnlir-human-validated) publishes only the 972 survivors; this dataset is a **strict superset of it** with identical schema. ## Agreement | | pairs | share | |---|---|---| | human majority **matches** connector label | 972 | 49.3% | | human majority **differs** | 999 | 50.7% | **Cohen's κ = 0.325** (chance agreement 0.249), conventionally *fair* — far from the near-identity that treating connectors as ground truth would require. ### Where the disagreement is Rows are the connector-derived label, columns the human majority. The diagonal is agreement. | connector ↓ / human → | contrasting | entailment | neutral | reasoning | retained | |---|---|---|---|---|---| | **contrasting** | **184** | 78 | 126 | 108 | 37.1% | | **entailment** | 20 | **219** | 102 | 159 | 43.8% | | **neutral** | 29 | 44 | **362** | 48 | 74.9% | | **reasoning** | 26 | 143 | 116 | **207** | 42.1% | Three things are visible here, and none of them are noise: **Contrasting connectors are the least trustworthy** (37.1% retained). *Sin embargo* and *no obstante* frequently join sentences whose contrast is rhetorical rather than semantic; strip the connector and readers see two unrelated statements (126 → neutral) or a causal chain (108 → reasoning). **The reasoning/entailment boundary cuts both ways.** 143 reasoning-connector pairs read as entailment to humans, and 159 entailment-connector pairs read as reasoning. This is the same confusion every model in the paper exhibits — it is a property of the distinction, not of the classifier. **Neutral is the only stable class** (74.9%). Absence of a connector is easier to confirm than the presence of a specific relation. ### The sample was balanced; the human labels are not | | contrasting | entailment | neutral | reasoning | |---|---|---|---|---| | connector label | 496 | 500 | 483 | 492 | | human majority | 259 | 484 | **706** | 522 | Pairs were drawn near-uniformly by connector label, yet annotators returned a neutral-heavy distribution. The guidelines instructed them to judge strictly what is written and to choose `neutral` when a pair was ambiguous, so neutral absorbs exactly the cases where the deleted connector carried information the sentences do not. ## Genres | genre | annotated | retained | rate | | genre | annotated | retained | rate | |---|---|---|---|---|---|---|---|---| | theses | 905 | 427 | 47.2% | | legal | 125 | 71 | 56.8% | | books | 412 | 217 | 52.7% | | clinical | 77 | 51 | 66.2% | | comments | 153 | 67 | 43.8% | | talks | 76 | 34 | 44.7% | | news | 149 | 71 | 47.7% | | articles | 74 | 34 | 45.9% | 24 domains, 32 source datasets. Clinical and legal text — the most formulaic registers — retain best; web comments retain worst. ## Fields Identical to [`Flaglab/esnlir-human-validated`](https://huggingface.co/datasets/Flaglab/esnlir-human-validated), so the two join directly. | field | meaning | |---|---| | `connector_type` | label derived from the discourse connector (distant supervision) | | `connection_type` | **human majority label** — differs from `connector_type` on 999 rows | | `connector` | the connector that was removed, e.g. `sin embargo,` | | `sentence_1`, `sentence_2` | premise and hypothesis, connector stripped | | `genre`, `domain`, `dataset` | provenance | | `extraction_strategy`, `distance`, `*_paragraph`, `*_position` | how the pair was extracted | | `id` | source-article id — **not unique per pair**; join on `(sentence_1, sentence_2)` | ## Annotation 2,136 pairs were sampled across domains and labelled by **27 Spanish-speaking university students**, who produced **13,626 annotations** — a mean of **6.4 independent labels per pair**. Adjudication was by majority vote. **165 pairs (7.7%) reached no majority and are not in this file**, which is why it holds 1,971 rather than 2,136. Their absence means the agreement rate here is measured only over pairs where annotators could converge, and is therefore an *optimistic* estimate of connector reliability. ## Usage ```python from datasets import load_dataset ds = load_dataset("Flaglab/esnlir-human-majority", split="train") disagree = ds.filter(lambda r: r["connector_type"] != r["connection_type"]) print(len(disagree), "pairs where the connector misleads") # 999 # the pairs a 'sin embargo' reader would get wrong for r in disagree.filter(lambda r: r["connector_type"] == "contrasting").select(range(3)): print(f"{r['connector']!r} connector={r['connector_type']} human={r['connection_type']}") print(" ", r["sentence_1"][:70], "/", r["sentence_2"][:70]) ``` Training on the human labels instead of the connector labels, or using the 999 divergent pairs as a hard evaluation slice, are both straightforward from here. ## A caution **Do not read model scores on the 972 agreeing pairs as evidence that annotation improves models.** That subset is *defined* by agreement, so every classifier scores higher on it — the fine-tuned XLM-RoBERTa baseline gains 4.8 points under the same filter without any change to the model. The gain measures label reliability, not model quality. This dataset is published precisely so the selection effect is inspectable rather than implicit. ## Related | | | |---|---| | [`Flaglab/esnlir-human-validated`](https://huggingface.co/datasets/Flaglab/esnlir-human-validated) | the 972 agreeing pairs — a subset of this file | | [`Flaglab/esnlir-llm-predictions`](https://huggingface.co/datasets/Flaglab/esnlir-llm-predictions) | per-pair model outputs on these pairs | | [`Flaglab/esnlir-test`](https://huggingface.co/datasets/Flaglab/esnlir-test) | the 80,213-pair test split these were sampled from | ## Citation ```bibtex @InProceedings{perez2026llmspanishnlicausal, author = {P{\'e}rez, Nicol{\'a}s and Portela, Johan R. and Manrique, Ruben}, title = {An Analysis of the Performance of Large Language Models in Spanish NLI Datasets with Causal Relationships}, booktitle = {Advances in Artificial Intelligence -- IBERAMIA 2026}, year = {2026}, publisher = {Springer Nature Switzerland}, address = {Cham}, note = {To appear}, } ``` The pairs come from ESNLIR, released here under CC BY 4.0 — attribution to the source corpus is a condition of that licence: ```bibtex @InProceedings{portela2025esnlirspanishmultigenredataset, author = {Portela, Johan R. and P{\'e}rez-Ter{\'a}n, Nicol{\'a}s and Manrique, Rub{\'e}n}, editor = {Florez, Hector and Peluffo-Ordo{\~{n}}ez, Diego}, title = {{ESNLIR}: Expanding Spanish {NLI} Benchmarks with Multi-genre and Causal Annotation}, booktitle = {Applied Informatics}, year = {2026}, publisher = {Springer Nature Switzerland}, address = {Cham}, pages = {345--361}, isbn = {978-3-032-07175-0}, doi = {10.1007/978-3-032-07175-0_23}, } ```