A couple of days ago I visited a resources page belonging to Linas Vepstas, the maintainer of the Link Grammar package of the OpenCog project, and that took me to the WaCky website, with some nice examples of tagged corpora.

That in turn led to Giuseppe Attardi and Antonio Fuschetto’s Wikipedia Extractor, which generates plain text from a Wikipedia dump, discarding markup etc. I’ve tried this sort of thing before, and the results leave a lot to be desired, but this one worked wonderfully.

Download it, make it executable, and point it at the dump: ./WikiExtractor.py -cb 250K -o output < cywiki-20131230-pages-articles-multistream.xml The -cb switch compresses the output into files of 250K, and the -o switch stores them in the output dir.

Then combine all these into one big file: find output -name '*bz2' -exec bzip2 -d -c {} \; > text.xml (the command on the website is missing the -d switch).

This gave a very tidy file with each document in its own doc tag. About the only stuff it didn't strip was some English in includeonly tags - I'm not sure what these are for, but they may have been updates that hadn't been translated yet. So Wikipedia Extractor did exactly what it said on the tin!

Once the doc tags, the English, and blank lines were removed, the text was relatively easily split into sentences, and imported into PostgreSQL for more tidying. I spent a day on that, but I didn't like the results, so I redid it over another day, and the outcome is Kwici - a 200k-sentence, 3.9m-word corpus of Welsh.

Kwici is now added to Eurfa as a seventh citation source.

Korrect/Kywiro was a collection of English-Welsh software translations that I put together in 2004, when there was a lot of community activity going on. It fell off the web a few years ago, but I’ve now resurrected it. In the new version I’ve chopped up the longer strings into chunks of one or two sentences, and stripped out a lot of the HTML. I also spent some time fixing up character encoding issues, which is an occupation hazard of dealing with older text from various machines, not all of which will have been using UTF-8. Hopefully it should be fairly tidy now, so I’ve added it to the citation sources at Eurfa.

Going even further back, in the late 90s Bob Morris Jones and colleagues put together two corpora dealing with child language acquisition – CIG1 looked at children aged 18-30 months, and CIG2 looked at children aged 3-7 years. I’m not sure how many people know about or use these, but they’re an excellent resource that deserves to be more widely available, so I’ve imported the text into a database to provide a searchable interface to them at Kig. That site also includes the original .cha files, along with information from the original website (which is showing signs of bitrot).

Although other search parameters could be introduced (along the lines of the BangorTalk site), I’ve chosen for the moment just to split the searches between child and adult utterances, since I think that’s what most people would be interested in initially – what sort of output do the children in these corpora produce? (That in turn should probably be segmented by age, but the data to do that is now available in the downloadable versions of the database, and it can be added to the website if there is any demand).

The adult utterances, of course, can also be used as another citation source in Eurfa, so I’ve added them there too.

I’ve just added citations from the Siarad and Patagonia corpora to Eurfa. Both these corpora are GPL-licensed, and were put together by a team led by Prof. Margaret Deuchar. They contain transcriptions of actual spoken Welsh – and as with all languages, the spoken version can be a rather different beast from the formal written language – and Siarad in particular contains quite a few “codeswitches” (where the speaker uses a word from another language, in this case English, in utterances that contain the main language, in this case Welsh). In the past, people have criticised this sort of thing as a signal of poor Welsh, but in fact the evidence suggests that it is actually a marker of linguistic competence – only speakers who are equally capable in both languages tend to do this, while those who are less capable tend to stick to one language.

The corpora provide important data for looking at this and other bilingual phenomena, and among other things have been used to show that the degree to which a codeswitch becomes a loanword is related to its frequency, that “cognates” (words that are similar in both languages, eg siop/shop, or perswadio/persuade) tend to raise the likelihood of codeswitches being used for other words (especially nouns), and so on.

The Patagonia corpus tends to use codeswitching less than the Siarad one, with the speakers switching into the other language for entire utterances or passages, and of course the main codeswitching language in that corpus is Spanish rather than English.

For use in Eurfa, both corpora had to be stripped down a bit. If you look at the material on Bangortalk, you’ll see that it follows the CLAN transcription format, which includes various markers for pauses, hesitations, non-verbal interactions, and so on. To make things easier to read, it made sense to get rid of most of these, with the exception of the brackets to indicate elision, eg mae’n gets represented as mae (y)n.

Then utterances where the transcription was less than 20 characters were removed – about 45% of the original 78,000-odd utterances in Siarad, and about 53% of the original 37,500-odd in Patagonia. One of the things that is noticeable in these corpora is how short the average spoken utterance is compared to the written language.

The last thing was to remove the utterances that are entirely in a language other than Welsh. The CLAN format was changed a couple of years ago to incorporate what I consider a regression. Previously, a default language would be chosen for the transcription (usually the most frequent language in the conversation), and all words not in that language would be tagged. The new system allows a “precode” marker to be attached to utterances which are entirely in the non-default language, presumably on the grounds that this makes the transcriptions easier to read. However, the problem with this is that you can no longer tell simply from the word what language it is in: an untagged word could be Welsh if the the default language is Welsh, but it could also be English if there is a precode marker.

I chose the “easy” (aka “fast”) way to do this: deleting any items with precodes marking English or Spanish. Unfortunately, that still leaves items that were not marked with precodes because the default language of the transcription was English or Spanish! I looked at addressing this with Alberto Simões’ Lingua::Identify, but, as I expected, the results are variable: the utterances are mostly too short to give a good “fingerprint”, and this is compounded by the brackets, the lack of punctuation, and the codeswitches.

For Siarad, some 20 languages (cs, da, de, en, es, fi, fr, hr, hu, id, it, la, nl, pl, pt, ro, sl, sq, sv, tr) were given for about 31% of the utterances remaining after removing those with less than 20 characters. Most of these were in fact Welsh, and even the ones marked English were about 50% Welsh. I therefore left them all, giving a total of around 42,300 citations.

For Patagonia, the same 20 languages (cs, da, de, en, es, fi, fr, hr, hu, id, it, la, nl, pl, pt, ro, sl, sq, sv, tr) were given for about 32% of the remaining utterances. Again, most of these were in fact Welsh, but the ones marked as Spanish and Portuguese were predominantly Spanish, so I removed them, leaving a total of around 15,500 citations.

The upshot is that if in Eurfa you search for an English word, and then click to get citations from Siarad or Patagonia, there is a small chance you will get a “Welsh” equivalent of the English citation that is actually in English or Spanish!