Event argument extraction means reading a passage, finding the word that signals an event (say, "arrested"), and pulling out who did it, to whom, where and when. It is hardest when the answer sits sentences away from the event word. In one benchmark, nearly 40% of events have at least one argument outside the trigger's sentence (Multi-Sentence Argument Linking). Large language models used without training, either from instructions alone (zero-shot) or from a handful of worked examples placed in the prompt (in-context learning), do badly here. On a 220-role benchmark, zero-shot GPT-3.5 scores 22.73% F1 against 39.43% for a supervised model (GENEVA). Asking the model to "think step by step" (chain-of-thought prompting) is the obvious fix, but it fails in a known way on extraction. Free-form reasoning consumes 15–55× more tokens and still often loses to structured pipelines (DiCoRe), and it tends to wander into roles and spans that are not in the text.
We propose replacing free-form reasoning with a short, typed chain whose every step must be a span copied verbatim from the document. For each role, the model writes a chain of zero, one or two hops: event word, then an optional intermediate "clue" mention, then the argument. A decoder that may only emit document substrings enforces the copying. The chain shape borrows the two-hop logic rule that a supervised model used to win on cross-sentence arguments (Chain Reasoning Paradigm), and turns it into a prompt-time reasoning format for frozen models. In two sentences: make the model's chain-of-thought for argument extraction a path of quoted spans linking the event to its distant argument, rather than prose. We expect this to recover most of chain-of-thought's gain on long-range arguments at a small fraction of its token cost and without its hallucinated arguments.
Core Claim
Our claim is that chain-of-thought helps argument extraction mainly by forcing explicit linking through intermediate mentions, such as a coreferent noun or a co-participant, and not through extra computation or free deliberation. If that is true, three consequences follow. Grounded clue chains should match or beat unconstrained chain-of-thought at equal or lower token budgets. Their advantage should concentrate on arguments one or more sentences from the trigger. Replacing the clue spans with random document spans of equal length should erase the gain. If the gain survives shuffled clues, the mechanism is just extra tokens and the claim is wrong.
Relation to Prior Work
Four lines of work supply the context, each in a different role.
The supervised chain model of Liu et al. motivates the design. It learns rules of the form "role of T is X if T relates to clue B and B relates to X" using fuzzy logic over BERT. It reaches 56.1% F1 on RAMS, with up to 15.7 F1 gains on cross-sentence arguments. However, it is gradient-trained, needs pre-identified entity candidates, and fixes the chain at two hops. We keep the hop structure, drop the training, let the model choose zero, one or two hops, and let clues be any document span. Allowing any span matters for GENEVA's abstract, non-entity roles.
Two lines supply in-context methods. Both use direct input-output demonstrations with no intermediate reasoning, and we build on rather than replace them:
- TISE chooses demonstrations by balancing similarity, diversity and event-type relevance, and formats them as code.
- Code4Struct renders the event ontology as Python classes.
HERO attributes the gap between prompted and supervised models to missing extraction rules. It learns a static, global rule set from labeled error feedback. Our chains are per-instance rather than global, and our zero-shot variant needs no labels.
DiCoRe constrains the design. It shows that open-ended reasoning is expensive and overloads the model on large schemas. Its decoupled propose–ground–verify pipeline with grammar-constrained decoding wins on event detection, but it never extracts arguments. We take its constrained-decoding lesson as given. We treat a direct port of its pipeline to arguments as our most serious competitor, discussed below.
Research Program
Chain format. For each role in the event's schema, the model emits a chain in a fixed format with role, hops and answer, or NONE. Each hop is a document span, enforced by finite-state constrained decoding that only allows substrings of the input. The role name is enforced against the schema. The chain therefore cannot introduce text that is absent from the passage or roles that are absent from the ontology. These were the two failure modes the file attributes to free chain-of-thought.
Few-shot demonstrations. We build demonstrations by asking the model to produce a clue chain that ends at the known gold argument. We keep only chains that pass span validity and whose answer reproduces the gold span. Demonstrations are then selected with TISE's scoring, so the selection signal is held fixed.
Zero-shot demonstrations. Here we bootstrap from unlabeled documents. The model generates several chains per role. We keep those that are span-valid and whose answers agree across samples, and use them as demonstrations. This tests whether grounded chains can replace labeled examples entirely.
Analysis. Results are stratified by trigger–argument sentence distance, by whether the role is entity-like or abstract (using GENEVA's distinction), and by tokens generated per instance. We will additionally run the analysis on two frozen open-weight models of different size, because constrained decoding requires access to the logits.
Baselines
Naive reference. Zero-shot direct prompting with a code-style schema prompt and no reasoning, in the style of Code4Struct, on the same frozen model. The only reported figure for this regime in the file is GENEVA's 22.73% F1 for zero-shot GPT-3.5 (GENEVA). We rerun it on our models rather than compare across models.
Strongest in-context competitor. TISE demonstrations combined with HERO rules, answer-only. This combines the best published selection and the best published rule guidance.
- No reported result was found on Lacuna for TISE's absolute RAMS score.
- HERO reports a 3.18% F1 gain over prompting baselines on RAMS, but no absolute value in the file.
- We therefore reimplement both and give them the same demonstration pool, the same number of demonstrations, and the same labeled data for HERO's rule search that we use for building chains.
- Beating this baseline means a statistically significant Arg-C F1 gain on RAMS and WIKIEVENTS, with the gain at least doubled on arguments outside the trigger sentence.
Unconstrained chain-of-thought. The same demonstrations with free-text rationales, run at matched and at unlimited token budgets. This isolates the grounding constraint.
DiCoRe-style port. The model first proposes all candidate spans liberally, then assigns roles under constrained decoding, then a verification pass filters them. This uses the same model and comparable token budget.
Supervised references. The supervised chain model at 56.1% F1 on RAMS (Liu et al.) serves as a ceiling, not a target. Ebner et al.'s 68.3/73.3 F1 (Multi-Sentence Argument Linking) uses a candidate-linking setup that is not directly comparable, and we will report it with that caveat.
Ablations.
- Clue spans replaced by random spans of equal length. This is the decisive mechanism test.
- Hop count fixed at two.
- Answer-only demonstrations padded to the same token length as the chained demonstrations.
Why not just…?
Why not just use free chain-of-thought or a reasoning model? DiCoRe found these spend 15–55× more tokens and often underperform structured decoding on event detection. Arguments are a harder, multi-span output, which should make drift worse. Still, a strong reasoning model might close the gap on long-range arguments. The unlimited-budget free chain-of-thought baseline settles this: we report F1 against tokens, and grounded chains must dominate that curve rather than merely win at one point.
Why not just add HERO's rules and better exemplars? This is a genuine threat, because HERO already diagnoses the missing arguments and boundary errors we target. Global rules, however, cannot say which intermediate mention links this trigger to this distant argument. That is a per-instance fact. If TISE plus HERO matches us on cross-sentence arguments, our claim fails. We also test the combination, since rules and chains are complementary.
Why not just port DiCoRe's propose–ground–verify pipeline to arguments? It might work as well. It handles schema overload and hallucination through constrained grounding, which is half of what we do. What it lacks is explicit linking, so if linking is the real bottleneck, its errors should concentrate on non-local arguments. The distance-stratified comparison is designed to show exactly this. If the port ties us across all distance strata, we will report that the clue structure adds nothing beyond constrained grounding.
Why not just fine-tune? Supervised models remain stronger with full data: 56.1% F1 on RAMS for the supervised chain model. Our target regime is few- and zero-shot, where metric-learning approaches reach only 11–12% F1 on document-level arguments (Few-Shot Document-Level EAE). We report learning curves to locate the crossover point.
Proposed Experiments
Benchmarks and metrics.
- RAMS (139 types, 65 roles, five-sentence windows) is the primary benchmark because the trigger–argument distance is explicit.
- WIKIEVENTS (Li et al.) adds longer documents and coreference-aware "most informative mention" scoring.
- GENEVA's zero-shot and few-shot splits test abstract roles and schemas of hundreds of roles.
- ACE05 serves as a sentence-level sanity check, where we expect little gain.
Metrics are argument identification and classification F1, stratified by sentence distance and role abstractness. We also report the fraction of predicted spans not present in the text or not allowed by the schema, and the number of generated tokens per event.
Decisive checks. The proposal stands if three conditions hold:
- Grounded chains beat TISE plus HERO and the DiCoRe-style port on cross-sentence arguments.
- The shuffled-clue ablation removes most of that gain.
- Grounded chains reach free chain-of-thought's accuracy with substantially fewer tokens.
A secondary check is whether label-free bootstrapped chains in the zero-shot setting move GENEVA's zero-shot score meaningfully above the direct-prompting reference we reproduce. That would show grounded reasoning can partly substitute for labeled demonstrations.
References
- Multi-Sentence Argument Linking
- GENEVA: Benchmarking Generalizability for Event Argument Extraction with Hundreds of Event Types and Argument Roles
- DiCoRe: Enhancing Zero-shot Event Detection via Divergent-Convergent LLM Reasoning
- Document-Level Event Argument Extraction With a Chain Reasoning Paradigm
- TISE: A Tripartite In-context Selection Method for Event Argument Extraction
- Code4Struct: Code Generation for Few-Shot Structured Prediction from Natural Language
- Rule-Guided Extraction: A Hierarchical Rule Optimization Framework for Document-Level Event Argument Extraction
- Few-Shot Document-Level Event Argument Extraction
- Document-Level Event Argument Extraction by Conditional Generation