search

tibetan-nlp / annotation-docs

Tibetan annotation docs
1 stars 0 forks source link

Verbal compounds with adverbs? #12

Closed solmsdorf closed 7 years ago

solmsdorf commented 7 years ago

How do we deal with adverb-verb compounds, i. e. any compound with kun tu, kun nas, nges par, mngon par, mchog tu, rjes su, nye bar, rnam par, yang dag par, yongs su, rab tu, legs par, shin tu, so sor, lhag par, and the like?

Some of these are derived from verbs and marked as such with POS tags. However, they clearly function as adverbs, either altering the meaning of the main verb or—especially in canonical literature translated from Sanskrit where they represent verbal prefixes like anu-, abhi-, pra-, prati-, etc. pp.—without changing the core meaning of the verb. Preferably, we would draw advmod relations from the verb to the adverb; however, since the adverbs are not identified as such in the POS tagging, it is not possible to do so.

E. g. Mila 12b: rje btsun mthu dang ser ba slob pa'i nye bar len pa

=> Many more examples will come up in mDzangs blun

heacu commented 7 years ago

This is a case where the POS tag was chosen for etymological reasons but is misleading in terms of grammatical function. Your suggestion for how to correct this in the dependency makes sense: I've expanded the advmod relation to allow VERB targets. Since a VERB in this case won't have any argument structure, there is nothing else that you need to do.

I'll keep this issue open until I've made a note in the documentation.

samyorode commented 7 years ago

We also need to link verbs to adverbs that do not derive from verbs, e. g. so sor gyes pa; phyir log

heacu commented 7 years ago

I'm having second thoughts about this proposal. The UD guidelines state that advmod should be reserved for adverbs, and that obl should be used for noun phrases and adpositional phrases which function adverbially. In other words, adverbial function is not sufficient to make something an advmod. But if this is the case, then it seems that the proposal won't work, unless nye bar is reanalyzed as being one word with the part-of-speech ADV. So we must mark nye ba as obl.

Would there be any mileage in using a sub-typed obl:adv relation? If we could reliably distinguish between plain obl (and not obl:arg) and obl:adv, then I could see the value.

What do you think?

heacu commented 7 years ago

To elaborate, at first I just thought we were finding a way to mark a relation from VERB to VERB where advcl wasn't appropriate. But I do not like the slippery slope to so sor, where so so is clearly nominal and so shouldn't be linked as advmod.

True, we are calling nye ba in nye bar a VERB, but we could equally have called it a NOUN, and indeed it shares case markers with nominals. In fact, the only reason for tagging verbal nouns VERB rather than NOUN that we are doing a project on verbs and it made it easier to track the verbs without relying on features, especially prior to lemmatisation. (Perhaps, however, this decision should be looked at.)

This is just a long way of saying that nye bar is like so sor. Syntactically they have nominal case marking and so should be considered obliques.

heacu commented 7 years ago

Shall I search the annotated corpus and replace all instances of advmod with obl:adv where the the dependent is a verbal noun?

solmsdorf commented 7 years ago

Just to get it right:

· We will still be using advmod for adverbs which are already tagged as such. · For adverb-verb compounds with nye bar, so sor, and the like, we will be using obl:adv from now on.

Right?

If so, no objections against obl:adv; go ahead, please replace advmod with obl:adv in the respective instances.

heacu commented 7 years ago

Where possible, we will follow the UD guidelines, using custom subtypes when we wish to supply additional information.

heacu commented 7 years ago

There is one case, however, that still needs to be considered: VERB-SCONJ-VERB - for example, phul du phyin pa (Mila 014b:T2468). How should we tag this?

heacu commented 7 years ago

@nikolaisolmsdorf or @samyorode can you look at one of the VERB-VERB compounds, namely 'phral dgos (Mila 013b:T1822)? 'phral is shown as a Nonverb lemma. What do you think about this example?

samyorode commented 7 years ago

As for phul du phyin pa: This is an adjective here, which is clear from the ind cig. phul has to be understood as N, not as V, so "the magical power (mthu) has reached (phyin pa) the highest degree/perfection (phul), i.e. is excellent." I suggest to annotate phul as obl: arg. What do you think?

solmsdorf commented 7 years ago

Concerning ’phral dgos: ’phral clearly is short for ’phral du, "immediate, at once", and should be tagged as AV. The relation from dgos thus should definitely be advmod.

heacu commented 7 years ago

ok, i defer to you all including @nh36 on 'phral dgos (at first line of http://178.62.106.226/~badw/#/mila/013b). the proposal is to change the POS tag of 'phral from VERB to ADV... with my apologies the BRAT notes are still not automatically going to slack or anywhere else yet.

heacu commented 7 years ago

re phul du phyin pa (first line of http://178.62.106.226/~badw/#/mila/014b), proposal above is to change POS tag of phul from VERB to NOUN. what do you think @nh36 ? if you want it to be an argument @samyorode then you could also do so by linking to the verb as argcl, w/o changing its POS category. but obl:arg is not an option for VERB.

nh36 commented 7 years ago

I am not sure about 'phral. it depends on how we understand 'phral in 'phral du. Is 'phral in 'phral du a noun? If so, then it makes sense to treat 'phral alone as an adverb. But if 'phral is not in the lexicon as a verb, then this means that seeing 'phral du as an oblique NP is probably right so seeing 'phral in isolation as an adverb is probably right. How do we propose to translate 'phral dgos? I would usually expect to see dgos come after a main verb. Are we sure 'phral dgos doesn't mean 'need to separate'?

As for phul du phyin pa, in isolation I would understand it as 'came to give', in which case phul is definitely a verb, but if it means "the magical power (mthu) has reached (phyin pa) the highest degree/perfection (phul), i.e. is excellent.", then there is no problem changing phul into a noun (and du into a case marker). @torma @samyorode

heacu commented 7 years ago

In that case @samyorode, you can use the BRAT UI to change the POS categories of phul and du, allowing phul to be linked to the verb as obl:arg. Please also make sure to remove the normalization reference to Hill.

heacu commented 7 years ago

Just inserting this conclusion also from slack for the record: we are removing advmod and advcl from the list of relations that humans are asked to hand-annotate.

solmsdorf commented 7 years ago

@nh36 ’phral dgos may very well be rendered with "need to separate", in this context (Mila 13b), however, I would lean towards Quintman's interpretation and translation, i. e. "immediate needs":

tshos tang (r. ltang) gnyis bur tang (r. ltang) gnyis khugs ’dug pa’i bur tang (r. ltang) gnyis ’phral dgos la spyad nas

"She also obtained two bundles of dye and two packs of dried molasses, of which the two packs of dried molasses were used for our immediate needs"

heacu commented 7 years ago

@nikolaisolmsdorf in that case i take it that you are able to change the POS category of 'phral and remove the compound relation? thanks

nh36 commented 7 years ago

It looks to me like Quintman is taking ’phral dgos 'immediate-need' as a noun. Incidentally, I have the Bacot and Lhalungpa translations which I can share if you want.

heacu commented 7 years ago

@nh36 if you feel comfortable doing so, you could share those translations to the Google Drive folder AHRC Lexicography in Motion > Corpus > Life of Milarepa which I just shared with you.

solmsdorf commented 7 years ago

Ok, once the need of the compound relation is decided upon the issue can be closed, right?

@nh36 Many thanks, yes, please go ahead and upload the other translations to the shared Google Drive.

nh36 commented 7 years ago

I have uploaded the translations.

heacu commented 7 years ago

I copy in some additional examples of compound reported by @nikolaisolmsdorf in slack, both from Mila 15b:

nga la mthu gza’ gdong dmar nag zhes bya ba hūṃ shi dang phaṭ ’gyel thongs pa gcig yod pa

I possess a black magic rite called Zadong Marnak (Dark Red Faced Dza) that, when cast, kills with the syllable hūṃ and causes unconsciousness with the syllable phaṭ. (Quintman, p. 55)

khong la ser ba ’dzub khrid thongs pa gcig ’dug pa

He possesses the practice of casting hailstorms by pointing one’s finger. (ibid.)

Butt (2010:49) defines a complex predicate as:

a construction that involves two or more predicational elements (such as nouns, verbs, and adjectives) which predicate as a single unit, i.e. their arguments map onto a monoclausal syntactic structure.

In contrast auxiliaries contribute merely TAME (tense, aspect, mood, evidentiality).

I will therefore close this issue and gradually build the case in the documentation for treating all of these examples as V-V "compounds". We shall erect a difficult to climb fence around the class of auxiliaries.