TL;DR
This paper surveys the last three decades of research on temporal reasoning for temporal information extraction (TIE) from text. It highlights the critical role of temporal reasoning in building coherent timelines of events and identifies key challenges and promising future directions, emphasizing the need for more expressive and efficient TR methods.
What Problem Does This Paper Address?
Humans effortlessly understand and communicate about time, weaving temporal cues into everyday language. However, automatically extracting and ordering events from text – Temporal Information Extraction (TIE) – remains a significant challenge for Natural Language Understanding (NLU). The core difficulty lies in combining diverse temporal cues (like verb tenses, dates, and durations) into a single, coherent timeline. This process, known as Temporal Reasoning (TR), is crucial for TIE but has not been comprehensively surveyed, leaving a gap in understanding its evolution, methodologies, and impact on TIE systems.
Methods
This survey provides a structured overview of TR in TIE, categorizing approaches based on their integration into the TIE pipeline:
- Dataset Annotation (TR-DA): TR can aid annotators by inferring relations, reducing effort, and ensuring consistency.
- Data Preprocessing (TR-DP): Transitive closure is used to expand sparsely annotated datasets, inferring new temporal links.
- Training or Prediction (TR-TP): TR is integrated to ensure temporal consistency among predicted relations and constrain the output space, improving accuracy. Methods include greedy inference (Best-First, Natural Reading Order), post-hoc conflict resolution, sieve-level inference, stacked inference, random restart hill climbing, and integer linear programming.
- Model Evaluation (TR-ME): TR is used to create more robust evaluation metrics like temporal awareness, addressing issues with simple accuracy measures.
The paper also delves into the theoretical underpinnings of TR, discussing Allen's interval algebra and its relation to point algebra, highlighting the trade-offs between expressiveness and computational efficiency.
Datasets
The paper discusses various corpora used for temporal information extraction, including:
- TimeML: The most widely used scheme for annotating temporal information.
- TimeBank corpus: Annotated with TimeML.
- TimeBank-Dense corpus: Explicitly annotates relations between all event pairs within a window.
- Clinical i2b2 corpus: Used for temporal information extraction in clinical texts.
- ECB+ corpus: Contains event co-reference and temporal information annotations across documents.
No new datasets are introduced in this survey.
Key Insights and Ideas
- TR is Fundamental: Temporal reasoning is not just a component but a critical enabler for effective temporal information extraction, impacting every stage from annotation to evaluation.
- Evolution of TR: Research has progressed from focusing solely on precedence relations (EE-R) to incorporating more complex temporal cues like inclusion, equivalence, and temporal expressions (TE-R, DCT-R, TT-R), moving towards a more holistic TIE.
- Point Algebra's Promise: While Allen's interval algebra is foundational, point-algebraic approaches offer a better balance of expressiveness and computational efficiency, making them more practical for real-world TIE systems.
- Integration is Key: Integrating TR not only during prediction but also during training and annotation consistently yields better performance and more robust systems.
- The Expressivity-Efficiency Trade-off: Achieving higher expressivity in TR often comes at the cost of computational complexity. Finding the right balance is crucial for practical applications.
- Unexplored Frontiers: Combining different types of temporal information (e.g., explicit durations with implicit cues) and developing TR for cross-document TIE remain significant open research questions.
Key Results
- TR-based prediction approaches generally outperform those that do not exploit TR.
- ILP-based approaches often outperform greedy methods in temporal graph construction.
- Integrating TR during training and inference leads to further performance improvements compared to inference-only approaches.
- The temporal awareness metric, which uses TR, has become a standard for evaluating TIE systems.
- While specific quantitative results vary across different methods and tasks, the consistent theme is that TR demonstrably improves TIE system performance.
Limitations and Future Work
- Incomplete Reasoning: Current TR systems often still address sub-fragments of the TIE problem, focusing on specific temporal cue types rather than extracting and combining all cues holistically.
- Combining Diverse Information: Effectively combining explicit event positions with implicit or background duration information, and integrating TR across multiple documents, are significant challenges.
- Efficiency vs. Expressivity: Developing TR frameworks that are both highly expressive and computationally efficient remains an ongoing quest.
- Future Directions:
- Point-based Reasoning: Further exploration of point-algebraic approaches for TR is recommended due to their favorable balance of efficiency and expressiveness.
- Deep Learning: Leveraging deep learning methods for their flexibility in model construction, multi-task learning, and representation sharing is a promising avenue.
- Cross-Document TIE: More research is needed in TR for cross-document TIE, especially considering computational efficiency and event co-reference.
- Integrating Diverse Data: Developing methods to combine various data sources (e.g., TimeML, event duration annotations) and explore temporal constraint networks for quantification.
Summary updated