Converting Jiayan's POS into UPOS of Universal Dependencies about jiayan HOT 3 CLOSED

Yes，z could be interpreted as ADV here. By the way, thank you for introducing the universal dependency project, which is very interesting, and an important missing piece of Jiayan. I will look into it and hopefully could make Jiayan support UD parsing in the future.

from jiayan.

Hi, yes you are right, PART should be u in this case. Let me answer your questions one by one:

1. Jiayan takes the same pos tag set as LTP, which is a popular NLP toolkit for modern Chinese, whose POS tag category is here. This POS tag set is defined in an important modern Chinese information processing project back in 2003 (let's call it 863 tagset). I tried to find the official page for you but unfortuantely it is obsolete now, so I haven't found any detailed documentations for the tag set for now.

2. From my understanding, the auxiliary/u in 863 POS tagset is not auxiliary verb in English, and that might be the confusing point for you. It is more a functional particle than a VERB. So you are right, PART in UPOS is suitable for 'u' in Jiayan. In modern Chinese, the words for particle（助词） and auxiliary verb（助动词） only have one character difference, I guess the tagset maker confused with these two words at first.

3. For PRON 之 doesn't go to 'r' but 'u' instead in Jiayan, that's the propagation issue in the model. Since I didn't find any annotated Classical Chinese data when implementing this feature, it became the most difficult task in this project. The way I resolved it is to tokenize training data with the CharHMMTokenizer from Jiayan, and use LTP mentioned above to postag the tokenized data, so the model is in fact trained in a modern Chinese fashion, which for most of the words is acceptable, but for some words, like 之, is not. 之 is polysemous in Classical Chinese, which can be either PART or PRON, however, in modern Chinese, it can only be PART, so you can see all 之s are tagged as u even it should be a PRON with the model.

4. The same issue applies to 地. On the contrast of 之, 地 is a polysemous word in modern Chinese, which can either be NOUN or PART, but can only be NOUN in Classical Chinese. Therefore with the modern Chinese fashion tagging, it doesn't only go to NOUN, but also goes to PART, which is u in Jiayan. In this case, you could convert all u tagged 地 to NOUN with post processing.

In conclusion, the modern Chinese tagging fashion of the Jiayan POS model could mistag the words that are polysemous in either Classical Chinese or modern Chinese. I will look deeper into the POS tagging feature to see what can be improved in the future.

Sorry for the long answer, hope my answer could help. Thank you very much!

from jiayan.

Thank you @jiaeyan for the information about LTP. I've just written a tentative table to convert POS of LTP into UPOS of Universal Denendencies:

In addition z (descriptive words) should be converted into... ah well, into ADV?

from jiayan.