Eduard Hovy · page 2
Papers
- Recursive Deep Models for Discourse Parsing (2014 · 206 citations)
The paper addresses document-level discourse parsing in the RST framework. In this task, a system must decide how clause-sized units join into a hierarchical tree and which rhetorical relation links each pair. The authors argue that hand-crafted lexical and syntactic features miss the semantic and intentional signals… - Probing the Probing Paradigm: Does Probing Accuracy Entail Task Relevance? (2021 · 144 citations)
Probing classifiers are often used to argue that a model relies on a linguistic property. A high probe score, however, only shows that the property can be decoded, not that the model needs it. Ravichander, Belinkov and Hovy test this gap on Natural Language Inference, using the SentEval setup of Conneau et al.… - Major Life Event Extraction from Twitter based on Congratulations/Condolences Speech Acts (2014 · 129 citations)
The paper tackles a problem earlier Twitter event work had largely ignored: finding the private milestones of ordinary users, such as weddings, graduations, new jobs or births, and describing them in structured form. - EQUATE: A Benchmark Evaluation Framework for Quantitative Reasoning in Natural Language Inference (2019 · 123 citations)
EQUATE is a benchmark of five test sets for quantitative reasoning in entailment: RTE-Quant (166 pairs from RTE2–4), NewsNLI (968 crowdsourced pairs from news summaries), RedditNLI (250 expert-labelled headline pairs), a synthetic stress test (7,500 pairs) and AwpNLI (722 algebra word problems turned into pairs). - Supervised Within-Document Event Coreference using Information Propagation (2014 · 90 citations)
The paper addresses within-document event coreference, which means deciding when two event mentions in one text describe the same real-world event. - Maximum Entropy Models for FrameNet Classification (2003 · 87 citations)
The paper tackles automatic semantic role labeling on FrameNet, where each annotated sentence has a target predicate whose frame elements must be assigned one of 120 semantic roles. - A Typology of Near-Identity Relations for Coreference (NIDENT) (2010 · 67 citations)
Coreference corpora such as ACE and OntoNotes sometimes annotate the same kind of example differently. The authors attribute this to a strict binary choice between identity and non-identity, which fails when two mentions denote almost the same discourse entity. Examples include a town before and after demographic… - Detecting Subevent Structure for Event Coreference Resolution (2014 · 65 citations)
Most work on event coreference only asks whether two event mentions denote the very same event. Araki et al. argue that this misses partial identity, especially subevent relations, where several events (e.g., capturing, killing, destroying) are parts of a larger parent event (an attack). They propose a two-stage… - On the Systematicity of Probing Contextualized Word Representations: The Case of Hypernymy in BERT (2020 · 55 citations)
Earlier probing work, notably Ettinger (2020), reported that BERT retrieves correct noun hypernyms in cloze prompts such as "A robin is a [MASK]". - An Extension of BLANC to System Mentions (2014 · 40 citations)
BLANC is a link-based metric for evaluating coreference resolution. It scores coreference links and non-coreference links separately and averages the two F-measures. In its original form (BLANC-gold, Recasens and Hovy, 2011), it assumes that system mentions match gold mentions exactly. Real systems must detect… - Coreference Resolution across Corpora: Languages, Coding Schemes, and Preprocessing Information (2010 · 40 citations)
The paper shows that a coreference score says as much about the corpus as about the system. The same entity-based system, CISTELL, is trained and tested under three corpus choices: language (English against Spanish, using OntoNotes and AnCora, both news), coding scheme (ACE against OntoNotes on the same 100,000-word… - Ontology-Aware Token Embeddings for Prepositional Phrase Attachment (2017 · 35 citations)
A type-level embedding gives one vector per word form, so "pool" as a place and "pool" as a game share a vector. - Approximating an Interlingua in a Principled Way (1992 · 31 citations)
The paper gives a method for building the ontology of an interlingua one language at a time. For each language, a taxonomy of the linguistic generalisations needed by its analyser and generator is built first. Then the domain entities are placed in that taxonomy, and the result is merged into the existing interlingua… - Unsupervised Ranking Model for Entity Coreference Resolution (2016 · 26 citations)
Entity coreference resolution groups the noun-phrase mentions in a text that point to the same real-world entity. - A Structured Distributional Semantic Model for Event Co-reference (2013 · 24 citations)
Standard distributional semantic models treat context as a bag of words, so they cannot tell "Booth killed Lincoln" apart from "Lincoln killed Booth." They also compose meaning poorly. - Automatic Interpretation of the English Possessive (2013 · 24 citations)
Tratz and Hovy address a gap in computational semantics: the English ’s construction (with possessive pronouns counted as the same phenomenon) is frequent and expresses many relations, such as ownership, kinship, part-whole, time and measure, yet had seen little NLP work and no public data. - Event Detection Using Frame-Semantic Parser (2017 · 17 citations)
The paper addresses event nugget detection and type/subtype classification under the ACE 2005 scheme. The authors argue that deep learning detectors built on shallow lexical cues and embeddings tend to be precise but miss many events. Their alternative trades precision for coverage. The frame-semantic parser Semafor… - EVOTOOL: Self-Evolving Tool-Use Policy Optimization in LLM Agents via Blame-Aware Mutation and Diversity-Aware Selection (2026 · 6 citations)
LLM agents that call external tools often fail in long, multi-step tasks. Feedback usually arrives only at the end, so it is hard to tell which step caused a failure. Prior optimizers either rewrite the whole agent prompt, which can break unrelated behaviors, or tune a single component and miss errors that spread… - Handling Biographical Questions with Implicature (2005 · 4 citations)
Feng and Hovy address a weakness of standard pipeline QA systems: they take a question's hidden assumptions for granted. - PANGLOSS: Knowledge-Based Machine Translation (1994 · 3 citations)
PANGLOSS is a project report on a knowledge-based interlingual system that combines symbolic and statistical methods for newspaper text in several languages. - CommonMorph: Participatory Morphological Documentation Platform (2026)
Building inflectional morphology datasets usually requires trained linguists, a lot of manual work and substantial funding, so most low-resource and endangered languages lack them. - MT at the paragraph level: improving English synthesis in SYSTRAN (1997)
Most machine translation systems handle one sentence at a time. Many output errors, however, come from phenomena that cross sentence boundaries. Hovy (USC/ISI) and Gerber (SYSTRAN) describe a research-to-industry project that adds a light discourse layer to the English synthesis stage of SYSTRAN's Japanese-to-English…
This page in Markdown → https://md.research.elamur.ai/free/en/authors/eduard-hovy/page/2/