This project uses mainly regular expressions with some additional tweaking functions to extract temporal information (dates, durations, time, etc) from, mainly, news articles or any other text.

Everything here should be viewed as a proof on concept. To code is not pretty and it is only targeted against the SemEval-2010 conference competition and not to be used in a real product.

The code was evaluated against SemEval-2010 with the following results against the training set:

true positives:   1053
true negatives:   51246
false positives:  87
false negatives:  1064

attribute type: +546.0 -17.0
attribute value: +494.0 -69.0

precision   0.92
recall      0.50
f1-measure  0.65
accuracy    0.98

attribute type       0.97
attribute value      0.88

Due to time constraints, I decided to quit when reaching 50% - but I believe that this could be greatly improved if I had more time to work on it. This is only a proof of concept but my belief is that this could work very good given enough time and dedication.

After reaching 50% on the training set and deciding that I was done with it, I ran it against the test set with the following results:

true positives:   165
true negatives:   9336
false positives:  8
false negatives:  104

attribute type: +94.0 -4.0
attribute value: +86.0 -12.0

precision   0.95
recall      0.61
f1-measure  0.75
accuracy    0.99

attribute type       0.96
attribute value      0.88

A demo exists written in Play Framework which you can try here:

http://tempex.antonfagerberg.com

And also browse the source:

https://github.com/AntonFagerberg/Temporal-Information-Extraction-Demo