Abstract
Previous studies on temporal relation extraction focus on mining sentence-level information or enforcing coherence on different temporal relation types among various event mentions in the same sentence or neighboring sentences, largely ignoring those discourse-level temporal relations in nonadjacent sentences. In this paper, we propose a discourse-level global inference model to mine those temporal relations between event mentions in document-level, especially in nonadjacent sentences. Moreover, we provide various kinds of discourse-level constraints, which derived from event semantics, to further improve our global inference model. Evaluation on a Chinese corpus justifies the effectiveness of our discourse-level global inference model over two strong baselines.
Introduction
Temporal Relation Extraction (TRE) is the task of determining the chronological order of events in a text. While crucial for applications like summarization and question answering, TRE faces two major hurdles: sparse data and the limitations of local context. Popular datasets like TimeBank only annotate a small subset of "easy" event pairs, often ignoring relations between events in nonadjacent sentences. This leads to fragmented timelines. Furthermore, in discourse-driven languages like Chinese—which features flexible sentence structures and frequent ellipsis (omitted words)—local context is often insufficient. Many temporal cues are implicit and require global inference across the entire document to resolve.
To address these issues, the authors developed a new event-driven fully-annotated Chinese temporal relation corpus based on the ACE 2005 Chinese corpus. This dataset includes 163 documents and 21,132 event pairs, providing a dense graph for document-level analysis that is three times the density of previous datasets.
Methodology: Discourse-level Global Inference Model (DGIM)
The proposed DGIM system operates in two stages. First, a Maximum Entropy (MaxEnt) classifier serves as the baseline, predicting relations (Before, After, Overlap, Unknown) using lexical, syntactic, and semantic features. Second, a Global Inference Engine utilizes Integer Linear Programming (ILP) to optimize the temporal graph of the entire document, enforcing consistency across all event pairs rather than treating them in isolation.
The core of the DGIM model lies in five types of discourse-level constraints derived from event semantics:
- Argument Time: Normalizing explicit time expressions (e.g., "yesterday" vs. "today") to anchor events.
- Event Relevance: Learning typical event sequences from data (e.g., an "Attack" usually precedes a "Die" event).
- Tense: Using the annotated tense of Chinese verbs to bound temporal possibilities.
- Connectives: Identifying causal or temporal markers (like "because" or "after") that link adjacent clauses.
- Coreference: Ensuring that different mentions of the same event are marked as "overlapping" in time.
Experimental Results
The DGIM model was evaluated against two strong baselines: a standard classifier (Baseline 1) and a basic global inference model using only transitivity and reflexivity (Baseline 2).
As shown in the results, DGIM reached 68.36% accuracy on gold events, significantly outperforming the classifier baseline (62.17%) and the basic GIM (64.12%). The model demonstrated particular strength in resolving relations between events in nonadjacent sentences, which comprised 56.3% of the dataset. Analysis of the constraints revealed that the Argument Time constraint provided the largest individual accuracy boost (+2.23%), followed by Event Relevance (+1.26%), as detailed in the ablation study.
Conclusion
This research demonstrates that global inference with semantic constraints significantly improves temporal relation extraction in Chinese. By moving beyond adjacent sentences and leveraging document-wide consistency, the DGIM model achieves a 6.19% accuracy boost over standard classification. Future work will explore deeper semantic roles and common-sense reasoning to improve event relevance modeling, as well as joint modeling of event and relation extraction to reduce error propagation.
Summary updated