capitalized or non-capitalized lemmas in Spanish

[AngledLuffa]

Example:

(base) [john@localhost UD_Spanish-AnCora]$ grep "       Año     " *
es_ancora-ud-test.conllu:1      Año     año     NOUN    ncms000 Gender=Masc|Number=Sing 12      obj     12:obj  Entity=(CESSCASTA2000122317865c38--1|SpaceAfter=No
es_ancora-ud-train.conllu:10    Año     Año     PROPN   np0000a _       0       root    0:root  MWE=Año_del_Gato|MWEPOS=PROPN|ArgTem=arg2:atr
es_ancora-ud-train.conllu:17    Año     Año     PROPN   np0000a _       10      conj    10:conj MWE=Año_del_Conejo|MWEPOS=PROPN
es_ancora-ud-train.conllu:25    Año     Año     PROPN   np0000a _       10      conj    10:conj MWE=Año_del_Buitre|MWEPOS=PROPN
es_ancora-ud-train.conllu:22    Año     Año     PROPN   np00000 _       18      nmod    18:nmod MWE=Año_Internacional_del_Voluntariado_2001|MWEPOS=PROPN|ArgTem=arg1:pat
es_ancora-ud-train.conllu:15    Año     Año     PROPN   np00000 _       13      nmod    13:nmod MWE=Año_Nuevo|MWEPOS=PROPN|ArgTem=argM:tmp|Entity=(NOCOREF:Spec.other-other-1-gstype:spec
es_ancora-ud-train.conllu:23    Año     Año     PROPN   np00000 _       21      nmod    21:nmod MWE=Año_Nuevo|MWEPOS=PROPN|ArgTem=argM:tmp|Entity=(NOCOREF:Spec.other-other-1-gstype:spec
es_ancora-ud-train.conllu:64    Año     Año     PROPN   _       _       62      flat    62:flat Entity=CESSCASTP1999070188c7)|SpaceAfter=No
es_ancora-ud-train.conllu:19    Año     Año     PROPN   np00000 _       17      nmod    17:nmod MWE=Año_Nuevo|MWEPOS=PROPN|Entity=(NOCOREF:Spec.other-other-1-gstype:spec
es_ancora-ud-train.conllu:5     Año     Año     PROPN   np00000 _       2       obj     2:obj   MWE=Año_Nuevo|MWEPOS=PROPN|ArgTem=arg1:pat|Entity=(NOCOREF:Spec.other-other-1-gstype:spec
(base) [john@localhost UD_Spanish-AnCora]$ grep "       Año     " ../UD_Spanish-GSD/*
../UD_Spanish-GSD/es_gsd-ud-dev.conllu:8        Año     año     NOUN    _       _       6       nmod    _       _
../UD_Spanish-GSD/es_gsd-ud-dev.conllu:5        Año     año     NOUN    _       Gender=Masc|Number=Sing 2       nmod    _       _
../UD_Spanish-GSD/es_gsd-ud-train.conllu:2      Año     año     NOUN    _       Gender=Masc|Number=Sing 5       nmod    _       _
../UD_Spanish-GSD/es_gsd-ud-train.conllu:10     Año     año     PROPN   _       Gender=Masc|Number=Sing 7       nmod    _       SpaceAfter=No
../UD_Spanish-GSD/es_gsd-ud-train.conllu:21     Año     año     NOUN    _       Gender=Masc|Number=Sing 18      nmod    _       _
../UD_Spanish-GSD/es_gsd-ud-train.conllu:28     Año     año     NOUN    _       Gender=Masc|Number=Sing 25      nmod    _       SpaceAfter=No
../UD_Spanish-GSD/es_gsd-ud-train.conllu:23     Año     año     NOUN    _       _       21      nmod    _       _
../UD_Spanish-GSD/es_gsd-ud-train.conllu:26     Año     año     NOUN    _       _       23      nmod    _       _
../UD_Spanish-GSD/es_gsd-ud-train.conllu:12     Año     año     NOUN    _       _       10      nmod    _       _
../UD_Spanish-GSD/es_gsd-ud-train.conllu:31     Año     año     PROPN   _       Gender=Masc|Number=Sing 29      obj     _       _

[Stormur]

In my opinion, the better practice for lemamtisation should be to put everything lowercase, unless an orthographic case difference is used to represent different graphemes/phonemes, which is not the case for Spanish.

(By the way, UniDive has a group dedicated to standardising lemmatisation, but I do not know exactly about their results.)

[AngledLuffa]

Does UniDive have a Spanish section? If so, I should contact them for these treebanks rather than reinventing the wheel.

[dan-zeman]

In my opinion, the better practice

It is no secret (from various other issues over the years) that @Stormur's and my opinion on this go in the exactly opposite ways :-)

Lemma should be the canonical form that one would expect as the headword in a dictionary; as such, it is also a form that can be used somewhere in text. If for a concrete lexeme the orthography requires that this form is capitalized in text (because of the word itself, not because it is first word of sentence or part of a multiword named entity), then the lemma should be capitalized as well, otherwise it would not be a valid word form. If it occasionally disambiguates two lexemes as a side effect, it is a bonus.

AnCora has capitalized lemmas for PROPN while GSD seems to prefer lowercase. I would be in favor of standardizing the capitalized approach (i.e., Madrid, not madrid).

However, both treebanks have a serious issue in identifying PROPN in the first place. Both have apparently relied on a word being a part of a multiword named entity, which is not a factor for UPOS (it could be added as a separate layer of annotation over UD morphosyntax, but not as its part). I recall that years ago I had to implement some heuristics to even tokenize multiword expressions in AnCora properly. Guessing the annotation for the individual tokens failed in many cases and I think it is the main reason behind most UPOS errors in AnCora: If you see MWE in MISC at the word or one of the immediately preceding words, you know that the morphology of these words is not manual, unlike the rest of AnCora. For GSD the tokenization was not a problem, but still the Google annotators obviously had a non-UD definition of PROPN (observable in GSD treebanks across multiple languages).

For example, I believe that año should have always lowercase lemma and upos NOUN, not PROPN.

As for UniDive, there are no sections for individual languages. There is one section for a language family (Turkic), otherwise the working groups are focused on tasks across languages. I'm in the lemmatization task but the final writeup is not out yet.