TL;DR
This paper introduces the Temporal Expression Anchorer (TEA), a system designed to understand temporal expressions in emails, a genre notoriously underspecified compared to news text. By leveraging a novel constraint-based representation called Time Calculus for Natural Language (TCNL) and a recency-based focus tracking mechanism, TEA achieves significantly improved accuracy in anchoring temporal information in emails.
What Problem Does This Paper Address?
Extracting and understanding temporal information from text is crucial for many Natural Language Processing (NLP) applications, such as scheduling or information retrieval. While previous research has focused heavily on newswire texts, temporal expressions in emails present a unique challenge. Emails are often informal, underspecified, and use temporal references in a more fluid and context-dependent manner than formal news articles. This makes it difficult for existing systems, primarily designed for newswire, to accurately anchor these temporal expressions. The paper aims to bridge this gap by developing a system specifically tailored for the email genre.
Methods
The core of the system is the Temporal Expression Anchorer (TEA), which takes English text with pre-identified temporal expressions and outputs their normalized temporal representations. TEA utilizes Time Calculus for Natural Language (TCNL), a constraint-based representation that models human calendars and temporal units. TCNL's key features include:
- Calendar-agnosticism: It can incorporate any temporal unit defined in its calendar model, allowing for flexibility.
- Constraint Graph: A model of the Gregorian calendar (or other calendars) as a constraint graph, enabling reasoning about temporal relationships.
- NL-motivated Operators: A set of operators and a type system to represent the intensional meaning of temporal expressions.
TEA also incorporates a recency-based focus tracking mechanism. This mechanism attempts to dynamically update the "temporal focus" (the current point of reference in the discourse) based on newly anchored temporal expressions, aiming to better capture the shifting context in emails. The system first identifies the speech time and initial focus, then parses temporal expressions into TCNL formulae, and finally evaluates these formulae to produce a normalized time string.
Datasets
The paper uses a subset of the CSpace email corpus, collected from MBA students at Carnegie Mellon University. Specifically, they selected 1,196 emails related to scheduling. These emails were then randomly divided into sets for baseline establishment (email1), development (email2 and email5), and testing (part of email4). The paper notes that explicit temporal expressions are significantly less frequent in this email corpus (around 9.5%) compared to newswire texts (around 25%).
Key Insights and Ideas
- Emails are a distinct genre for temporal expression: The underspecified and context-dependent nature of temporal expressions in emails requires specialized handling compared to newswire texts.
- TCNL's flexibility is key: The constraint-based nature of TCNL, particularly its ability to model calendars and handle underspecified expressions, makes it well-suited for the email domain.
- Focus tracking is critical but challenging: Accurately tracking the temporal focus in emails is a major contributor to anchoring errors, highlighting the difficulty of capturing the dynamic discourse context.
- Recency-based focus tracking is a naive but useful heuristic: While the implemented recency-based focus tracking is acknowledged as a simplistic approach, it proves to be a significant factor in improving performance over a system without focus tracking.
Key Results
The paper reports the following accuracies for their Temporal Expression Anchorer (TEA):
- email5 (dev set): 85.45% accuracy
- email4 (test set): 76.34% accuracy
These results are compared favorably to a baseline of 50% (achieved by a system without focus tracking) and are presented as a closer match to results on transcribed phone conversations than to other works on newswire genre. The paper also categorizes errors into parsing errors, human errors, and anchoring errors, with anchoring errors being the most significant.
Limitations and Future Work
The authors acknowledge several limitations and suggest future research directions:
- Focus Tracking Mechanism: The recency-based focus tracking is identified as a major source of anchoring errors. Future work could involve re-examining this mechanism, potentially treating it as a classification problem or migrating it to a separate discourse module.
- Disambiguation Procedure: The current disambiguation procedure is noted as needing improvement.
- Coverage of Temporal Expressions: The system currently does not cover all types of temporal expressions, such as recurrence expressions (e.g., "every 4 years"). Expanding this coverage is a key future goal.
- Noun-Modifying Expressions: The co-referencing tendency of noun-modifying temporal expressions could lead to better anchoring strategies.
- Handling of "Human Errors": Emails contain more "human errors" like misspellings and date confusions, which pose challenges for robust anchoring.
Summary updated