TIMEN An Open Temporal Expression Normalisation Resource H.Llorens, L.Derczynski, R.Gaizauskas, E. Saquete
Jan 26, 2015
TIMENAn Open Temporal Expression
Normalisation Resource
H.Llorens, L.Derczynski, R.Gaizauskas, E. Saquete
Outline ● Introduction: Timex normalisation● Related work● Problem: reinventing the wheel once and again
● Proposal: TIMEN● Evaluation● Conclusions● Further Work
Timex NormalisationTemporal information extraction subtask.
Timex: linguistic expression of a time point or interval.
Normalisation: semantic interpretation of timexes.
Temporal Expression (TIMEX)Linguistics/Variability/RelativityJune 2012, next month, 06/2012this morning 7 a.m.3 days and 3 hoursweekly
Timex normalizationISO 8601/Invariable interpretation2012-062012-05-24T07:00PT3D3HXXXX-XX-WXX
Timex Normalisation (II)Useful for a variety of NLP applications: IR, QA, Summarization, etc. I went to the cinema yesterday. When did he go to the cinema? 2012-05-23 The main advantage of normalisation is having timexes in standard time representations (e.g., gregorian calendar).
event timex
Value: 2012-05-23
Related Work There are many approaches to timex normalisation ● Pre TempEval-2
○ TempEx (2000), GUTime (2005), Chronos (2004), TERSEO (2005), TimexTag (2005), TEA (2006), DANTE (2007)...
● TempEval-2 (2010)○ HeidelTime, TRIPS/TRIOS, TIPSem/TIPSemB...
Similarities and differences● Approaches have slightly different architectures and
show slightly different performances on tests.
● But all the approaches are rule-based and in general they use the same normalization strategies.
● & also require the same parameters to perform the task.○ DCT: document creation time (deictic) (2 days ago: 2012-05-22)○ Reference time: time talked about (anaphoric)
(2 days before: 2012-05-20)○ Tense: Resolution direction (October)
Past (2011-10), Present/Future (2012-10)
The problemReinventing the wheel once and again● Implementation of high-performance approaches is
costly and it is done all the times from the scratch.● all the approaches are similar: rule-based with similar
normalization rules and strategies.● none is meant to be reused and refined by others.
Proposal: TIMENCharacteristics:
● Open philosophy: meant to be reused and refined (even across languages)
● Not only meant for computer scientists:
○ the algorithms (source code) and normalisation rules (db of user-friendly rules with a documented syntax) are separated.
● Independent from other timex processing tasks
● Multi-platform and easy integration
TIMEN Library ArchitectureExample:timex: three days agoDCT:2012-05-24normtext: 3_day_agopattern: Num_TUnit_agoonly 1 rule matches.normalized value: 2012-05-21
Example2:timex: October 202 rules matchingdisambiguation20 probably a dayrather than a yearbecause <32
Rule base sample (English)
TIMEN integration
TIMEN community ● Open-source software:
http://code.google.com/p/timen/ ● Crowd extension of the rule set (interactive
web interface to upload and check new rules): http//timen.org
* new rules only accepted if they improve the performance on the current dataset or new examples (human reviewed). Eg: New Year's Eve
EvaluationExperiments:● Normalization accuracy of TIMEN
● Performance gain in s-o-a approaches by integrating TIMEN
Datasets:● TempEval-2 test-set
(already known for approaches, mainly common dates and duration)
● TimenEval dataset (new, unknown for appr., balanced among different timex types)
Normalisation accuracy
yesterday2012Octoberdailymorning...
TIMEN
2012-05-2320122012-10xxxx-xx-xx2011...
correctcorrectincorrectcorrectincorrect...
normalisationgold timexes
e.g. TOTAL: 100 timexes to normalise e.g. TOTAL: 90 correct normalizations
RESULT: 90/100 --> 90% ACCURACY
Normalisation accuracy ● TIMEN shows a high performance even in this first
version (only 76 rules). ● TimenEval accuracy is lower. This corpus is more
heterogeneous (times/sets) and normalization is more difficult.
TEST SET NORMALISAION ACC
TempEval-2 0.90
TimenEval 0.68
Performance gain
Approach X recognized timexes
built-innormalisationof Approach X
Originalnormalisation
Performance gain = New accuracy - Original accuracy
TIMEN Newnormalisation
Performance gain (TempEval-2) "known data" ● Replacing built-in normalization approaches of the
systems by TIMEN generally improves their performance in TE2 testset.
● Tested (current) versions of the systems may have been developed/updated being aware of this data. What does it happen with data which is new for them?
System built-in norm. TIMEN norm. Err. Redution
TIPSemB 0.83 0.89 35%
HeidelTime 0.94 0.94 0%
TERNIP 0.76 0.92 66%
Performance gain (TimenEval) "new data" ● Using new data, the built-in approaches performance
decreases in general.● TIMEN favours the normalization performance for all the
systems.
System built-in norm. TIMEN norm. Err. Redution
TIPSemB 0.57 0.67 23%
HeidelTime 0.72 0.74 7%
TERNIP 0.70 0.72 66%
Conclusions ● We presented an open tool for timex normalisation:
TIMEN. ● ADVANTAGES:
○ High performance (above recent approaches).○ Easily integrated in any timex recognition
approach.○ Can be improved by the community (open philosophy),
and avoids re-development from scratch.○ Available: http://timen.org and Google code
Further Work ● Community-based extension and refinement
of TIMEN (rulebase). ● Extensive evaluation of TIMEN in various
languages (Spanish, Chinese, Italian and Danish).
TIMEN: An Open TIMEX Normalisation Resource
THANK YOU!QUESTIONS?
http://timen.org
H.Llorens, L.Derczynski, R.Gaizauskas, E. Saquete