What is Language Identification?

Language identification is the problem of classifying a sample of characters based on its language. This is a critical pre-processing stage in many applications that apply language-specific modeling. For instance, a search engine might use different tokenizers based on the language being stored.

How does LingPipe Perform Language ID?

LingPipe's text classifiers learn by example. For each language being classified, a sample of text is used as training data. LingPipe learns the distribution of characters per language using character language models. Character language models provide state-of-the-art accuracy for text classification. Character-level models are particularly well-suited to language ID because they do not require tokenized input; tokenizers are often language-specific.

How Many Languages are there?

The short answer is a whole lot of them. The more complex answer is that it depends how you count dialect variation, etc. A rough estimate is between 5000 and 10,000. Not all of these languages even have written forms. For more information, check out:

Running Language ID

We start with an example of running language identification from a pre-built model. This can be carried out from the command-line using the following command (for Windows: replace colons in classpath with semicolons, and remove backslashes and put the command all on one line).

> cd demos/tutorial/langid
> java \
  -cp languageId.jar:../../../lingpipe-4.1.2.jar \
  RunLanguageId ../../models/langid-leipzig.classifier \
  "TextToClassify1" \

This uses a small model distributed in the $LINGPIPE/demos/models directory. Later, we describe how to build larger and more accurate models.

Language ID may also be run from Ant, with target run:

> ant run
Reading classifier from C:\mycvs\lingpipe\demos\models\langid-leipzig.classifier

Input=Per poder jutjar l'efectivitat d'una novetat Acs imprescindible deixar
Rank  Category  Score  P(Category|Input)   log2 P(Category,Input)
0=cat -2.023136071289025 1.0 -145.6657971328098
1=fr -4.321846555117093 1.504464337461871E-50 -311.17295196843065
2=it -4.581659915472809 3.516783524040892E-56 -329.87951391404226
3=nl -4.696851550631136 1.1206338610900442E-58 -338.1733116454418
4=en -4.766148642045668 3.52782891712393E-60 -343.16270222728804
5=no -4.958001178925975 2.4505580508882693E-64 -356.9760848826702
6=se -4.987497723277155 5.622794125159236E-65 -359.0998360759552
7=dk -5.104818791491201 1.6110781779891577E-67 -367.54695298736647
8=tr -5.219139511312786 5.361805775146849E-70 -375.7780448145206
9=de -5.265340898319162 5.344841592730491E-71 -379.10454467897966
10=sorb -5.518926633655108 1.704814443523678E-76 -397.3627176231678
11=ee -5.579930924219301 8.118211859258627E-78 -401.75502654378965
12=fi -6.07538483408664 1.4822218818370195E-88 -437.4277080542381
13=jp -10.067560120929357 4.404215405114351E-175 -724.8643287069137
14=kr -10.728373897912512 2.0954882281132754E-189 -772.4429206497009

Rank  Category  Score  P(Category|Input)   log2 P(Category,Input)
0=de -2.4763103634840435 0.7510836287966632 -22.28679327135639
1=cat -2.7980617047077603 0.10091995592652597 -25.18255534236984
2=fr -2.922969589040629 0.04629860154204275 -26.30672630136566
3=en -2.937075498441145 0.042398564875480625 -26.433679485970305
4=it -3.033600088032701 0.023218811490292642 -27.30240079229431
5=ee -3.137058174682278 0.012177106355447293 -28.233523572140502
6=se -3.141665660082584 0.01183208218028833 -28.274990940743255
7=dk -3.2402194566918268 0.006398119344121462 -29.16197511022644
8=nl -3.2940283974733138 0.0045737180617331724 -29.646255577259826
9=sorb -3.525396435862732 0.0010800178006085642 -31.728567922764586
10=fi -4.202790233233854 1.5782952392567225E-5 -37.825112099104686
11=no -4.439911553666941 3.5955271788398965E-6 -39.95920398300247
12=tr -5.316690003944057 1.514722435945023E-8 -47.85021003549652
13=jp -8.529346597929084 2.9948774318501194E-17 -76.76411938136177
14=kr -9.181964473001777 5.108119495312997E-19 -82.63768025701599

Rank  Category  Score  P(Category|Input)   log2 P(Category,Input)
0=cat -3.052651519527544 0.34111151235773624 -21.36856063669281
1=dk -3.19525946387983 0.17076210305468095 -22.366816247158813
2=se -3.212108777343314 0.15735713996594727 -22.484761441403197
3=it -3.212646751837158 0.15694693118670108 -22.488527262860107
4=no -3.4648140448407854 0.046172500475394666 -24.253698313885497
5=fr -3.4922600806686943 0.040415581714115904 -24.44582056468086
6=fi -3.511941223577881 0.03673470305405576 -24.583588565045165
7=en -3.6670872518921986 0.017304190037373053 -25.66961076324539
8=de -3.709920120332009 0.014057025366728032 -25.969440842324065
9=ee -3.8734110543633595 0.006358924994944445 -27.113877380543517
10=nl -3.8744412865885054 0.006327217836117602 -27.121089006119536
11=sorb -3.8749770727403914 0.00631079064205628 -27.12483950918274
12=tr -4.657859758912603 1.4137921681172042E-4 -32.605018312388225
13=kr -7.645064681009673 7.173276613510208E-11 -53.515452767067714
14=jp -7.857388955447441 2.5603878072330347E-11 -55.00172268813209

After reading the classifier from the demos/models directory, there are a series of test cases and their output. The first test case is a sentence fragment from Catalan. The output is presented as a rank-ordered of predicted categories plus statistics. The language predicted by the classifier is cat (Catalan); the second-best match is fr (French). After the ranks and categories, there are three numbers per line. The second is perhaps the most useful, as it is a conditional probability estimate of the category given the input. For the input given, the conditional estimate is that (within rounding error), the choice of Catalan is 100% certain. The chance of the input being French, according to the classifier's estimate, is very very very low (roughly 1/1050). The last column is the log (base 2) joint probability estimate of the category and the input. The first column is the score, which is a kind of entropy rate, which is roughly the log joint probability estimate divided by the length of the input.

For the short inputs, consisting of single first names, the classifier is far less certain about its categorization. Although it's 75% sure that Michael is German, it holds out a 10% chance that it's Catalan, a 4.2% chance it's English, and so on. The name Maria, in the third example, is even more confusible, with Catalan being the most likely guess, but with a confidence estimate of only 34%.

The Leipzig Corpora Collection

The example model in the last section was derived from training data provided as part of the Leipzig Corpora Collection, which is available from:


The collection consists of a corpus of texts collected from the web for 15 different languages: Catalan (cat), Danish (dk), English (en), Estonian (ee), Finnish (fi), French (fr), German (de), Italian (it), Japanese (jp), Korean (kr), Norwegian (no), Sorbian (sorb), Swedish (se), and Turkish (tr).

If you are not able to acquire the Leipzig corpus, this demo can be carried out with any available language samples by simply putting them into the same format to which we will convert the Leipzig corpora.

Downloading and Unpacking

Make a place to put the Leipzig files. We'll call it:

> mkdir leipzig
> mkdir leipzig/dist

Download the relevant text zip files to leipzig/dist. After download, these should look as follows:

> cd leipzig
> ls dist
cat300k.zip  ee300k.zip  fr100k.zip  kr300k.zip  se100k.zip
de1M.zip     en300k.zip  it300k.zip  nl100k.zip  sorb100k.zip
dk100k.zip   fi100k.zip  jp100k.zip  no300k.zip  tr100k.zip

The numbers in the suffixes indicate how many sentences are provided for the specified language; e.g. one million German sentences, one hundred thousand Turkish, etc. First, these files need to be unzipped and placed in a second directory. This can be done by hand or scripted. We provide an ant target to do the job in the build.xml file. This can be called, from within the langid/ directory, as follows:

ant -Ddir.dist=leipzig/dist -Ddir.unpacked=leipzig/unpacked unpack

Here's what the run should look like. Note that it takes several minutes to unzip this much data.

> ant -Ddir.dist=leipzig/dist -Ddir.unpacked=leipzig/unpacked unpack
Buildfile: build.xml

    [unzip] Expanding: C:\mycvs\data\leipzig\dist\cat300k.zip into C:\mycvs\data\leipzig\unpacked
    [unzip] Expanding: C:\mycvs\data\leipzig\dist\de1M.zip into C:\mycvs\data\leipzig\unpacked

Data Format

Each zipped corpus unpacks into a directory with the same name as the zip file. These directories mostly contain derived data. The raw text data is provided in a single file named sentences.txt in each directory. Furthermore, there are meta-information files meta.txt in each directory, which we will use to extract the character encoding, which varies by corpus.

C:\mycvs\lingpipe\demos\tutorial\langid>cat c:\mycvs\data\leipzig\unpacked\en300k\meta.txt
1       number of sentences     300000
1       average sentence length in characters   128.5906
1       content encoding        iso-8859-1

Each meta.txt file contains a line with the content encoding, as shown above (as well as a number of statistics derived from the corpus).

The raw data itself is organized with one sentence per line, starting with a sentence number, a tab character, and then the sentence text.

> less c:\mycvs\data\leipzig\unpacked\en300k\sentences.txt
1       A rebel statement sent to Lisbon from Jamba said 86 government soldiers and 13 guerrillas were killed in the fighting that ended Jan. 3. It said the rebel forces sill held Mavinga.
2       Authorities last week issued a vacate order for a club in Manhattan and closed another in the Bronx.
3       At the first Pan Am bankruptcy hearing, for example, at least five airlines were represented.
4       Mr. Neigum, poker-faced during the difficult task, manages a 46-second showing.


We provide a program src/Munge.java that extracts the character encoding from the meta.txt files, uses it to read the sentences.txt files, removing the line numbers, tabs and replacing line breaks with single space characters. The output is uniformly written using the UTF-8 unicode encoding. This program may be run from Ant using target munge.

> ant -Ddir.unpacked=leipzig\unpacked -Ddir.munged=leipzig\munged munge

reading from=leipzig\unpacked\cat300k\sentences.txt charset=iso-8859-1
writing to=leipzig\munged\cat\cat.txt charset=utf-8
total length=37055486

reading from=leipzig\unpacked\de1M\sentences.txt charset=iso-8859-1
writing to=leipzig\munged\de\de.txt charset=utf-8
total length=110907216


Final Data Format

The final result is a single directory leipzig/munged which contains one subdirectory per language, with one sample file per language of the same name with suffix .txt.

> ls leipzig\munged
cat  de  dk  ee  en  fi  fr  it  jp  kr  nl  no  se  sorb  tr

> ls leipzig\munged\en

Training Language ID

Now that the corpora are in a uniform format, it's easy to train them.

Training Code

We provide a training program in the form of a single main() method in src/TrainLanguageId.java. We repeat the code here.

public static void main(String[] args) throws Exception {
    File dataDir = new File(args[0]);
    File modelFile = new File(args[1]);
    int nGram = Integer.parseInt(args[2]);
    int numChars = Integer.parseInt(args[3]);

    String[] categories = dataDir.list();

    DynamicLMClassifier classifier
        = DynamicLMClassifier

    char[] csBuf = new char[numChars];
    for (int i = 0; i < categories.length; ++i) {
        String category = categories[i];
        File trainingFile = new File(new File(dataDir,category),
                                     category + ".txt");
        FileInputStream fileIn
            = new FileInputStream(trainingFile);
        InputStreamReader reader
            = new InputStreamReader(fileIn,Strings.UTF8);
        String text = new String(csBuf,0,numChars);
        Classification c = new Classification(category);
        Classified<CharSequence> classified
            = new Classified<CharSequence>(text,c);

The command takes five arguments, the name of the directory in which to find the data (in our case, leipzig/munged), the name of the file to which we will write the compiled model, the n-gram order to use for training, and the number of characters to use for training each language. The first few lines of code simply read in the command-line parameters.

Next, the names of the directories in the data directory are used to provide the categories. These are used to create a dynamic (trainable) classifier, along with the n-gram length. The flag is set to false, meaning that the language models used for classification will be process models rather than boundary models.

The character array csBuf is used to hold the training data. It is allocated to be the same size as the number of characters used for training. The program then just creates a reader to read the training characters from the specified file into the buffer. Once the characters are in hand, they're passed to the classifier for training. All that is necessary is the category and the character slice for training.

Calling the Training Command

After training the classifier on each category, the classifier is written to the specified model file using the utility method compileTo in com.aliasi.util.AbstractExternalizable.

We supply an ant target train for calling the training command (in Windows DOS, remove the backslashes and put the command all on one line).

> ant -Dcorpus.dir=leipzig/munged \
      -Dmodel.file=../../models/langid-leipzig.classifier \
      -Dtraining.size=100000 \
      -DnGram=5 \

nGram=5 numChars=100000
Training category=cat
Training category=tr

Compiling model to file=..\..\models\langid-leipzig.classifier

Evaluating Language ID

With a model in hand, evaluating a classifier is straightforward. In this case, we'll evaluate a specified number of samples of a specified length from the portions of the corpora outside of the training set.

Evaluation Code

The code for evaluation is provided in src/EvalLanguageId.java. We repeat that code here, with the sections cut-and-pasted from the training code greyed out.

public static void main(String[] args) throws Exception {
    File dataDir = new File(args[0]);
    File modelFile = new File(args[1]);
    int numChars = Integer.parseInt(args[2]);
    int testSize = Integer.parseInt(args[3]);
    int numTests = Integer.parseInt(args[4]);

    String[] categories = dataDir.list();

    BaseClassifier<CharSequence> classifier
                = (BaseClassifier<CharSequence>) AbstractExternalizable.readObject(modelFile);
    BaseClassifierEvaluator<CharSequence> evaluator
        = new BaseClassifierEvaluator<CharSequence>(classifier,categories);

    char[] csBuf = new char[testSize];
    for (int i = 0; i < categories.length; ++i) {
        String category = categories[i];
        File trainingFile = new File(new File(dataDir,category),
                                         category + ".txt");
        FileInputStream fileIn
            = new FileInputStream(trainingFile);
        InputStreamReader reader
            = new InputStreamReader(fileIn,Strings.UTF8);

        reader.skip(numChars); // skip training data

        for (int k = 0; k < numTests; ++k) {
            Classification c = new Classification(category);
            Classified<CharSequence> cl
                = new Classified<CharSequence>(new String(csBuf),c);


The first step is reading in five command line arguments. These specify the directory in which to find the data, the file in which to find the model, the number of characters used for training (so they are not used for testing), the size of each test in characters, and the number of test samples to run per language.

Next, the classifier is reconstituted using the utility method readObject in com.aliasi.util.AbstractExternalizable. This classifier is then used to construct the evaluator.

For each category, we first skip the number of characters used for training. Then for each test, we read the appropriate number of characters into the character buffer csBuf. Note that the buffer is now sized to fit a single test instance. The critical code is where we add the test case to the evaluator using the addCase method. This method simply requires the reference (true) category and the text.

Finally, the results are printed by simply converting the evaluator to a string.

Running the Evaluation

There is an ant target eval which runs an evaluation. The command-line arguments are specified on the command-line using the -D arguments.

> ant -Dcorpus.dir=leipzig/munged -Dmodel.file=../../models/langid-leipzig.classifier -Dtraining.size=100000 -Dtest.size=50 -Dtest.num=1000 eval

Reading classifier from file=..\..\models\langid-leipzig.classifier
Evaluating category=cat
Evaluating category=tr
Categories=[cat, de, dk, ee, en, fi, fr, it, jp, kr, nl, no, se, sorb, tr]
Total Count=15000
Total Correct=14797
Total Accuracy=0.9864666666666667
95% Confidence Interval=0.9864666666666667 +/- 0.001849072921311086
Confusion Matrix
reference \ response

This provides accuracy statistics for a run of 1000 tests per category, with test lengths fixed at 50 characters. The overall accuracy is 14797/15000, or 98.647%. The 95% confidence interval reported is +/- 0.185%. The confidence interval is determined by a normal approximation to a binomial, and the 95% confidence interval is +/- 0.18%. If we run fewer tests, the confidence interval will be broader (98.733% +/- 0.566% or so at 100 tests), and if we run more, it will be tighter (98.472% +/- 0.062% at 10,000 tests).

Perhaps the most interesting report for classification is first-best confusion, reported in matrix form. What this shows is the number of times a given language was misidentified as another. Reading the first line, Catalan (cat), was correctly identified 991/1000 times, with 1 error identifying it as Danish, 5 errors confusing it as English, 1 error confusing it with French, and 2 confusing it with Turkish. Reading the last line, Turkish was correctly identified 1000/1000 times. That is, of 1000 Turkish examples, all of them were classified as Turkish. Japanese, Korean and French also had perfect scores. The worst performer was Danish at 92%, followed by Norwegian at 93%, with the two languages very often being confused for one another (58 Norwegian cases were mistakenly classifed as Danish; 74 Danish cases were mistakenly classified as Norwegian)

These global reports are followed with a variety of other global statistics, all of which are explained in the class documentation to the classifier evaluator, com.aliasi.classify.ClassifierEvaluator, and in the documentation to which it points.

After the global reports, there are per-category reports. These reports begin with a one-versus all report, the performance of which is often substantially superior to n-way classification results. An interesting report is the rank histogram, which provides the number of times the correct answer was at a specified rank. For instance, for Catalan, this is reported as:

 Rank Histogram=

This only looks at cases which should've been classified as Catalan. 991/1000 of these had Catalan as their first-best category. In 6/1000 cases, Catalan was the second guess. 2/1000 times it was 3rd-best, and 1/1000 times 5th-best.

Tuning and Evaluation Parameters

Input Length Sensitivity

Language identification is highly sensitive to length. While this is to some degree true of topic identification, the effect is dramatic for language identification. The following table reports accuracies of the model for different test lengths.

5-grams, 100K training
Test Size (characters)Accuracy

N-gram Length Sensitivity

Language identification performance varies based on the n-gram length used.

32 char test, 100K training
N-gram Order Unpruned Model Size Train/Compile Time Accuracy

Training Data Sensitivity

Language identification performance varies based on the size of the training data used.

5-gram, 32 char test
Training Data Unpruned Model Size Train/Compile Time Accuracy

LingPipe could scale beyond 10M characters/language without pruning, but Sorbian only provides 9.325M characters of training data. It would also be interesting to see if these learning curves would be shaped differently for different length n-grams.


The most effective models of any given size are constructed by building larger models and then pruning them. Much smaller models than the unpruned ones reported above could be used effectively. Pruning can be carried out by either specifying the language models in the classifier constructor, or by retrieving the models using lmForCategory(String) method of DynamicLMClassifier, casting the result to the appropriate class (NGramProcessLM in this case). For instance, the following code will prune all substring counts below 5 from the models underlying the classifier.

String[] categories = ...;
DynamicLMClassifier classifier = ...;
for (int i = 0; i < categories.length; ++i) {
  NGramProcessLM lm
    = (NGramProcessLM) classifier.lmForCategory(categories[i]);
  TrieCharSequenceCounter counter
    = lm.substringCounter();

Language Set Selection

Another important consideration in language identification is the set of languages. As we saw above, some languages are simply more confusible than others. This approach would probably not even work at all for separating dialect variation (British versus American English, for instance).

Tuning Cross-Entropy per Language

Different languages have different per-character entropy rates. For instance, Chinese packs much more information into a character than Catalan. It would be possible using the finer-grained constructor for dynamic language model classifiers to allocate different n-gram lengths to different languages.

Language ID in the Wild

Language identification in realist applications is made much more difficult by a number of factors. In the ideal case, each language would use a completely distinct set of characters. Unfortunately, this isn't even true of very distant pairs such as English/Japanese.


In addition to underlying character set overlaps, trouble is caused by the borrowing of words, phrases or names.

Non-linguistic Noise

In realistic applications of language identification, matters are made even more difficult by non-language characters such as spaces, line-breaks, hyphens used as separators, tables in running text, HTML, scripts, etc. All of these may confound a language identifier if not carefully accomodated. One approach is to simply strip all non-language characters and normalize all whitespace sequences to single space characters.

Genre Mismatch

Yet another problem may be caused by mismatches between training and test data. If the training data comes from newswire and the test data is technical manuals, blog entries, etc., the genre mismatch may cause confusion by increasing cross-entropy against the model. For instance, highly technical medical reports might not resemble newswire English very closely.

Unknown Encodings

Perhaps the most difficult challenge is faced when the character encoding of the underlying text is not known. This is often the case for plain text documents or HTML documents without character set specifications.

We can use LingPipe to build classifiers that figure out encoding as well as language. LingPipe's classifiers are over character sequences, not byte sequences. Luckily, we can cast a byte to a character without loss of information. This means we can simply convert arrays of bytes to equal-length arrays of characters by casting each byte. Now all that is required is examples of the various character sets. In some cases, it's not possible to determine the character set. For instance, text that only contains ASCII characters is also valid Latin1 and UTF-8.

Multilingual Documents

Multilingual texts pose special difficulties. The assumption underlying the classifiers is that each text being classified has a unique language. The best way to handle multi-lingual documents is to segment them into sections containing a single language. Although building a segmenter is not difficult, it is not yet built into LingPipe.


Language identification has been a fairly widely studied problem. In fact, there's a vast literature. It's almost impossible to compare approaches because everyone's using different languages and different kinds of data.

Language ID on the Web

Language ID Papers