Language Detection Identification Generation (NLD, NLI, NLG)
- 3.Word based vs char based - Word-based LMs display higher accuracy and lower computational cost than char-based LMs. However, char-based RNN LMs better model languages with a rich morphology such as Finish, Turkish, Russian etc. Using word-based RNN LMs to model such languages is difficult if possible at all and is not advised. Char-based RNN LMs can mimic grammatically correct sequences for a wide range of languages, require bigger hidden layer and computationally more expensive while word-based RNN LMs train faster and generate more coherent texts and yet even these generated texts are far from making actual sense.
Full results for every language that I tested are in table at the end of blog post & on Github. From them I can make following conclusions:
- all detectors are equally good on some languages, such as, Japanese, Chinese, Vietnamese, Greek, Arabic, Farsi, Georgian, etc. - for them the accuracy of detection is between 98 & 100%;
- CLD is much better in detection of "rare" languages, especially for languages, that are similar to more frequently used - Afrikaans vs Dutch, Azerbaijani vs. Turkish, Malay vs. Indonesian, Nepali vs. Hindi, Russian vs Bulgarian, etc. (it could be result of imbalance of training data - I need to check the source dataset);
- for "major" languages not mentioned above (English, French, German, Spanish, Portuguese, Dutch) the fastText results are much better than CLD's, and in many cases lingid.py's & OpenNLP's;
- for many languages results for "compressed" fastText model are slightly worse than results from "full" model (mostly only by 1-2%, but could be higher, like for Kazakh when difference is 33%), but there are languages where the situation is different - results for compressed are slight better than for full (for example, for German or Dutch);
OpenNLP has many misclassifications for Cyrillic languages - Russian/Ukrainian, ...
Rafael Oliveira posted on FB a simple diagram that shows what languages are detected better by CLD & what is better handled by fastText
Here are some additional notes about differences in behavior of detectors that I observe during analyzing results:
- fastText is more reliable than CLD on the short texts;
- fastText models & langid.py detect language as Hebrew instead of Jewish as in CLD. Similarly, CLD uses 'in' for Indonesian language instead of standard 'id' used by fastText & langid.py;
- fastText distinguish between Cyrillic- & Latin-based versions of Serbian;
- CLD tends to incorporate geographical & person's names into detection results - for example, blog post in German about travel to Iceland is detected as Icelandic, while fastText detects it as German;
- In extended detection mode CLD tends to select more rare language, like, Galician or Catalan over Spanish, Serbian instead of Russian, etc.
- OpenNLP isn't very good in detection for short texts.
The models released by fastText development team provides very good alternative to existing language detection tools, like, Google's CLD & langid.py - for most of "popular" languages, these models provides higher detection accuracy comparing to other tools, combined with high speed of detection (drawback of langid.py). Even using "compressed" model it's possible to reach good detection accuracy. Although for some less frequently used languages, CLD & langid.py may show better results.
Performance-wise, the langid.py is much slower than both CLD & fastText. On average, CLD requires 0.5-1 ms to perform language detection. For fastText & langid.py I don't have precise numbers yet, only approximates based on speed of execution of corresponding programs.
“[BLEU] looks at the presence or absence of particular words, as well as the ordering and the degree of distortion—how much they actually are separated in the output.”
BLEU’s evaluation system requires two inputs: (i) a numerical translation closeness metric, which is then assigned and measured against (ii) a corpus of human reference translations.
BLEU averages out various metrics using an n-gram method, a probabilistic language model often used in computational linguistics.
The result is typically measured on a 0 to 1 scale, with 1 as the hypothetical “perfect” translation. Since the human reference, against which MT is measured, is always made up of multiple translations, even a human translation would not score a 1, however. Sometimes the score is expressed as multiplied by 100 or, as in the case of Google mentioned above, by 10.
a BLEU score offers more of an intuitive rather than an absolute meaning and is best used for relative judgments: “If we get a BLEU score of 35 (out of 100), it seems okay, but it actually has no correlation to the quality of the output in any meaningful sense. If it’s less than 15, we can probably safely say it’s very bad. If it’s greater than 60, we probably have some mistake in our testing! So it will generally fall in there.”
“Typically, if you have multiple [human translation] references, the BLEU score tends to be higher. So if you hear a very large BLEU score—someone gives you a value that seems very high—you can ask them if there are multiple references being used; because, then, that is the reason that the score is actually higher.”
- 1.General talk about FAMG (fb, ama, micro, goog) and research direction atm, including some info about BLUE scores and the comparison issues with reports of BLUE (boils down to diff unmentioned parameters)
Named entity language transliteration