Semantic role labeling asks who did what to whom in a sentence. Two major annotation schemes answer that question differently. PropBank uses verb-specific numbered slots such as A0 and A1. VerbNet uses general thematic roles such as Agent and Theme. A model that predicts both at once can share what it learns, but only if its two sets of predictions agree. Today that agreement comes from SEMLINK, a hand-built dictionary that lists which PropBank slot may pair with which VerbNet role for each verb sense. Where SEMLINK has an entry, enforcing it works very well. Where it has none (new verbs, new domains, senses nobody mapped), the joint model has nothing to enforce.
We propose to predict that compatibility relation instead of only looking it up. Both schemes already come with short natural-language descriptions of every label: a PropBank frame file says A0 of eat.01 is "eater", and VerbNet says an Agent is the participant who initiates the event. We train a cheap matching model over these descriptions. It scores how compatible any PropBank slot is with any VerbNet role, and we use those scores as soft cross-task constraints inside a globally constrained decoder. Hard structural rules, such as one core role per predicate and no overlapping arguments, are written declaratively and compiled into that decoder. In two sentences: the system learns to say which PropBank labels can go with which VerbNet labels by reading their definitions, then decodes both label layers jointly under rules it is not allowed to break. If this works, joint labeling keeps its consistency guarantees exactly where the hand-built mapping runs out.
Where this sits
Two papers come closest, and each supplies half of the method. Learning Semantic Role Labeling from Compatible Label Sequences predicts PropBank and VerbNet from a shared RoBERTa encoder. It uses a joint conditional random field, a sequence model that scores whole label sequences, over the 685 SEMLINK-admissible label pairs. It reports a 0% violation rate against 3–10% for ordinary multi-task heads, and in-domain gains of 3.5 F1 on VerbNet and 0.8 F1 on PropBank over prior systems. That paper motivates our work, because it shows that cross-task constraints matter. It also constrains our work, because any replacement must keep violations near zero. Its compatibility table is fixed and extensional: a pair is admissible only if a human listed it.
Label Definitions Improve Semantic Role Labeling supplies the label-semantics half. It matches each candidate argument against PropBank definitions with a RoBERTa cross-encoder and reports +4.4 F1 on unseen predicates, +3.2 F1 with 1,000 training examples, and +1.1 F1 on the out-of-domain Brown corpus. It is single-task and labels each argument independently, with no global adjudication. Because it runs one full transformer pass per candidate and definition, it is about 3.1× slower than a classification head. It also found no gain on Chinese, where definitions were generic. We take that negative result seriously: compatibility learned from definitions can only be as informative as the definitions are.
The constrained-inference machinery is older and well understood. Semantic Role Labeling Via Integer Linear Programming Inference cast SRL decoding as an integer linear program (ILP) with ten linguistic constraints. It reached 66.39 F1 on CoNLL-2004, about 1 F1 of which came from global inference. Structured learning with constrained conditional models generalized this: constraints are declared in first-order logic, kept separate from the scoring model, and can act as cheap supervision. We reuse that separation. We differ in one respect: our constraint weights between label pairs are produced by a model that reads label text, not authored by hand. Multilingual Semantic Role Labelling with Markov Logic shows the failure mode we target. Its declarative templates keyed on surface lemma identifiers left it 10.76 points behind the best systems on Czech, where frame identifiers are opaque. Templates over symbols do not carry over to unseen symbols; templates over descriptions might.
Core Claim
A compatibility function learned from PropBank and VerbNet label descriptions can stand in for SEMLINK on predicates that SEMLINK does not cover. Used as soft constraints in a joint decoder alongside declared hard rules, it keeps cross-task violations close to the SEMLINK-constrained joint model and keeps that model's accuracy on covered predicates. On uncovered ones it beats both unconstrained multi-task heads and simpler fallback mappings. A second, practical claim: because the definitions are encoded once and cached in a dual encoder rather than re-read for every candidate in a cross-encoder, the system recovers most of ACD's unseen-predicate gain without its 3.1× slowdown.
Research Program
The system has three parts.
The first is the scoring model. A shared RoBERTa encoder produces a vector $h_a$ for each candidate argument span $a$. PropBank and VerbNet labels are represented by encodings of their definitions, $u_p$ for the PropBank slot $p$ under the current predicate sense and $w_v$ for VerbNet role $v$. Both are computed once per sense and cached.
The second is the compatibility model. A small cross-encoder reads the pair of definitions, together with the predicate's frame description, and outputs $\kappa(p,v) \in (0,1)$. Because it runs over label pairs rather than over sentence–label pairs, its cost is paid once per predicate sense, not once per sentence. It is trained on SEMLINK pairs from training predicates: positives are listed pairs, negatives are unlisted pairs for the same sense.
The third is the joint decoder. It assigns each span a PropBank–VerbNet pair $(p_a, v_a)$ to maximize
$$\sum_{a} \big[ h_a^\top u_{p_a} + h_a^\top w_{v_a} + \lambda \log \kappa(p_a, v_a) \big] \quad \text{subject to } A\mathbf{y} \le \mathbf{b},$$
where $\lambda$ weights the learned compatibility term and $A\mathbf{y} \le \mathbf{b}$ collects the hard constraints. These are written in a small declarative rule file over label attributes and compiled to linear inequalities, in the style of constrained conditional models. Examples are "no two spans overlap", "each core PropBank slot appears at most once per predicate", and "if SEMLINK lists this sense, only listed pairs are allowed". When a sense is SEMLINK-covered, the last rule makes the decoder reduce to the 685-pair joint model. When it is not covered, only the learned $\kappa$ couples the two layers. The rule file also states candidate filters, which prune pairs with $\kappa$ below a threshold before decoding. This is where the declarative layer pays for itself: it keeps the joint search space near the size of the SEMLINK-pruned one.
Training runs both layers jointly. Sentences that have only PropBank labels are handled as the compatible-label-sequences paper does: we marginalize over VerbNet roles, here weighted by $\kappa$ instead of by a 0/1 mask.
Proposed Experiments
We use CoNLL-2005 with VerbNet roles attached through SEMLINK, following Learning Semantic Role Labeling from Compatible Label Sequences. Evaluation covers the WSJ test set and the out-of-domain Brown test set. To get a controlled coverage gap, we build a mapping-holdout split. We remove the SEMLINK entries for a random 20% of predicate lemmas, choosing them by lemma so that no sense of a held-out verb leaks. Models never see those entries during training, and the held-out pairs serve as gold for evaluating $\kappa$. We report three things separately for covered, held-out, and naturally unmapped predicates: PropBank F1, VerbNet F1, and the violation rate $\rho$ against the full SEMLINK. We also report throughput in sentences per second against ACD on the same hardware. For the low-resource regime we repeat this with 1,000 training examples, the setting where ACD reports +3.2 F1.
The file contains no absolute CoNLL-2005 F1 values for either the joint CRF or ACD, only gains relative to earlier systems. We will therefore reproduce both under our splits rather than quote targets.
Baselines
The naive reference point is a shared RoBERTa encoder with two independent classification heads, one for PropBank and one for VerbNet, decoded independently. The compatible-label-sequences paper reports that this setup violates SEMLINK on 3–10% of predictions, and we expect to reproduce that range. Beating it means lowering $\rho$ on held-out predicates while matching or improving F1.
The strongest existing method is the joint CRF with SEMLINK constraints from Learning Semantic Role Labeling from Compatible Label Sequences. It reports 0% violations and +3.5 VerbNet F1 and +0.8 PropBank F1 over prior work on WSJ. We run it with the same encoder, data, mapping-holdout split, and hyperparameter search budget as our system. On held-out predicates it must fall back to the unconstrained product of labels, since it has no table entry. Beating it means three things: statistically indistinguishable F1 on covered predicates; higher VerbNet F1 and lower $\rho$ on held-out predicates; and no significant loss on Brown.
The strongest single-task label-semantics competitor is ACD from Label Definitions Improve Semantic Role Labeling. It reports +4.4 F1 on unseen predicates at 3.1× the cost of a classification head. We train it on the same PropBank data and give it the same tuning budget. We compare PropBank F1 on unseen predicates and wall-clock throughput. A win here is at least 2× ACD's throughput while staying within 1 F1 of it on unseen predicates, since our dual encoder trades some matching power for speed.
Two ablations isolate the key ingredient. In the first, $\kappa$ is replaced by a uniform constant, so the hard rules stay and cross-task coupling disappears on uncovered predicates. In the second, $\kappa$ is computed from label identifiers alone, with no definition text. If neither ablation loses ground on held-out predicates, the definitions are not doing the work.
Why not just…?
Why not just run ACD and then fix its outputs with an ILP? This post-hoc repair is in the spirit of the projection approach in Locally Coherent, Globally Incoherent. It enforces the within-task rules cheaply, and we include it as a baseline because it will probably close most of ACD's overlap and duplicate-role errors. What it cannot do is couple PropBank to VerbNet on unmapped verbs, because no constraint exists there to project onto. The held-out-predicate VerbNet F1 and $\rho$ comparisons settle whether the learned coupling adds anything beyond repair. Post-hoc repair also keeps ACD's 3.1× cost.
Why not just back off to the VerbNet class mapping? Verbs in the same VerbNet class usually share SEMLINK mappings. For a held-out verb, one could copy the mapping of its classmates. This is the most serious threat, because it may recover held-out pairs as well as $\kappa$ does with no learning at all. We run it as a full baseline: a joint CRF with class-level backoff tables. We also report precision and recall of each method's recovered compatibility pairs against the withheld SEMLINK entries. If class backoff matches learned compatibility on held-out predicates, we will narrow the claim to verbs with no VerbNet class membership, or with unhelpful membership, where backoff is undefined. We will report that honestly rather than argue for the general case.
Why not just prompt a large language model with both sets of definitions and ask for consistent labels? Locally Coherent, Globally Incoherent reports structural incoherence on over 97% of partition tasks for frontier models under naive composition. It also finds that richer prompting and aggregator models often fail to fix such incoherence. Adjudicating LLMs as PropBank Adjudicators reports that unconstrained language models struggle with role boundaries. We will still measure $\rho$ and F1 for a prompted model on a 500-sentence sample of held-out predicates, so the claim rests on our own numbers and not only on those reports.
Why not just add the definitions as input features to the existing joint CRF? This tests whether definitions help scoring rather than coupling. It corresponds to our system with $\kappa$ removed but definition-based $u_p$ and $w_v$ kept, which is the uniform-$\kappa$ ablation above. If that variant matches the full model, the compatibility model is unnecessary.
Decisive Checks
The idea stands or falls on held-out predicates. First, $\kappa$ must recover the withheld SEMLINK pairs clearly better than class backoff. We judge this by pair-level F1 against the withheld entries, with the comparison restricted to verbs where backoff is defined. Second, the full decoder must cut $\rho$ on held-out predicates well below the 3–10% of independent heads. It must do so while losing nothing on covered predicates relative to the SEMLINK-constrained joint CRF. Third, the uniform-$\kappa$ ablation must lose VerbNet F1 on held-out predicates. If $\kappa$ recovers pairs no better than class backoff, the project reduces to an engineering combination of known pieces and we will say so. As a final check, applying the same pipeline to CoNLL-2009 Chinese, where ACD found definitions too generic, should show the effect shrinking. That prediction ties the method's value directly to how informative the definitions are.
References
- Learning Semantic Role Labeling from Compatible Label Sequences
- Label Definitions Improve Semantic Role Labeling
- Semantic Role Labeling Via Integer Linear Programming Inference
- Structured learning with constrained conditional models
- Multilingual Semantic Role Labelling with Markov Logic
- Locally Coherent, Globally Incoherent: Detecting Compositional Incoherence in Multi-Component LLM Agents
- Adjudicating LLMs as PropBank Adjudicators