Applied for a visiting position, I was asked to implement a method for language identification of a document. Though simple, it works well. So, I put it here.

My approach for identifying the language of a text is to utilize character n-grams. Fortunately, it is easy to collect training data for this task. For each language we only need a number of documents. Once we have training texts we can use them and extract features. In my method, features are character n-grams. N-grams are extracted for each language and their frequencies are calculated. There are several machine learning methods which can be applied for language identification. I exploit Näive Bayes, a simple yet very effective method. Näive Bayes uses the probability of n-grams to identify the language of a new text.

For language identification a model of languages should be created. In the code, a list of text files, each containing training texts for a language, is read at the begining. The name of the files indicates languages. To add a language, one can simply add another text file and put some texts in it. These files are used to create the language identification model. For each language, a list of character n-grams and their frequencies are extracted and saved into a file. The default maximum length of n-grams is set to 3.

In the language identification phase, files containing character n-grams and their frequencies are read into a Map. As the total frequency of n-grams are required to calculate probabilities, the total frequencies are also saved into a Map. To identify the language of an input text, first it is converted to a list of n-grams. Afterwards, p(Language|T ext) is computed using Näive Bayes. The prior probabilities of languages are set equal to each other as there is no information about the domain the code may be used in. I also apply Laplace method to smooth p(n-gram|Language). The language with the maximum probability p(Language|T ext) is considered as the language of the given text.

Parameters of the code are the maximum length of n-grams, the path to the language text files, and the path to the language n-grams files. The default value of the parameters is set in a property file called properties.prop. The test code performs on four languages: English, French, Persian and Arabic.

MIT