Identification and extraction of explicit and im­plicit temporal information has become a very im­portant field of research within the computational linguistics area since some years ago (Allen, 1983) (Allen, 1984). Moreover, a large number of NLP applications are exploiting this extracted informa­tion, such as question answering and summariza­tion systems, allowing these applications to per­form in a more complex level.

When dealing with temporal information iden­tification and normalization, different approaches can be taken, depending on the available resources of the target language and the requirements of the system being developed. The most extended ap­proaches to the problem are: a) rule-based ap­proaches, such as Chronos (ITC-irst): recognizes and normalizes temporal expressions in English and Italian (Negri, 2007); TERSEO (University of Alicante, the system used for this work): a knowl­edge based system for Spanish that has been auto­matically extended to other languages, such as En­glish, Italian and Catalan (Saquete et al., 2006), b) machine learning approaches, such as TimexTag (University of Amsterdam): applies data-driven methods for recognition and normalization tasks (Ahn et al., 2005) (Ahn, 2006); CU-TMP (Uni­versity of Colorado): uses machine learning for automatic annotation (Bethard and Martin, 2007), and c) mixed combination of rules and ML ap­proaches, such as, TempEx (MITRE Corpora­tion): combines hand-coded patterns with ma­chine learning rules to tag documents (TempEx, 2008) (Mani and Wilson, 2000); TARSQI (Bran­deis University): currently uses GUTime (2008) for temporal expression annotation, which extends the capabilities of the TempEx tagger while gener­ating TIMEX3 annotations (Verhagen et al., 2005). However, whatever the approach, the output of these systems is a standardized annotation scheme.

The most popular annotation schemes are TIDES (Mani et al., 2001) and TimeML (Puste-jovsky et al., 2003b). TIDES program followed the efforts started in the context of the Message Understanding Conference, MUC (1998), and de­fined the TIMEX2 tag, with the goal of interpret­ing temporal expressions within a normalized rep­resentation of the times they denote, adopting the ISO 8601 standard (Technical Committee ISO/TC

Time Expression Recognition and Normalization (TERN, 2004) evaluation workshop was held, re­quiring by the participation systems to detect and normalize the temporal expressions mentioned in the source data, according to the TIDES annota­tion standard . In spite of the widespread use of this annotation scheme within NLP researchers, it is necessary to identify other types of tempo­ral information such as events or the relations be­tween events and temporal expressions. Motivated by such considerations, the TimeML annotation scheme (Pustejovsky et al., 2003a) (Pustejovsky et al., 2005) (Lee et al., 2007) was developed, speci­fying four major data structures (elements) for annotation: EVENT, TIMEX3, SIGNAL and LINK (Pustejovsky et al, 2005).

The system presented in this paper is a combi­nation of two separated systems. First of all, TERSEO system, which is a knowledge-based system for Spanish automatically extended to En­glish, performs an identification and normaliza­tion of all the temporal expressions in the text, an­notating them with TIMEX2 tags. Once the text has been annotated with TIMEX2, the T2T3 trans­ducer applies a set of translation rules to convert this TIMEX2 output to a TIMEX3 output.

In this section all the evaluation results for Task A in English are presented. Table 2 shows the results of the system using the trial corpus provided by the organization, the results of the system using the first delivered training corpus and the whole training data, and finally, the score of the system with the test corpus. Accuracy value is not given in the test results and it can not be calculated from the results data provided.

As shown in the results of the different evalu­ations, test results are very similar to training re­sults, what means that the system is performing steadily. Besides, in the test evaluation, the type attribute result is the best one obtained, being close to 100%. It would be interesting to have the corpus annotated also with TIMEX2 in order to determine which errors derive from TERSEO and which er­rors derive from the Transducer.

Our participation in Tempeval 2 competition was only in Task A, due to the fact that the system pre­sented is a extension of TERSEO system, which only performs identification and normalization of temporal expressions generating TIMEX2 annota­tion output. Events and links are out of the scope of this system currently.

However, our motivation for participating in Tempeval 2 competition was the possibility to de­termine the performance of the extension applied to TERSEO, by means of a transducer that is able to convert TIMEX2 annotation to TIMEX3, only using the information of the TIMEX2 tags as in­put. The transducer applies a set of rules, in order to transform the extent of the temporal expression according to TimeML annotation guidelines, and a set of rules to translate the TIMEX2 attributes to the attributes established by TimeML also. It is important to consider that TERSEO system is a knowledge-based system, with hand-made rules for Spanish. These rules were automatically ex­tended to other languages (English is one of them) using only automatic resources and without man­ual revision. This automatic extension is very in­teresting since it is possible to create a new knowl­edge for the system very fast and with satisfactory results.

The results of the evaluation of this combination (TERSEO + T2T3 Transducer) are 76% precision, 66% recall and 71% F1-Measure. For the case of the attributes, it obtained 98% for type and 65% for value.

This research has been partially supported by the Spanish government, projects TIN-2009-13391-C04-01 and PROMETEO/2009/119. Further­more, I want thank James Pustejovsky for being the co-author ofT2T3 Transducer.