📄 Refining Event Extraction through Cross-Document Inference

Summary

Sentence-level event extractors judge each event mention in isolation. As a result, they miss triggers that appear in unfamiliar contexts and accept spurious ones that happen to match a pattern. Ji and Grishman apply Yarowsky's "One Sense Per Discourse" idea to ACE event extraction and widen the "discourse" from one document to a cluster of topically related documents.

The pipeline works in four steps:

• A baseline English IE system (Grishman et al., 2005), which combines patterns with Maximum Entropy classifiers, tags each test document and assigns confidences.
• An INDRI query built from the detected triggers and arguments retrieves related documents from unlabeled corpora.
• The same baseline tags those related documents.
• Weighted frequency and margin statistics are pooled within the document and across the cluster. A cascade of threshold rules then deletes weak mentions, spreads the dominant event type or role, and adds missed triggers and arguments.

Without any additional labeled data, the method raises F-measure by 7.6% for trigger labeling and by 6% for argument labeling. The work shows that redundancy across related texts can act as free, dynamically built background knowledge for IE.

Key points

• The task is ACE 2005 event extraction, with 8 event types and 33 subtypes treated as 33 separate types. Event mention coreference is not handled.
• The baseline extracts events sentence by sentence. It uses trigger-to-argument patterns plus MaxEnt classifiers for trigger labeling, argument detection, role assignment and reportable-event decisions.
• For each test text, the top N=25 related documents are retrieved with INDRI. The query weights triggers and arguments by local confidence and frequency, and it is expanded with coreferent or ACE-related names.
• Error analysis motivates the method. For example, "fire" is an End-Position trigger in only 7% of 81 training occurrences but in 100% of the occurrences in test plus related documents. Two hypotheses follow: "one trigger sense per cluster" and "one argument role per cluster".
• Inference uses nine rules with thresholds δ1–δ13. The rules first act within the document and then across the cluster.

Interesting ideas

• I1. Retrieve your own context at test time. Fetching documents on the same topic and running the same tagger over them creates a topic-specific distribution of labels. This kind of self-supervision could help any low-resource extractor, for example a Ukrainian NER or event tagger with little annotated data.
• I2. Let easy mentions help hard ones. A phrasing that matches a known pattern with high confidence in one document can justify labeling an ambiguous phrasing of the same event elsewhere. The confident evidence can also veto a locally confident but globally rare label.
• I3. Use margins as well as counts. The method propagates a label only when the top type or role clearly beats the runner-up. This guards against spreading errors when the evidence is split, and the same safeguard suits any label-propagation or majority-vote scheme.
• I4. Cascade from local to global. Rules apply first within the sentence and document, and then across the cluster, with cluster-level decisions overriding document-level ones on conflict. This ordering is a simple, interpretable alternative to joint inference.
• I5. Reuse high-confidence output. The authors suggest using consistent, high-confidence cluster results as bootstrapped training data. The same aggregation could also produce event-driven summaries or coreference hints.