Time plays a key role in narrative. However, cor­rectly recognizing temporal order among events is a challenging task. As a follow-up to the first TempEval competition, TempEval-2 addresses this challenge. Among the three proposed tasks of TempEval-2, we address the temporal rela­tions task consisting of four subtasks: predicting temporal relations that hold between events and time expressions in the same sentence (Task C), events and the document creation time (Task D), main events in adjacent sentences (Task E), and main events and syntactically dominated events, such as those in subordinated clauses (Task F). We are primarily concerned with Task C, E, and F, because D is not relevant to our application domain. However, rather than eliminating Task D altogether, we build a very simple model for this task by using only those features that are shared with other task models (i.e., the document creation time data are not used because none of the other task models need them as features). It was expected that this approach would support more interesting comparisons with other systems that take a more sophisticated approach to the task. Further, we experiment with a joint model­ing technique to examine if the communication with other task models brings a boost to a per­formance of the simple model.

Taking a supervised machine-learning ap­proach with Markov Logic (ML) (Richardson and Domingos, 2006), we constructed two systems, NCSU-indi and NCSU-joint. NCSU-indi con­sists of four independently trained classifiers, one for each task, whereas NCSU-joint models all four tasks jointly. The choice of ML as learn­ing technique for temporal relations is motivated both theoretically and practically. Theoretically, it is a statistical relational learning framework that does not make the i.i.d. assumption for the data. This is a desirable characteristic for com­plex problems such as temporal relation classifi­cation, as well as many other natural language problems, in which the features representing a given problem are often correlated with one an­other. Practically, ML allows us to build both individual and joint models in a uniform frame­work; individual models can be easily combined together into a joint model with a set of global formulae governing over them.

In previous work (Yoshikawa et al., 2009), ML was successfully applied to temporal relation classification task. Our approach is different from this work in two primary respects. First, we introduce new lexical relation features derived from English lexical ontologies. Second, our model addresses a new task introduced in Tem-pEval-2, which is to identify temporal relations between main and syntactically dominated events in the same sentence. We also employ phrase-based syntactic features (Bethard and

Martin 2007) rather than dependency-based syn­tactic features.

We consider three types of features: basic, syn­tactic, and lexical relation features. Basic fea­tures represent the information directly available from the original data provided by the task orga­nizer; syntactic features are extracted from syn­tactic parses generated by Charniak parser (Charniak, 2000); and lexical semantic relations that are derived from two external lexical data­bases, VerbOcean (Chklovski and Pantel, 2004) and WordNet (Fellbaum, 1998).

ML is a probabilistic extension of first-order logic that allows formulae to be violated. it as­signs a weight to each formula, reflecting the strength of the constraint represented by the for­mula. A Markov logic network (MLN) is a set of weighted first-order clauses, which, together with constants, defines a Markov network. We constructed two systems, NCSU-indi and NCSU-joint using an off-the-shelf tool for ML (Riedel, 2008).

The predicted outputs from our systems exhibit mixed results. NCSU-indi achieves the highest precision score on the test data for Task F by a relatively large margin (6%) from the second-place system, as well as the second highest preci­sion score on Task C, tied with three other sys­tems. Given the encouraging result for Task F, we would preliminarily conclude that the Ve-bOcean relations are effective predictors of temporal relations between main and syntacti­cally dominated events. However, the same sys­tem does not achieve the same level of accuracy for Task E, even though it is closely related to Task F. The major difference between the mod­els of Task E and F is that the Task E model uses weighted VerbOcean relations along with a WordNet feature, while the Task F model uses unweighted VerbOcean relations without the WordNet feature. We suspect these two features might negatively impact the classification deci­sions on the test data, even though they prelimi­narily appeared to be effective predictors on the development data.

NCSU-joint also yields mixed results. The performance on both Task D and F dramatically drops with the joint modeling approach, while there is a modest improvement on Task E. Man­ual examination of the results on the test data revealed that the majority of the relations in Task D and F were classified as OVERLAP, which may be due to overly strict global constraints; rather than violating global constraints, the system re­sorted to rather neutral predictions.

Temporal event order recognition is a challeng­ing task. Using basic, syntactic, and lexical rela­tion features, we built two systems with ML: NCSU-indi models each subtask independently, and NCSU-joint models all four tasks jointly. NCSU-indi was most effective in predicting temporal relations between main events and syn­tactically dominated events (66% precision), as well as temporal relations between time expres­sions and events (63% precision). Future direc­tions include conducting a more rigorous exami­nation of the predictive power of the features, as well as the impact of global formulae for the joint model.

This research was supported by the National Sci­ence Foundation under Grant IIS-0757535. Any opinions, findings, and conclusions or recom­mendations expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation.