13 wrzenia 2012
Task denition
Our task
To recognize semantic relations between proper names in running text.
Proper names are recognized a priori 56 categories. Relations between proper names within one sentence. Existance of a relation must be stated in the sentence it is possible to point out a pattern which indicate the relation.
Examples
1 2 3
Facebook was created by Mark Zuckerberg. Mark Zuckerberg is a creator of Facebook. Mark Zuckerberg has prole on Facebook.
13 wrzenia 2012
3 / 16
Our goal
Major goal
To create a supervised method that will learn to recognize semantic relations of given categories on the basis of positive and negative examples.
Current goal
To what extent can we automate rule creation on the basis of positive and negative examples and existing tools and resources (i.e., morphological analysis, wordnet). We wanted to create a method that will be able to identify some repeating patterns in the data indicating the existence of relations.
13 wrzenia 2012
4 / 16
13 wrzenia 2012
5 / 16
base form
hypheronym
synonym
13 wrzenia 2012
6 / 16
Data representation
#code #software #make #game #freeware #name
Title BusTycon
N John
V create
DT a
N freeware
N game
V call
N BusTycon
13 wrzenia 2012
7 / 16
Detailed relation statistics aliation alias composition creator location nationality neighbourhood origin
M. Marciczuk and M. Ptak
61 33 50 31 156 6 28 12
135 67 58 62 169 7 33 26
8 / 16
Baseline results
Set Relation aliation alias composition creator location nationality neigbourhood origin
Training R [%] 9.56 13.01 5.56 6.25 11.16 28.57 5.36 67.12
Testing R [%] 11.49 7.69 18.87 1.49 9.36 28.57 2.78 40.74
good precision but very low recall, best results for origin relation.
13 wrzenia 2012
10 / 16
Title BusTycon
N John
V create
DT a
N freeware
N game
V call
N BusTycon
13 wrzenia 2012
11 / 16
ILP conguration
a general-purpose ILP systems: FOIL, GOLEM, Progol, Aleph, the values of the parameters were experimentaly obtained on the held-out set, parameters:
breadth-rst search strategy, i=8 upper bound on layers of new variables, clauselength=8 upper bound on number of literals in clause, nodes=320000 upper bound on the nodes to be explored, minpos=2 lower bound on the number of positive examples to be covered by a clause, noise=10 upper bound on number of negative examples covered by clause.
13 wrzenia 2012
14 / 16
Sample rule
Sample rule for origin relation
relation(A,B,origin) :annotation_range(B,C,D), prev_orth(E,C,word_w), // in annotation_range(A,F,G), next_orth(G,H,meta_BRACKET_LEFT), next_orth(H,I,word_ur), // born next_orth(I,J,meta_DOT), next_pattern(J,K,PATTERN_NUM), next_pattern(K,L,PATTERN_LOWERCASE).
Evaluation
Comparision with WCCL rules on the testing set:
Set Relation aliation alias composition creator location nationality neigbourhood origin WCCL Rules P [%] R [%] F [%] 80.95 100.00 100.00 100.00 100.00 66.67 100.00 84.62 11.49 7.69 18.87 1.49 9.36 28.57 2.78 40.74 20.12 14.29 31.75 2.94 17.11 40.00 5.41 55.00 P [%] 43.79 62.86 40.63 30.77 31.30 0.00 12.00 48.28 ILP R [%] 49.63 32.84 44.83 12.90 42.60 0.00 9.38 53.85 F [%] 46.53 43.14 42.62 18.18 36.09 0.00 10.53 50.91 F +/[%] +26.41 +28.84 +10.85 +15.24 +18.98 -40.00 +5.12 -4.09
higher recall but lower precision, no rules for neighbourhood only 14 examples in the training set, better but not good as expected.
M. Marciczuk and M. Ptak 13 wrzenia 2012 16 / 16
Summary Conclusions
Conclusions
Generic ILP system can be eectively used to discover the rules in the task of semantic relation recognition. Search space control is an option to improve the performance (denition of prunning rules). Creating rules with * operator is inecient needs more attention. The weak point rules consisting of two disjoint sequences. Insucient number of data for some relation categories.
4 5
13 wrzenia 2012
17 / 16
The background knowledge was extended by dependency information of tokens to reduce word order variety. The rules were used as features for classiers. For some categories we reached the level of 5060% of F-measure (i.e., composition, origin, nationality and aliation).
2 3
Future work
1 2
Make use of shallow parsers. Force the continuity of patterns or introduce the maximum distance between the parts of the patterns. Apply bootstrapping to extend the training set.
13 wrzenia 2012 18 / 16
The end
13 wrzenia 2012
19 / 16