Package info
The version S7b (previous S7)
- each beam gets expanded by its own feature extractor
The version S7 (previous S6)
- added new feature to the pos tagger for suffixes of previous and following words
- speeded up the completion model by checking if we need the done array in the State
The version S6 is derived from S5
- revised the feature selection of the completion model entirely.
The version S4 is derived from S3Default
- fixed errors in the feature selection such as s1P_s1Pp1_s0P_s0Pp1_p1 , s1P_s0Pp2_s0P_p1
The version S3Default is derived from S2
- experiments with the beam such as different beam sizes during training
The version S1 is derived from R9m9u
The R9m8 derived from R9m4f
- Made it specific for Finnish to select to lemmata and to use xptags instead for xgtags in the training.
The R9m8 derived from R9m4f
- Added the possibility to use predicted pos tags and morphology
The R9m4f derived from R9m4
- Included a lexicon into the morphology.
The R9m4f derived from R9m4
- Just corrected the normailization in the parameters divided the pred now by the number of pred steps and the gold by the number of gold steps.
The R9m3 derived from R9m2
- Added some feature selection capabilities.
- Added one morphological feature analog to the pos features.
The R8m6 derived from R9m7
- Included the possibility to do the cross product of morphological features.
The R8m5 derived from R9m4
- cleaned up the code of the tagger and morphological tagger
- moved the model of the taggers into model that is stored.
The R8m6 derived from R9m5
- Improved just the memory usage of the parser while training as this had lead to much lower training times. I drop now all taggers that are not needed.
The R8m5 derived from R9m4
- cleaned up the code of the tagger and morphological tagger
- moved the model of the taggers into model that is stored.
The R8m4 derived from R9m3
- Answer: well its pos 97.51 and 97.72 and 89.5 vs 89.8 for the joint model. So the joint model pos+morph works better. Unclear is now why did the pos tags improve!
- Question: It remains unclear if for pos tagging while training the initializing with previous pos tags helps or with gold only is better.
The R8m1 derived from R6m9j
- Initialize the morphological tagger with the good morphology for the first round
- Improved the morphological tagger again one step and moved in the repetition what should move the results from 88.2 to 88.6
The R6m9j derived from R6m9h
- Found now in the morphological tagger that what makes it most successful -> inject the previous morphological tags into the tagger.
- remove the fix mapping of the morphology form2morph since its seems that it has a negative impact on the accuracy of the parser.
- Changed the range of e.g. d1.a1 to the size of feats. results for 4k on the cross validated set 78.93 (possibly no improvement)
The R6m9h derived from R6m9g
- Changed training of the tagger to include the first word: results for 4k moved 78.61% to 78.9.
The R6m9g derived from R6m9f
- replaced for the morphology the feature extraction with that of the morphological tagger. Results for 4k are 78.61%
- Question: why starts training with 1 and classification with 0 ?
- removed some dead code in the method Tagger2.tag.
The R6m9f derived from R6m9e
- included the direct morphological mapping from words to morphological features. Note that the lower case words are now included in the word list what might have an effect on parsing accuracy. This seems not to improve the accuracy but the speed for the morphological tagger.
- included the pos tag patterns into the morphological features prediction that improve the accuracy from 81.5 (81.9 on the test set) to 83.0 (for 4k it resulted in 78.76% on the jackknifed part)
- for 4k, LAS is 80.95 and lower as the 81.5, I got for R6m9c
R6m9e
- Accuracy : 88.72/90.87; pos 97.54 morp 83.33
The R6m9b derived from R6m9
- changed decoder use the n-best in a beam and the next best where the pos tags are different or where morphology is best
The R6m9 derived from R6m8
- took the morphological features in combination with b0-morphology out
- combined interaction features of R6m8 with info about the selection process of the morphology of s0 and s1
The R6m8 derived from R6m7
- included four new features that model the interaction between s1 and s0 and left-most s0 and s1
- fixed another a bug with s1==-1 .. s0F=s_end => s0==-1 .. s_end
The R6m6 derived from R6m5
- fixed a bug with s1==-1 .. S0=s_end => s0==-1 .. s_end
The R6m5 derived from R6m3
- re-included the morphology as linear features and not as cross product
- the result is promising: baseline (using 1st-best morphological tag and 0.0 threshold ) 88.5 LAS / 97.64 T-ACC / 81.75 M-ACC
- using the 2 best morphological tags and 0.1 threshold : 88.6 LAS / 97.51 T-ACC / 82.8 M-ACC
- this seem to mean that the pos tags and morphological tags compete
The R6m3 (derived from R6m1)
- I have been running an experiment with mthreshold 0.1 and cross product: 88.7? / 97.7 / 83.2 ???
- sadly the result dropped to 88.29 LAS / 97.5 ACC / 83.56 ACC of Morphology => therefore, I added the cross product of morphological features
- does the accuracy improve compared to the base-line on the full set ?
- removed now from ExtractorPet.static the cross product of the features
- included the morphology into the selection process
- which seems to work this time: both improves pos tags and morphology
- removed gold features and used predicted ones
The R6m1 (derived from R6m)
- checked with gold morphologic features, pleae remove those later and replace those with the predicted ondes !!!
- build a morphologic tagger
- found in the tagger m7 the use of pos stacking which is problematic as a different tagger is used to adapt
The R6m (derived from R6j)
- Changed to FFEATS and checked it
- Designed some new morphologic features
The R6j (derived from R6f)
- changed again the setting of s_rel, s_rel1 and s_pos (see Extractor)
- cleaned up ExtractorR.encodeCat method
- changed the feature calculation from c to d methods in class ExtractorR, e.g. c4 -> d4
The R6f0 (derived from R6f)
- Feature extraction for sets of states in ExtractorR - found the bug of version R6h.
- In a final test, we reached on the penn tree (penn-to-malt conversion) LAS/UAS/Tag: 92.31/93.35/97.23 .
- settings: beam: 40; threshold=0.2; pos tag considered 2; hsize 500 000 001; tagger-hsize 90000001 ; -tt 25 ; -tx 2; -ti 10; -tnumber 10
- measured a speed improvement of about 20% compared to version R6f.
The R6f (derived from R6b)
- Corrected or cleaned up cases in completion model
- Changed else value in tagger setting in static feature model of the transition-based parser from else v+=2 -> +0
The R6c (derived from R6b)
- I moved the label creation outside of the main loop so that it is done in a uniform way.
The R6b (derived from R6a)
- In this version, I measured the time consumption of different components and carried out improvements on the feature extractions:
- transition-based features
The R6a version was derived from ysp7b
- I improved the speed of the version. I made the class state more slim
- I fixed two bugs:
(1) One in the tagger. Several tags of the same pos-tags have been provided in the n-best list of the tags. I wonder still of this case that it did not cause more harm to the scores.
(2) The swap operation causes a second shift. Because of this, we get a second time the same state and same pos tag combination with the same score. Now, we avoid to get the same variant again.
In order to solve this two cases, we did not add equal case when the score was as high as the top ranked. I guess we lost a couple of good results.
adapt tagger for usage in shared task to parse web text
- use predicted pos tags in the completion model
- use model 7 and retag
fixed bug in Tagger class moved best tag inside loop of the static tagger
Compare between only complete scoring and incomplete scoring
This version does an incomplete scoring.