SemanticParsingPMB

Mono-lingual Semantic parsing

Preparation

1. Copy sskip.100.vectors to PMB2/gold
2. Copy train_input.txtRaw to PMB2/gold
3. Copy dev_input.txtRaw to PMB2/gold
4. Copy test_input.txtRaw to PMB2/gold

Pre-processing

cd PMB2/gold

python replaceCard.py

python replaceCard.py -src dev_input.txtRaw -trg dev_input.txt

python replaceCard.py -src test_input.txtRaw -trg test_input.txt

3. Generate the global condition tag file tag.txt from the training data

python globalRel.py

Run

cd seq2tree (or cd tree2tree or cd tree2treePos or cd seqtree2tree)

python main.py

	-s  if the model will be saved?  default=False
	-t  if only test the saved model?  default=False
	-r  if reload the model when continuing to train?  default=False
	-mp  if delete the Universal POS tags as features?  default=False
	-md  if delete the Universal Dependency tags as features?  default=False
	-mw  if delete the word embeddings as features?  default=False
	-model the model name   default='output_model/1.model'

3. use Ctrl+c to stop the training

Evaluation

The devlopment and test results after each epoch are listed in output_dev and output_tst respectively. Firstly, choose the epoch which performs best on the devlopment set. Then get the test results on the test set.

Evaluate on the development set.

cd seq2tree/output_dev (or cd tree2tree/output_dev or cd tree2treePos/output_dev or cd seqtree2tree/output_dev)

2. Transform the formation of the result Discourse Representation Structure to lines which fit to Counter, and generate the rough scores by comparing each line of the two files. Choose the epoch (file) i with the highest f-score as the i of the next step.

python convertAndRoughTest.py

	-r1 startpoint of the tested file range (epoch number)  default=1
	-r2  endpoint of the tested file range (epoch number)  default=2
	-src  the gold development file   default='dev.gold'
	-trg  transformation results of the gold development file  default='dev.test'
	-gold  the gold development file after transformation  default='dev.test'

python ../../DRS_parsing/counter.py -f1 i.test -f2 dev.test -pr -prin -ms (> i.results)

Evaluate on the test set. (Roughly the same as the evaluation on the development set. For example, cd seq2tree/output_tst)

Error analysis

jupyter notebook

2. Open analysis_dmatch.ipynb
3. Change the file name in the 3rd and 4th cell an run the first 4 cells

Multi-lingual Semantic parsing

Preparation

1. Copy wiki.multi.de.vec.txt to PMB_multi
2. Copy wiki.multi.it.vec.txt to PMB_multi
3. Copy wiki.multi.nl.vec.txt to PMB_multi
4. Copy wiki.multi.en.vec.txt to PMB_multi
5. Copy de.input to PMB_multi/PMB_de_v2/PMB/gold
6. Copy nl.input to PMB_multi/PMB_nl_v2/PMB/gold
7. Copy it.input to PMB_multi/PMB_it_v2/PMB/gold

Pre-processing

cd PMB2/gold

python replaceCard.py

python replaceCard.py -src dev_input.txtRaw -trg dev_input.txt

3. Generate the global condition tag file tag.txt from the training data

python globalRel.py

cd PMB_multi/PMB_de_v2/PMB/gold

python replace1.py

python replace2.py

Replace step 4 , 5 and 6 on nl and it.

Run

cd seq2tree_multi (or cd tree2tree_multi or cd tree2treePos_multi or cd seqtree2tree_multi)

python main.py

	-s  if the model will be saved?  default=False
	-t  if only test the saved model?  default=False
	-r  if reload the model when continuing to train?  default=False
	-mp  if delete the Universal POS tags as features?  default=False
	-md  if delete the Universal Dependency tags as features?  default=False
	-mw  if delete the cross-lingual word embeddings as features?  default=False
	-model the model name   default='output_model/1.model'

3. use Ctrl+c to stop the training

Evaluation

The devlopment and test results after each epoch are listed in output_dev and output_tst respectively. Firstly, choose the epoch which performs best on the devlopment set. Then get the test results on the test set.

Evaluate on the development set.

cd seq2tree_multi/output_dev (or cd tree2tree_multi/output_dev or cd tree2treePos_multi/output_dev or cd seqtree2tree_multi/output_dev)

2. Transform the formation of the result Discourse Representation Structure to lines which fit to Counter, and generate the rough scores by comparing each line of the two files. Choose the epoch (file) i with the highest f-score as the i of the next step.

python convertAndRoughTest.py

	-r1 startpoint of the tested file range (epoch number)  default=1
	-r2  endpoint of the tested file range (epoch number)  default=2
	-src  the gold development file   default='dev.gold'
	-trg  transformation results of the gold development file  default='dev.test'
	-gold  the gold development file after transformation  default='dev.test'

python ../../DRS_parsing/counter.py -f1 i.test -f2 dev.test -pr -prin -ms (> i.results)

Evaluate on the nl/de/it test set. (Roughly the same as the evaluation on the development set. For example, cd seq2tree_multi/output_it_tst)

Error analysis

jupyter notebook

2. Open analysis_dmatch.ipynb
3. Change the file name in the 3rd and 4th cell an run the first 4 cells

Acknowledgments

The code is based on the code in EdinburghNLP/EncDecDRSparsing.