Hey everyone,

I wanted to share some good news on the language detection front. I've been messing around with TextCat, an n-gram-based language ID tool that's been around for a good while. (The version I've been using is in Really Old Perl, but as David pointed out, it's been ported to other languages (see below) and it would be easy for us to port to Python or PHP.)

I trained fresh models on 59 languages based on queries (not random non-query text, but fairly messy data), and made some improvements to the code (making it Unicode-aware, making it run much faster, and improving the "unknown n-gram" penalty logic) and ran lots of tests with different model sizes. I got decent improvements over baseline TextCat.

After doing some analysis on my English Wikipedia sample, I limited the languages to those that are really good (Thai pretty much only looks like Thai) and really necessary (English, or course, and Spanish). I dropped languages that performed poorly and wouldn't help much anyway (Igbo doesn't come up a lot on enwiki), and dropped some nice-to-have languages that performed too poorly (French, alas). I got down to 15 relevant languages for enwiki: Arabic, Bulgarian, Bengali, Greek, English, Spanish, Farsi, Hindi, Japanese, Korean, Portuguese, Russian, Tamil, Thai, and Chinese.

The improvement over the ES plugin baseline (with spaces) is phenomenal. Recall doubled, and precision went up by a third. F0.5 is my preferred measure, but all these are waaaay better:

• ES-sp is the ES-plugin with spaces added at the beginning and ending of each string.

• ES-thr is the ES-plugin (with spaces) with optimized thresholds for each language (more fine grained than choosing languages, and over trained because the optimization data and test data is the same, and brittle because the sample for many languages is very small). Also, this method targets precision, and in theory could improve recall a tiny bit, but in practice took it down a bit.

• TC-lim is TextCat, limiting the languages to the relevant 15 (instead of all 59).

I still have some other things to try (equalizing training set sizes, manually selected training sets, trying wiki-text training sets), and a more formal write up is coming, so this is just a preview for now.

Generally, I'm just really glad to get some improved results!

If we want to try an A/B test with this, we'll also need to turn it into something ES can use. Stas noted that it doesn't have to be hosted by ES - since we have a separate call to language detection (which goes to ES now, but can be anything), it can be implemented separately if we wanted to.

David found several other implementations in Java and PHP:

David also pointed out that TextCat uses somewhat the same technique used by Cybozu (ES plugin), so maybe we could try to train it with the same dataset used for TextCat and generate custom profiles. We could reuse the same PHP code in Cirrus. There's a question so whether it will be possible to implement the new "unknown n-gram" penalty logic. And unfortunately the es-plugin does not support loading multiple profiles making it hard to run A/B tests. So we will have to write some code anyways...

And David noted that solutions based on scripts will suffer from loading the ngram files for each request, and asked about overhead for the current models.

Erik noted that if we are going to rewrite it in something, a PHP library would make the most sense for reuse in our application (being PHP and all). Otherwise, we could tie the existing Perl together with a Perl based HTTP server and set it up somewhere in the cluster, perhaps the services misc cluster would make sense. The content translation team was also interested in the work we are doing with language detection and may find a microservice more accessible than a PHP library. 

As Erik pointed out, the proof is in the pudding, so we'll have to see how this works out when we start sending live user data to it.

• Stas pointed out that it's surprising that Bulgarian is on the short list because it's so similar to Russian. Actually Bulgarian, Spanish, and Portuguese aren't great (40%-55% F0.5 score) but they weren't obviously actively causing problems, like French and Igbo were.

• I should have gone looking for newer implementations of TextCat, like David did. It is pretty simple code. But that also means that using and modifying another implementation or porting our own should be easy. The unknown n-gram penalty fix was pretty small—using the model size instead of the incoming sample size as the penalty. (more detail on that with my write up.)

• I'm not 100% sure whether using actual queries as training helped (I thought it would, which is why I did it), but the training data is still pretty messy, so depending on the exact training method, retraining Cybozu could be doable. The current ES plugin was a black box to me—I didn't even know it was Cybozu. Anyone know where the code lives, or want to volunteer to figure out how to retrain it? (Or, additionally, turning off the not-so-useful models within it for testing on enwiki.)

• I modified TextCat to load all the models at once and hold them in memory (the previous implementation was noted in the code to be terrible for line-by-line processing because it would load each model individually for each query processed—it was the simplest hack possible to make it run in line-by-line mode). The overhead isn't much. 3.2MB for 59 models with 5000 n-grams. Models seem to be around 70KB, but range from 45K to 80K. The 15 languages I used are less then 1MB. Right now, a model with 3000 n-grams seems to be the best, so the overhead would go down by ~40% (not all n-grams are the same size, so it might be more). In short, not too much overhead.

• I think it would make sense to set this up as something that can keep the models in memory. I don't know enough about our PHP architecture to know if you can init a plugin and then keep it in memory for the duration. Seems plausible though. A service of some sort (doesn't have to be Perl-based) would also work. We need to think through the architectural bits.

More questions and comments are welcome! I'm going to spend a few more days trying out other variations on models and training data now that I have an evaluation pipeline set up.

There's also some work involved in using this on other wikis. The best results come from using the detectors that are most useful for the given wiki. For example, there really is no point in detecting Igbo on enwiki because Igbo queries are very rare, and Igbo incorrectly grabs a percentage of English queries, and it's extra overhead to use it. On the Igbo wiki—which made it onto my radar for getting 40K+ queries in a day—an Igbo model would obviously be useful. (A quick check just now also uncovered the biggest problem—many if not most queries to the Igbo wiki are in English. Cleaner training data will definitely be needed for some languages.)

—Trey