PaperlayerView exact revision on arXiv

Source-linked AI summary

Fact over Fiction: Detection of Pathological Hallucinations in Sinhala-to-English Neural Machine Translation

Navam Obeysekara, Nevidu Jayatilleke

arXiv:2610.11389v1cs.CL

TL;DR

Sinhala-to-English NMT is vulnerable to fluent translations unrelated to their sources, and this language pair lacks established hallucination detection evidence. The paper builds a synthetic-data detector and evaluates its source dependence and signals across NMT systems. Detection firings vary by nearly an order of magnitude across architectures, while source ablation reduces sentence AUROC to chance.

  • Problem

    Sinhala-to-English NMT can produce fluent translations unrelated to their sources, while labelled data, benchmarks, and detection baselines for the language pair are absent.

    Source: §1 · p. 1

  • Method

    The paper builds a synthetic corpus and reference-free mDeBERTa-v3 token detector, then evaluates a three-signal ensemble and benchmarks eight NMT systems.

    Source: §1 · pp. 1–2

  • Results

    Detector firings vary nearly an order of magnitude across architectures, and removing or replacing the source reduces sentence AUROC from 0.970 to chance.

    Source: §4.2 · p. 7, §4.4 · p. 8

  • Takeaways & Limitations

    Det Only detections show that fluent, confident hallucinations exist that log-probability alone misses, while LaBSE adds a small ranking gain on synthetic data.

    Source: §5 · p. 9

    Generalisation to naturally occurring translation errors remains unverified because training and evaluation use synthetic corruptions that may under-represent authentic Sinhala-origin failure modes.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Neural Machine Translation (NMT) models, while capable of producing highly fluent outputs, remain vulnerable to hallucinations, which are translations that are natural yet semantically unrelated to the source. This vulnerability is acute in low-resource settings like Sinhala-to-English, where weak cross-lingual alignment leads to hallucinations. This paper introduces a framework for reference-free hallucination detection in this language pair. We present a 45,000-sample synthetic dataset generated through a probabilistic chain of five linguistically motivated corruption strategies, with a semantic rescue mechanism that uses character-level similarity to distinguish hallucinations from morphological variants. We fine-tune mDeBERTa-v3 for token-level sequence labelling, reaching a token-level F1 of 0.841 +/- 0.001 over three seeds on a source-disjoint test set, and study a three-signal ensemble integrating neural risk scores, sequence log-probabilities, and cross-lingual semantic embeddings (LaBSE). A source-ablation control shows that the detector relies on the Sinhala source rather than on surface artefacts of the corruption process: shuffling or removing the source reduces sentence-level AUROC from 0.970 to chance. We benchmark eight NMT systems spanning five model families and find that detector firings vary by an order of magnitude across architectures.

1 Introduction

Hallucinations are fluent translations that can fabricate information unrelated to the source, a risk heightened for low-resource Sinhala-to-English NMT. This paper introduces a reference-free detector and synthetic-data framework for the under-explored language pair.

1 Introduction

  • Hallucinations can be fluent falsehoods that mimic valid output while fabricating information, and low-resource settings heighten this risk.The reverse Sinhala-to-English direction is under-explored relative to English-to-Sinhala translation.Source: §1 · p. 1
  • The Sinhala-to-English direction lacks labelled data, evaluation frameworks, and detection baselines, while colloquial Sinhala and diglossia pose additional challenges for translation.Models may insert English agents or auxiliary verbs absent from the source when inputs omit subjects or pronouns.Source: §1 · p. 1
  • The paper creates a 45,000-sample synthetic dataset and trains a reference-free token-level mDeBERTa-v3 detector, reaching F1 = 0.841±0.001 on a source-disjoint test set.The dataset uses five linguistically motivated corruption strategies and a semantic rescue mechanism.Source: §1 · p. 1
  • A source-ablation control tests whether the detector uses the Sinhala source, alongside a three-signal ensemble study and a benchmark of eight NMT systems across five model families.The ensemble combines neural detection, sequence log-probability, and LaBSE cross-lingual semantic similarity.Source: §1 · p. 2

2 Related Work

Prior work addresses hallucination detection through synthetic data, critique of synthetic errors, and confidence, quality-estimation, and cross-lingual embedding signals. The paper positions its approach against these strategies and the concern that synthetic detectors may exploit construction artefacts.

2.1 Synthetic Data Approaches

  • Synthetic hallucination datasets address scarce natural labels by corrupting clean text and training token-level detectors, including approaches using BART reconstruction and bilingual alignment.Zhou et al. (2021) used bilingual lexicon induction and alignment heuristics to support XLM-R supervision.Source: §2.1 · p. 2

2.2 Natural Hallucinations and Critique

  • Guerreiro et al. (2023) argue synthetic errors differ from real NMT hallucinations, while sequence log-probability can perform comparably to complex reference-based metrics.They also proposed DEHALLUCINATOR for rewriting flagged translations.Source: §2.2 · p. 2

2.3 Extended Detection Strategies

  • Detection strategies include fluency and semantic similarity, uncertainty-based quality estimation, COMET, and ALTI+, which measures generated-token impact.For Sinhala-English, Fomicheva et al. (2020) report weaker correlation with human judgments than for higher-resource pairs.Source: §2.3 · p. 2
  • LaBSE provides cross-lingual sentence embeddings, while sequence log-probability may reflect target-language prior rather than source faithfulness; synthetic detectors may also exploit construction artefacts.The paper adopts a source-ablation control to examine the latter concern.Source: §2.3 · p. 2

3 Methodology

The methodology builds a Sinhala-to-English hallucination detector using a synthetic benchmark, token-level labels, and source-hypothesis inputs. It defines hallucination categories, generates and filters corrupted references, and creates source-disjoint detector splits.

3 Methodology

  • The framework generates an English hypothesis and confidence score, then uses a fine-tuned cross-lingual encoder to assign binary hallucination labels to hypothesis tokens given the Sinhala source.The formal task maps source S and hypothesis H to a binary label sequence Y.Source: §3 · pp. 2–3

3.1 Hallucination Taxonomy

  • Hallucinations are grouped as detached or fluent: detached outputs diverge from source meaning, while fluent ones remain coherent but introduce subtle semantic errors.Detached outputs range from strongly detached fabrications to fully detached outputs with no semantic link to the source.Source: §3.1 · p. 3

3.2 Dataset Construction

  • The benchmark corrupts English references from a Sinhala-English parallel corpus, filters for grammaticality, and releases 45,000 labelled rows, with a balanced 15,000-row subset for detector experiments.The source corpus contains approximately 25,000 sentence pairs; 7,500 source sentences were sampled to construct the released rows.Source: §3.2 · p. 3, §3.2.1 · p. 3

3.2.2 Synthetic Hallucination Generation

  • Five corruption strategies include named-entity swaps, antonym substitutions, subject-object swaps, numeric distortions, and high-temperature NLLB generation filtered by BERTScore.The dependency swap requires a non-pronominal subject and object of the same verb; high-temperature outputs with BERTScore F1 below 0.92 were retained.Source: §3.2.2 · pp. 3–4

3.2.3 Probabilistic Chain Generation

  • A probabilistic three-phase chain plans a high-temperature or deterministic first corruption, generates and filters batches with deterministic fallback, then applies a distinct secondary strategy.The initial choice is high-temperature sampling with p = 0.8 or a deterministic strategy with p = 0.2; secondary mutation is applied with p = 0.8.Source: §3.2.3 · p. 4

3.2.4 Grammar Validity Filter

  • Grammar filtering rejected 109 of 32,686 hallucinated rows (0.33%), including meaning-altering samples with grammatical errors; rejected rows were not manually verified.High-temperature sampling accounted for 78 of the 109 rejections.Source: §3.2.4 · p. 4

3.2.5 Corpus Composition and Splits

  • The released corpus contains 32,577 hallucinated and 12,423 faithful rows, with 42.9% of hallucinated rows compound and counted under their first-applied strategy.Faithful rows include references and reverted rows retained as duplicates of their references.Source: §3.2.5 · p. 4, Table 2 · p. 4
  • Detector experiments use a balanced 15,000-row subset split 70/10/20 by source sentence, preventing any Sinhala source from appearing across partitions.The train, validation, and test sets contain 10,486, 1,497, and 3,017 rows, respectively.Source: §3.2.5 · p. 4

3.3 Token-Level Label Construction

Token-level labels are derived by aligning corrupted hypotheses with references, then selectively rescuing likely morphological variants. The rescue rule reduces some label noise but misses paraphrases and has a threshold trade-off.

3.3 Token-Level Label Construction

  • Alignment labels equal and delete operations as faithful and replace and insert operations as hallucinated, leaving omissions unlabelled.Word-level labels are propagated to subword tokens; source and special tokens receive label −100.Source: §3.3 · pp. 4–5
  • For high-temperature rows, candidate words are compared with words in the aligned reference span, extended three words on either side, using character-level similarity.A score of at least θr = 0.80 downgrades a candidate label to faithful; deterministic corruptions receive no rescue.Source: §3.3 · pp. 4–5
  • Surface-dissimilar paraphrases such as “stated” and “said” remain labelled hallucinations even when meaning is preserved, creating label noise and possible deployment false positives.Rescue does not detect lexical paraphrases that lack surface similarity.Source: §3.3 · p. 5
  • At θr = 0.80, 13.3% of candidate label-1 words were rescued; lower thresholds rescued more words, but sampled rescues increasingly included different words.This sensitivity analysis concerns labels only; the detector was not retrained at other thresholds, and short words were the main source of false rescues.Source: §3.3 · p. 5

3.4 Translation and Confidence Scoring

Sequence confidence is measured from the generating model’s log-probability for the selected translation. The scoring procedure also specifies how to score supplied hypotheses and normalize scores.

3.4 Translation and Confidence Scoring

  • Beam-selected translations are scored by summing the decoder’s transition log-probabilities, matching the score used to select the hypothesis.For synthetic evaluation with a supplied hypothesis, the same quantity is computed by teacher forcing under NLLB-200-distilled-1.3B.Source: §3.4 · p. 5
  • Log-probability scores are min-max normalized per model.For benchmark data, the normalization uses per-model values; for synthetic evaluation, it uses validation-set values.Source: §3.4 · p. 5

3.5 Hallucination Detection Architecture

The detector uses mDeBERTa-v3 for token-level classification of translation hypotheses conditioned on Sinhala sources. Its input, training procedure, and sentence-level risk scoring are specified across three subsections.

3.5 Hallucination Detection Architecture

  • mDeBERTa-v3’s Replaced Token Detection pretraining mirrors the task of identifying replaced tokens, while disentangled attention separates content interactions from positional distance.This is relevant to the concatenated Sinhala-source and English-hypothesis input, where cross-segment positional proximity carries no semantic information.Source: §3.5 · pp. 5–6

3.5.1 Input Representation and Training

  • Hypotheses are tokenized word by word for label alignment, with the source truncated first when the pair exceeds 512 tokens.The model is trained with cross-entropy; the checkpoint with the highest validation token F1 is retained, using three seeds.Source: §3.5.1 · p. 6

3.5.2 Inference and Risk Scoring

  • The Hallucination Ratio aggregates flagged hypothesis tokens into a sentence-level score, and the sentence threshold is calibrated on validation data.All three seeds select τ = 0, so any flagged hypothesis token flags the sentence; maximum token probability is also reported as a threshold-free score.Source: §3.5.2 · p. 6

3.6 Three-Signal Ensemble Framework

The ensemble combines neural detector risk, inverted sequence log-probability, and LaBSE similarity as complementary translation-quality signals. Its weights and decision threshold are evaluated under validation-tuned configurations.

3.6 Three-Signal Ensemble Framework

  • The neural risk score, inverted log-probability, and LaBSE similarity capture different aspects of translation quality, with log-probability and neural risk correlations ranging from −0.103 to −0.437 across eight systems.Their ensemble score combines the neural risk, inverted log-probability, and one minus LaBSE similarity.Source: §3.6 · p. 6
  • The original ensemble weights are 0.5 for the detector, 0.3 for log-probability, and 0.2 for semantic similarity, alongside four alternatives and validation-tuned weights.Each configuration’s sentence-flagging threshold is tuned on validation data.Source: §3.6 · p. 6
  • LaBSE encodes source and hypothesis as mean-pooled, L2-normalized vectors in a shared multilingual space; cosine similarity below 0.80 is treated as semantic divergence.This criterion is independent of reference translations.Source: §3.6 · p. 6
  • The inverted log-probability score is higher when model confidence is lower.Normalization bounds are taken per model on the benchmark or from validation data on the synthetic evaluation.Source: §3.6 · p. 6

3.7 Evaluation Protocol

The evaluation protocol measures token-level detector performance and tests whether detection depends on the Sinhala source. It also evaluates the ensemble on synthetic splits and records detector signals across eight NMT systems.

3.7 Evaluation Protocol

  • Token-level precision, recall, and F1 are measured over hypothesis tokens on the source-disjoint test split and averaged across three seeds.Table 3 covers 3,017 test sentences.Source: §3.7 · p. 6, Table 3 · p. 7
  • The source-ablation control tests original, mismatched, and removed sources, a hypothesis-only detector, and faithful hypotheses paired with mismatched sources.In the detached-pair case, a source-aware detector should flag the pair, unlike one relying on perturbation artefacts.Source: §3.7 · pp. 6–7
  • On synthetic validation and test splits, the ensemble combines detector risk, teacher-forced log-probability, and LaBSE similarity, with weights and thresholds tuned on validation.Per-signal AUROC is also reported on test.Source: §3.7 · p. 7
  • Eight Sinhala-to-English systems spanning five model families are evaluated on 1,485 unique test sources, and signal firings and log-probability–risk correlations are recorded.All systems use beam search with four beams and a maximum of 256 new tokens.Source: §3.7 · p. 7

4 Evaluation

Evaluation tests the detector on synthetic data, checks source dependence, compares ensemble signals, and benchmarks eight NMT systems before examining false positives.

4.1 Detector Performance

  • 0.841±0.001 F1 lets mDeBERTa-v3 outperform XLM-RoBERTa by 5.2 points, with its advantage consistent across corruption strategies.Dependency Swap is hardest (0.804 vs. 0.752), while Semantic Drift (0.934 vs. 0.857), Numeric Distortion (0.933 vs. 0.855), and Entity Swap (0.916 vs. 0.860) are easiest.Source: §4.1 · p. 7

4.2 Source-Ablation Control

  • Sentence AUROC falls from 0.970 to chance (0.54 and 0.51) when the source is replaced or removed, showing reliance on the source-hypothesis relationship.Faithful hypotheses are flagged 4.5% of the time with their own source and 100% with a mismatched source.Source: §4.2 · p. 7, Table 4 · p. 7

4.3 Ensemble Evaluation

  • 0.972 AUROC for RH exceeds inverted log-probability (0.880) and LaBSE (0.882); the validation-tuned ensemble reaches 0.979.The tuned weighting assigns zero weight to log-probability and 0.1 to LaBSE, matching detector F1 while slightly improving ranking.Source: §4.3 · pp. 7–8, Table 5 · p. 8

4.4 NMT Benchmarking Results

  • 183–412 log-probability-only sentences per system contrast with 12–67 LaBSE-only detections, indicating distinct signal coverage.Detector-only cases number 22–97 per system and have normal confidence with high LaBSE similarity; LaBSE-only detections are consistent with a confirmatory role.Source: §4.4 · p. 8
  • 0.960–0.961 validation F1 holds for τ ≤0.03 but falls to 0.571 at 0.20; OPUS-MT and mT5 remain highest-flagging at tested thresholds.Detector firings decrease monotonically as τ rises for every system.Source: §4.4 · p. 8

4.5 Sources of False Positives

  • Precision on natural translations cannot be measured; among 129 M2M 1.2B sentences flagged by the detector, 36 (27.9%) also fell below the LaBSE threshold, while the remainder were fluent hallucinations or false positives.Source: §4.5 · p. 8
  • Surface-dissimilar synonyms and valid agents or auxiliaries inserted for colloquial Sinhala can be flagged as hallucinations under the label and alignment scheme.High-temperature training rows label some synonyms as hallucinations, while reference alignment penalizes syntactically forced insertions.Source: §4.5 · p. 8

4.4 NMT Benchmarking Results

  • 51–433 detector firings across systems reveal nearly an order-of-magnitude architectural spread, but these counts indicate relative risk, not verified hallucination rates.OPUS-MT and the mT5 fine-tune have the most detections; NLLB systems have the fewest. These patterns suggest family and training data may matter more than parameter count, pending controlled experiments.Source: §4.4 · p. 8, Table 6 · p. 9

5 Conclusion

The paper presents a synthetic corpus, reference-free detector, ensemble evaluation, and benchmark for Sinhala-to-English hallucination detection. It concludes that natural-output precision and paraphrase-sensitive supervision require further work.

5 Conclusion

  • Detector firings vary by an order of magnitude across architectures, and Det Only detections reveal fluent, confident hallucinations missed by log-probability alone.On synthetic data, the detector carries nearly all of the ensemble’s signal, while LaBSE adds a small ranking gain.Source: §5 · p. 9
  • The study provides a 45,000-sample synthetic corpus, a source-disjoint mDeBERTa-v3 detector with F1 = 0.841 ± 0.001, and an eight-system hallucination benchmark.It also reports a source-ablation control and evaluates a three-signal ensemble.Source: §5 · p. 9
  • Future work should annotate natural translation errors, replace difflib supervision with cross-lingual word alignment, and recalibrate ensemble weights and thresholds on annotated outputs.Thresholds tuned on balanced synthetic data do not transfer to the much lower hallucination rates of real output.Source: §5 · p. 9

Limitations

The detector and benchmark rely on synthetic corruptions, with natural-error generalization and precision unverified. Label construction, validation choices, and single-seed benchmark evaluation further constrain interpretation.

Limitations

  • The detector is trained and evaluated on synthetic data, and generalization to natural translation errors remains unverified.English-only reference edits and English re-encoding may under-represent authentic Sinhala-origin failure modes; a hypothesis-only baseline also reaches F1 0.622.Source: Limitations · p. 9
  • Difflib labels miss lexical paraphrases and omissions, while the rescue and BERTScore thresholds were not fully validated through alternative-threshold retraining.At θr = 0.80, some short, different words are wrongly rescued; the grammar filter also disproportionately rejects high-temperature samples.Source: Limitations · p. 9
  • The ensemble and NMT benchmark use a single detector seed, and benchmark firing counts lack human ground truth, so they do not establish natural-output precision.Thresholds calibrated on balanced synthetic data over-flag real outputs, so the benchmark reports independent signal firings rather than ensemble decisions.Source: Limitations · p. 9
Loading 2610.11389v1…