Understanding and Detecting Hallucinations in Neural Machine Translation via Model Introspection

Weijia Xu, Sweta Agrawal, Eleftheria Briakou, Marianna J. Martindale, Marine Carpuat
1/18/2023
cs.CL

No official code found. Best implementation: weijia-xu/hallucinations-in-nmt (6 stars, last commit 2023-03-06).

Abstract

Neural sequence generation models are known to "hallucinate", by producing outputs that are unrelated to the source text. These hallucinations are potentially harmful, yet it remains unclear in what conditions they arise and how to mitigate their impact. In this work, we first identify internal model symptoms of hallucinations by analyzing the relative token contributions to the generation in contrastive hallucinated vs. non-hallucinated outputs generated via source perturbations. We then show that these symptoms are reliable indicators of natural hallucinations, by using them to design a lightweight hallucination detector which outperforms both model-free baselines and strong classifiers based on quality estimation or large pre-trained models on manually annotated English-Chinese and German-English translation test beds.

View on arXivView PDF

Code Implementations(1)

weijia-xu/hallucinations-in-nmtCommunity100%

Understanding and Detecting Hallucinations in Neural Machine Translation via Model Introspection

61PythonFeb 24, 20233 years ago

Similar Papers

On Early Detection of Hallucinations in Factual Question Answering

91% match2 repos152023

Scalable Token-Level Hallucination Detection in Large Language Models

90% match2026

Building Reliable Long-Form Generation via Hallucination Rejection Sampling

90% match1 repo12026

RAGognizer: Hallucination-Aware Fine-Tuning via Detection Head Integration

90% match1 repo52026

HaluEval: A Large-Scale Hallucination Evaluation Benchmark for Large Language Models

89% match1 repo5952023

DynHD: Hallucination Detection for Diffusion Large Language Models via Denoising Dynamics Deviation Learning

89% match2026

Cite this paper

@article{xu2023understanding,
  title  = {Understanding and Detecting Hallucinations in Neural Machine Translation via Model Introspection},
  author = {Weijia Xu and Sweta Agrawal and Eleftheria Briakou and Marianna J. Martindale and Marine Carpuat},
  year   = {2023},
  eprint = {2301.07779},
  archivePrefix = {arXiv},
  url    = {http://arxiv.org/abs/2301.07779v2}
}

Add a badge to your README

Shows whether official code is linked for this paper.

Code of Paper badge
[![Code of Paper](https://codeofpaper.com/badge/2301.07779.svg)](https://codeofpaper.com/paper/2301.07779)

Discussion