RTE6 - Ablation Tests

The following table lists the results of the ablation tests (a mandatory track since the RTE5 campaign), submitted by participants to RTE6 .

Participants are kindly invited to check if all the inserted information is correct and complete.

Ablated Component	Ablation Run	Resource impact - F1	Resource Usage Description
WordNet	BIU1_abl-1	0.9	No Word-Net. On Dev set: 39.18% (compared to 40.73% when WN is used)
CatVar	BIU1_abl-2	0.63	No CatVar. On Dev set achieved about 40.20% (compared to 40.73% when CatVar is used)
Coreference resolver	BIU1_abl-3	-0.88	No coreference resolver On Dev set 41.62% (Compared to 40.73% when Coreference resolver is used). This ablation test is an unusual ablation test, since it shows that the co-reference resolution component has a negative impact.
DIRT	Boeing1_abl-1	3.97	DIRT removed
WordNet	Boeing1_abl-2	4.42	No WordNet
Name Normalization	budapestcad2_abl-2	0.65	no name normalization was performed (e.g. George W. Bush -> Bush).
Named Entities Recognition	budapestcad2_abl-3	-1.23	no NER
WordNet	budapestcad2_abl-4	-1.11	No WordNet. (In the original run, WordNet was used to find the synonyms of words in the triplets, and additional triplets were generated from all possible combinations.)
WordNet	deb_iitb1_abl-1	8.68	Wordnet is albated in this test.No change of code required only wordnet module is removed while matching.
VerbOcean	deb_iitb1_abl-2	1.87	VerbOcean is albated in this test.No change of code required only VerbOcean module is removed while matching.
WordNet	deb_iitb2_abl-1	7.9	Wordnet is albated in this test.No change of code required only wordnet module is removed while matching.
VerbOcean	deb_iitb2_abl-2	0.94	VerbOcean is albated in this test.No change of code required only VerbOcean module is removed while matching.
WordNet	deb_iitb3_abl-1	11.43	Wordnet is albated in this test. No change of code required only wordnet module is removed while matching.
WordNet	deb_iitb3_abl-2	2.54	VerbOcean is albated in this test.No change of code required only VerbOcean module is removed while matching.
POS-Tagger	DFKI1_abl-4	4.99	No wordform/POS-tags included for the comparison.
POS-Tagger	DFKI1_abl-6	2.22	No named entity recognition for the comparison.
WordNet	DFKI1_abl-7	-0.23	No WordNet similarity for the comparison.
Coreference resolver	DFKI1_Main	-1.54	Coreference resolution used for the comparison.
WordNet	DirRelCond23_abl-1	8.43	WordNet removed. Only basic word comparison used instead of word relations.
Wikipedia	FBK_irst3_Main	-23.91	This run is produced by the system configuration for run3 and uses rules extracted from Wikipedia
Wikipedia	FBK_irst3_Main	-3.58	This run is produced by the system configuration for run3 and uses rules extracted from Wikipedia with probability above 0.7
Proximity similarity dictionary of Dekang Lin	FBK_irst3_Main	-7.79	This run is produced by the system configuration for run3 and uses rules extracted from proximity similarity dictionary of Dekang Lin
WordNet	FBK_irst3_Main	-3.21	This run is produced by the system configuration for run3 and uses rules extracted from WordNet
WordNet	FBK_irst3_Main	-2.08	This run is produced by the system configuration for run3 and uses rules extracted from WordNet with probability above 0.7
VerbOcean	FBK_irst3_Main	-4	This run is produced by the system configuration for run3 and uses rules extracted from Verbocean
Dependency similarity dictionary of Dekang Lin	FBK_irst3_Main	-13.56	This run is produced by the system configuration for run3 and uses rules extracted from dependency similarity dictionary of Dekang Lin
Dictionary of Named Entities Acronyms and Synonyms	IKOMA2_abl-3	-0.76	Remove synonym dictionaries: as acronym dictionary constructed automatically from the corpus and a synonym dictionary that contains geographical terms.
WordNet	JU_CSE_TAC1_abl-1	13.29	The Run-1 is based on the composition of lexical based RTE methods and Syntactic RTE Method. The lexical based RTE methods are: WordNet based unigram match, bigram match, longest common sub-sequence, skip-gram and stemming. Here we ablated the WordNet based unigram match only.
WordNet	JU_CSE_TAC2_abl-1	10.19	The Run-2 is based on the composition of lexical based RTE methods, Syntactic RTE Method, Chunk and Named Entity Methods. The lexical based RTE methods are: WordNet based unigram match, bigram match, longest common sub-sequence, skip-gram and stemming. Here we ablated the WordNet based unigram match only.
WordNet	JU_CSE_TAC3_abl-1	3.86	The Run-3 is based on the Support Vector Machine that uses twenty five features for lexical similarity, the output tag from a rule based syntactic two-way TE system as feature, and output from a rule based Chunk Module and Named Entity Module. The important lexical features that are used in the present system are: WordNet based unigram match, bigram match, longest common sub-sequence, skip-gram, stemming and lexical distance (17 features). Here we ablated the WordNet based unigram match only.
LingPipe co-reference	PKUTM2_abl-1	0.17	Lingpipe co-reference are removed, the experiment was based on named-entity, wordnet, verbocean
VerbOcean	PKUTM2_abl-2	1.02	Verbocean are removed, the experiment was based on named-entity, wordnet, co-reference
LingPipe Named Entities	PKUTM2_abl-3	13.84	Lingpipe named-entity are removed, the experiment was based on wordnet, co-reference, verbocean
WordNet	saicnlp1_abl-1	-0.02	Ablation run, with WordNet stubbed.
Shalmanesar Parser	Sangyan1_abl-1	-1.76	This ablation run was executed with the Shalmanesar parser ablated from our system. Our system uses the Shalmanesar Parser for Framenet frame assignment. Our system required minor tweaking of the source code in order ablate Shalmanesar Parser.

Footnotes

   Return to RTE Knowledge Resources