Benjamin Van Durme · page 2
Papers
- LLM-Rubric: A Multidimensional, Calibrated Approach to Automated Evaluation of Natural Language Texts (2024 · 88 citations)
Using LLMs as automatic judges of text is cheap and scalable, but their ratings often diverge from human ones. - FrameNet+: Fast Paraphrastic Tripling of FrameNet (2015 · 75 citations)
FrameNet is a manually built lexical-semantic resource of nearly 1,000 frames. Its practical value is limited because many words in real text are not listed as lexical units (LUs) for any frame. In the authors' New York Times sample, fewer than half of the targets could be matched to a correct frame. To address this… - Few-Shot Semantic Parsing with Language Models Trained on Code (2022 · 72 citations)
This paper asks whether language models pre-trained partly on source code are better few-shot semantic parsers than models trained mainly on prose. - mmBERT: A Modern Multilingual Encoder with Annealed Language Learning (2025 · 59 citations)
Multilingual encoder-only models have seen little progress in recent years, and XLM-R remains the default choice despite being more than six years old. - Temporal Reasoning in Natural Language Inference (2020 · 59 citations)
The paper turns four temporal datasets (TempEval3, TimeBank-Dense, RED and UDS-Time) into NLI pairs about two kinds of temporal reasoning: event order and event duration. - Incremental Neural Coreference Resolution in Constant Memory (2020 · 55 citations)
The paper turns a strong all-pairs coreference model (SpanBERT, Joshi et al. 2020) into an incremental one that reads a document segment by segment and keeps a fixed-size list of entities. Each new mention is scored against entity representations (not individual earlier mentions); it joins the best entity if the… - “According to . . . “: Prompting Language Models Improves Quoting from Pre-Training Data (2024 · 46 citations)
Large language models often invent facts even though they were trained on reliable text. The authors test whether a simple cue in the prompt, such as "According to Wikipedia", can make a model reproduce more material it memorized during pre-training. To measure this, they introduce QUIP-Score (Quoted Information… - FollowIR: Evaluating and Teaching Information Retrieval Models to Follow Instructions (2025 · 40 citations)
Neural retrievers are built on language models that can obey elaborate prompts, yet almost no retrieval system lets users attach detailed instructions to a query. - Adapting Coreference Resolution Models through Active Learning (2022 · 35 citations)
Neural coreference resolution (CR) systems trained on OntoNotes 5.0 often perform poorly in new domains, where text style, entity meanings and annotation guidelines differ. - The Universal Decompositional Semantics Dataset and Decomp Toolkit (2020 · 18 citations)
Older semantic annotation schemes rely on large, mutually exclusive category inventories. They need expert annotators and fail on borderline cases, on items that fit several categories, and on phenomena the ontology left out. Decompositional semantics replaces categories with many simple questions that untrained… - Augmenting FrameNet Via PPDB (2014 · 17 citations)
FrameNet, the lexico-semantic resource based on frame semantics, has limited lexical coverage. Many words that should trigger a frame are missing, so parsers such as SEMAFOR need test-time heuristics for out-of-vocabulary items. Rastogi and Van Durme instead enlarge the training data. They take lexical rules from… - Joint Universal Syntactic and Semantic Parsing (2021 · 16 citations)
Earlier attempts to parse syntax and semantics jointly usually improved one level at the expense of the other. - Zero and Few-shot Semantic Parsing with Ambiguous Inputs (2023 · 15 citations)
Semantic parsing benchmarks usually assume each utterance has exactly one correct formal meaning, so it has been unclear whether parsers built on large language models represent ambiguity at all. - Is That Your Final Answer? Test-Time Scaling Improves Selective Question Answering (2025 · 11 citations)
Benchmarks of test-time scaling for reasoning LLMs usually assume the model must always answer, so wrong answers cost nothing. - Multilingual Coreference Resolution in Multiparty Dialogue (2023 · 11 citations)
Most coreference resolution work targets single-author text such as newswire. Conversations with more than two speakers are poorly covered, even though there "you" or "we" may refer to any subset of the people present. Existing resources such as Friends CI label only people and are English-only. - Enhancing Systematic Decompositional Natural Language Inference Using Informal Logic (2024 · 7 citations)
Entailment-tree reasoners answer a question by recursively splitting a hypothesis into premises that should jointly support it. - WikiVideo: Article Generation from Multiple Videos (2026 · 5 citations)
Video is a natural evidence source for retrieval-augmented generation, yet most article-generation work relies on text. - Predicting the Argumenthood of English Prepositional Phrases (2018 · 4 citations)
The paper asks whether a model can tell if a prepositional phrase attached to a verb is an argument, meaning it is selected by the verb, or an adjunct, meaning an optional modifier. - arXiv2Table: Toward Realistic Benchmarking and Evaluation for LLM-Based Literature-Review Table Generation (2026 · 1 citation)
The paper argues that current evaluations of LLM-generated literature-review tables are too idealized. They assume every supplied paper belongs in the table, and they use the gold table's caption as the target, which is either too vague or reveals the expected columns. The authors build arXiv2Table, a benchmark… - Synthetic Function Demonstrations Improve Generation in Low-Resource Programming Languages (2026 · 1 citation)
LLMs struggle with programming languages that have little natural training data, such as Excel formulas. Even GPT-4o scores only 58.2% on a central Excel test set. The authors propose a recipe for building finetuning data for such low-resource languages that does not require a teacher already skilled in them. They…
This page in Markdown → https://md.research.elamur.ai/free/en/authors/benjamin-van-durme/page/2/