langsci / 25

Grammatical theory files
10 stars 2 forks source link

Discuss CCG-TAG aquivalence more carefully #183

Open stefan11 opened 1 year ago

stefan11 commented 1 year ago

Rui Cahves: 07.05.2022 The notion of `limited cross-serial dependencies', as Joshi himself noted, lacks a precise formulation. More recent definitions of MCS, like that of Genkin et al. 2010 and Kallmeyer (2010) are more explicit, but they are not equivalent definitions. There isn't a broadly accepted definition of MC.

Besides, the TAG--CCG equivalence proof is no longer as relevant as it once was. As Kuhlmann et al. (2018, 449) note, modern CCG includes combinatory rules (the substitution and type-raising rules) that were absent from the original version of CCG that was proved to be weakly equivalent to TAG. These rules are not supported by any known polynomial-time parsing algorithms. Modern CCG also does not allow the assignment of lexical rules to the empty string, and it is not known if the weak equivalence proof can be re-written without them. A similar question arises in connection with more modern versions of TAG.

TAG seems to wrongly predict islands:

Rui Chaves: ps - For examples of islands that more modern TAGs allegedly predict see the attached work.

https://onlinelibrary.wiley.com/doi/pdfdirect/10.1207/s15516709cog2805_3 https://direct.mit.edu/books/book/2521/Phrase-Structure-Composition-and-Syntactic (wh-questions, basically. Which are among the weakest islands imaginable)