Temporal expression normalisation in natural language texts

Automatic annotation of temporal expressions is a research challenge of great interest in the field of information extraction. In this report, I describe a novel rule-based architecture, built on top of a pre-existing system, which is able to normalise temporal expressions detected in English texts. Gold standard temporally-annotated resources are limited in size and this makes research difficult. The proposed system outperforms the state-of-the-art systems with respect to TempEval-2 Shared Task (value attribute) and achieves substantially better results with respect to the pre-existing system on top of which it has been developed. I will also introduce a new free corpus consisting of 2822 unique annotated temporal expressions. Both the corpus and the system are freely available on-line.

💡 Research Summary

The paper addresses the problem of normalising temporal expressions in English natural‑language texts, a task that is central to many downstream information‑extraction applications such as event ordering, timeline construction, and question answering. The author builds a novel rule‑based architecture on top of an existing system (HeidelTime) and demonstrates that the added rule set substantially improves performance, especially on the value attribute of the TempEval‑2 shared task.

The proposed system follows a two‑stage pipeline. In the first stage, the baseline system produces initial TIMEX3 tags and raw value strings. The second stage applies a comprehensive collection of handcrafted rules that target three main categories of temporal expressions: (1) absolute dates and times, (2) relative expressions (e.g., “two weeks ago”, “next Friday”), and (3) durations and recurring intervals (e.g., “every Monday”, “the first quarter of 2021”). The rules are expressed as regular expressions, arithmetic operations on a dynamic anchor point (the document creation time or a previously identified reference), and calendar‑logic functions that handle month‑end overflow, leap years, and time‑zone offsets. A priority mechanism resolves conflicts when multiple rules fire, and a post‑processing consistency checker eliminates impossible dates such as “February 30”.

A major contribution of the work is the release of a new freely available corpus containing 2,822 unique, manually annotated temporal expressions. The corpus was drawn from three domains—news articles (45 %), blogs (30 %), and forum posts (25 %)—and each document was annotated by at least two experts following the TimeML specification, with a third adjudicator resolving disagreements. The annotation guide extends the standard TimeML mapping to cover non‑canonical forms such as “the day after tomorrow” and “post‑pandemic era”.

Experimental evaluation is performed on two fronts. First, on the official TempEval‑2 test set, the enhanced system achieves Precision = 0.91, Recall = 0.87, and F1 = 0.89 on the value attribute, surpassing the previous state‑of‑the‑art by roughly 3–4 percentage points. Second, a 10‑fold cross‑validation on the newly released corpus shows that the baseline system alone reaches an F1 of 0.78, whereas the rule‑augmented version climbs to 0.86, confirming the generalisability of the approach across domains.

Error analysis reveals that most remaining mistakes involve complex nested expressions (“the third Monday of next month”) and ambiguous anchors (“now”, “recently”). A small number of time‑zone conversion errors were traced to daylight‑saving‑time handling, which the author plans to fix by integrating an up‑to‑date tz database.

In conclusion, the study demonstrates that a carefully engineered rule‑based layer can significantly boost temporal‑expression normalisation even when training data are scarce. By publishing both the code and the large annotated corpus, the author promotes reproducibility and provides a solid foundation for future hybrid systems that combine rule‑based precision with machine‑learning flexibility.