Iryna Gurevych · page 2
Papers
- The Argument Reasoning Comprehension Task: Identification and Reconstruction of Implicit Warrants (2018 · 217 citations)
Arguments in everyday discussion rarely state the warrant, the link explaining why a premise supports a claim. - Sentence and Expression Level Annotation of Opinions in User-Generated Discourse (2010 · 216 citations)
Most expression-level opinion annotation has targeted newswire (MPQA) and meeting dialogs. The few annotated review corpora, such as Hu & Liu (2004) and Zhuang et al. (2006), record only targets and polarities. They do not mark opinion spans, modifiers or anaphoric links, and they report no agreement figures. Toprak… - A broad-coverage collection of portable NLP components for building shareable analysis pipelines (2014 · 196 citations)
Building NLP pipelines from heterogeneous tools is difficult, and sharing a finished pipeline is harder still. - UNKs Everywhere: Adapting Multilingual Language Models to New Scripts (2021 · 176 citations)
Multilingual models such as mBERT and XLM-R transfer well between well-resourced languages. They perform poorly on low-resource languages missing from pretraining, and worst when the script itself was never seen, because the tokenizer then maps most of the text to UNK. The authors confirm that the original tokenizer… - SemEval-2017 Task 7: Detection and Interpretation of English Puns (2017 · 172 citations)
Puns break a core assumption of word sense disambiguation: that each word in context carries a single intended meaning. - What makes a convincing argument? Empirical analysis and detecting attributes of convincingness in Web argumentation (2016 · 165 citations)
The paper asks why one Web argument persuades better than another, not just which one does. Earlier quality work on essays, Reddit and argument pairs gave only holistic judgments. The authors start from UKPConvArg1, which has 11,650 argument pairs from 16 online debates. Each pair comes with crowd-written free-text… - Mind the Trade-off: Debiasing NLU Models without Degrading the In-distribution Performance (2020 · 159 citations)
Fine-tuned language models such as BERT often succeed on NLU benchmarks by exploiting dataset biases, such as word overlap or indicative words, and they then fail on out-of-distribution test sets. - GPL: Generative Pseudo Labeling for Unsupervised Domain Adaptation of Dense Retrieval (2022 · 155 citations)
Dense retrievers trained on large labeled collections such as MS MARCO lose much of their quality on specialized domains, where labeled data is rarely available. - Investigating Pretrained Language Models for Graph-to-Text Generation (2021 · 151 citations)
Graph-to-text generation turns structured inputs, such as AMR graphs or knowledge graphs, into fluent natural-language text. - A Richly Annotated Corpus for Different Tasks in Automated Fact-Checking (2019 · 147 citations)
Training machine learning systems for automated fact-checking requires large, reliably annotated data covering the whole verification pipeline. - Text Processing Like Humans Do: Visually Attacking and Shielding NLP Systems (2019 · 147 citations)
People often replace letters with look-alike symbols to write slang such as "1337", hide insults, or spoof domain names, and readers decode these easily. - A Retrospective Analysis of the Fake News Challenge Stance-Detection Task (2018 · 134 citations)
The FNC-1 shared task (2017) asked systems to decide whether a news article agrees with, disagrees with, discusses, or is unrelated to a given headline. - What to Pre-Train on? Efficient Intermediate Task Selection (2021 · 132 citations)
Training on an intermediate task before the target task can boost performance, but some source tasks help and others hurt. - MathDial: A Dialogue Tutoring Dataset with Rich Pedagogical Properties Grounded in Math Reasoning Problems (2023 · 131 citations)
Research on automatic dialogue tutors has been limited by a shortage of large, pedagogically sound datasets. - Before Name-Calling: Dynamics and Triggers of Ad Hominem Fallacies in Web Argumentation (2018 · 128 citations)
Ad hominem attacks are among the best-known fallacies, yet empirical evidence about them is scarce. Theoretical treatments usually rely on a handful of invented examples. This paper studies personal attacks in real dialogical argumentation at three levels of context: isolated comments, direct replies to an original… - Better Rewards Yield Better Summaries: Learning to Summarise Without References (2019 · 127 citations)
RL-based summarisers usually optimise ROUGE. ROUGE ranks whole systems reasonably well, but it is weak at telling a strong individual summary from a mediocre one, and it needs reference summaries. The authors replace it with a reward learned directly from human judgements. They use the ratings collected by Chaganty… - Wiktionary: A new rival for expert-built lexicons? Exploring the possibilities of collaborative lexicography (2012 · 120 citations)
The chapter examines whether an online dictionary compiled by volunteers can compete with lexicons produced by professional lexicographers. - Cross-Genre and Cross-Domain Detection of Semantic Uncertainty (2012 · 114 citations)
The paper addresses how to detect lexical cues of semantic uncertainty, meaning propositions whose truth cannot be determined given the speaker's mental state, across different genres and domains. - Aspect-Controlled Neural Argument Generation (2021 · 103 citations)
This work addresses argument generation that users can steer precisely. Earlier neural generators tend to produce long passages and give little say over which angle an argument takes. The authors fine-tune the CTRL conditional language model into Arg-CTRL. Given a control code made of topic, stance (PRO/CON) and… - Culturally Aware and Adapted NLP: A Taxonomy and a Survey of the State of the Art (2025 · 90 citations)
Interest in culture within NLP has grown quickly, yet the field still lacks a common idea of what "culture" means, so progress is hard to measure. - Automatic disambiguation of English puns (2015 · 90 citations)
Standard word sense disambiguation assumes every word carries one intended meaning. Puns break that assumption because the writer wants two meanings active at once. Miller and Gurevych treat pun interpretation as a new task: given a word already known to be a pun, recover both of its senses. They limit the scope to… - Out-of-domain FrameNet Semantic Role Labeling (2017 · 82 citations)
FrameNet role labeling is usually tested on one in-domain set (das-test). The authors test how it generalises. They build YAGS, a new test set of 1,415 sentences from Yahoo! Answers with 3,091 frame and 6,081 role annotations. Inter-rater agreement is Krippendorff's alpha 0.76 for frames and 0.62 for roles given… - Behind the Article: Recognizing Dialog Acts in Wikipedia Talk Pages (2012 · 81 citations)
Wikipedia Talk pages are where editors decide how to improve an article. Before this work, no one had applied machine learning to classify these contributions, and no corpus captured this kind of coordination. The authors built a dialog act schema focused on article improvement. They based it on a manual study of… - Analyzing Dataset Annotation Quality Management in the Wild (2024 · 76 citations)
Even popular NLP benchmarks contain many mislabeled items, yet no one had systematically checked how dataset authors actually handle annotation quality. - Annotation Error Detection: Analyzing the Past and Present for a More Coherent Future (2023 · 73 citations)
The authors survey methods for finding annotation errors, reimplement 18 of them, and test them on 9 English datasets for text classification, token labelling and span labelling.
This page in Markdown → https://md.research.elamur.ai/free/en/authors/iryna-gurevych/page/2/