News from the MG frontier

🕑 3 min • 👤 Aniello De Santo • 📆 June 24, 2019 in Discussions • 🏷 MGs, parsing, NLP

True to my academic lineage, I’m a big fan of Minimalist grammars (MGs): they are a pretty malleable formalism, their core mechanisms are very easy to grasp on an intuitive level, and they are close enough to current minimalist syntax to allow for interesting computational insights into mainstream syntax. However, I often find that MGs’ charms don’t work that well on my more NLP-oriented colleagues — especially when compared to some very close cousins like TAGs or CCGs. There are very practical reasons for this, of course, but two in particular come to mind right away: the lack of any large MG corpus (and/or automatic ways to generate such corpora) and, relatedly, the lack of efficient, state-of-the-art, probabilistic parsers.

This is why I’m very excited about this upcoming paper by John Torr and co-authors (henceforth TSSC), on a (the first ever?) wide-coverage MG parser. The parser is implemented by smartly adapting the \(A^*\) search strategy developed by Lewis and Steedman (2014) for CCGs to MGs (basically, a CKY chart + a priority queue), and coupling it with a complex neural network supertagger trained on an MG treebank.

Truth to be told, John Torr has been working towards this result for a while, by building his wide coverage MG treebank (MGbank), and had already provided some interesting parsing results. However, it is the \(A^*\) algorithm that does the trick here, hopefully putting this parser (and thus MGs) on the NLP map.

The paper does, imho, a very good job at clarifying the importance of the result. So I’ll spare you a super technical summary post (just go read it!). Just a few thoughts:

  1. How does the parser perform? The worst case time complexity is \(O(n^{28} \text{ } log \text{ } n)\), in the length of the sentence. “What?! You told us this was an interesting result! We already knew MG parsing was slow”. As people have mentioned on twitter: Yes, that is insanely high (CKY is \(O(n^3)\), and recent RNN-based constituency parsers are linear). However, the cool thing about the \(A^*\) component is that the parser performs way better (\(O(n^3)\)) in practice. I personally find the high variance still a bit concerning, but it is definitely an encouraging result!

    Note also that Milos Stanojevic (who has been producing so. many. parsers!) has recently shown that the worst case time complexity for parsing MGs with head movement can be dropped to \(O(n^{13})\) (so I’m assuming \(O(n^{13} \text{ } log \text{ } n)\) for their CKY+\(A^*\) implementation). My guess is that this might also reduce the expected time complexity, but of course one would have to try and see.

  2. Although cross-formalism comparisons are tough (as the authors mention: are we comparing the quality of the parser or the quality of the grammar?), having lifted the CKY+\(A^*\) strategy directly from CCGs allows this paper to do a few interesting comparisons (cf. Rimell, Clark, and Steedman (2009)). In terms of accuracy, results are not exactly comparable to those of the CCG algorithm. However, this might be due to the MG parser requiring more training than the CCG one (due to the increased complexity of the supertags?). Thus, the authors argue that performance might improve in parallel with the MG treebank. I find this pretty convincing, although in terms of practical applications it opens questions about training/performance trade-offs.

  3. There are a lot of interesting syntactic considerations one could make based on the test cases the authors report. I’ll just point out that the MGbank makes very specific syntactic choices. While it’s possibly a bit too early for this, it would be interesting to look into how different theoretical choices could impact the quality of the grammar with respect to parsing performance.

  4. A side note. MG researchers have long known that Move is a powerful operation, which needs to be restricted quite a bit. The MGbank does this by following the usual suggestions in the literature (SMC, SpIC, etc). SMC (or equivalents) aside, the appeal to syntactically motivated constraints is interesting (although not new), and re-opens questions about the relation between syntactic constraints and parsing performance even on the more NLP/engineering side.

References

Lewis, Mike, and Mark Steedman. 2014. A* ccg parsing with a supertag-factored model. Proceedings of the 2014 conference on empirical methods in natural language processing (emnlp).

Rimell, Laura, Stephen Clark, and Mark Steedman. 2009. Unbounded dependency recovery for parser evaluation. Proceedings of the 2009 conference on empirical methods in natural language processing: Volume 2-volume 2. Association for Computational Linguistics.