Finite state automata and transducers are computational models that were widely used decades before in natural language processing for morphological parsing and other nlp tasks. I wonder if these computational models are still used in NLP nowadays for significant purposes. If these models are in use, can you give me some examples ?
$\begingroup$
$\endgroup$
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
4
Both are used, for example, in the GATE framework, which is still widely used. I suspect that this also applies to many other applications.
I would think that many recent academic publications are now on other approaches, as FSAs and FSTs are fairly established and mature technologies, but I've been out of academia for a while now.
-
$\begingroup$ GATE is a good example for discussion. The framework provides a processing layer including named entity recognition, co-reference resolution, etc. Surely, machine learning models could be more efficient than FSA and FSTs for these tasks, but the GATE would still use the latter computational models. One reason for doing this is that the framework was started in the mid-1990s and many human hours have spent building and maintaining it. It is difficult to change methods and technologies from a mature framework. But, I wonder if there is a reason for using these computational models nowadays. $\endgroup$ Oct 27, 2020 at 10:30
-
$\begingroup$ @DimitrisDimitriadis If something works well, why replace it? Just because there are other frameworks available, that seem more 'trendy'. And why would ML necessarily be more efficient? $\endgroup$ Oct 27, 2020 at 10:50
-
$\begingroup$ About your first question, I think that a reason for the replacement is the lack of expertise from the new generations. If the current trend becomes an established technology then people will prefer to work with this technology rather than to work with (e.g. FSA) only for the GATE framework. About your second question, that's why I posted this question :) I search for a reason why to prefer these obsolete methods (my opinion) rather than ML methods, which have a lot of benefits. $\endgroup$ Oct 27, 2020 at 11:11
-
$\begingroup$ @DimitrisDimitriadis This is not the place for an extended discussion, but I disagree: they are not obsolete. ML is mostly a black box, ie you cannot explain its outcomes or fine-tune it, unlike other methods, which are not obsolete at all. $\endgroup$ Oct 27, 2020 at 11:35