This sub-system is employed for finding tempor­al expressions in the text. It takes a sentence as input and looks for temporal expressions in it.

Pre-processing: At this step the input text un­dergoes syntactic analysis. Sentence detection, tokenization, part-of-speech tagging and parsing are applied.

Candidate selection: Since only certain lexi­cal categories can be temporal expressions and they are defined in the TIDES standard (Ferro et al., 2003), in our implementation we consider the following chunk-phrases as candidates for tem­poral expressions: nouns (week, day), proper names (Tuesday, May), noun phrases (last Tues­day), adjectives (current), adjective phrases (then current), adverbs (currently), adverbial phrases (a year ago), and numbers (2000). As input it takes the sentences with provided syntactic in­formation and marks phrases in the parse tree belonging to the above types for temporal ex­pressions.

Annotation alignment: If the system is used for training classifiers, all the candidates in a sentence are examined against the available an­notations. The candidates, whose parse and anno­tation extents aligned, are taken as positive ex­amples and the rest is considered as negative.

Feature Design: To produce a feature-vector we use most valuable features extracted for phrase-candidate. After a number of experiments the following features were selected:

• Last token in the phrase, most probable token to be a temporal trigger;

• Lemma of the last phrasal token;

• Part-of-speech of the last phrasal token;

• Character pattern of the last phrasal to­ken as introduced in (Ahn et al., 2007);

• Neighbor POS's. The concatenated part-of-speech tags of the last phrasal token and its preceding token;

• Character pattern of the entire phrase;

• Phrase surface. A concatenated string of sub-parse types for the phrase;

• A Boolean feature indicating nested complex phrasal parses, such as noun verb, adverbial, adjective or preposition­al phrase;

• Depth of the phrase. The number of the nested sub-parses to the deepest pre-terminal sub-parse.

All the features are considered as Boolean.

Classification: Once the classifiers are trained they can be used for recognition of temporal ex­pressions on test sentences. A preprocessed sen­tence is taken as input and starting from its parse-tree root the candidate-phrases are classi­fied. The most probable class will be assigned to the candidate under consideration. Once the phrase is classified as temporal expression no further classification of nested phrases is per­formed, since no embedded timexes are allowed in the corpus. After a series of experiments with different machine learning techniques on the training data the maximum entropy classifier was chosen.

Extending positive instances: Sparseness of annotated corpora is the biggest challenge for any supervised machine learning technique. To overcome this problem we hypothesize that knowledge of semantic similar words could be found by associating words that do not occur in the training set to similar words that did occur in the training set. Furthermore, we would like to learn these similarities automatically in order to be as much as possible independent of know­ledge sources that might not be available for all languages or domains. For example, there is in TimeBank a temporal expression "last summer" with the temporal trigger summer, but there is no annotation of temporal expressions built around the temporal trigger winter, and this means that no temporal expression with the trigger winter can be recognized. Something similar usually happens to any annotated corpus and we want to find a way how to find other temporal expres­sions outside the available data, which can be used for training. On the other hand, we want to avoid a naïve selection of words as, for example, from a gazetteer with temporal triggers, which may contradict with grammatical rules and the lexical context of a timex in text, e.g.:

on Tuesday said....

But grammatically wrong by naïve replacement from a gazetteer:

... on week said*... ... on day said*... ... on month said* ...

In order to find these words, which are legiti­mate at a certain position in a certain context we use the latent word language model (LWLM) (Deschacht & Moens, 2009) with a Hidden Mar­kov Model approach for estimating the latent word parameters.

Complementary, we use WordNet (Miller, 1995) as a source that can provide a most com­plete set of words similar to the given one. One should note that the use of WordNet is not straight-forward. Due to the polysemy, the word sense disambiguation (WSD) problem has to be solved. Our system uses latent words obtained by the LWLM and chooses the synset with the highest overlap between WordNet synonyms and coordinate terms, and the latent words. The over­lap value is calculated as the sum of LWLM probabilities for matching words.

Having these two sets of synonyms and after a series of preliminary tests we found the setting, at which the system produces the highest results and submitted several runs with different strate­gies:

• Baseline (no expansion) (KUL Run 1)

• 3 LWLM words with highest probabili­ties (KUL Run 2)

• 3 WordNet coordinate terms; WSD is solved by means of LWLM (KUL Run

For each available annotation in the corpus a positive instance is generated. After that, the to­ken at the most probable position for a temporal trigger is replaced by a synonym from the syn­onym set found to the available token.

Normalization of temporal expressions is a process of estimating standardized temporal val­ues and types. For example, the temporal expres­sion "summer 1990" has to be resolved to its value of 1990-SU and the type of DATE. In contrast, for the expression "lastyear" the value cannot be estimated directly, rather it gets a mod­ified value of another time expression.

Due to a large variance of expressions denot­ing the same date and vagueness in language, rule-based systems have been proven to perform better than machine-learning ones for the norma­lization task. The current implementation follows a rule-based approach and takes a pre-processed document with recognized temporal expressions (as it is described in Section 2.1) and estimates a standardized ISO-based date/time value. In the following sections we provide implementation details of the system.

Before the temporal value is estimated, we employ a classifier, which uses the same feature sets and classify the temporal expression among type classes DATE, TIME, DURATION and SET.

Labeling: Labeling text is a process of provid­ing tags to tokens of chunk-phrases from a defined set of tags. We carefully examined availa­ble annotated temporal expressions and annota­tion standards to determine categories of words participating in temporal expressions. The fol­lowing set of categories with labels based on se­mantics of temporally relevant information and simple syntax was defined: ordinal numbers (first, 30th etc.), cardinal numbers (one, two, 10 etc.), month names (Jan., January etc.), week day names (Mo., Monday etc.), season names (summer, winter etc.), parts of day (morning, afternoon etc.), temporal directions (ago, later, earlier etc.), quantifiers (several, few etc.), mod­ifiers (recent, last etc.), approximators (almost, nearly etc.), temporal co-references (time, period etc.), fixed single token timexes (tomorrow, to­day etc.), holidays (Christmas, Easter etc.) and temporal units (days, months, years etc.). Also fine-grained categories are introduced: day num­ber, month number and year number. For each category we manually construct a vocabulary, in which each entry specifies a value of a temporal field or a final date/time value, or a method with parameters to apply.

As input, the normalization takes a recognized temporal expression and its properties, such as the temporal type and the discourse type. During labeling each token in a temporal expression is tagged with one or multiple labels corresponding to the categories defined above. For each of the categories a custom detector is implemented. The detector declares the method to run and the ex­pected type of the result. The rules that imple­ment the logics for the detector are inherited from an abstract class for this specific detector, so that if a new rule needs to be implemented its realization is limited to the development of one class, all the rest the detector does automatically. Besides, the order, in which detectors have to be run, can be specified (as for example, in case of fine-grained detectors). As output, the module provides labels of the categories to the tokens in the temporal expression. If there is no entry in the vocabulary for a token, its part-of-speech tag is used as the label.

Value estimation: Value estimation is implemented in the way of aggregating the values defined for entries in the vocabulary and/or executing instructions or methods specified. Also a set of predefined resolution rules is provided and can be extended with new implementations of resolution strategies.

For resolution of complex relative temporal expressions, the value for which cannot be esti­mated directly, we need to rely on additional in­formation found at the recognition step. This in­cludes the semantic type of the timex, discourse type and contextual temporal information (speech or document creation time, or previously mentioned timexes). Let's consider the following temporal expression as an example: 10 days ago. In this example the temporal expression receives a modified value of another timex, namely the value of the document creation time. The tem­poral expression is recognized and classified as a date (SEM TYPE : DATE), which refers to another timex (DISCOURSE TYPE: DEIC­TIC). It takes the value of the referenced timex and modifies it with respect to the number (10), magnitude (days) and temporal direction (ago). Thus, the final value is calculated by subtracting a number of days for the value of the referenced timex.