Gendered Labels and Logic Repair in Cross-Lingual Ontology Matching

Ontologies are formal vocabularies that say what kinds of things exist in a domain and how they relate, for example that every reviewer is a person and that no paper is a person. Linking two of them is called ontology matching. Increasingly it is done by multilingual language models that score how likely a label in one language ("Author") names the same concept as a label in another ("Autora"). Those models learned from text in which masculine role nouns dominate. We therefore expect them to match feminine or gender-inclusive labels ("Autora", "Autor*in") less reliably than masculine generic ones. Nobody has measured this. This note proposes to measure it and to test a specific remedy. The remedy is logic-based alignment repair: a reasoner checks whether the combined ontologies contradict themselves, and a solver removes the fewest, least-trusted correspondences that cause the contradictions. The repair works only on the ontologies' axioms, such as "papers and people are disjoint", and never looks at how a label is spelled. So in principle it corrects mistakes regardless of whether a label is masculine, feminine or inclusive.

In two sentences: we build a cross-lingual matcher by fine-tuning a multilingual encoder only on English matching data and applying it zero-shot to other languages. We then ask whether its errors concentrate on feminine and inclusive person-role labels, and whether logic repair that also recovers alternative correspondences closes that gap without the gender-blind step of rewriting every label into its masculine form.

Core Claim

We make three claims, and each can fail on its own. First, zero-shot cross-lingual matchers have a measurable gender label gap: on classes naming people in roles, recall and precision drop when labels are given in feminine or inclusive form. The drop is larger than for gender-irrelevant surface changes of the same edit size. Second, the gap grows when the scorer is fine-tuned on English and transferred zero-shot, compared with using the pretrained encoder directly. This mirrors the finding that zero-shot transfer can amplify demographic bias in sentiment classification (Cross-lingual Transfer Can Worsen Bias in Sentiment Analysis). Third, a repair step does more than delete conflicting correspondences. It re-selects among candidates under logical constraints, and this recovers much of the lost precision and part of the lost recall on gendered labels. Deletion-only repair recovers precision alone.

What is new is the pairing of two things. Formal diagnosis has never been applied to demographic failure modes; section 2 of the literature review we drew on finds no bridge between alignment repair and gender imbalance. And gender imbalance has never been treated as a stress test for symbolic alignment systems.

Where This Sits

The repair machinery is supplied by prior work. Improving Automatically Created Mappings Using Logical Reasoning treats the source ontologies as trusted and the automatically generated correspondences as suspects. It finds minimal conflict sets, meaning the smallest groups of correspondences that jointly make some class unsatisfiable (a class that, given the axioms, can have no members). It then removes a minimal hitting set, a smallest set of correspondences that touches every conflict. On OntoFarm, every correspondence it removed was confirmed wrong. Resolving Logical Contradictions in Description Logic Ontologies Based on Integer Linear Programming recasts the hitting-set search as a 0-1 integer linear program (an optimization over yes/no variables). Where tree search took over 100 seconds or timed out, the program found optimal repairs in about 0.4 seconds. The LogMap family (LogMap and LogMapLt results for OAEI 2013) instead repairs greedily inside local modules. It reached an incoherence degree of roughly 0.015% on the Large BioMed track, against 34.1% for LogMapLt, which disables repair. We reuse this machinery. Our change is to make it choose among alternatives rather than only delete.

The motivating gap comes from the MultiFarm track, where the same conference ontologies are labeled in different languages. There, lexical matchers like LogMap lose recall because their candidate generation relies on string similarity (Results of the Ontology Alignment Evaluation Initiative 2013). Multilingual encoders are the obvious replacement: English-only fine-tuning of such models retains most of their performance in other languages (Zero-shot Reading Comprehension by Cross-lingual Transfer Learning with Multi-lingual Language Representation Model). But the gender literature says these same representations carry masculine-default priors. Europarl is only 30% female speech (Gender Bias in Machine Translation). Translation systems default to masculine and fall from 80% to 54% accuracy on anti-stereotypical French roles (Evaluating Gender Bias in Machine Translation). mBERT carries gender bias through cross-lingual transfer, sometimes more than static embeddings do (Gender Bias in Multilingual Embeddings and Cross-Lingual Transfer).

Two lines of prior work constrain the design. Examining Gender Bias in Languages with Grammatical Gender shows that grammatical gender and social gender bias are distinct directions in embedding space. Our perturbations must therefore separate a feminine form that simply names a female role-holder from a label that is merely grammatically feminine. Twists, Humps, and Pebbles: Multilingual Speech Recognition Models Exhibit Gender Performance Gaps reports, for speech recognition, that models which encode gender more clearly show smaller female performance gaps (correlation −0.65). That result is in a different modality and we do not assume it transfers. It does make us wary of treating "erase the gender marking" as the default fix, and our experiments test that fix directly.

Research Program

We will build three things.

The first is a gender label perturbation suite for MultiFarm. For every MultiFarm translation into a grammatically gendered language, native speakers will write feminine and, where a convention exists, inclusive variants of each label that names a person in a role, such as author, reviewer, chair or participant. The reference alignments stay unchanged, because the concepts are identical. As a control, the same annotators will make meaning-preserving surface edits of similar character edit distance to labels of classes that do not name people, such as synonyms or inflectional changes. This control lets us tell "gender breaks matching" apart from "any label change breaks matching." Where the languages overlap with English, Spanish, German and French, we will also compute the intrinsic bias score from Gender Bias in Multilingual Embeddings and Cross-Lingual Transfer on the role terms. This asks whether an encoder's measured bias on a language predicts its matching gap there.

The second is a zero-shot cross-lingual matcher. We will fine-tune a multilingual encoder as a correspondence scorer using only English-English conference-ontology reference alignments. We will then apply it unchanged to cross-lingual MultiFarm pairs, taking the top-k candidates per class. A non-fine-tuned variant, which uses cosine similarity of the pretrained encoder's embeddings, isolates how much of any gap the English fine-tuning adds.

The third is repair with recovery. Let $x_m \in \{0,1\}$ indicate whether candidate correspondence $m$ is kept, let $s_m$ be its scorer confidence, and let $\mathcal{C}$ be the set of minimal conflict sets. The program is

$$\max_{x} \sum_m s_m x_m \quad \text{subject to} \quad \sum_{m \in C} x_m \le |C| - 1 \;\; \text{for all } C \in \mathcal{C}, \qquad \sum_{m \ni a} x_m \le 1 \;\; \text{for each class } a.$$

The first constraint forbids keeping every member of any conflict. The second makes the alignment one-to-one. Conflict sets over all top-k candidates cannot be enumerated in advance. We therefore solve, check the chosen alignment with a Description Logic reasoner, add any newly found conflict sets, and repeat. Unlike the removal-only formulation of Resolving Logical Contradictions in Description Logic Ontologies Based on Integer Linear Programming, rejecting a wrong top candidate frees the solver to pick the second-best candidate, which the reasoner then re-checks. This is where recall on perturbed labels could come back.

The analysis asks three things. Do conflicts involve person-role correspondences at a higher rate under feminine and inclusive labels? Is repair precision (the fraction of removed correspondences that are truly wrong) equal across label variants? Or does repair wrongly discard correct feminine-labeled matches, which would add a second, symbolic layer of disparity?

Baselines

The naive reference point is the zero-shot scorer with no repair: each class takes its top candidate above a threshold tuned on English data. It shows the raw gap and how incoherent unrepaired embedding matching is. The OAEI results give an order-of-magnitude expectation: on Large BioMed, removing repair left LogMapLt at 34.1% incoherence versus about 0.015% for LogMap (LogMap and LogMapLt results for OAEI 2013).

The strongest existing repair method is LogMap's greedy local repair (LogMap and LogMapLt results for OAEI 2013). It is the most widely validated coherence-preserving matcher in the file and fast enough for any MultiFarm pair. We will run it twice. The first run is off the shelf, with its own lexical candidates. The second feeds it exactly the same zero-shot encoder candidates and confidences our ILP receives, so the only difference is the repair algorithm. The file gives no reported MultiFarm F-measure for LogMap or any other system (no reported result found on Lacuna for MultiFarm), so we will reproduce LogMap's numbers ourselves on the original labels before perturbing anything.

We also include removal-only weighted ILP, in the style of Alg2 in Resolving Logical Contradictions in Description Logic Ontologies Based on Integer Linear Programming, applied to the top-1 alignment. It is the ablation that isolates our key ingredient, re-selection among alternatives. All repair variants get the same reasoner, the same candidate lists, and the same threshold chosen on English development data only.

Beating the baselines means this: on feminine and inclusive person-role labels, recall and F-measure significantly exceed both greedy repair and removal-only ILP under paired bootstrap over correspondences; the reasoner reports zero unsatisfiable classes; and the gap between original and perturbed labels shrinks by more than the gap on the non-gender control edits.

Why not just…?

Why not just normalize every label to its masculine or lemma form before matching? This is the real threat. For regular feminine suffixes, a morphological analyzer may close the gap outright, and we will include it as a baseline. It has costs. It needs a per-language analyzer, which works against the zero-shot setting. It probably fails on inclusive forms that analyzers do not handle. And it is gender-blind, rewriting a user's ontology into masculine defaults, which is analogous to the scrubbing that Gender Bias in Multilingual Embeddings and Cross-Lingual Transfer found insufficient for embeddings. If normalization closes the gap on both feminine and inclusive labels, we will say so plainly. The gender study then stands as a measurement result, and the repair contribution must justify itself on unperturbed MultiFarm.

Why not just translate all labels to English and run LogMap? Translation into English naturally neutralizes gender, so translate-then-match could erase the gap. It is a serious competitor and a planned baseline. Two things argue against assuming it wins. Isolated ontology labels give the translator no context. And translation systems are documented to mishandle gender, near or below 50% accuracy on non-stereotypical roles in many languages (Evaluating Gender Bias in Machine Translation). The comparison on inclusive labels will settle it. If translation wins everywhere, the finding is that the gap is specific to encoders, which is still useful to practitioners adopting encoder matchers.

Why not just raise the confidence threshold for person-role classes? Thresholding trades recall for precision. It cannot tell a wrong feminine-labeled match from a right one when both score low, which is exactly what the bias predicts. Our English-tuned-threshold baseline and a per-variant oracle threshold, tuned on test data as an upper bound for this approach, will show whether any threshold matches repair with recovery. If the oracle threshold matches it, the logical constraints are not earning their place.

Why not just debias the encoder, for example by re-anchoring alignment into a gendered language? Re-anchoring English into Spanish space cut intrinsic bias by 23% with little loss in lexicon translation for static embeddings (Gender Bias in Multilingual Embeddings and Cross-Lingual Transfer). We will run it on a static-embedding variant of the matcher. It works alongside repair rather than replacing it, so we will report it alone and combined with repair. The question is whether they fix the same errors.

Proposed Experiments

The first decisive check is whether the gap exists. We compare person-role recall and precision for original, feminine and inclusive labels against the non-gender control edits, per language, for the pretrained and English-fine-tuned scorers. If feminine and inclusive labels degrade no more than the control edits, the gender hypothesis is false. We would then report the null and continue only with the repair-with-recovery question on standard MultiFarm.

The second check is whether zero-shot fine-tuning amplifies the gap. We compare the gap of the English-fine-tuned scorer with that of the pretrained-similarity scorer, and correlate the per-language gap with intrinsic bias scores where they are available.

The third check is whether repair is gender-fair. We compute repair precision separately for each label variant, following the debugging-precision criterion of Improving Automatically Created Mappings Using Logical Reasoning. If greedy or ILP repair removes correct feminine-labeled correspondences more often, the symbolic stage is adding disparity, and that matters for anyone deploying repair downstream.

The fourth check tests whether recovery matters. Repair with recovery versus removal-only ILP on perturbed labels is the ablation that determines whether the proposal's main technical idea earns anything. We also report solver time per ontology pair, expected to stay near the sub-second regime reported for ILP repair. As a final sanity check, the full pipeline must not lose F-measure on the original MultiFarm labels relative to our reproduced LogMap run. If it does, any fairness gain comes at an accuracy cost we must report.