> Maybe the term unification is misleading here. Well, that's how they call it ;). They are extremely useful when working with grammatical formalisms which make use of them (HPSG, LFG, TAG, etc. ). If you're doing statistical NLP they are not all that useful, but they are at the core of symbolic NLP.
> But it could unify with [Agr [number sg, gender f]]? yes > My system narrow the search space. So indeed, if you want to know everything that can unify with [Agr [number sg]] you *could* query every every facts that are an Agr (instead of all the facts) and then filter manualy those that don't have [number sg] or have something else and do that recursively... In theory it's possible to do such thing I guess. Not sure why you think the search would have to be that deep and complex. To me it feels like it should be like the `appendo` example, where once you define a relational append, it can run backwards just fine. Once one defines `unifyo` relationally it should be able to do the search backwards, no? > I will come back to this subject at some point, but the lack of real-world problem to solve doesn't make it appealing. Well, it would be useful for people like me working on symbolic NLP, who want to write relational parsers. I will work on it using your representation, I was trying to do it using lists: (Agr (number sg) (person 3)), but I couldn't figure it out. El sáb., 1 dic. 2018 a las 18:19, Amirouche Boubekki (< [email protected]>) escribió: > > > Le sam. 1 déc. 2018 à 17:00, Matías Guzmán Naranjo <[email protected]> > a écrit : > >> Hi Amirouche, >> >> yes, those are the feature structures I am talking about. Your idea for >> representing them seems interesting, I am wondering whether you have a >> solution for unification? >> > > Maybe the term unification is misleading here. That's why I asked the > question initially. I also asked the question because I am interested in > NLP and feature structures seems to be thing in this field. But I failed to > find more ressources to better understand the use cases. > > Quick answer is that it doesn't do unification in terms of feature > structures. It does unification of triples with a semantic that *looks* > like SPARQL (without support for OPTIONAL). > > Say I have [Agr [number sg]] and want to unify it with [Agr [person 1]], >> it should produce: [Agre [number sg, person 1]] (where order doesn't >> matter). >> > However, it should fail if I want to unify [Agr [number sg]] and [Agr >> [number pl]]. >> > > But it could unify with [Agr [number sg, gender f]]? > > >> Ideally, it should also run in reverse: I should be able to ask what [Agr >> [number sg]] needs to unify with to produce [Agr [number sg, person 1]]. >> > > That is an interesting problem. > > >> Can your system do this? >> > > My system narrow the search space. So indeed, if you want to know > everything that can unify with [Agr [number sg]] you *could* query every > every facts that are an Agr (instead of all the facts) and then filter > manualy those that don't have [number sg] or have something else and do > that recursively... In theory it's possible to do such thing I guess. > > Like I said previously, it can not do feature-structure unification as-is. > To support that last query it requires a minikanren with constraints > support (and possibly to support bigger than RAM dataset it would require > streamable solutions). > > I will come back to this subject at some point, but the lack of real-world > problem to solve doesn't make it appealing. > > -- > You received this message because you are subscribed to the Google Groups > "minikanren" group. > To unsubscribe from this group and stop receiving emails from it, send an > email to [email protected]. > To post to this group, send email to [email protected]. > Visit this group at https://groups.google.com/group/minikanren. > For more options, visit https://groups.google.com/d/optout. > -- You received this message because you are subscribed to the Google Groups "minikanren" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To post to this group, send email to [email protected]. Visit this group at https://groups.google.com/group/minikanren. For more options, visit https://groups.google.com/d/optout.
