[Corpora-List] Parallel corpora and word alignment, WAS: American and British English spelling converter

[Merle Tenney]

Ramesh Krishnamurthy wrote:

>

> ...and there is no obvious parallel corpus of Br-Am Eng to consult...

> Do you know of one by any chance...

>

> And Mark P. Line responded:

>

>Why would it have to be a *parallel* corpus?



In a word, alignment.  The formative work in parallel corpora has come from the machine translation crowd, especially the statistical machine researchers.  The primary purpose of having a parallel corpus is to align translationally equivalent documents in two languages, first at the sentence level, then at the word and phrase level, in order to establish word and phrase equivalences.  A secondary purpose, deriving from the sentence-level alignment, is to produce source and target sentence pairs to prime the pump for translation memory systems.



Like you, I have wondered why you couldn't study two text corpora of similar but not equivalent texts and compare them in their totality.  Of course you can, but is there any way in this scenario to come up with meaningful term-level comparisons, as good as you can get with parallel corpora?  I can see two ways you might proceed:



The first method largely begs the question of term equivalence.  You begin with a set of known related words and you compare their frequencies and distributions.  So if you are studying language models, you compare sheer, complete, and utter as a group.  If you are studying dialect differences, you study diaper and nappy or bonnet and hood (clothing and automotive).  If you are studying translation equivalence in English and Spanish, you study flag, banner, standard, pendant alongside bandera, estandarte, pabellón (and flag, flagstone vs. losa, lancha; flag, fail, languish, weaken vs. flaquear, debilitarse, languidecer; etc.).  The point is, you already have your comparable sets going in, and you study their usage across a broad corpus.  One problem here is that you need to have a strong word sense disambiguation component or you need to work with a word sense-tagged corpus to deal with homophonous and polysemous terms like sheer, bonnet, flat, and flag, so you still have some hard work left even if you start with the related word groups.



The second method does not begin, a priori, with sets of related words.  In fact, generating synonyms, dialectal variants, and translation equivalents is one of its more interesting challenges.  Detailed lexical, collocational, and syntactic characterizations is another.  Again, this is much easier to do if you are working with parallel corpora.  If you are dealing with large, nonparallel texts, this is a real challenge.  Other than inflected and lemmatized word forms, there are a few more hooks that can be applied, including POS tagging and WSD.  Even if both of these technologies perform well, however, that is still not enough to get you to the quality of data that you get with parallel corpora.



Mark, if you can figure out a way to combine the quality and quantity of data from a very large corpus with the alignment and equivalence power of a parallel corpus without actually having a parallel corpus, I will personally nominate you for the Nobel Prize in Corpus Linguistics.  J



Merle



PS and Shameless Microsoft Plug:  In the last paragraph, I accidentally typed "figure out a why to combine" and I got the blue squiggle from Word 2007, which was released to manufacturing on Monday of this week.  It suggested way, and of course I took the suggestion.  I am amazed at the number of mistakes that the contextual speller has caught in my writing since I started using it.  I recommend the new version of Word and Office for this feature alone.  J
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://listserv.linguistlist.org/pipermail/corpora/attachments/20061109/583663d4/attachment.htm>