A language identification tool which aims for both speed and accuracy. Mostly an efficient HeLI-OTS port to Rust, achieving 25x speedups while maintaining same accuracy levels.
Install it in your environment
pip install heliport
then download the binarized model
heliport download
Install the requirements:
Clone the repo, build the package and binarize the model
git clone https://github.com/ZJaume/heliport
cd heliport
pip install .
heliport binarize
Just run the heliport identify
command that reads lines from stdin
cat sentences.txt | heliport identify
eng_latn
cat_latn
rus_cyrl
...
Identify languages of input text
Usage: heliport identify [OPTIONS] [INPUT_FILE] [OUTPUT_FILE]
Arguments:
[INPUT_FILE] Input file, default: stdin
[OUTPUT_FILE] Output file, default: stdout
Options:
-j, --threads <THREADS> Number of parallel threads to use.
0 means no multi-threading
1 means running the identification in a separated thread
>1 run multithreading [default: 0]
-b, --batch-size <BATCH_SIZE> Number of text segments to pre-load for parallel processing [default: 100000]
-c, --ignore-confidence Ignore confidence thresholds. Predictions under the thresholds will not be labeled as 'und'
-s, --print-scores Print confidence score (higher is better) or raw score (higher is better) in case '-c' is provided
-m, --model-dir <MODEL_DIR> Model directory containing binarized model or plain text model. Default is Python module path or './LanguageModels' if relevant languages are requested
-l, --relevant-langs <RELEVANT_LANGS> Load only relevant languages. Specify a comma-separated list of language codes. Needs plain text model directory
-h, --help Print help
>>> from heliport import Identifier
>>> i = Identifier()
>>> i.identify("L'aigua clara")
'cat_latn'
Remember to download or binarize the model first!
use std::path::PathBuf;
use heliport::identifier::Identifier;
use heliport::lang::Lang;
let identifier = Identifier::load(
PathBuf::from("/path/to/model_dir",
None,
);
let lang, score = identifier.identify("L'aigua clara");
assert_eq!(lang, Lang::cat);
Although heliport
currently uses the same models as HeLI-OTS 2.0 and the
identification algorithm is almost the same, there are a few differences
(mainly during pre-processing) that may cause different results.
However, in most case, these should not deacrease accuracy and should not happen frequently.
Note: Both tools have a pre-processing step for each identified text to remove all non-alphabetic characters.
The implementation differences that can change results are:
HeLI
during preprocessing removes urls and words beginning with@
, whileheliport
does not.- Since 1.5, during preprocessing, HeLI repeats every word that does not start with capital letter, This is probably to penalize proper nouns. However, in our tests, we have not find a significant improvement with this. Therefore,to avoid multiplying the cost of prediction by almost x2, this has not been implemented. In the future it might end up being implemented if there is need for it and can be implemented efficiently.
- Rust and Java sometimes have small differences on the smallest decimals in a float, so the stored n-gram probabilities are not exactly the same. But this is very unlikely to affect predicted labels.
Speed benchmarks with 100k random sentences from OpenLID, all the tools running single-threaded:
tool | time (s) |
---|---|
CLD2 | 1.12 |
HeLI-OTS | 60.37 |
lingua all high preloaded | 56.29 |
lingua all low preloaded | 23.34 |
fasttext openlid193 | 8.44 |
heliport | 2.33 |