[Paper Review] Structured Interpretation of Temporal Relations

Yuchen Zhang, Nianwen Xue|arXiv (Cornell University)|Aug 23, 2018

Natural Language Processing Techniques18 references3 citations

TL;DR

This paper proposes a structured interpretation of temporal relations by modeling events and time expressions as a dependency tree, where each node depends on a single antecedent for temporal interpretation via temporal anaphora. The approach achieves high and stable inter-annotator agreement (F1 > 0.72) across news and narrative genres, and the resulting corpus is publicly available for temporal reasoning research.

ABSTRACT

Temporal relations between events and time expressions in a document are often modeled in an unstructured manner where relations between individual pairs of time expressions and events are considered in isolation. This often results in inconsistent and incomplete annotation and computational modeling. We propose a novel annotation approach where events and time expressions in a document form a dependency tree in which each dependency relation corresponds to an instance of temporal anaphora where the antecedent is the parent and the anaphor is the child. We annotate a corpus of 235 documents using this approach in the two genres of news and narratives, with 48 documents doubly annotated. We report a stable and high inter-annotator agreement on the doubly annotated subset, validating our approach, and perform a quantitative comparison between the two genres of the entire corpus. We make this corpus publicly available.

Motivation & Objective

• To address the limitations of unstructured, pairwise temporal relation modeling, which leads to inconsistent and incomplete annotations.
• To provide a linguistically grounded structural framework for temporal interpretation by modeling dependencies between events and time expressions.
• To improve computational modeling of temporal relations by reducing combinatorial explosion and resolving transitivity conflicts.
• To validate the approach through high inter-annotator agreement on a 235-document corpus annotated in news and narrative genres.
• To release a publicly available, genre-diverse corpus for temporal relation analysis and downstream NLP tasks.

Proposed method

• Construct a dependency tree where each node (event or time expression) has exactly one parent, representing its reference time for temporal interpretation.
• Define each dependency as a temporal anaphora instance, with the parent as the antecedent and the child as the anaphor that depends on it.
• Use a single antecedent per anaphor to ensure tree structure and avoid ambiguity in temporal grounding.
• Annotate 235 documents (48 doubly annotated) in news and narrative genres, with explicit labeling of reference time and edge relations.
• Apply the transitive property to infer additional temporal relations from the dependency structure, avoiding pairwise conflict.
• Use F1 scores to evaluate inter-annotator agreement on recognition, classification, and relation labeling across genres.

Experimental results

Research questions

• RQ1 Can a dependency tree structure based on temporal anaphora improve consistency and completeness in temporal relation annotation?
• RQ2 How does inter-annotator agreement vary across news and narrative genres when using the proposed structured approach?
• RQ3 What are the differences in temporal structure between news and narrative texts in terms of event and time expression distribution?
• RQ4 How do the distributions of temporal relation types (e.g., before, overlap, depend-on) differ between news and narratives?
• RQ5 To what extent do statives and eventives contribute to structural differences in dependency trees across genres?

Key findings

• The proposed dependency tree approach achieved high and stable inter-annotator agreement, with F1 scores above 0.72 for labeled relations in both genres.
• News texts showed a higher proportion of time expressions (20%) and overlapping relations (54%), while narratives had more eventive events (51%) and 'before' relations (53%).
• Stative events in news were more likely to have other stative events as parents (57%), forming deep branches, whereas in narratives, 49% of statives directly depended on eventive parents, forming short branches.
• Time expression types differed significantly: narratives had a higher proportion of vague time expressions (67%) compared to news (22%).
• Event recognition and classification showed lower agreement in narratives (F1 = 0.75 and 0.72) than in news (F1 = 0.94 and 0.95), indicating greater annotation difficulty in narrative texts.
• The corpus of 235 documents, including 48 doubly annotated, is publicly available for research in temporal reasoning and story timeline construction.