`Number` of the adjective "olik(a)"

[harisont]

In all three Swedish treebanks, the adjectiv "olik(a)" consistently has no feature Number. Is this deliberate, and if so, are there other words that should be handled similarly?

[LarsAhrenberg]

When discussing FEATS for Swedish ADJs a year ago @jnivre and I could agree on some principles. A basic principle was to base feature values on the morphology rather than context. However for adjectives that have three different forms such as stor, stort, stora the forms ending in -a which can be either definite or plural we cannot express the disjunction well in the UD format so we decided to use the feature Definite=Def and no value for Number when in a definite context and Definite=Ind|Number=Plur when in a plural context (in addition to values for Case, and Degree).

Adjectives that have no form variation such as bra or äkta should have one feature setup only, ie have values for Case and Degree but no values for Definite and Number.

For some reason, olika has been placed in the latter group while it actually belongs to the first group. In LinES there are a few instances of olik and olikt that actually have the Definite and Number features but there seem not to be any instances of those in Talbanken. This may explain why they ended up in the second group, as non-alternating forms were automatically generated (if I remember correctly). But @jnivre may have a better memory?

[jnivre]

This sounds like a very plausible explanation. It is still odd that this wasn't caught, since the lemma is "olik" also in Talbanken. It should be straightforward to fix, but it will require some manual validation, so I cannot promise that I can do it for the next release because there are other revisions to do first.

[harisont]

Thank you for the explanation, then I know what to do with new sentences!

Happy to do the validation if that helps.

[jnivre]

I have now added features to instances of the lemma "olik" in Talbanken and PUD. The overwhelming majority were Definite=Ind|Number=Plur. However, while doing this, I realized that, although the forms "olik" and "olikt" exists, they are not normally used in attributive position (cf. "av olika uppfattning"). Hence, it is probably only in predicative position that the opposition exists, and I noticed that the one occurrence of "olikt" that occurs in Talbanken was already fully specified for gender, number and definiteness. Nevertheless, since the opposition exists in the paradigm, I am inclined to keep the features now that I have added them. Or do you think I should revert the change? @LarsAhrenberg @harisont

[LarsAhrenberg]

No, don't revert the change. While I have not implemented the feature updates yet I plan to do so before the deadline.

[jnivre]

Thanks @LarsAhrenberg. Agreed.

[harisont]

Sorry for adding to this already closed thread. It's not necessarily that I disagree with the decision that has been taken, but I was trying to explain to someone else why this is how it is and found myself struggling:

A basic principle was to base feature values on the morphology rather than context. However for adjectives that have three different forms such as stor, stort, stora the forms ending in -a which can be either definite or plural we cannot express the disjunction well in the UD format so we decided to use the feature Definite=Def and no value for Number when in a definite context and Definite=Ind|Number=Plur when in a plural context (in addition to values for Case, and Degree).

Why is it ok to look at the context to determine that stora is plural cases like de stora husen, några stora hus, but not to infer that it is singular in cases like det stora huset or den stora boken?

[jnivre]

I agree that this is a bit strange, and I am not sure it's the best solution, but it's the compromise that we settled on at the time. I will try to explain the reasoning that led to it.

Our basic approach is to base features on overt morphology rather than context. In the case of the -a forms, what we would then like to say is something like "definite or plural (or both)", that is, the form is compatible with any feature combination except indefinite singular. However, the feature representation is UD is not expressive enough to say this. It allows disjunctions of values for a single feature (like "Case=Nom,Acc") but not disjunctions of feature specifications (like "Definite=Def,Number=Plur"). Therefore, we were faced with a choice of either (a) having no features at all for these forms (which would be incorrect in that it would imply that they can be used in the indefinite singular case) or (b) use context to arrive at a more specific feature specification than the disjunction we would like to have. We decided that solution (a) would lose to much information and therefore settled on (b) as the least evil alternative given the constraints, even though it is in conflict with the general approach.

@LarsAhrenberg may have more to say about it, but this is how I remember the discussion.

[harisont]

Thank you for the speedy reply, the thought process is very clear to me now!

[LarsAhrenberg]

My memory is the same as @jnivre's. Ultimately we were faced with two alternatives, the one we chose, ie Definite=Def vs Definite=Ind|Number=Plur, or Number=Plur vs. Definite=Ind|Number=Sing. I think the choice is largely arbitrary but we need to pick one for consistency. For someone familiar with traditional Swedish grammar it also means that we signal "svag böjning" with the feature Definite=Def. Note also that the choice depends on context, whichever pair we would use.