On 07/23/2013 11:23 AM, Mathieu Stumpf wrote:
Here is what I would like to do : generating reports which give, for a given language, a list of words which are used on the web with a number evaluating its occurencies, but which are not in a given wiktionary.
How would you recommand to implemente that within the wikimedia infrastructure?
Some years back, I undertook to add entries for Swedish words in the English Wiktionary. You can follow my diary at http://en.wiktionary.org/wiki/User:LA2
Among the things I did was to extract a list of all Swedish words that already had entries. The best way was to use CatScan to list entries in categories for Swedish words. Even if there is a page called "men", this doesn't mean the Swedish word "men" has an entry, because it could be the English word "men" that is in that page.
Then I extracted all words from some known texts, e.g. novels, the Bible, government reports, and the Swedish Wikipedia, counting the number of occurrencies of each word. Case significance is a bit tricky. There should not be an entry for lower-case stockholm, so you can't just convert everything to lower case. But if a sentence begins with a capital letter, that word should not have a capitalized entry. Another tricky issue is abbreviations, which should keep the period, for example "i.e." rather than "i" and "e". But the period that ends a sentence should be removed. When splitting a text into words, I decided to keep all periods and initial capital letters, even if this leads to some false words.
When you have word frequency statistics for a text, and a list of existing entries from Wiktionary, you can compute the coverage, and I wrote a little script for this. I found that English Wiktionary already had Swedish entries covering 72% of the words in the Bible, and when I started to add entries for the most common of the missing words, I was able to increase this to 87% in just a single month (September 2010).
Many of the common words that were missing when I started were adverbs such as "thereof", "herein", which occur frequently in any text but are not very exciting to write entries about. This statistics-based approach gave me a reason to add those entries.
It is interesting to contrast a given text to a given dictionary in this way. The Swedish entries in the English Wiktionary is a different dictionary than the Swedish entries in the German or Danish Wiktionary. The kinds of words found in the Bible are different from those found in Wikipedia or in legal texts. There is not a single, universal text corpus that we can aim to cover. Google has released its ngram dataset. I'm not sure if it covers Swedish, but even if it does, it must differ from the corpus frequencies published by the Swedish Academy.
It is relatively easy to extract a list of existing entries from Wiktionary. But to prepare a given text corpus for frequency and coverage analysis needs more preparation.
The request is to create a web-based text corpus[1] from which to derive frequencies and then compare with existing wiktionaries. Not a light undertaking, but one which has been proposed and implemented previously (e.g. Connel's Gutenberg project[2])
Generically speaking, someone would need to determine the appropriate size of the corpus sample, it's temporal currency, and the method of creating and maintaining it. This isn't easy to do, and having no strictures results in unwieldy and mostly irrelevant products like Google's n-grams[3] (on the other hand, if someone can figure out how to filter n-grams usefully it would mean we don't have to build our own.)
Amgine
[1] https://en.wikipedia.org/wiki/Linguistic_corpus [2] https://en.wiktionary.org/wiki/User:Connel_MacKenzie/Gutenberg [3] http://storage.googleapis.com/books/ngrams/books/datasetsv2.html
On 26/07/13 09:18, Lars Aronsson wrote:
On 07/23/2013 11:23 AM, Mathieu Stumpf wrote:
Here is what I would like to do : generating reports which give, for a given language, a list of words which are used on the web with a number evaluating its occurencies, but which are not in a given wiktionary.
How would you recommand to implemente that within the wikimedia infrastructure?
Some years back, I undertook to add entries for Swedish words in the English Wiktionary. You can follow my diary at http://en.wiktionary.org/wiki/User:LA2
Among the things I did was to extract a list of all Swedish words that already had entries. The best way was to use CatScan to list entries in categories for Swedish words. Even if there is a page called "men", this doesn't mean the Swedish word "men" has an entry, because it could be the English word "men" that is in that page.
Then I extracted all words from some known texts, e.g. novels, the Bible, government reports, and the Swedish Wikipedia, counting the number of occurrencies of each word. Case significance is a bit tricky. There should not be an entry for lower-case stockholm, so you can't just convert everything to lower case. But if a sentence begins with a capital letter, that word should not have a capitalized entry. Another tricky issue is abbreviations, which should keep the period, for example "i.e." rather than "i" and "e". But the period that ends a sentence should be removed. When splitting a text into words, I decided to keep all periods and initial capital letters, even if this leads to some false words.
When you have word frequency statistics for a text, and a list of existing entries from Wiktionary, you can compute the coverage, and I wrote a little script for this. I found that English Wiktionary already had Swedish entries covering 72% of the words in the Bible, and when I started to add entries for the most common of the missing words, I was able to increase this to 87% in just a single month (September 2010).
Many of the common words that were missing when I started were adverbs such as "thereof", "herein", which occur frequently in any text but are not very exciting to write entries about. This statistics-based approach gave me a reason to add those entries.
It is interesting to contrast a given text to a given dictionary in this way. The Swedish entries in the English Wiktionary is a different dictionary than the Swedish entries in the German or Danish Wiktionary. The kinds of words found in the Bible are different from those found in Wikipedia or in legal texts. There is not a single, universal text corpus that we can aim to cover. Google has released its ngram dataset. I'm not sure if it covers Swedish, but even if it does, it must differ from the corpus frequencies published by the Swedish Academy.
It is relatively easy to extract a list of existing entries from Wiktionary. But to prepare a given text corpus for frequency and coverage analysis needs more preparation.
On 07/26/2013 08:26 PM, Amgine wrote:
Google's n-grams[3] (on the other hand, if someone can figure out how to filter n-grams usefully it would mean we don't have to build our own.)
Exactly. And nothing stops us from going both ways, compare the results and let the best frequency list win. If it was a good idea to arrive at the one and true list, then linguists would have done so long ago.
Since the 1960s, Gothenburg University collects word frequencies for Swedish based on newspaper text, where the text is copyrighted but the frequency lists are made openly available, http://spraakbanken.gu.se/pub/statistik/
I'm sure you can find similar resources for many other languages.
What WMF could do is to compile its own frequency lists based on Wikipedia and Wikisource, and publish them at regular intervals (annually?) along with XML dumps.
Le 2013-07-26 20:26, Amgine a écrit :
The request is to create a web-based text corpus[1] from which to derive frequencies and then compare with existing wiktionaries. Not a light undertaking, but one which has been proposed and implemented previously (e.g. Connel's Gutenberg project[2])
Generically speaking, someone would need to determine the appropriate size of the corpus sample, it's temporal currency, and the method of creating and maintaining it. This isn't easy to do, and having no strictures results in unwieldy and mostly irrelevant products like Google's n-grams[3] (on the other hand, if someone can figure out how to filter n-grams usefully it would mean we don't have to build our own.)
Actually, I think it would be interesting to have a trend history of words usage over centuries (current trend would also be interesting but probably harder to implement). Wikisource may be used in order to achieve that.
Amgine
[1] https://en.wikipedia.org/wiki/Linguistic_corpus [2] https://en.wiktionary.org/wiki/User:Connel_MacKenzie/Gutenberg [3] http://storage.googleapis.com/books/ngrams/books/datasetsv2.html
On 26/07/13 09:18, Lars Aronsson wrote:
On 07/23/2013 11:23 AM, Mathieu Stumpf wrote:
Here is what I would like to do : generating reports which give, for a given language, a list of words which are used on the web with a number evaluating its occurencies, but which are not in a given wiktionary.
How would you recommand to implemente that within the wikimedia infrastructure?
Some years back, I undertook to add entries for Swedish words in the English Wiktionary. You can follow my diary at http://en.wiktionary.org/wiki/User:LA2
Among the things I did was to extract a list of all Swedish words that already had entries. The best way was to use CatScan to list entries in categories for Swedish words. Even if there is a page called "men", this doesn't mean the Swedish word "men" has an entry, because it could be the English word "men" that is in that page.
Then I extracted all words from some known texts, e.g. novels, the Bible, government reports, and the Swedish Wikipedia, counting the number of occurrencies of each word. Case significance is a bit tricky. There should not be an entry for lower-case stockholm, so you can't just convert everything to lower case. But if a sentence begins with a capital letter, that word should not have a capitalized entry. Another tricky issue is abbreviations, which should keep the period, for example "i.e." rather than "i" and "e". But the period that ends a sentence should be removed. When splitting a text into words, I decided to keep all periods and initial capital letters, even if this leads to some false words.
When you have word frequency statistics for a text, and a list of existing entries from Wiktionary, you can compute the coverage, and I wrote a little script for this. I found that English Wiktionary already had Swedish entries covering 72% of the words in the Bible, and when I started to add entries for the most common of the missing words, I was able to increase this to 87% in just a single month (September 2010).
Many of the common words that were missing when I started were adverbs such as "thereof", "herein", which occur frequently in any text but are not very exciting to write entries about. This statistics-based approach gave me a reason to add those entries.
It is interesting to contrast a given text to a given dictionary in this way. The Swedish entries in the English Wiktionary is a different dictionary than the Swedish entries in the German or Danish Wiktionary. The kinds of words found in the Bible are different from those found in Wikipedia or in legal texts. There is not a single, universal text corpus that we can aim to cover. Google has released its ngram dataset. I'm not sure if it covers Swedish, but even if it does, it must differ from the corpus frequencies published by the Swedish Academy.
It is relatively easy to extract a list of existing entries from Wiktionary. But to prepare a given text corpus for frequency and coverage analysis needs more preparation.
Wiktionary-l mailing list Wiktionary-l@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wiktionary-l
DEAR SENDERS OF WIKIMEDIA ETC. PLAEASE STOP SENDING ME REQUEST OR QUESTION ABOUT PARTICIPATING IN WIKIPEDIA DELETE MY E-MAIL-ADRESS I AM NOT INTERESTED!
Op 30 jul. 2013, om 17:15 heeft Mathieu Stumpf psychoslave@culture-libre.org het volgende geschreven:
Le 2013-07-26 20:26, Amgine a écrit :
The request is to create a web-based text corpus[1] from which to derive frequencies and then compare with existing wiktionaries. Not a light undertaking, but one which has been proposed and implemented previously (e.g. Connel's Gutenberg project[2])
Generically speaking, someone would need to determine the appropriate size of the corpus sample, it's temporal currency, and the method of creating and maintaining it. This isn't easy to do, and having no strictures results in unwieldy and mostly irrelevant products like Google's n-grams[3] (on the other hand, if someone can figure out how to filter n-grams usefully it would mean we don't have to build our own.)
Actually, I think it would be interesting to have a trend history of words usage over centuries (current trend would also be interesting but probably harder to implement). Wikisource may be used in order to achieve that.
Amgine
[1] https://en.wikipedia.org/wiki/Linguistic_corpus [2] https://en.wiktionary.org/wiki/User:Connel_MacKenzie/Gutenberg [3] http://storage.googleapis.com/books/ngrams/books/datasetsv2.html
On 26/07/13 09:18, Lars Aronsson wrote:
On 07/23/2013 11:23 AM, Mathieu Stumpf wrote:
Here is what I would like to do : generating reports which give, for a given language, a list of words which are used on the web with a number evaluating its occurencies, but which are not in a given wiktionary.
How would you recommand to implemente that within the wikimedia infrastructure?
Some years back, I undertook to add entries for Swedish words in the English Wiktionary. You can follow my diary at http://en.wiktionary.org/wiki/User:LA2
Among the things I did was to extract a list of all Swedish words that already had entries. The best way was to use CatScan to list entries in categories for Swedish words. Even if there is a page called "men", this doesn't mean the Swedish word "men" has an entry, because it could be the English word "men" that is in that page.
Then I extracted all words from some known texts, e.g. novels, the Bible, government reports, and the Swedish Wikipedia, counting the number of occurrencies of each word. Case significance is a bit tricky. There should not be an entry for lower-case stockholm, so you can't just convert everything to lower case. But if a sentence begins with a capital letter, that word should not have a capitalized entry. Another tricky issue is abbreviations, which should keep the period, for example "i.e." rather than "i" and "e". But the period that ends a sentence should be removed. When splitting a text into words, I decided to keep all periods and initial capital letters, even if this leads to some false words.
When you have word frequency statistics for a text, and a list of existing entries from Wiktionary, you can compute the coverage, and I wrote a little script for this. I found that English Wiktionary already had Swedish entries covering 72% of the words in the Bible, and when I started to add entries for the most common of the missing words, I was able to increase this to 87% in just a single month (September 2010).
Many of the common words that were missing when I started were adverbs such as "thereof", "herein", which occur frequently in any text but are not very exciting to write entries about. This statistics-based approach gave me a reason to add those entries.
It is interesting to contrast a given text to a given dictionary in this way. The Swedish entries in the English Wiktionary is a different dictionary than the Swedish entries in the German or Danish Wiktionary. The kinds of words found in the Bible are different from those found in Wikipedia or in legal texts. There is not a single, universal text corpus that we can aim to cover. Google has released its ngram dataset. I'm not sure if it covers Swedish, but even if it does, it must differ from the corpus frequencies published by the Swedish Academy.
It is relatively easy to extract a list of existing entries from Wiktionary. But to prepare a given text corpus for frequency and coverage analysis needs more preparation.
Wiktionary-l mailing list Wiktionary-l@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wiktionary-l
-- Association Culture-Libre http://www.culture-libre.org/
Wiktionary-l mailing list Wiktionary-l@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wiktionary-l
On 30/07/13 08:15, Mathieu Stumpf wrote:
Actually, I think it would be interesting to have a trend history of words usage over centuries (current trend would also be interesting but probably harder to implement). Wikisource may be used in order to achieve that.
Not really. Or, more fairly, the available texts are probably not a valid sample though they could be used for an informal guideline.
"Full documentation: The sobering examples of the research experiences of Timberlake and Ruppenhofer (mentiolned above) show that even 100,000,000 words is at least an order of magnitude too small to capture phenomena that, though of low frequency, are in the competence of ordinary native speakers. That would represent at least 20,000 recorded hours, and it is too low by an order of magnitude."[1]
Of course this is referencing spoken language which, in most cases, differs significantly from written language, but a running word corpus of 100,000,000 seems a useful target, with samples weighted between transcripts, periodicals, and texts from a delimited time and region. Lemmatized corpus of 6,000-10,000.
Amgine
[1] http://emeld.org/school/classroom/text/lexicon-size.html
On 07/30/2013 07:17 PM, Amgine wrote:
Of course this is referencing spoken language which, in most cases, differs significantly from written language, but a running word corpus of 100,000,000 seems a useful target, with samples weighted between transcripts, periodicals, and texts from a delimited time and region. Lemmatized corpus of 6,000-10,000.
If you want to compare one year or decade to the next, you need a similar sample from both years. One way to get this is to narrow down to a corpus of just one journal or newspaper. Wikisource can do this with Popular Science Monthly, https://en.wikisource.org/wiki/PSM
You'll get popular science and only that for every year. You won't have romantic poetry for one year, and theological texts for the next year. You can spot trends in the use of words like engine/motor or steam/electricity, just because that is what this journal is about, and you get the same number of issues and pages each year.
Some assembly required: Most volumes of PSM are not complete yet. Lots of proofreading remains.
Relatedly: * an old effort to list missing entries for en.wikt: http://thread.gmane.org/gmane.org.wikimedia.wiktionary/784, * a recent bugzilla report to get lists of search queries which gave no results/no title matches, to identify requested entries: https://bugzilla.wikimedia.org/show_bug.cgi?id=56830.
Nemo
wiktionary-l@lists.wikimedia.org