Ling, X. and Weld, D. Temporal Information Extraction. AAAI-10

Citation

Temporal Information Extraction, by X. Ling, D.S. Weld. In Proceedings of the Twenty Fifth National Conference on Artificial Intelligence, 2010.

Online version

This paper is available online [1].

Summary

This paper addresses the problem of Temporal ordering. Unlike other works that focus on time-agnostic extraction of facts, this paper presents TIE, a pipeline system that tries to extract as many facts from text while also inducing as much temporal information as possible. Temporal relations between events and times in a sentence are identified while transitivity is enforced to bind the start and ending time points for each event. The paper also presents temporal entropy as a way to evaluate recall of temporal information extraction systems.

Brief description of the method

The objective of this paper is given a sentence, output (1) a maximal set of temporal elements that are either events (e.g. "officers were dispatched") or temporal references (e.g. "1999") and (2) the tightest set of temporal constraints as directly implied by the text.

The constraints are linear inequalities of the form $p_{1}\leq p_{2}+d\,\!$ where $p_{1}\,\!$ and $p_{2}\,\!$ denotes either the beginning ($^{\triangleleft }e\,\!$) or ending time point ($e^{\triangleright }\,\!$) of a temporal element $e\,\!$.

For example, the sentence 'Steve Jobs revealed the iPhone in 2007' might produce these constraints:

$^{\triangleleft }Year-2007\leq ^{\triangleleft }Reveal(Jobs,iPhone)\,\!$

$Reveal(Jobs,iPhone)^{\triangleright }\leq Year-2007^{\triangleright }\,\!$

The paper introduces a pipeline system called TIE with these components:

• Preprocessing: (1) parse the sentence using a syntactic parser (the Stanford parser), (2) detect semantic roles for verbs in the sentence using SRL, (3) use TARSQI to find temporal elements (events and times) in the sentence, (4) generate features for each temporal element as well as between elements.

The features include descriptional features (tense, grounded time values, etc) generated for each element by TARSQI and syntactic features (dependency features and SRL features).

• Classification: use probabilistic model trained on TimeBank Corpus combined with transitivity rules to classify each pair of points ($^{\triangleleft }e\,\!$ or $e^{\triangleright }\,\!$) of elements the point-wise relation.

Experimental Result

The experiment focuses on events about people as entities. A list of entities involving people are obtained from Wikipedia. Two evaluation approach are done: (1) entity-based and (2) list-based.

In the entity-based evaluation, a sample of 30 entities is taken from the list of entities. Gold-set of events about these entities are constructed by examining various web-based sources (Wikipedia, official home pages, news search and web search) to identify all events involving each entity in the sample. For each discovered event, the (1) time period for the event and (2) all other entities involved in the event are recorded. Precision and recall values are measured for matching DROP and Gold-set events: i.e. events that occur in the same time period and have at least one entities in common.

Precision are measured by the fraction of entities in DROP events that are participants of the gold events. Recall are measured by the fraction of entities in gold events that are found in the DROP events.

In the list-based evaluation, a sample of 35 events produced by each method (LTC, GTC, TRJ: simple, time-agnostic clustering of connected vertices in the PTG graph) is taken. For each sample event and its participating entities, a real-world event that best explains the co-occurrence of these entities for the specified time period is identified. A gold event is a subset of entities within the sample event that actually participated in the real-world event. Precision and recall is again measured like in the entity-based evaluation.

In the entity-based evaluation, TRJ has good precision and recall because it produces large, time-agnostic clusters (events) that basically involves most of the entities. Therefore, its recall is high because its large events contain all the entities in the gold events. Its precision is also high because it produces large events and some very small ones. Although the large events have low precision, the small ones have high precision; taking the average results in an overall high precision. However, the events produced by TRJ are not meaningful as basically it simply dumps most of the entities which may or may not be temporally related in a cluster.

In the list-based evaluation, LTC and GTC outperform TRJ in both precision and recall. Both LTC and GTC exhibit around 21% higher precision and 47% and 24% higher recall respectively than that of TRJ that has lower precision due to its large clusters: hence the fraction of participants that are actually involved in real-world events are low. TRJ has also lower recall because it produces fewer events.

Discussion

This paper presents an interesting take on the problem of temporal information extraction by finding dynamic events that are not predefined by any schema, rather is a result of entities that co-bursting together in a time period. However, a reader cannot help but question if the paper's assumption is valid that entities which co-burst together in time indeed have a relationship. In the case of celebrities for example, as they mention in the paper, some celebrities appear to be co-occurring a lot together in documents even though they are not involved in any relationships. Buzzy entities (entities that appear a lot in documents no matter when) may also cause an issue because they will appear to co-burst with many other unrelated entities. A better evaluation of the approach may be necessary. The fact that TRJ, which produces less than useful events, results in higher precision and recall in their entity-based evaluation also highlights this need to have a better evaluation approach and measure.

Another drawback to this paper is the lack of standardized data set for training and testing the approach. The choice of entities to list and events to sample seem arbitrary: take a random 30 entities and a random 35 entities (why 30 and 35?). The novelty of the task that this paper is addressing may cause finding standard data set difficult. In future a better and more standardized data set could be created to better train, test, and compare approaches in this area.

Related papers

Due to the novelty of the problem that this paper tries to address, there are still few works that are closely related to it. Perhaps the closest work is by Q. Zhao, P. Mitra, and B. Chen. Temporal and information flow based event detection from social text streams. In AAAI, 2007 where events are detected using keywords and a variety of signals including temporal and social constraints.

Although the paper is novel in terms of its take on temporal information extraction, the idea of using co-occurrence to identify relationships among entities are widespread in other applications. This idea of using distributional clustering has been used in Banko_2007_Open_Information_Extraction_from_the_Web to mine co-occurring entities from large collection of documents while being agnostic to the real world relations causing the co-occurrence. In this paper, instead of document co-occurrence, temporal co-occurrence is used to mine related entities.

Other related work that uses distributional clustering to discover structure in documents is Chambers, N. and Jurafsky, D. Template-based information extraction without the templates, ACL 2011. Although the relationship is not immediately clear (due to the non-temporal nature of the paper), this paper uses very similar ideas of co-occurrence and clustering to discover event templates in documents. It may be interesting to draw the similarities and differences between the two papers.