Verb-List: Curated Verb Inventories

• Verb-List is a curated inventory of verbs paired with semantic annotations like classes, paraphrase weights, and sense labels, used to drive various NLP tasks.
• These inventories employ diverse annotation strategies—from closed-class semantic tagging to graded multi-label assignments—to model verb behavior in contexts like arithmetic problem solving and syntactic bias.
• Practical applications include predicting Hungarian verb–object order, enhancing video action recognition, and facilitating cross-lingual sense mapping in deep semantic lexicons.

In current computational linguistics and multimodal learning, a “Verb-List” denotes a curated inventory of verbs paired with semantic classes, paraphrase weights, constructional preferences, ontological types, or sense labels. Such inventories are used as operational resources rather than as simple lexicographic lists: they determine arithmetic operations in Hindi word problems, condition verb–object order in Hungarian, encode ambiguity in video action recognition, parameterize dative-alternation benchmarks, paraphrase noun compounds, populate deep semantic lexicons, and define cross-lingual sense inventories for polysemous French verbs (Sharma et al., 2023, Demszky et al., 2020, Wray et al., 2019, Hawkins et al., 2020, Nakov, 2019, Allen et al., 2020, Rakho et al., 27 May 2026).

1. Major resource types and their formal roles

Across the literature, verb inventories vary primarily along three axes: the unit being listed, the annotation attached to each verb, and the downstream inference task. Some resources enumerate lemmas and assign each lemma to a small closed set of semantic classes; others preserve open-ended paraphrase distributions or sense inventories; still others treat verbs as dimensions in a dense or sparse label space.

Resource	Inventory	Primary use
Hindi word-problem resource	~200 unique Hindi verb lemmas; six-way labels	Addition/subtraction operation detection
Hungarian verb-order study	380 transitive verbs; 11 semantic classes	Predicting OV vs. VO order
Visual action resources	90 verbs; 47 manner and 43 result verbs	Recognition and cross-modal retrieval
DAIS	200 English verbs	Modeling dative verb bias
Noun-compound paraphrase lexicon	250 compounds; 17,821 filtered verb tokens	Compound interpretation and RTE-style pairs
French sense corpus	20 French polysemous verbs	LG, translation, and LG#TRSL labeling

The Hindi resource is extracted from 1,713 addition/subtraction word problems in the HAWP corpus and uses six verb categories, including a catch-all na label for POS-tagging noise (Sharma et al., 2023). The Hungarian study groups 380 transitive verbs into 11 semantic classes such as Stative, Activity, Achievement, Psych, and Possession/Transfer, with class membership determined by dictionary definitions, argument-structure diagnostics, and corpus-based aspectual tests (Demszky et al., 2020). In egocentric video understanding, a 90-verb vocabulary is used as a shared label space over three datasets, and the vocabulary is manually split into 47 manner verbs and 43 result verbs (Wray et al., 2019). DAIS selects exactly 200 English verbs—100 alternating and 100 non-alternating—by starting from Beth Levin’s 1993 classification and taking the top 100 most frequent verbs in each class (Hawkins et al., 2020).

Other resources prioritize distributional richness over closed-class categorization. The noun-compound paraphrasing dataset stores, for each of 250 noun–noun compounds, the set of human-proposed paraphrasing verbs together with counts of distinct annotators who proposed each verb (Nakov, 2019). COLLIE-V is a broad-coverage deep semantic lexicon for verbs “with the coverage of WordNet,” but augments lexical entries with ontological concepts, semantic role preferences, linking templates, and entailment axioms (Allen et al., 2020). The French sense-tagged corpus annotates instances of 20 polysemous verbs with three labels simultaneously: an English translation, a Lexicon-Grammar entry, and a fine-grained concatenation of the two (Rakho et al., 27 May 2026).

2. Verb categorization as arithmetic control in Hindi word-problem solving

In Hindi arithmetic word-problem solving, verb categorization is treated as a direct mechanism for identifying state change. The resource defines six categories: Observation, Positive, Negative, Positive Transfer, Negative Transfer, and na. Observation states the presence or measurement of entities in a container without change; Positive denotes creation or addition; Negative denotes removal, destruction, or loss; Positive Transfer denotes a one-way transfer from Container B to Container A; Negative Transfer denotes a one-way transfer from Container A to Container B; and na is reserved for tokens tagged as verbs but not functioning as semantic verbs (Sharma et al., 2023).

The solver processes each sentence token by token and maintains a list of states represented as entity–container–quantity triples. State extraction identifies a container, quantity, and entity, and instantiates a new state when a fresh entity–container–quantity is found. Positive verbs add the quantity to the relevant state, while Negative verbs negate the quantity before storing it. Transfer verbs update two containers: for Positive Transfer, the system subtracts $X$ from the source and adds $X$ to the destination; for Negative Transfer, it subtracts $X$ from the first container and adds $X$ to the second (Sharma et al., 2023).

Final answer extraction combines verb-derived state updates with question indicators. If no transfer occurs, negative indicators such as “mukable,” “pehle,” and “chahiye” trigger subtraction of all subsequent states from the first, whereas positive indicators such as “kul” and “milakar” trigger summation of all relevant states. Additional heuristics include circumscription, under which a bare numeral inherits the last seen entity/container; normalization of monetary entities such as “paisa” and “rupay” to the ₹ symbol; and an absolute-value rule under which negative final answers are returned as positive (Sharma et al., 2023).

The annotation study reports Fleiss’ $\kappa = 0.89$ on a 225-verb, 100-sentence sample, and the full inventory contains ~200 unique Hindi verb lemmas annotated by two computational-linguist annotators. In five-fold cross-validation for verb categorization, MuRIL fine-tuning attains an $F_1$-score of $0.962$, outperforming SVM at $0.904$, 1-NN on FastText at $0.895$, Random Forest at $0.883$, and Logistic Regression at $X$0. Using predicted categories, the end-to-end rule-based solver reaches $X$1 accuracy on one-operation HAWP problems, compared with a comparable baseline at $X$2. The reported error modes include irrelevant information not being filtered, entity/container mismatches due to POS or dependency parsing mistakes, set-completion problems being out of scope, and shallow-parser tagging errors propagating into incorrect state updates (Sharma et al., 2023).

A recurrent misconception is that arithmetic reasoning in such problems is driven solely by numerals. The Hindi system explicitly treats verbs as indispensable for addition/subtraction contexts because they determine whether the sentence contributes an initial state, an increment, a decrement, or a transfer (Sharma et al., 2023).

3. Verb classes in syntactic alternation and constructional bias

Verb lists also function as explanatory variables for syntactic choice. In Hungarian verb–object order, a mixed-effects logistic regression predicts $X$3 using semantic class, object definiteness, object NP weight, and random intercepts for verb lemma and document. The 11 semantic classes are Stative, Activity, Accomplishment, Achievement, Psych, Creation, Consumption, Motion-with-Object, Contact, Perception/Communication, and Possession/Transfer (Demszky et al., 2020).

The reported fixed effects show that Stative verbs are strongly OV-preferring, with coefficient $X$4 and odds ratio $X$5, whereas Psych verbs are very strong VO-preferring, with coefficient $X$6 and odds ratio $X$7. Achievement verbs are also strongly VO-preferring at $X$8, and Perception/Communication verbs are VO-preferring at $X$9. Definiteness has a substantial positive effect on OV order, with coefficient $X$0 and odds ratio $X$1, while each additional word of object weight reduces OV probability, with coefficient $X$2 and per-word odds ratio $X$3. All semantic effects are significant at $X$4 except for Contact and Consumption (Demszky et al., 2020).

The study’s substantive conclusion is that semantic class is the single strongest predictor of Hungarian verb–object order, with Stative verbs exhibiting the highest OV rate and Psych verbs the strongest VO rate. Object definiteness and object weight also reliably shift order, but to a lesser degree than lexical semantics (Demszky et al., 2020). This is a strong counterpoint to analyses that treat Hungarian word order as determined only by information structure.

A related but distinct use of verb lists appears in DAIS, a benchmark for English dative alternation. DAIS contains 50K human judgments for 5K distinct sentence pairs and uses 200 unique verbs, selected as the top 100 most frequent alternating verbs and the top 100 most frequent non-alternating verbs from Levin’s 1993 classification. Here “verb bias” is operationalized not through a semantic class label but as the log-likelihood ratio, or difference in surprisal, between the double-object and prepositional-dative realizations of the same sentence pair. Human verb-bias scores are mean slider ratings aggregated across participants and conditions, and model performance is evaluated by rank-correlating model log-likelihood ratios with human preferences (Hawkins et al., 2020).

Taken together, these studies show two distinct modeling strategies. One represents verbs through a manually defined semantic taxonomy and estimates class-level effects on a syntactic alternation; the other keeps a flat 200-verb inventory and measures lexically specific constructional preference directly. This suggests that “Verb-List” resources can support either abstraction over verbs or fine-grained lexical profiling, depending on the target phenomenon.

4. Multi-verb label spaces for visual actions

In visual action recognition and retrieval, verb lists are used to represent semantic ambiguity rather than eliminate it. The core resource is a 90-verb vocabulary collected over the BEOID, CMU, and GTEA+ egocentric datasets. In one formulation, every video segment is mapped to a 90-dimensional label vector, with single-verb labels defined by majority vote, multi-verb hard labels defined by a threshold of $X$5, and soft-assigned multi-labels defined by normalized annotator counts in $X$6. These labels are learned with a two-stream fusion CNN using sigmoid binary cross-entropy over all 90 verb dimensions (Wray et al., 2018).

The later multi-verb study uses the same 90-verb union and reports dataset-specific usage statistics. BEOID contains 732 clips and 42 unique verbs under Multi-Verb hard labels; CMU contains 404 clips and 31 unique verbs; GTEA+ contains 1,001 clips and 34 unique verbs. Under Soft-Assigned Multi-Verb labeling, BEOID uses all 90 verbs, with only 42 exceeding the $X$7 threshold for hard binarization; CMU uses 89 of the 90 verbs; and GTEA+ uses all 90 verbs in the soft annotations (Wray et al., 2019).

The vocabulary is manually partitioned into 47 manner verbs and 43 result verbs. Manner verbs describe how the action is carried out, with examples such as “pull,” “rotate,” and “stir,” while result verbs describe the end state, with examples such as “open,” “fill,” and “crack.” Supplementary verbs, such as “grab” and “hold,” mark sub-actions and appear naturally in the manner set (Wray et al., 2019). The five most frequently selected verbs across all datasets are “move,” “touch,” “hold,” “pick up,” and “pull out.” Figure-level statistics show that “crack” has very high annotator agreement, with median soft score approximately $X$8, whereas generic verbs such as “move” often have low median scores, approximately $X$9 (Wray et al., 2019).

The central methodological claim is that single-verb labels neglect legitimate semantic ambiguities and class overlaps between verbs. Multi-verb hard labels capture common co-occurrences, while soft-assigned multi-verb labels preserve graded semantic relevance and improve cross-modal retrieval. The papers explicitly state that WordNet and Word2Vec were insufficient to discover result–manner or supplementary relationships, which motivated direct human annotation (Wray et al., 2019). This is important because it frames the verb list not as a closed taxonomy derived from lexical resources, but as an empirically elicited semantic space.

Examples in the action literature illustrate both overlap and disambiguation. “Open door” and “open cupboard” are annotated with overlapping verbs such as {open, hold, turn, rotate}, while “open bottle” shares “open” and “hold” but adds “twist,” thereby separating it from hinge-based actions without using object nouns (Wray et al., 2018). A plausible implication is that verb lists in multimodal settings are most useful when they encode partial agreement and graded salience rather than forcing a single canonical label.

5. Paraphrase lexicons, ontological verb entries, and cross-lingual sense inventories

Another major use of verb lists is semantic interpretation. For noun-compound interpretation, paraphrasing verbs are collected for 250 noun–noun compounds derived from Levi’s 1978 list after removing non-predicate NPs and pure concatenations. Annotation on Amazon Mechanical Turk used HITs of five compounds, with workers instructed to provide at least three verbs, optionally with a following preposition, that fit the paraphrase template “$X$0 that $X$1 $X$2.” The corpus comprises 174 distinct workers, 19,018 raw verbs, and 17,821 verb tokens after filtering and normalization (Nakov, 2019).

Each compound is associated with verbs weighted by the number of distinct annotators who proposed them. For a fixed compound $X$3 and verb $X$4, the raw weight is $X$5, the normalized weight is $X$6, and the add-1 smoothed weight is $X$7. The resulting lexicon is strongly Zipfian: top verbs across all compounds include be, have, make, use, cause, result in, contain, carry, produce, spread, and transmit, while the majority of verbs appear for only one or two compounds (Nakov, 2019).

COLLIE-V extends the notion of a verb list into a deep semantic lexicon. It organizes verbs under top-level ontological categories such as ONT :: CAUSE-EFFECT, ONT :: DESTROY, ONT :: DIE, ONT :: CONSUME, ONT :: EAT, ONT :: DRINKING, ONT :: MOVE, ONT :: BECOME, and ONT :: CONFINE. Each entry contains a gloss, syntactic frames, semantic roles and preferences, and entailment axioms. For example, kill is associated with the gloss “cause to die,” transitive and intransitive frames, AGENT and AFFECTED roles, and axioms such as $X$8. The lexicon is built by bootstrapping from the TRIPS lexicon and WordNet synset-to-TRIPS-type mappings, parsing glosses, identifying roles via IMPRO nodes, inheriting preferences, extracting entailment axioms, proposing linking templates, classifying new concepts, and running consistency checks (Allen et al., 2020).

The French semantically annotated corpus takes a third approach: each instance of a verb receives three linked labels. The corpus covers 20 French polysemous verbs, and every instance is annotated with the actual English translation in a parallel corpus, a Lexicon-Grammar entry, and a fine-grained sense produced by concatenating the two. For comprendre, the paper reports 8,240 instances, 8 Lexicon-Grammar entries, 183 English translations, and 308 fine-grained LG#TRSL senses. For porter, the reported figures are 3,301 instances, 20 Lexicon-Grammar entries, 219 translations, and 495 fine-grained senses. The example labels for comprendre include V_12_17#understand, V_12_17#appreciate, V_12_17#sympathise, V_12_17#realise, V_12_17#comprehend, and V_12_17#know (Rakho et al., 27 May 2026).

These three resources illustrate distinct semantics-oriented interpretations of a verb list. In the noun-compound lexicon, the list is a weighted paraphrase set; in COLLIE-V, it is an ontology-linked lexical entry with roles and axioms; in the French corpus, it is an aligned sense inventory spanning syntax, lexical semantics, and translation.

6. Recurring design choices, limitations, and open interpretive issues

Several recurring design choices appear across these resources. First, exhaustive verb inventories are often externalized to appendices or repositories rather than printed in full. The Hindi paper states that the paper does not print the entire list of ~200 verbs, but that the released GitHub repository contains the full inventory (Sharma et al., 2023). The Hungarian paper places the full lemma-by-class assignments for its 380 verbs in Appendix A and supplemental materials (Demszky et al., 2020). DAIS likewise makes the full 200-verb inventory available in a repository file verbs.csv rather than in the main body (Hawkins et al., 2020). This suggests a stable division between article-level exposition and resource-level distribution.

Second, direct human annotation is repeatedly preferred when lexical ambiguity is central. The visual-action studies reject forced disambiguation into a single verb or verb–noun pair and instead rely on crowdsourced multi-verb selection because generic lexical resources were insufficient to discover result–manner or supplementary relationships (Wray et al., 2019). The noun-compound study also treats the space of plausible verbs as empirically recoverable from annotator proposals rather than reducible to a small manually fixed relation set (Nakov, 2019).

Third, the usefulness of a verb list depends on the error model of the downstream system. In Hindi arithmetic solving, POS-tagging noise is severe enough to justify an explicit na category, and shallow-parser errors can propagate into incorrect state updates (Sharma et al., 2023). In Hungarian word order, semantic classes have large effects, but not every class effect reaches conventional significance: Contact and Consumption are near-neutral and do not reach the standard threshold (Demszky et al., 2020). In action recognition, highly generic verbs such as “move” are frequent yet weakly agreed upon, while highly specific verbs such as “crack” receive much stronger consensus (Wray et al., 2019). In the French corpus, the excerpted statistics are incomplete for several verbs—such as mettre, passer, présenter, répondre, and tirer—even though the paper indicates that each has its own figures in Table 2 (Rakho et al., 27 May 2026).

A common misconception is that verb lists are intrinsically simple lexical enumerations. The evidence instead shows that they are often structured objects whose entries can carry state-change semantics, syntactic alternation biases, graded annotator relevance, paraphrase weights, semantic roles, entailment axioms, or cross-lingual sense identifiers. A plausible implication is that the most informative verb lists are those that bind lexical items to an explicit formal interpretation rather than treating verbs as unanalyzed labels.