MRP to UCCA write codec

[danielhers]

codec/ucca.py currently allows converting UCCA XML files to the MRP format, but not the other way around. Adding the reverse conversion will allow

• Training the TUPA baseline (and any other UCCA parser) on all MRP data (Baseline model #2)
• Applying the UCCA scorer to MRP data (quality control of UCCA scorer #32)
• Round-trip conversion for evaluation and testing of the conversion

[wangxinyu0922]

Hi, when will reverse conversion (MRP to UCCA XML format) be available?

[oepen]

over time, we want all mtool converters to be bidirectional, but this is no priority. there ismots more to do that is directly needed for the 2019 shared, so bidrectional codecs may not come until after the evaluation phase.

[JZXXX]

When I converted UCCA XML files to the MRP format, some files succeeded, but some failed with reporting '' 'Linkage' object has no attribute 'start_position' ''. All of the files come from the English Wiki corpus (https://github.com/UniversalConceptualCognitiveAnnotation/UCCA_English-Wiki), with passages being split into sentences. The converted files are:
test-id-sentence.zip. I can't find the reason, thanks!

[longxudou]

@JZXXX Could you tell me how to convert UCCA XML files (English-Wiki) to the MRP format. I want to compare my model with TUPA on this dataset.
In data/sample/ucca, I found that the wsj.ids and wsj.txt are required, while the original UCCA Wiki Corpus doesn't provide those resource.

[danielhers]

@DreamerDeo as said on #45, you are not allowed to use UCCA data available on GitHub for developing your parser. This includes the public pre-trained TUPA models, as they are trained on this UCCA data. It is understandable that you want to compare your current system to a baseline. We are working on a baseline model (#2), but unfortunately it is not ready yet.

[danielhers]

You can, however, use the sample data as you wish (it is also part of the training data you can download under the LDC license). wsj.ids and wsj.txt are available in data/sample, where you can also find the sample UCCA data in MRP format.

[JZXXX]

@DreamerDeo I wrote scripts to get the .txt files.

[longxudou]

@danielhers @JZXXX Got it! Thanks a lot for your timely response.

[JZXXX]

@DreamerDeo @danielhers Did you convert UCCA XML files which were parsed from tupa to the MRP format successfully? I find that if there is one token in a sentence that can't be anchored, this sentence will not be output. And there are lots sentences can't be anchored successfully. For example, dev.sent_119: failed to anchor |but| in |okay, I dont have a review, but why in the hell would you name your business something that has the initials KKK....is there something behind the scenes at this place?| (26). I extract these sentences from the original files to get .txt files. I don't know why.
sample.zip

[danielhers]

@JZXXX we haven't tried converting TUPA outputs to MRP since anyway pre-trained TUPA models are not allowed to be used for system development. They were trained on older versions of the corpora, which use a different UCCA version.

[danielhers]

python main.py --read mrp data/sample/ucca/wsj.mrp wsj.xml --write ucca
mkdir output
cd output
csplit -zk ../wsj.xml '/^<root/' -f wsj -b '%02d.xml' {99}
cp -r ../data/sample/ucca/xml ../xml
rename 's/\.(\d)\.xml/.0$1.xml/' ../xml/*
rm -f ../xml/wsj_0002.01.xml ../xml/wsj_0003.04.xml ../xml/wsj_0007.01.xml ../xml/wsj_0012.01.xml ../xml/wsj_0012.03.xml
pip install ucca
python -m scripts.evaluate_standard . ../xml -I
>>> Average labeled F1 score: 1.000