Native Language Identi cation is the problem of identifying the rst language of speakers based on his/her writings in another language. The proposed approach is a deep learning based methodology using convolutional neural networks. Convolutional neural networks are a class of neural networks that have proven very e ective in areas such as pattern recognition and classi cation. They are able to capture the local texture within the text and can be used to nd the representative patterns in a text document. The proposed system consists of a language identi cation model, which is trained by a corpus of 1233 documents. The experiments were conducted using the dataset provided for INLI@FIRE2018. The results indicate that the system is capable of giving performance comparable to the methods employing more sophisticated approaches.
Native Language Identi cation is the process of distinguishing the native
language of a writer from his/her writings in the second language(English) [
Hindi is by far the most widely spoken language in India. Even though
roughly 40% of the population speak Hindi, people use English as their
major second language. English is spoken natively by around 375 million people
across the globe. It is the second o cial language of India and is used for
business, teaching, learning, and trade on a day to day basis. Around 10% of Indias
population speak English and use it in their day to day activities. But it is only a
rst language for 0.019% people in the Country and becoming a second language
for around 125 million people all over the world [
The structure of this paper is as follows. Section 2 brie y reviews the similar works in this area. Section 3 discusses the task description and details about the dataset. Section 4 explains the methodology and Section 5 demonstrates the results and evaluation metrics. Section 6 concludes the article along with some routes for the future works. 2
Native Language Identi cation has a lot of importance in di erent areas of
Natural Language Processing [
Syntactic features of the text are also focussed on the recent works. Wong and
Dras [
It has been found that the semantic features are the least experimented
one for NLI. Gamon extracted semantic features from semantic dependency
graphs[
The task is focused on identifying the rst language of an author from the given Text/XML le which includes a set of Facebook comments in the English language. Six Indian languages are considered for this study. They are Tamil, Hindi, Kannada, Malayalam, Bengali, and Telugu. Spoken forms of English shows signi cant variations across the di erent states of India and it is relatively easy to recognize the native language of the speaker using his English accent. But nding the rst language of a writer based on his comments or posts in English is a di cult task in the present scenario.
The shared dataset contains data from six di erent Indian languages. The
training data is a set of les in XML format. Each language has around 200 les
of facebook comments. Each le contains around 150 words as the comment.
Sentence segmentation is carried out using the regular expression. Statistics of
the training data is shown in Table 1. The testing data contains two folders say
test1 and test2. Test1 consists of 783 les and test2 contains 1185 les from the
above-mentioned languages.
The proposed system is a CNN-based language identi cation model which
predicts the native language of a writer from his scripts. CNN's are responsible for
the important breakthroughs in Image Classi cation problems and are the core
of the most Computer Vision systems today. But they are not common in text
analytics. CNN's have been proved to be successful in various text classi cation
problems in recent years [
The problem is shaped as a text classi cation task with language names as labels(classes). The number of classes is the same as the number of languages considered for the study. Text from each language in the training data is sent to a sentence segmentation module where the raw text is converted to a set of sentences using regular expressions. Providing sequences of raw human-alike words will make no sense to computers. For that reason, the raw words are converted into numeric values using dictionaries. For that, we create a vocabulary of words, an array which stores all the words in the training data, but each word appears only once. Two dictionaries which map from word to its corresponding index value and reverse are also created. Two special words-'ZERO' and 'UNKNOWN' are added to the dictionary. 'ZERO' is used to make all the sequences of unique length and 'UNKNOWN' is used for out of vocabulary words. Then the sequences of strings are converted into sequences of numbers using the aforementioned dictionaries. The sentences may have a di erent length. But CNN training requires sequences of uniform length. So we padded the sentences with less number of words with 'ZEROS' to make them of unique length(ZERO padding). That is why we added the word ZERO to the dictionary. Each sentence in the training data is labeled with a corresponding language label. Hence our nal training data contains a lot of sentences and their corresponding labels. Identifying the patterns within the sentences is our ultimate goal.
Sequential model of keras is used for implementation [
Di erent con gurations of the network are attempted. Experiments are
conducted using deep and shallow convolutional neural networks. The performance
of di erent CNN architectures on the test data is given in Table 3. The best
results are given by the above-described architecture. In our experiments, we
selected the rst 90% of the data as training data and the remaining data as
testing data. The batch size is xed at 64. 'Categorical cross entropy' is used
as the loss function. We used 'Adam', the e cient gradient descent algorithm
as the optimizer because it is an e cient one for optimization. Dropout is used
to prevent over tting [
Experiments are also conducted to measure the e ect of training data size on the system performance. It is observed that the performance of the system increases Name Network Con guration Accuracy CNN1 1 Conv,1 Maxpool,1 Dense,1 Dropout 17.5% CNN2 2 Conv,2 Maxpool,2 Dense,1 Dropout 21.2% CNN3 3 Conv,2 Maxpool,2 Dense,2 Dropout 22.2% CNN4 4 Conv,3 Maxpool,2 Dense,2 Dropout 25.7% CNN5 4 Conv,4 Maxpool,2 Dense,2 Dropout 25.3%
CNN6 5 Conv,5 Maxpool,2 Dense,2 Dropout 24.7% with the increase in training data size. Figure 2 shows the e ect of training data size on our best performed CNN architecture. Hence it is better to have a larger sized training corpus when dealing with deep learning based classi cation methodologies. We used accuracy to quantify the performance of our model. Accuracy computes the degree to which the result of a prediction conforms to the true value. The proposed system was tested with the test data sets provided by the task organizers. Our system predicts the tag for each sentence in the post(comment). But our goal is to predict the tag for each XML le(post). So we labeled each post according to the maximum number of predictions for that particular post. Table 4 demonstrates the results of our experimentation on both the datasets. It is clear from the table that test2 results are far better than test1 results. Di erent runs correspond to di erent architecture of the proposed network. In this article, we have discussed a deep learning based native language identi cation system. The exclusive feature of our approach is the use of the Convolutional neural network for this task. The main reason we preferred a CNN rather than traditional feature-based methods is its ability to capture local texture in a sequence. It has been found that the accuracy of the system increases with the increase in training data size. Hence it is better to have a larger sized training corpus to get improved performance. The accuracy of the system can also be improved by using trained word embeddings. Due to insu cient system requirements, we could not perform this activity. Apart from NLI, Convolutional Neural Networks can be applied e ciently for various language processing applications. We hope to apply CNN based methods to di erent language processing applications such as text classi cation, sentiment analysis, etc.