Online fashion retailers lose enormous sums to returns, and most of those returns happen because a garment did not fit. Customer reviews ("runs small, take one size up") are one of the best early signals that an item is mis-sized. Systems that read reviews for fit, however, exist only for English, and the low-resource languages spoken in fast-growing markets have no labeled fit data at all. This proposal starts from a simple observation. The platform already holds language-independent labels for every review, in the customer's own order history. Someone who buys a size M, returns it, and keeps the L a week later has told us the M ran small, in a way no translator can garble. We propose to train a bidirectional sequence model over each customer's timeline of orders, keeps, returns and reorders. A bidirectional model reads the timeline both forwards and backwards, so a later reorder can clarify an earlier purchase. This model infers a fit label for each purchase. We then use those inferred labels to train a text classifier for reviews written in any language, without human annotation and without a cross-lingual teacher.
The payoff is a fit-reading model for reviews in languages where cross-lingual transfer is known to collapse. Its scores can serve as an early prior that flags a new article as running small or large after far fewer orders than return counts alone would need. In short, return behavior becomes the annotator, and bidirectional context is what makes that annotator accurate.
Core Claim
We claim that per-purchase fit labels inferred from complete order-and-return timelines are a better source of supervision for low-resource-language fit classifiers than the two available alternatives. The first alternative is labels transferred from an English model through translation or a multilingual encoder. The second is the small human-labeled sets a platform can afford. We further claim that inferring those labels well requires reading the timeline in both directions, because the event that disambiguates a purchase, such as an exchange, a reorder in another size, or a later keep of the same cut from another brand, usually comes after it.
The closest prior work sets the boundaries of this claim. Incorporating Customer Reviews in Size and Fit Recommendation systems for Fashion E-Commerce (FFC) fuses customer features, product metadata and BERT-encoded English reviews to predict Small/Fit/Large. It beats the tabular SFNet by 1.37% to 4.31% Macro-F1. It assumes English text and human-provided fit labels, treats each review in isolation, and its own review-only variant performed poorly. This motivates the text signal and warns us that text should complement purchase context rather than replace it.
MultiFiT supplies the student recipe. It trains a monolingual model with a per-language subword vocabulary on pseudo-labels from a cross-lingual teacher (LASER), and the student beat its teacher on Amazon review sentiment. We keep the student and replace the teacher. Ours is a behavioral model that never reads text, so its errors are uncorrelated with the language-specific errors a translation or multilingual teacher makes.
SizeNet supplies the general pattern of a statistical teacher built from return skew that trains a perceptual student. It applied that pattern to packshot images at the article level (AUC 0.74). We apply it to text at the purchase level, with a learned sequence teacher rather than aggregate skew.
SizeFlags provides the deployment target: Beta priors that let a Bayesian flag fire with fewer orders.
Research Program
The first component is the timeline teacher. Each customer's history becomes a sequence of events. Each event records the article, brand, category, ordered size, outcome (kept, returned, exchanged) and the time gap to the neighboring events. Many returns carry a size-related reason code selected from a menu ("too small", "too large"). These codes are independent of the review language, but they are missing or set to "other" for a large fraction of returns and are absent by construction for kept items.
We train the teacher by masked event prediction. We hide the reason code of a coded purchase, and the model predicts it from the rest of the timeline. This is the event-sequence analogue of masked language modeling. The trained model is then applied to every purchase, including uncoded returns and kept items that may have been tight or loose but tolerated.
For the backbone we use BLUR. BLUR runs forward and backward linear recurrences in parallel with linear cost in sequence length, which matters when labeling hundreds of millions of purchases. The teacher runs offline over completed histories, so the streaming limitation that BLUR's authors note does not apply. We expect the choice within the bidirectional family to matter less than bidirectionality itself, and we test that directly.
The second component is the language student. For each target language we train a MultiFiT-style monolingual model, with a language-specific tokenizer pretrained on the platform's unlabeled reviews in that language. We fine-tune it on reviews paired with the teacher's soft fit distribution for the reviewed purchase. Following FFC's finding, the deployed classifier fuses the review vector with customer and product features. We also report the text-only student to measure what the text alone carries. The tokenizer choice responds to evidence from NGLUEni and the Catalan infrastructure study that shared multilingual vocabularies fragment low-resource words. Domain cleaning of crawled monolingual text follows the filtering in Large-scale Machine Translation for Indian Languages in E-commerce.
The third component is the cold-start prior. Student scores from the first few reviews of a new article are averaged into the Beta prior of a SizeFlags-style article flag, in the same slot where SizeFlags injected SizeNet's visual prior.
The intended setting is Indian-language fashion markets, where the MT paper documents scarce data in languages such as Assamese and Hindi and colloquial e-commerce phrasing. This requires a retail partner holding order-return logs and reviews. For public replication of the text pipeline we use the Modcloth fit dataset under FFC's protocol. We will also commission a human-annotated fit test set of roughly two thousand reviews per target language. This set is the only human labeling in the project, and it is used for evaluation only.
Baselines
The naive reference point is no text at all: SizeFlags' Binomial-Beta article flag with uninformative priors, run on the same order logs. SizeFlags reports that its full model, with visual priors and tuned thresholds, needed 17% fewer orders to flag an item and cut returns in flagged items by 40% relative to its own baseline. Our text prior beats this reference if, at matched false-flag rate, it reduces orders-to-flag on new articles in a replay over held-out months.
The strongest English-transfer competitor is translate-then-classify. We translate target-language reviews into English with the best domain-adapted system we can build following the Indian e-commerce MT paper, which reached 41.32 BLEU English-to-Hindi against a 32.3 baseline and outperformed public APIs. We then apply an FFC model trained on the partner's English reviews with human fit labels. FFC's reported gain is 1.37% to 4.31% Macro-F1 over SFNet. No reported result was found on Lacuna for fit classification on low-resource-language reviews, so the absolute target comes from our own annotated test set.
The strongest multilingual competitor is a multilingual or region-centric encoder, such as the AfroXLMR-style models with language- and task-adaptive pretraining that won AfriSenti or an Indic equivalent. It is fine-tuned on English human labels and evaluated zero-shot, and also fine-tuned on 100, 500 and 2,000 in-language human labels.
The MultiFiT reference keeps our student but uses a cross-lingual teacher's pseudo-labels. This isolates the teacher swap.
All text models receive the same unlabeled in-language corpus for adaptive pretraining, the same hyperparameter search budget (a fixed number of trials on a shared dev split drawn from the annotated set), and comparable GPU hours. The behavioral teacher's training compute is counted against our method.
Two ablations isolate the key ingredient. The first replaces the bidirectional teacher with an identical causal one that sees only past events. The second replaces it with hand rules: use the reason code if present; mark "too small" if a return is followed within 14 days by a larger size of the same article; otherwise use "fit" for keeps. A further check swaps BLUR for a bidirectional GRU to show whether the specific architecture matters.
Why not just…?
Why not just translate reviews into English and reuse an English fit model? This is the most serious competitor, and it is run at full strength. Two findings in the file make us expect it to fall short. First, the Indian e-commerce MT work found public APIs fail on colloquial shopping language, and it reports its own pipeline breaking down for languages under a million monolingual tokens. Second, in cross-lingual e-commerce retrieval, translate-then-match reached 79.8% R10@1 against 86.8% for direct monolingual encoding. Fit phrases are short, idiomatic and polarity-critical ("thoda tight", or a regional word for "loose at the shoulder"), and one mistranslation flips the label. If translate-then-classify lands within one Macro-F1 point of our student in Hindi, we will conclude that translation suffices for mid-resource languages. The claim then narrows to the lowest-resource languages, where we expect the gap to widen.
Why not just label a few hundred reviews per language and fine-tune a multilingual encoder? MultiFiT showed that 100 labels with a good monolingual model can go far, so this is a credible threat. Our label-efficiency curves answer it directly, by locating how many human labels a fine-tuned encoder needs to match a student trained on zero human labels and millions of behavior-derived ones. The AfriSenti zero-shot score of 58.15 F1, against 71.31 supervised, shows how much transfer alone leaves on the table. It does not tell us how quickly small in-language label sets close that gap for fit, and we will report it either way.
Why not just skip text and flag articles from return counts? SizeFlags already does this well, and the logs feed it directly. Text adds information only if a review is worth more than a return event. That can happen because reviews come from keepers too, and because they carry direction and body region where many returns carry no code. The orders-to-flag replay against uninformed and SizeNet-style priors measures exactly this. If text priors do not reduce orders-to-flag, the student is still useful for size advice in the product page but not for cold-start flagging.
Why not just use rules on bracketing and exchanges instead of a learned bidirectional teacher? Rules capture the cleanest cases, such as an exchange to the next size. We expect them to miss cross-brand evidence and tolerated-but-poor keeps. This is the ablation most likely to embarrass the sequence model. If rule labels train students within noise of the learned teacher's labels, we drop the BLUR teacher and present the simpler pipeline. The low-resource contribution survives either way.
Decisive Experiments
The first check validates the teacher on its own terms. On held-out purchases whose reason code we mask, we compare macro-F1 for the bidirectional, causal and rule teachers. On the annotated reviews we report agreement between each teacher's label for the purchase and the human reading of the text. Bidirectionality earns its place only if it beats the causal teacher specifically on purchases followed by a reorder or exchange, which is where the theory says the future matters. We stratify results by that condition.
The second check is the main result. Macro-F1 on the human-annotated test sets in each target language compares the outcome-supervised student against translate-then-FFC, multilingual zero-shot and few-shot fine-tuning, and the MultiFiT cross-lingual-teacher student, all under the shared budget. Results are reported separately for mid-resource and very low-resource languages. We expect the advantage over transfer baselines to grow as monolingual data shrinks. If it does not, transfer quality is not the bottleneck we think it is.
The third check is downstream. In a time-split replay of new-article launches, the student-informed Beta prior is compared with uninformed and visual priors on orders-to-flag and on simulated return reduction among flagged articles. We report results per category, since SizeFlags found shoes and textiles behave differently.
The final check addresses noise. We inject synthetic corruption into the teacher's labels to test MultiFiT's observation that a monolingual student filters teacher noise. If it holds here, the student should outperform the teacher on the annotated set for purchases the teacher found ambiguous.
References
- Incorporating Customer Reviews in Size and Fit Recommendation systems for Fashion E-Commerce
- MultiFiT: Efficient Multi-lingual Language Model Fine-tuning
- SizeNet: Weakly Supervised Learning of Visual Size and Fit in Fashion Images
- SizeFlags: Reducing Size and Fit Related Returns in Fashion E-Commerce
- Bidirectional linear recurrent models for sequence-level multisource fusion
- NGLUEni: Benchmarking and Adapting Pretrained Language Models for Nguni Languages
- Building a Data Infrastructure for a Mid-Resource Language: The Case of Catalan
- Large-scale Machine Translation for Indian Languages in E-commerce under Low Resource Constraints
- SemEval-2023 Task 12: Sentiment Analysis for African Languages (AfriSenti-SemEval)
- Cross-Lingual Low-Resource Set-to-Description Retrieval for Global E-Commerce