Talk:Finite-state transducer

Regular Relations and Closure Properties
The article should mention that finite state transducers correspond to regular relations, which are a language class on their own. The section about "Operations on finite state transducers" is a also a bit misleading (though correct) in this respect. Although there is no notion of the intersection of two FSTs it should be pointed out, that this is due to the fact that the language class of regular relations is not closed under intersection (Proof sketch: the intersection of {(a^n b^m, c^n)} and {a^n b^m, c^m} is {a^n b^n, c^n}. The input language is trivially non-regular). Also, the concept of an FST can be understood to be a recognizer or generator for complex Symbols I x O. Under this conception FSTs _are_ in fact closed under intersection as they do no longer differ from FSAs. FSTs without "feasible pairs" for insertion/deletion can also be intersected. --Reas0n85 21:20, 29 November 2006 (UTC)

Similarity to Literature
Furthermore some formulations in this article are very similar to those in ''Jurafsky, Daniel, James H. Martin (2000). Speech and Language Processing. Prentice Hall, 71-83. ISBN 0-13-095069-6''. --Reas0n85 21:20, 29 November 2006 (UTC)
 * Do you mean formulas or WP:close paraphrasing of text? JMP EAX (talk) 13:52, 20 August 2014 (UTC)

Mealy machines
The concept of FSTs is similar if not equivalent to Mealy machines. Probably the articles should be merged, but as a beginning I have added it to the "see also" section. --Reas0n85 21:20, 29 November 2006 (UTC)

I disagree. Mealy machines are a particular kind of transducer, contrasted with Moore machines. For some applications (e.g. speech recognition) the Mealy interpretation is most common, but for others (e.g. bioinformatics) the Moore machine view is more common, and the word "transducer" is more commonly used than either Moore or Mealy machines. It so happens that the definitions presented here are more in line with Mealy machines, but in general "finite state transducers" are well-understood to be more general than "Mealy machines". Several bioinformatics articles refer to the definition of "finite-state transducer" given here, and I would strongly favor keeping it intact. Ian Henty Holmes (talk) 19:03, 21 March 2011 (UTC)


 * This article needs to cover interpretations of FSTs. Presently the two automata you mentioned are only linked as "see also". JMP EAX (talk) 13:56, 20 August 2014 (UTC)

Missing terms in context
--Reas0n85 21:20, 29 November 2006 (UTC)
 * feasible pair
 * insertion
 * deletion

Implementations
Does it make sense to have links to implementations? Or is that outside the scope of the Wikipedia? I have in mind xfst (documented in Beesley and Karttunen's book, with an older version of the software on an included CD), SFST (freely available), PC-KIMMO, and a number of others.

I should write up something on practical uses, the fact that there are implementations that do two-level phonology/ morphology and other implementations that have rewrite rules, and so forth... Life is too short. Mcswell (talk) 03:10, 12 May 2009 (UTC)
 * Many articles do have lists of implementations, or at least lists of notable implementations. The general policy is at External links.—greenrd (talk) 07:30, 13 May 2009 (UTC)

Wikipedia is so bad at describing these academic concepts to laypersons, even the layperson who is interested and educated in a related field. So many Wikipedia articles on academic topics read like pages torn out a 600 page textbook, written by a LaTeX-babbling automaton, to borrow a term. Is there any hope that these highly-specialized articles can mean something to anyone but the highly specialized? Ivionday (talk) 06:41, 24 February 2010 (UTC)
 * Indeed. You will also discover that rational series is just as bad. It has everything to do with certain editors who write such articles. JMP EAX (talk) 13:57, 20 August 2014 (UTC)

Operations on finite state transducers
Isn't the case to list the "local extension" of a FST under this section? "Local Extension" is defined (primarily?) on E. Roche, Y. Schabes. Deterministic Part-of-Speech Tagging with Finite-State Transducers. Computational Linguistics, Vol. 21, Number 2, 1995. --Rslemos (talk) 15:57, 21 February 2012 (UTC)

Weighted FST
I'm not really impressed with the way these are presented intermingled with the rest. They are generally much less notable. The overwhelming majority of textbooks don't cover them, only specialized books (in weighted stuff!) do. It's fine to mention them as a generalization and perhaps treat them in a separate article but I could not help the feeling that his article is/was being WP:COATRACKed. Also this page doesn't say WTF you do with the weights (besides assigning them to the edges). JMP EAX (talk) 00:35, 19 August 2014 (UTC)

I suppose the above isn't an entirely accurate assessment because this is the FST not the FSA page. FSTs themselves are rather obscure, so it may be the case that the weighted ones are predominant in applications. It's hard to tell from the present state of the article. JMP EAX (talk)


 * I agree. In fact, weighted automaton redirects here but a weighted automaton is not the same as a weighted FST (and I didn't even known weighted FST were studied, weighted automata are far more common). So what we need to do is create a new page for weighted automata, explain the concept there, and if weighted FST deserve a mention it would be as a subsection of that article. Introducing a semiring in this article seems heavy-handed and I'm not sure the notability is justified. Caleb Stanford (talk) 17:21, 9 November 2021 (UTC)