Most tools that clean up who-wrote-what-where in science assume the record is in English. A Chinese, Persian or Korean paper whose authors list their departments in their own language is hard to match to a global institution registry such as ROR, or to the same authors' English-language papers. That matching breaks author disambiguation (deciding which "Wang L." is which person), institutional credit and citation counts. The usual fix would be to train a cross-lingual matcher on parallel text, meaning pairs of sentences that are translations of each other. Parallel text for affiliation strings ("Sch. of Mech. Eng., Xi'an Jiaotong Univ.") does not exist in any curated form, and general-domain parallel text looks nothing like it.
The proposal is to stop treating this as a data-scarcity dead end. The scholarly record already contains a large, free parallel corpus for exactly this text. Many national journals print every author's affiliation twice, once in the native language and once in English, in the same author order. We propose to mine those dual-language metadata blocks into affiliation-level parallel pairs. We would align the individual affiliation segments with a parallel-data-free matching method, train a compact cross-lingual affiliation encoder on the result, and test whether it links native-language affiliations to canonical institutions better than the obvious alternatives. The key test is on institutions the mined pairs never covered.
Core Claim
Dual-language paper metadata is an in-domain parallel corpus that comes with each paper. An encoder trained on it should beat two alternatives at linking native-script affiliation strings to registry entries: general translation followed by an English linker, and encoders trained on general-domain or filtered parallel text. The gain should hold for institutions absent from the mined pairs. If the gain exists only for seen institutions, the method is a glorified lookup table, and the proposal is designed to detect that.
The claim draws on both directions unevenly. Attribution supplies the task and the evaluation. Parallel-data scarcity supplies the methods: segment alignment without parallel data, and in-domain data selection, which here serves mainly as a competitor.
Prior Work and Where This Sits
Three works motivate the gap. The end-to-end affiliation framework in An end-to-end entity recognition and disambiguation framework for identifying Author Affiliation from literature publications works as follows. It detects organization and location mentions with GliNER, a transformer entity tagger. It then ranks DBpedia and GeoNames candidates using GloVe word vectors plus geographic coordinates. On English PubMed records it reaches 0.758 F1, against 0.639 for rules. It covers only English, and it needs hard-coded tables for aliases like "UK". We keep its candidate-ranking and geographic-prior design but replace the English-only text side.
S2AND: A Benchmark and Evaluation System for Author Name Disambiguation shows two things. Affiliation and co-author context matter more than name strings, and models fail to transfer across fields. Its pipeline works within monolingual, mostly Latin-script records. We supply the missing cross-script affiliation signal as a new pairwise feature.
Attribution in Scientific Literature: New Benchmark and Methods shows that adding structured metadata cuts LLM citation hallucination from 32.3% to 0.39% for o1. That is evidence that clean metadata is worth getting right. It is also a warning that LLMs left to recall attribution facts unaided do badly.
Two works supply the method. SimAlign: High Quality Word Alignments without Parallel Training Data using Static and Contextualized Embeddings aligns words with multilingual BERT embeddings and bipartite matching, and beats an aligner trained on 100k sentence pairs. We reuse its matching step one level up. We align whole affiliation segments rather than words, because the two language versions of a metadata block often differ in segment order and superscript numbering. SimAlign was tested only on running sentences, so whether it works on terse comma-separated metadata is itself an open question we measure. CCMatrix: Mining Billions of High-Quality Parallel Sentences on the WEB mines parallel text from the web by distances between multilingual sentence embeddings. We borrow its margin-based filtering to discard bad mined pairs.
Two works constrain the design. Domain Adaptation via Pseudo In-Domain Data Selection found that selecting 1% of a general corpus by cross-entropy difference beat the full corpus (42.22 vs 40.51 BLEU). Neural Machine Translation Training in a Multi-Domain Scenario found the opposite for neural models: training on all data and then fine-tuning in-domain beat data selection. Staged transfer of the kind used in UCAM Biomedical Translation at WMT19: Transfer Learning Multi-domain Ensembles improved Bio translation from 36.1 to 41.3 BLEU. This conflict shapes our training schedule. We would not assume the mined data alone suffices. We test it against selected general data and against a staged general-then-mined schedule.
Research Program
The first step is building the corpus. We would collect metadata from journals that publish bilingual author blocks in Chinese, Persian and Korean. These languages were chosen for different scripts and transliteration habits; the scientific domain of Esposito: An English-Persian Scientific Parallel Corpus for Machine Translation suggests Persian scientific publishing is a workable source. For each paper we split both blocks into affiliation segments. We then align segments across languages by Hungarian matching on mBERT middle-layer similarities, using author order as a soft constraint, and drop low-margin pairs in the style of CCMatrix.
We would measure alignment quality on a hand-checked sample before anything else. If segment alignment is poor on metadata, the rest of the program would need a supervised aligner, and we would report that.
The second step is training the encoder. We fine-tune a multilingual encoder contrastively, so that each native-language segment lands near its English counterpart. Registry names and multilingual registry labels serve as additional anchors. Linking then works as in the PubMed framework: retrieve registry candidates by embedding similarity, and rerank with the geographic prior from the detected location mention.
The third step is the downstream use. The cross-lingual affiliation similarity becomes one more pairwise feature inside the S2AND pipeline, so that native-script and Latin-script records of the same person can be merged.
Baselines
The naive reference point is transliteration plus fuzzy string matching against registry names and aliases. This is the cross-lingual analogue of the rule-based baseline that reached 0.639 F1 in the PubMed affiliation study. It shows how much the problem is solved by spelling alone.
The strongest existing system for English is the GliNER-plus-geographic-embedding linker, at 0.758 overall F1 on English PubMed affiliations (An end-to-end entity recognition and disambiguation framework for identifying Author Affiliation from literature publications). We run it in a translate-then-link configuration. A general translation system and, separately, a strong prompted LLM translate each native affiliation to English, and the unchanged English linker takes over. This is the real competitor. No reported result for cross-lingual affiliation linking was found on Lacuna, so all its numbers will be our own.
Fairness requires that every system sees the same test strings, retrieves from the same registry snapshot, and uses the same geographic reranker. Any tunable thresholds receive the same validation-set budget. The LLM gets a few-shot prompt built from the same training-split pairs our encoder uses.
The ablations isolate the key ingredient, which is mined in-domain metadata bitext. We train the identical encoder in four ways: on general parallel data; on a bilingual Moore-Lewis-selected subset of that data, scored against a small affiliation reference set; on mined metadata pairs only; and on general data followed by mined pairs. A further ablation removes the margin filter. Beating the translate-then-link baseline means a gain of at least three points of linking accuracy at rank one on the unseen-institution split, with bootstrap confidence intervals that exclude zero. On the seen-institution split we expect parity at best and would not count it as evidence.
Why not just…?
Why not just prompt a large LLM to translate or canonicalize the affiliation? This is the most serious threat. LLMs have seen many bilingual institution names. Against it, Attribution in Scientific Literature: New Benchmark and Methods shows that unaided recall of scholarly facts hallucinates heavily, at 32.3% even for o1, and long-tail provincial institutes are exactly where recall should fail. The translate-then-link baseline with an LLM settles this directly, and we report results by institution frequency. If the LLM matches us on rare institutions, the honest conclusion is that mined bitext is unnecessary. In that case its remaining value is as a cheap distillation target for a small model, which we would then evaluate on cost.
Why not just build a lookup table from the mined pairs? For institutions that appear in the mined corpus, a table may be as good as anything. That is why the decisive split holds out institutions entirely. The table scores zero on that split by construction, so any learned generalization shows up there. We also report a hybrid that uses the table when it hits and the encoder otherwise, since that is what a practitioner would deploy.
Why not just use an off-the-shelf multilingual encoder, or general bitext filtered for relevance? The generic encoder is the "general parallel data" arm. The filtered variant is the Moore-Lewis arm. The conflicting results of Domain Adaptation via Pseudo In-Domain Data Selection and Neural Machine Translation Training in a Multi-Domain Scenario mean we cannot predict the outcome, so we test it rather than assume it. If selected general data matches mined pairs, mining is not worth its cost.
Proposed Experiments
We would build a test set of roughly 1,500 native-language affiliation strings per language, hand-linked to ROR entries. It is split by journal and by institution, so that test institutions never appear in the mined training pairs. The primary metric is accuracy at rank one and F1 on that unseen-institution split, broken down by institution frequency. As a sanity check, the English linker is rerun on the English PubMed affiliation dataset to confirm our reimplementation is close to the reported 0.758.
Downstream, we construct a cross-script author disambiguation set from the mined papers. Here the bilingual blocks give gold identity links between native and romanized names. We measure B³ F1 for S2AND with and without the new affiliation feature. B³ F1 scores clusters by how many of each record's true co-identity mentions share its predicted cluster. S2AND reports over 50% error reduction relative to the Semantic Scholar production system but gives no absolute figure in the file, so this comparison will also be our own.
The decisive check is the unseen-institution comparison against LLM translate-then-link. A clear win supports the claim that the scholarly record is its own parallel corpus. A tie redirects the work toward cost and distillation.
References
- An end-to-end entity recognition and disambiguation framework for identifying Author Affiliation from literature publications
- S2AND: A Benchmark and Evaluation System for Author Name Disambiguation
- Attribution in Scientific Literature: New Benchmark and Methods
- SimAlign: High Quality Word Alignments without Parallel Training Data using Static and Contextualized Embeddings
- CCMatrix: Mining Billions of High-Quality Parallel Sentences on the WEB.
- Domain Adaptation via Pseudo In-Domain Data Selection
- Neural Machine Translation Training in a Multi-Domain Scenario
- UCAM Biomedical Translation at WMT19: Transfer Learning Multi-domain Ensembles
- Esposito: An English-Persian Scientific Parallel Corpus for Machine Translation