Entailment Diagnostics for Finding Nouns That Act as Verbs

Many nouns describe events rather than things. "The investigation of the mayor lasted months" reports that someone investigated the mayor. In "the building collapsed", by contrast, the noun names an object, even though it comes from a verb. Systems that extract "who did what to whom" handle verbs well and nouns poorly. The first obstacle is deciding which noun tokens are event-like predicates in the first place. When this step is automated, Chinese nominal semantic role labeling falls from 72.67 to 55.90 F1 (Li et al., 2009). The QANom resource handled the same decision by asking crowd workers an explicit "is this noun verbal here?" question (QANom).

We propose to turn the tests linguists use by hand to tell event nouns from object nouns into automatic entailment checks. A large language model, prompted with a few examples, writes candidate rewrites of the sentence. One kind spells the noun out as a verb clause using words from the sentence ("Someone investigated the mayor"). The other kind substitutes an explicit event phrase ("The act of investigating lasted months"). An off-the-shelf natural language inference model then judges whether the original sentence entails each rewrite. Natural language inference (NLI) is the task of deciding whether one sentence logically implies another. A noun counts as a predicate when its rewrites survive this check. Running the check over unlabeled text produces automatically labeled training data, and that data trains a fast tagger.

Core Claim

Our hypothesis is that nominal predicate identification is best done by having a generator propose and an entailment model verify. The rewrites are grounded in the sentence: they reuse its own argument spans, and they substitute an event description back into the sentence's own frame. Two simpler routes look weaker. Asking a language model directly whether the noun is an event lets its errors pass unchecked. Asking an NLI model generic questions such as "an event of building occurred" cannot separate "the building collapsed" from "the building of the dam took years", because both sentences arguably entail that some building happened.

The distinction matters most on ambiguous lemmas, which appear as predicates in some sentences and as plain nouns in others. It also matters on lemmas never seen in any training data. If the claim holds, a tagger distilled from verified labels on raw Wikipedia text should beat direct prompting on these subsets. It should also beat lexicon-based distant labels. And it should beat supervised taggers on held-out lemmas, while costing a fraction of the inference that the verifier itself requires.

Relation to Prior Work

The closest work runs in the opposite direction. NounAtlas prompts Gemini-Pro to rewrite verbal SemCor sentences into nominal ones and projects role labels with a bi-encoder. The result is 28,055 synthetic sentences and 72.4 F1 on nominal test sets. That pipeline generates nominal text without verifying it, and it never decides predicate status for nouns in natural text. Our pipeline starts from naturally occurring nouns, maps them back to verb clauses, and keeps only rewrites that an entailment model accepts.

Automatic Nominalization of Clauses through Textual Entailment supplies our method's central bet. It shows that an entailment model ranks candidate nominalizations of clauses better than GPT-2 likelihoods: 0.630 versus 0.476 nominalization accuracy. That work ranks surface forms produced from known clauses and never identifies predicates in running text. We reuse its premise-hypothesis framing for the identification problem.

Two papers supply the case for entailment as a cleaner signal than raw distant labels. LITE recasts ultra-fine entity typing as NLI and keeps 32.9 F1 on unseen types, where a classifier baseline falls to 10.8. It also finds that labels derived from Wikipedia entity links were actively harmful. Indirectly Supervised NLP surveys this style of transfer and notes how sensitive it is to hypothesis wording. LITE types nouns that name entities, whereas we target nouns that name events, where the hypothesis must encode argument structure.

Generated Knowledge Prompting both motivates and constrains the design. Having a model generate intermediate statements improves reasoning, but 86% of the resulting errors traced back to hallucinated statements. This is why our generated rewrites never vote directly and must pass an independent check.

BTZSC constrains our choice of verifier. It reports that rerankers (Qwen3-Reranker-8B, macro F1 0.72) now beat classical NLI cross-encoders, whose gains from scale have plateaued. We therefore treat the verifier as a swappable component and test both.

Three further papers bear on the problem from the argument side:

  • Spiliopoulou et al. map nominal arguments to verbal labels without supervision, reaching 58.82 F1 on CoNLL-2009.
  • RoleQ generates role questions that recover implicit arguments.

Both assume the predicate is already identified, which is exactly the step we automate. Conia and Navigli find that pretrained models encode nominal and verbal predicates differently. This warns against simply reusing verb-oriented tools on nouns. It is part of why we test nouns through explicit verbal rewrites rather than assuming the encoder treats them alike.

Research Program

Candidate nouns. Every noun token found by a part-of-speech tagger is a candidate. We do not first filter with a deverbal dictionary, because nouns like war or trip are eventive without an obvious verbal root, and lexical resources have large coverage gaps on web text (Textual Coverage of Eventive Entries).

Generating rewrites. An instruction-tuned model, prompted with a handful of examples, produces two families of rewrites for each candidate.

The argument-grounded clause uses a verb or light-verb paraphrase of the noun. Each filler must be either a span copied verbatim from the sentence or the placeholder "someone"/"something". This constraint makes hallucinated arguments mechanically detectable. It also lets an eventive noun whose arguments go unstated still produce a hypothesis. That case is common: 43% of arguments in QANom are implicit.

The event substitution replaces the noun phrase in the original sentence with "the act/process of V-ing …" and leaves the rest of the sentence unchanged. Result and object readings fail this test ("the act of building collapsed" is not entailed). Event readings pass it.

Scoring. The verifier assigns each rewrite a probability that the original sentence entails it. A candidate's score combines the best argument-grounded clause with the best substitution. A rewrite that the verifier judges a contradiction counts as negative evidence. On raw text we keep only candidates on which both families agree confidently, in either direction.

Distillation. The kept labels, drawn from a large Wikipedia and news sample, train a RoBERTa token tagger. This addresses the cost problem LITE identified: cross-encoders must score every candidate pairing, which becomes prohibitively expensive at scale. We also test the verifier as a filter on the NounAtlas synthetic corpus, discarding rewritten sentences whose nominal form does not entail the source verbal clause.

Baselines

Naive reference points. The weakest is the majority-class rule, which says no noun is a predicate; it is relevant because only about 17.5% of Chinese NomBank nouns are predicates. The second is a lexicon rule that marks every noun with a NomLex or WordNet derivational verb link as a predicate. Both run on identical test tokens with the same tagger output.

Direct prompting. This is the strongest zero-label competitor. The same instruction-tuned model that writes our rewrites is asked, zero-shot and few-shot, whether the marked noun describes an event or action in this sentence. To match compute, it gets the same number of model calls per token our pipeline spends: multiple samples aggregated by majority vote, which is self-consistency. It uses the same demonstrations, drawn from the same development pool, and the same tuning budget for prompt wording.

Supervised taggers. The strongest existing method with labels is a token classifier fine-tuned on QANom's training split, and separately on NomBank. It sets the in-domain ceiling, and it is the comparison that matters on held-out lemmas. For Chinese, the tree-kernel recognizer of Li et al. reports 90.96 F1 on gold trees. Under fully automatic parsing, their downstream semantic role labeling drops to 55.90 F1. For automatic predicate detection on English QANom we found no reported result on Lacuna, so we will train and report the supervised tagger ourselves. A model trained on NounAtlas synthetic data provides the strongest distant-supervision competitor.

Ablations. Four ablations isolate the key ingredients:

  • generic hypotheses such as "an event of V-ing occurred", testing whether argument grounding matters;
  • the model judging its own rewrites without an entailment check, testing whether verification matters;
  • verbs taken from NomLex instead of generated, testing whether the language model contributes beyond the lexicon;
  • the NLI cross-encoder swapped for Qwen3-Reranker-8B.

What counts as winning. On the ambiguous-lemma and unseen-lemma subsets, the proposal must beat direct prompting and the generic-hypothesis ablation by more than the spread across three seeds and prompt variants. The distilled tagger must stay within a few F1 of the supervised tagger on seen lemmas and exceed it on unseen ones.

Why not just…?

Why not just ask a strong language model whether the noun is an event? This is the real threat. On fine-grained distinctions in BTZSC, instruction-tuned models (macro F1 up to 0.67) trailed cross-attention rerankers (0.72). Generated Knowledge Prompting shows unchecked generations drive most errors. Neither result concerns token-level predicate status, so the compute-matched direct-prompting baseline settles the question. If direct prompting ties our pipeline on the ambiguous and unseen subsets, we will say so. The contribution would then reduce to cheaper distillation and to the diagnostic analysis of which rewrite family carries the signal.

Why not just fine-tune a tagger on QANom or NomBank? In-domain it will likely win, and we expect that. Its weakness is predicates it never saw. In LITE's analogous test, a supervised classifier fell from competitive to 10.8 F1 on held-out types while entailment held 32.9. Our lemma-disjoint split measures whether the same gap appears for nominal predicates.

Why not just label raw text with NomLex links (classic distant supervision)? A dictionary assigns predicate status by lemma, not by sentence, so building is always labeled the same way. It also misses eventive nouns without a recorded verbal root. Cumulative pipeline noise already prevented unlabeled data from helping NomBank systems (Liu and Ng). We train the same tagger on equal-size lexicon-labeled and verifier-labeled data to compare the two label sources directly.

Why not just score generic hypotheses with an NLI model or reranker? This is the cheapest entailment design, and it is our main ablation. Our prediction is that it fails on result nominals, where both readings entail that some building happened. If generic hypotheses do as well, the argument-grounding idea is wrong.

Proposed Experiments

The primary evaluation is token-level predicate-identification F1 on three English benchmarks: QANom's verbal-noun labels, the NomBank predicates in CoNLL-2009, and the NomBank partitive and "%" subsets. Results are reported overall, on lemmas that occur with both labels, and on a split whose test lemmas are disjoint from any demonstration or supervised training lemma.

A cross-lingual check runs the pipeline with a multilingual verifier on Chinese NomBank, under both gold-tree and automatic settings, against the reported tree-kernel numbers.

A downstream test measures whether identification quality carries over to role labeling. We train a nominal role labeler from (i) NounAtlas synthetic data, (ii) the verifier-filtered version of it, and (iii) our verified natural-text data at the matched 28,055-sentence size. We evaluate on NounAtlas's nominal test set, where the reported figure is 72.4 F1.

Cost is reported as model calls and wall-clock time per thousand tokens, for the verifier and for the distilled tagger.

Two checks decide the project. First, the argument-grounded and substitution scores must beat generic hypotheses and direct prompting on ambiguous lemmas; if they do not, the central claim fails. Second, verifier-labeled natural text must outperform lexicon labels and unverified synthetic data at equal size; if it does not, the distant-supervision half adds nothing.