Hi, Firstly, you should write to the CG mailing list instead of CC'ing us all - https://groups.google.com/forum/#!forum/constraint-grammar / [email protected]<mailto:[email protected]> - I have done so with this reply.
Sorry for my ignorance. I didn’t know about this e-mail list as I am not attached to this community. Anssi, are you on that list? This seems to be a good way of discussing issues related to the project we are planning. It’s probably good if you join in case you are not part of it yet. I will probably try to avoid to join yet another mailing list but you could keep me in the cc with relevant messages. Thanks! Anyway, I have been saying this for a long time. The past decade of machine learning has simply approached the hand-written method. E.g., when Google published https://research.googleblog.com/2016/05/announcing-syntaxnet-worlds-most.html my immediate comment was: Their induced rules smell an awful lot like constraint grammar, just expressed in vector fields. There is nothing controversial about it. Humans learn languages from data and experience and then write down their rules in terms of grammars and lexica. Machines do the same and optimise parameters of their models (no matter whether they are rules, lexica or vector-based representations). However, humans are not good in expressing uncertainty and quantifying subjective interpretation. That’s why I would not go so far and claim that ML models approach hand-written methods but hopefully they are able to include the regularities that humans are capable of discovering. That’s exactly what we would like to study - in what way machine learning is capable of discovering regularities that humans identify (e.g. constraints in CG). We do not expect that they will match exactly especially because CG grammars and neural LMs are developed/trained for different tasks but we hope to see some correlations. So, any help and suggestions on how to study this question would be very much appreciated. Other papers in the field even explicitly say that their models look a lot more like classic systems, with separate source language analysis, transfer, and target language generation. And this holds for any text-to-text transformation, not just translation. So I am not the least bit surprised that more advanced models look more and more like rule-based systems. True, encoder-decoder models are exactly using that kind of architecture and the success of pre-trained neural language models shows the importance of generic language learning before applying it to other downstream tasks. This is nothing new, even n-gram LMs have been generic tools that came in handy for all kinds of downstream applications. However, even though the overall architecture might look the same, I would not claim that they look more and more like rule-based systems. Statistical machine translation is a rule-based approach with large probabilistic phrase-based or even tree-based translation rules. There are no more rules in neural MT and other neural models. So, the shift to non-rule-based systems just happened with the deep learning wave and not before. In any case, I really welcome further discussions about conceptual and internal similarities / differences as we really need this to understand what is going on in the models we develop no matter whether they are hand-written, rule based or neural. Thanks in advance for any feedback and comments in that direction. All the best, Jörg -- Tino Didriksen On Wed, 26 Feb 2020 at 14:26, Tiedemann, Jörg <[email protected]<mailto:[email protected]>> wrote: Dear CG community, I am reaching out to you because we have the idea to follow-up on Anssi Yli-Jyrä’s ideas on comparing CG to transformer models to see whether there is some commonalities between expert-made linguistic grammars and learned neural language models. This is some kind of fascinating question and we would like to carry out some empirical studies to find possible correlations and patterns. It would be great to get an update about available CG resources to get started and it would also be interesting to hear whether anyone of you would be interested to even collaborate in that study. What I had in mind was to look into the disambiguation process done on real-world data using CG-based parsers and compare that with the activations triggered in trained neural language models. It would be excellent to know whether there are some (hopefully freely available) wide-coverage grammars and parsers available that we can study. Most likely, we need to look into high-resource languages (including Finnish( to also make proper comparisons to neural models but other scenarios are possible as well. Please, let me and Anssi know whether you have any suggestions. Thanks a ot! All the best, Jörg -- You received this message because you are subscribed to the Google Groups "Constraint Grammar" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To view this discussion on the web visit https://groups.google.com/d/msgid/constraint-grammar/E715C266-CC5E-4CF0-AE57-4D812BBD62EC%40helsinki.fi.
