[Corpora-List] Q: Classification performance across languages and language families

Sat Jun 2 16:43:52 UTC 2012

Dear Albert and Adam,

I can confirm that the EuroVoc descriptor (class) labels are NOT used in the classification process, meaning that the translation of the thesaurus labels themselves is irrelevant. Our experiments showed that using the label during the classification process did not help, probably because class names such as ‘equality between men and women’ do not usually occur in the text. For English, only 31% of the names (labels) of the manually assigned descriptors actually occur in the indexed document. 

Impact of translation quality: The idea that the text translation quality (and especially the consistency of the translation of terms) may have an impact on the  classification performance could indeed be a reason. When the EU went from 15 to 25 (and then to 27) member states, large numbers of legal documents needed to be translated in a short time. And indeed, the 5 Slavic languages (all ‘new’ member states) are all among the less well performing ones. However, Hungarian and Lithuanian (also ‘new’ member states) are in first and second position! For Maltese, I agree: Being a Semitic language with strong Romance influences may be a challenge for consistent term translation. 

Does anybody have comparative experience with document classification for Slavic, Finno-Ugric or Baltic languages? 

Maltese stop words: Our classification tool JEX automatically weighs words depending on how specifically they occur in texts indexed with one class and it does thus automatically discard most high-frequency words. Using a manually compiled list of stop words (including words such as ‘paragraph’ and ‘decision’) nevertheless increases the performance. We did not use more Maltese stop words because we had never looked at the language and we first did not have any stop words. We later added 296 Maltese stop words and the performance (F1) increased from 0.4366 to 0.4500, which is still far below the value for the other languages. 

Any other possible explanations or experiences?

Greetings,

Ralf

From: bertugatt at gmail.com [mailto:bertugatt at gmail.com] On Behalf Of Albert Gatt
Sent: 02 June 2012 15:26
To: Ralf Steinberger
Cc: Adam Kilgarriff; corpora at uib.no
Subject: Re: [Corpora-List] Q: Classification performance across languages and language families

Dear Ralf

I find the issue you've raised quite interesting and I too wonder why Maltese should behave so differently. Like Adam, wondered about the quality of the thesaurus at first. Perhaps that's not the reason, as you suggest. But another reason -- also related to the relatively recent development of vocabularies in certain technical areas in Maltese (Malta being bilingual, most such technical areas were written about in English) -- might be inconsistencies and/or variation in the way the documents in your set were translated, which would also affect the distribution of lexical features and the reliability with which they are associated with particular categories. I am aware of an initiative in recent years among Maltese translation bureaux to standardise some of the translations of technical terms/phrases. (One of the problems seems to have been that, because Maltese is Semitic, but has been heavily influenced by Romance, there is often more than one possible translation for a given term. Another problem is simply that translators, especially in the early days after Malta's accession to the EU, would have relied on circumlocution and similar "workarounds", before a vocabulary was gradually developed.) I guess the more recent the document collection, the more likely it would be to avoid such inconsistencies.

I've also taken a look at your LREC paper, mainly at Table 1, where your precision/recall and other stats are reported. Here too, there are some things which I find surprising. For example, why are there only 6 elements in your stop-word list for Maltese, compared to much bigger numbers for many other languages?

albert

On 2 June 2012 14:29, Ralf Steinberger <ralf.steinberger at jrc.ec.europa.eu> wrote:

Dear Adam,

Thanks for your proposal and for allowing me to clarify: EuroVoc is a classification scheme with exactly the same 6700 subject domain classes in all languages, i.e. each class has a numerical identifier and exactly one class label that has been translated into all 27 or so languages. Example EuroVoc categories are ‘nuclear materials’, ‘Austria’, ‘fishery management’, ‘xenophobia’, ‘budget’, ‘population statistics’, ...

I cannot see how such a classification scheme would favour one language over another, especially as the documents are parallel translations, as well: they have the same contents in all languages. EuroVoc is in no way comparable to a resource such as WordNet, which rather lists and organises existing words of a language, with varying coverage. 

Greetings from Italy to the UK.

Ralf

From: adam.kilgarriff at gmail.com [mailto:adam.kilgarriff at gmail.com] On Behalf Of Adam Kilgarriff
Sent: 02 June 2012 14:13
To: Ralf Steinberger
Cc: corpora at uib.no; clef at dei.unipd.it; ln at cines.fr
Subject: Re: [Corpora-List] Q: Classification performance across languages and language families

Ralf,

Please excuse scepticism, but what about the simple hypothesis that it all depends on thesaurus-quality.  My hunch would be that it started from a Germanic language, hence good performance there, and that Slavic lgs have been added more recently, so there have been less years for debugging/improving, and that there was a particularly inspired Hungarian translator!

Maltese has a special problem - Maltese hasn't ever had a technical vocabulary so there was nothing the Maltese thesaurus-translators could do except make things up.

(Of course I'll be happy to have my hypothesis quashed by someone who knows the history of Eurovoc)

Adam

On 2 June 2012 12:40, Ralf Steinberger <ralf.steinberger at jrc.ec.europa.eu> wrote:

A question and an invitation to discussion.

We recently carried out multi-label categorisation experiments <http://langtech.jrc.ec.europa.eu/Documents/2012_LREC-JEX-final.pdf>  on a mostly parallel set of documents in 22 languages, covering the language families Germanic, Romance, Slavic, Hellenic, Finno-Ugric, Baltic and Semitic. The document set is reasonably large (22K to 42K documents per language), using the thousands of subject domain categories from the EuroVoc thesaurus <http://eurovoc.europa.eu/> . The performance across languages was rather uniform, with the exception of the outlier Maltese, which performed considerably less well. The languages covered are Bulgarian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Slovak, Slovenian, Spanish and Swedish. 

To my great surprise, the highly inflected agglutinative language Hungarian produced the best results of all. The five Germanic languages ended up in the top ten positions, the five Slavic languages in the bottom half. The results for the other language families were less consistent. 

Q1: Does anyone have an intuition how these results could be explained?

Q2: Has anyone ran similar experiments with other types of classifiers or data? Are the results similar?

My initial expectation had been that highly inflected languages would perform less well and that feature space reduction using lemmatisation would improve the results. However, our experiments for Czech, English, Estonian and French (described in Ebrahim et al., forthcoming) showed the contrary, rather consistently for all four languages and language families: (1) lemmatisation reduces the performance and (2) adding part-of-speech (POS) information to the word form and/or to the lemma improves the performance. 

Q3: Can we conclude that: the scarcer the feature space, the better the classification performance? 

Q4: If that were the case, why did Slavic languages (and Maltese) perform less well in our experiments? 

I would be pleased if you could share your own experience and/or your opinions.

The classification tool (JRC EuroVoc Indexer JEX <http://langtech.jrc.ec.europa.eu/Eurovoc.html> ) and the multilingual document set can be downloaded from http://langtech.jrc.ec.europa.eu/Eurovoc.html . Details of our experiments are given in the two papers below.

Steinberger Ralf, Mohamed Ebrahim & Marco Turchi (2012). JRC EuroVoc Indexer JEX - A freely available multi-label categorisation tool. Proceedings of the 8th international conference on Language Resources and Evaluation (LREC'2012), Istanbul, 21-27 May 2012. (PDF <http://langtech.jrc.ec.europa.eu/Documents/2012_LREC-JEX-final.pdf> )

Ebrahim Mohamed, Maud Ehrmann, Marco Turchi & Ralf Steinberger (forthcoming). Multi-label EuroVoc classification for Eastern and Southern EU Languages. In: Cristina Vertan & Walther v. Hahn: Multilingual processing in Eastern and Southern EU languages - Low-resourced technologies and translation. Cambridge Scholars Publishing, Cambridge, UK.

Greetings,

Ralf

Ralf Steinberger 

European Commission – Joint Research Centre (JRC)

URL: http://langtech.jrc.ec.europa.eu/RS.html  

_______________________________________________
UNSUBSCRIBE from this page: http://mailman.uib.no/options/corpora
Corpora mailing list
Corpora at uib.no
http://mailman.uib.no/listinfo/corpora

-- 
========================================
Adam Kilgarriff <http://www.kilgarriff.co.uk/>                   adam at lexmasterclass.com                                             
Director                                    Lexical Computing Ltd <http://www.sketchengine.co.uk/>                 
Visiting Research Fellow                 University of Leeds <http://leeds.ac.uk>      

Corpora for all with the Sketch Engine <http://www.sketchengine.co.uk>                  

                        DANTE:  <http://www.webdante.com> a lexical database for English                  

========================================

_______________________________________________
UNSUBSCRIBE from this page: http://mailman.uib.no/options/corpora
Corpora mailing list
Corpora at uib.no
http://mailman.uib.no/listinfo/corpora

-- 

-----------------------------------------------------------------

Albert Gatt

Institute of Linguistics

Centre for Communication Technology Rm 402B

University of Malta

Tal-Qroqq Msida MSD2080

Malta

tel: (+356) 2340 2150

http://staff.um.edu.mt/albert.gatt/

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://listserv.linguistlist.org/pipermail/corpora/attachments/20120602/dc7a1c5b/attachment.htm>
-------------- next part --------------
_______________________________________________
UNSUBSCRIBE from this page: http://mailman.uib.no/options/corpora
Corpora mailing list
Corpora at uib.no
http://mailman.uib.no/listinfo/corpora