The Language Technology Group (LTG) at Edin­burgh has been active in the field of information extraction (IE) for a number of years. Up until re­cently our main focus has been in biomedical IE (Alex et al., 2008) but we have also been pursuing projects in other domains, e.g. digitised histori­cal documents (Grover et al., 2010) and we are currently participants in the EU-funded SYNC 3 project where our role is to analyse news arti­cles and establish spatio-temporal and other re­lations between news events. As a step towards this goal, we have been extending and adapting our IE pipeline to ground spatial and temporal en­tities. We have developed the Edinburgh Geop-arser for georeferencing documents and have eval­uated our system against the SpatialML corpus, as reported in Tobin et al. (2010). We are cur­rently in the process of developing a rule-based date and time grounding component and it is this component that we used for Task A, which re­quires systems to identify the extents of tempo­ral named entities and provide their interpreta­tion. The TempEval-2 data also contains event en­tities and we have adapted the output of our in-house chunker (Grover and Tobin, 2006) to iden­tify events for Task B, which requires systems to identify event denoting words and to compute a range of attributes for them. In future work we will adapt our machine-learning-based relation extraction component (Haddow, 2008) to recognise re­lations between spatial and temporal entities and event entities along the lines of the linking tasks.

Our IE system is a modular pipeline system built around the LT-XML2 and LT-TTT2 toolsets. Documents are converted into our internal doc­ument format and are then passed through a se­quence of linguistic components which each add XML mark-up. Early stages identify paragraphs, sentences and tokens. Part-of-speech (POS) tag­ging is done using the C&C tagger (Curran and Clark, 2003a) and lemmatisation is done using morpha (Minnen et al., 2000).

We use both rule-based and machine-learning named entity recognition (NER) components, the former implemented using LT-TTT2 and the lat­ter using the C&C maximum entropy NER tagger (Curran and Clark, 2003b). We are experiment­ing to find the best combination of the two dif­ferent NER views but this is not an issue in the case of date and time entities since we have taken the decision to use the rule-based output for these. The main motivation for this decision arises from the need to ground (provide temporal values for) these entities and the rules for the grounding are most naturally implemented as an elaboration of the rules for recognition.

Our IE pipeline also uses the LT-TTT2 chun­ker to provide a very shallow syntactic analysis. Figure 1 shows an example of the results of pro­cessing at the point where the rule-based NER and chunker have both applied. As can be seen from Figure 1, a positive feature for TempEval-2 is that the verb group analysis provides in­formation about tense, aspect, voice, modality and polarity which translate relatively straightfor­wardly into the Task B attributes. The noun group analysis provides verbal stem information (e.g.

'www.ltg.ed.ac.uk/software/ltxml2 www.ltg.ed.ac.uk/software/lt-ttt2

Figure 1 : Example of NER tagger and chunker output for the sentence "The announcement must not be made today."

vstem="announce") about nominalisations.

Various attributes are computed for <timex> elements and these are used by a temporal resolu­tion component to provide a grounding for them. The final output of the IE pipeline contains entity mark-up in "standoff" format where the entities point at the word elements using ids. The date and event entities for "made" and "today" are as follows:

The date entity has been grounded with respect to the date of writing (16th April 2010). To do the grounding we calculate a day-number value for each date where the day number count starts from 1st January 1 AD. Using this unique day number we are able to calculate the date for any given day number as well as the day of the week. We use the day number to perform simple arithmetic to ground date expressions such as "last Monday", "the day after tomorrow" etc. Grounding informa­tion is spread across the attributes for day, date, month and year. A fully grounded date has a value for all of these while an underspecified date, e.g. "2009", "March 13th", "next year", etc., only has values for some of these attributes.

Our system has been developed independently of TimeML or TempEval-2 and there is therefore a gap between what our system outputs and what is contained in the TempEval-2 data. In order to run our system over the data we needed to convert it into our XML input format while preserving the tokenisation decisions from the original. Certain tokenisation mismatches required that we extend various rules to allow for alternative token bound­aries, for example, we tokenise "wasn't" as was + n' t whereas the TempEval-2 data contains was + n + ' t or occasionally wasn + ' t.

Other adaptations fall broadly into two classes: extension of our system to cover entities in TempEval-2 that we didn't previously recognise, and mapping of our output to fit TempEval-2 re­quirements.

The recognition results for both timex and event extents are shown in Table 1. For Task A (timex) we achieved a close balance between precision and recall, while for Task B (events) we erred towards recall at some cost to precision.

Table 1 : Extent Results

For timex entities our false negatives were all entities of the vaguest kind, for example, "10-hour", "currently", "third-quarter", "overnight", "the week": these are ones which the original sys­tem did not recognise and for which we added ex­tra rules, though evidently we were not thorough enough. The false positives were mostly of the kind that would usually be a date entity but which were not considered to be so in the key, for exam­ple, "1969", "Oct 25", "now", "the past", "a few days". In two cases the system mistakenly identi­fied numbers as times ("1.02", "2.41").

For event entities we had 73 false negatives. Some of these were caused by verbs being mistagged as nouns ("complies", "stretch", "suit") while others were nouns which didn't occur in the WordNet derived lexicon as events. There were 143 event false positives. Some of these are clearly wrong, for example, "destruction" in "weapons of mass destruction" while others are a consequence of the subtle distinctions that the TempEval-2 guidelines make and which our shal­low approach cannot easily mimic.

The current phase of development of the Ed­inburgh IE system is supported by the SYNC3 project (FP7-231854).