There's a lot to catch up on, but some quick and easy stuff first, in response to David's comments.

For queries that are marked as "language" (775 queries), the distribution of token counts (word counts) up to 10 is below:

160 1 tokens
152 2 tokens
141 3 tokens
91 4 tokens
63 5 tokens
49 6 tokens
35 7 tokens
18 8 tokens
22 9 tokens
10 10 tokens

More detailed token count info, for all queries, language queries, and non-language queries, including longer queries (max 84 tokens) see [0].

I also quickly tested David's discovery that spaces help, and the short version is that it's worth a couple of percentage points in recall and precision, so it's an easy win. More details at [1].

And, just for grins, I scored the current default—assume everything is English—to see how that looks. Recall, precision, and F-Score are much better, but it doesn't help zero results rate (or general relevancy), since these are all queries that failed. So R&P aren't everything. Details at [2].

—Trey

[0] https://www.mediawiki.org/wiki/User:TJones_(WMF)/Notes/Language_Detection_Evaluation/Corpus_Info
[1] https://www.mediawiki.org/wiki/User:TJones_(WMF)/Notes/Language_Detection_Evaluation#ElasticSearch_language_detection_plugin.2C_with_spaces
[2] https://www.mediawiki.org/wiki/User:TJones_(WMF)/Notes/Language_Detection_Evaluation#Always_.22English.22_detector




Trey Jones
Software Engineer, Discovery
Wikimedia Foundation

On Mon, Sep 7, 2015 at 5:50 AM, David Causse <dcausse@wikimedia.org> wrote:
Thanks!

this is awesome.

Concerning soburdia: the typo is in the first 2 chars so our misspelling identification will fail, searching for sucurbia properly displays "suburbia" as a "did you mean" suggestion. This was one the enhancement we tried to implement but we are currently blocked by a bug in elasticsearch. I hope it's not a common pattern because we'll add a second error with language detection...

Is it possible to identify how many queries are 1 one/2 words/3 words?
I'm asking this question because there's another weakness in this language detector. Characters at word boundaries seems to bear some valuable informations concerning language features and the detector fails to make any benefit of them if it's a one word query. Running the detector with additional trailing spaces changed significantly the results.

For example граничащее (russian)
Detecting "граничащее" returns bg at 0.99
But detecting " граничащее " returns ru at 0.57 and bg at 0.42

But in the end I agree with your analysis in "Stupid language detection". Mainly because the detector does not weight its results on the wiki size (ru should be weighted higher because ruwiki is larger than bgwiki) because it's what we are looking for. We're looking for results, we don't care too much about the actual language of the query.



Le 05/09/2015 00:45, Trey Jones a écrit :
I've written up my analysis of the ElasticSearch language detection plugin that Erik recently enabled:

The short version is that it really likes Romanian (and Italian, and has a bit of a thing for French), and precision on English is great, but recall is poor (probably because of all the typos and other crap that go to enwiki that is still technically "English"). Chinese and Arabic are good.

I think we could do better, and we should evaluate (a) other language detectors and (b) the effect of a good language detector on zero results rate (i.e., simulate sending queries to the right place and see how much of a difference it makes).

Moderately pretty pictures included.

—Trey

Trey Jones
Software Engineer, Discovery
Wikimedia Foundation


_______________________________________________
Wikimedia-search mailing list
Wikimedia-search@lists.wikimedia.org
https://lists.wikimedia.org/mailman/listinfo/wikimedia-search


_______________________________________________
Wikimedia-search mailing list
Wikimedia-search@lists.wikimedia.org
https://lists.wikimedia.org/mailman/listinfo/wikimedia-search