Graham Neubig · page 2
Papers
- Neural Cross-Lingual Named Entity Recognition with Minimal Resources (2018 · 249 citations)
The paper tackles named entity recognition for languages with no labeled data, transferring knowledge from English annotations. - Politeness Transfer: A Tag and Generate Approach (2020 · 235 citations)
The authors frame politeness transfer as a new text style transfer task: rewriting non-polite action-directives, such as requests and commands, into polite versions without changing their meaning. - Prometheus 2: An Open Source Language Model Specialized in Evaluating Other Language Models (2024 · 234 citations)
Teams often use proprietary models such as GPT-4 to judge LM outputs, but these judges are opaque, hard to control, and costly. - Learning to Deceive with Attention-Based Explanations (2020 · 233 citations)
Attention weights are often presented as explanations of what drives a neural model's output, including in fairness audits such as checking whether occupation classifiers depend on gender. - Detecting Hallucinated Content in Conditional Neural Sequence Generation (2021 · 232 citations)
Neural generators for translation and summarization can produce fluent output that contains material the input does not support. - MTNT: A Testbed for Machine Translation of Noisy Text (2018 · 227 citations)
Modern MT systems, including neural ones, often break down on the informal, error-filled language of social media. - Beyond BLEU: Training Neural Machine Translation with Semantic Similarity (2019 · 223 citations)
Fine-tuning NMT systems toward an evaluation metric usually means optimizing sentence-level BLEU. BLEU is a weak training signal, though. Exact n-gram matching gives synonyms and paraphrases no credit, and many different hypotheses receive identical scores, which leaves the reward landscape flat. Wieting and… - FlowSeq: Non-Autoregressive Conditional Sequence Generation with Generative Flow (2019 · 218 citations)
Autoregressive seq2seq models produce output one token after another, which limits parallel decoding on GPUs. - Systematic Inequalities in Language Technology Performance across the World's Languages (2022 · 216 citations)
NLP has advanced quickly, but the gains are concentrated in a small share of the world's roughly 6,500 languages. - Contextual Parameter Generation for Universal Neural Machine Translation (2018 · 206 citations)
The paper addresses a trade-off in multilingual neural machine translation. Universal models share all parameters across languages and struggle when training data is scarce or the languages differ a lot. Per-language encoder-decoder setups share nothing between languages, so they end up overparameterized. The authors… - compare-mt: A Tool for Holistic Comparison of Language Generation Systems (2019 · 188 citations)
Corpus-level metrics such as BLEU show whether one generation system beats another, but not what actually changed in the output. - Guiding Neural Machine Translation with Retrieved Translation Pieces (2018 · 188 citations)
Neural machine translation often fails to recall and correctly render rare words and phrases, especially in narrow domains where similar sentences recur. - Learning Language Representations for Typology Prediction (2017 · 187 citations)
Typological databases such as WALS are sparsely filled, while NLP applications often need complete feature sets for the languages they handle. - VisualWebArena: Evaluating Multimodal Agents on Realistic Visual Web Tasks (2024 · 184 citations)
Most benchmarks for autonomous web agents rely on text alone. Many everyday web tasks, however, depend on what a page or a reference picture looks like. VisualWebArena addresses this gap. It is a benchmark that extends the WebArena framework with tasks that cannot be solved without visual understanding. - Generalized Data Augmentation for Low-Resource Translation (2019 · 179 citations)
Neural MT for low-resource languages (LRLs) is hampered by tiny parallel corpora. Ordinary back-translation also helps little here, because the reverse model is itself trained on scarce data. Xia et al. propose a unified augmentation framework that draws on every available resource: English monolingual text, parallel… - Stronger Baselines for Trustable Results in Neural Machine Translation (2017 · 165 citations)
NMT papers often report gains over weak "vanilla" systems, so it is unclear whether a new method would still help a production-grade model. - On Evaluation of Adversarial Perturbations for Sequence-to-Sequence Models (2019 · 162 citations)
Adversarial attacks on seq2seq models are usually judged only by how much they degrade the output. That ignores whether the altered input still means the same thing. If the source meaning changes, a different translation is expected, so this is no model failure. Using untargeted word-substitution attacks on machine… - Extreme Adaptation for Personalized Neural Machine Translation (2018 · 161 citations)
Standard NMT systems use one model for everyone, but people differ in vocabulary, topics, gender, and other traits that affect how a sentence should be translated. - FacTool: Factuality Detection in Generative AI — A Tool Augmented Framework for Multi-Task and Multi-Domain Scenarios (2023 · 160 citations)
Large language models such as ChatGPT write fluent long-form answers that may contain factual errors. Classic fact-checking setups rarely fit this case. They usually assume that the claim to check, or the evidence for it, is already given. LLM outputs instead come without either, and they are long, with no clear unit… - Bilingual Lexicon Induction with Semi-supervision in Non-Isometric Embedding Spaces (2019 · 159 citations)
Most bilingual lexicon induction (BLI) methods map one monolingual embedding space onto another with an orthogonal matrix. - The Devil Is in the Errors: Leveraging Large Language Models for Fine-grained Machine Translation Evaluation (2023 · 158 citations)
Automatic MT metrics, even strong learned ones, usually output a single number that does not show what went wrong in a translation. - ChatGPT MT: Competitive for High- (but Not Low-) Resource Languages (2023 · 153 citations)
LLM chatbots are now widely used by ordinary people for translation, yet for most of the world's languages there was no published evidence on how well recent models translate them. - MasakhaNER 2.0: Africa-centric Transfer Learning for Named Entity Recognition (2022 · 151 citations)
African languages have over a billion speakers but very few annotated NER resources, and it is unclear which source language to use when transferring models across languages. - Quality-Aware Decoding for Neural Machine Translation (2022 · 146 citations)
Standard NMT systems choose their output by approximating the most probable translation (MAP) with beam search. - Do LLMs Exhibit Human-like Response Biases? A Case Study in Survey Design (2024 · 140 citations)
The paper asks whether LLMs react to changes in survey question wording the way people do. Humans show systematic response biases in surveys, so a model's sensitivity to prompt phrasing could be useful if it matched those human patterns. To test this, the authors took expert-written questions from Pew Research's…
This page in Markdown → https://md.research.elamur.ai/free/en/authors/graham-neubig/page/2/