Skip to content

Latest commit

 

History

History
95 lines (75 loc) · 5.23 KB

README.md

File metadata and controls

95 lines (75 loc) · 5.23 KB
Raw

Whichlang

This is a language detection library, aiming for both precision and performance.

Why build this?

While building Quickwit, a search engine tailored for log and tracing data, we found ourselves needing a light, fast, and precise language detection library in Rust that works well with our high throughput requirement. The full story and how it works are detailed in this blog post.

Features

  • No dependency
  • Throughput above 100 MB/s for short and long strings.
  • Good accuracy (99.5% on my validation dataset, but it really depends on the size of your input.)
  • Supported languages: Arabic, Dutch, English, French, German, Hindi, Italian, Japanese, Korean, Mandarin, Portuguese, Russian, Spanish, Swedish, Turkish, and Vietnamese.

How does it work?

It uses a multiclass logistic regression model over:

  • 2, 3, 4-grams of letters on ASCII
  • codepoint / 128
  • a slightly smarter projection of codepoints over a given class.

We use the hashing trick and project these features over a space of size 4_096.

The logistic regression is trained in the python notebook attached, and used to generate weight.rs.

Comparison with Whatlang

The following compares the throughput using the simple benchmark found in this repository and the accuracy using whatlang-accuracy-benchmark benchmark. Overall, Whichlang is about 10x faster and slightly more accurate than Whatlang.

Throughput

To generate the throughput benchmark, we ported the benchmark available in this repository. Please, check this repository to see our changes.

Processing Time (µs) Throughput (MiB/s)
whatlang/short 16.62 1.66
whatlang/long 62.00 9.42
whichlang/short 0.26 105.69
whichlang/long 5.21 112.31

Accuracy

To generate the accuracy benchmark, we have changed the whatlang-accuracy-benchmark to add support for Whichlang. Given that Whatlang supports more languages, we have used its FilterList feature to restrict its analysis to only languages that are supported in Whichlang. We also use the trigram method in Whatlang. Please, check this repository to see our changes.

Crate: Whatlang
AVG: 91.69%

| LANG       | AVG    | <= 20   | 21-50  | 51-100 | > 100   |
|------------|--------|---------|--------|--------|---------|
| Arabic     | 99.68% | 99.51%  | 99.64% | 99.83% | 99.76%  |
| Mandarin   | 96.09% | 97.54%  | 96.92% | 95.45% | 94.43%  |
| German     | 88.57% | 70.00%  | 88.53% | 96.61% | 99.16%  |
| English    | 85.99% | 57.82%  | 88.37% | 97.97% | 99.78%  |
| French     | 90.88% | 72.84%  | 92.51% | 98.54% | 99.65%  |
| Hindi      | 99.80% | 100.00% | 99.83% | 99.78% | 99.61%  |
| Italian    | 87.75% | 66.67%  | 87.74% | 97.04% | 99.54%  |
| Japanese   | 94.37% | 93.97%  | 96.04% | 94.30% | 93.18%  |
| Korean     | 99.17% | 98.88%  | 99.69% | 99.44% | 98.66%  |
| Dutch      | 89.68% | 72.13%  | 89.78% | 97.40% | 99.40%  |
| Portuguese | 88.08% | 72.90%  | 85.76% | 95.22% | 98.44%  |
| Russian    | 99.98% | 100.00% | 99.96% | 99.98% | 100.00% |
| Spanish    | 82.91% | 55.45%  | 82.24% | 94.85% | 99.10%  |
| Swedish    | 84.16% | 58.33%  | 83.78% | 96.35% | 98.18%  |
| Turkish    | 86.73% | 61.01%  | 88.94% | 97.32% | 99.63%  |
| Vietnamese | 93.23% | 82.84%  | 92.96% | 97.88% | 99.24%  |
| AVG        | 91.69% | 78.74%  | 92.04% | 97.37% | 98.61%  |

Crate: Whichlang
AVG: 97.03%

| LANG       | AVG     | <= 20   | 21-50   | 51-100  | > 100   |
|------------|---------|---------|---------|---------|---------|
| Arabic     | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% |
| Mandarin   | 98.65%  | 98.69%  | 98.48%  | 98.55%  | 98.87%  |
| German     | 94.20%  | 80.00%  | 97.47%  | 99.49%  | 99.84%  |
| English    | 97.15%  | 91.84%  | 97.25%  | 99.57%  | 99.93%  |
| French     | 97.59%  | 93.83%  | 97.61%  | 99.20%  | 99.71%  |
| Hindi      | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% |
| Italian    | 97.20%  | 93.06%  | 97.33%  | 98.85%  | 99.57%  |
| Japanese   | 94.92%  | 88.95%  | 95.14%  | 97.74%  | 97.85%  |
| Korean     | 99.83%  | 99.44%  | 99.98%  | 99.97%  | 99.94%  |
| Dutch      | 97.08%  | 92.84%  | 96.98%  | 98.91%  | 99.60%  |
| Portuguese | 94.07%  | 83.87%  | 94.89%  | 98.18%  | 99.36%  |
| Russian    | 99.92%  | 99.69%  | 99.99%  | 100.00% | 100.00% |
| Spanish    | 92.12%  | 76.36%  | 93.78%  | 98.65%  | 99.70%  |
| Swedish    | 95.37%  | 90.28%  | 94.94%  | 97.76%  | 98.51%  |
| Turkish    | 95.51%  | 88.24%  | 98.11%  | 98.38%  | 97.33%  |
| Vietnamese | 98.79%  | 96.57%  | 98.87%  | 99.77%  | 99.96%  |
| AVG        | 97.03%  | 92.10%  | 97.55%  | 99.06%  | 99.39%  |