Native Language Identi cation (NLI) is the task of determining user's native language based on their second language. NLI is an important task that has applications in di erent areas such as social media analysis, authorship identi cation, second language acquisition, and forensic investigation. In this paper, we propose to use Hyperdimensional Computing (HDC) as supervised learning for identifying Indian Native Languages from the user's social media comments written in English. HDC represents language features as high dimensional vectors called hypervectors. Initially, comments are broken in character bi-grams and tri-grams which are used for generating comment hypervectors. These hypervectors are further combined to create di erent language pro le vectors. Pro le hypervectors are then used for classi cation of test comments. The proposed approach was validated on 70% of training instances with 65.13% of accuracy. The accuracy of classi cation on the test set for bi-gram and tri-gram based feature encoding is 29.2% and 31.5% respectively.
The diversity of languages used in India has given rise to a large number of challenging problems, both for the linguist and for the computer scientist. The 2001 Indian census recorded 30 languages which were spoken by more than a million native speakers in India. English plays prominent role in every sector especially in higher education in India. It is certainly the most preferred language for reading and in o cial communication. The number of second or third language speakers of English is increasing day by day. In uence of their regional and native languages is observed in their usage of English. It is possible to identify the native speaker based on their English pronunciation and accent. But, ? Indian Native Language Identi cation (INLI) is a shared task held in conjunction with FIRE 2018. Aim of the task is to identify Indian language from user's social media comments written in English. it is challenging to identify the native users based on their written comments in their non-native language. Native Language Identi cation (NLI) is important for a number of applications, such as authorship identi cation, forensic analysis, tracing linguistic in uence in multi-author texts, and to support Second Language Acquisition research. It can also be used in educational applications such as developing grammatical error correction systems. The extensive use of social media and online interactions can be a reason of violent threat, which is a common issue faced by commenters. If a comment or a post poses any type of threat, then identifying the native language of the person will be one of the signi cant feature in nding the source.
In the INLI track [
NLI challenge has gained a lot of attention and has been part of shared tasks
in various events [
Hyperdimensional Computing
The architecture of a computer that is in use since last 70 years, has been a
milestone in mechanizing arithmetic of symbols. E orts are reported in the form
of architectural proposals starting from the representation of entities, memory,
and processing units for computing [
Properties of High dimensional spaces such as holistic representation,
oneshot learning, and robustness against noise show similarity with the brain
behavior. Instead of computing with numbers, brain-inspired hyperdimensional
computing (HDC) emulates computing with hypervectors. In HDC the entities
are represented using high dimensional vectors called hypervectors. In
literature, there are di erent representation schemes proposed for HDC, for example
HRR [
INLI is a multiclass classi cation task. INLI@FIRE2018 shared task focuses on identifying native languages namely Bengali, Hindi, Kannada, Malayalam, Tamil, and Telugu, from XML les. Table 1 describes the training data that is used for the task. Training XML les are annotated as per their native language. These XML les contain extracted social media comments in English. 4
For Indian native language identi cation task, training instances of six languages
are available. For each language, the number of training instances (user
comments) are mentioned in Table 1. It is assumed that the comments are from
non-native English users. Since the comments are scribed in English (Latin)
alphabets, a preprocessing step sets rid of special characters. Preprocessing
involves removal of non-English characters, special characters and converting
the text in lowercase (alphabets). This ensures that the resulting string will
contain twenty-seven di erent symbols in total, i.e. `a,...,z' and a ` ' - blank
space. Each of these 27 symbols is assigned a hypervector, i.e. a bipolar
spatter code. Each hypervector is a random string of +1's and -1's of length 10000.
It is easy to show that with very high probability the generated 10k
dimensional hypervector will have equal number of +1's and -1's. Moreover the mean
of bitwise distance between any two randomly generated hypervectors will be
5000 bits. So the generated 27 hypervectors are all almost equidistant from
each other with almost same number of 1's and -1's For more on bipolar
spatter codes, refer to [
Algorithm 1 shows the training procedure, which takes comments (training set) as input, computes hypervectors for tri-grams and generates a class hypervector for each language. Tri-gram hypervector is generated by adding twice rotated hypervector of rst character, once rotated hypervector of second character and third character's hypervectors. Similarly we can create class hypervector using bi-gram.
Algorithm 2 is used for predicting the class of test instances. The query hypervector (Q) is generated by similar process as the language hypervector is generated, i.e. superposition of tri-gram hypervectors corresponding to test/query instance. This query hypervector is compared with all the language hypervectors. The similarity between bipolar hypervectors is calculated using the cosine of the angle between them. The language which has the highest similarity (lowest angle) with query hypervector is written in the output le with the test dataset name.
For validation of classi cation algorithm, we have considered 70% as training and 30% as validation instances for given training dataset. Fig. 2. shows the normalized confusion matrix corresponding validation dataset. Data is crossvalidated 10 times and average accuracy reported is 65.13%.
Normalized confusion matrix BE 0.48 0.25 0.00 0.16 0.07 0.05
Run1 and Run2 results are corresponding to the bi-gram and the tri-gram language pro le vectors respectively. The results are presented in Table 2 and Table 3 in which Table 2 and 3 represent result of TestSet-1 and TestSet-2. TestSet-1 is used for comparing results with previous year's shared task. In Run1 (bi-gram) 32.4% and Run2 (tri-gram) 32.1% accuracy is reported. For Run1 and Run2 for Testset-2 accuracy is 29.2% and 31.5% respectively. These experiments are conducted on 3.6 GHz Intel Core i7 PC with 4 GB of RAM. The python implementation of 3-gram training algorithm recorded 2.81667 minutes. The validation results di er from test results primarily due to the presence of unseen tri-gram (bi-gram) in the test data while training phase. 6
In this paper, we have proposed HDC approach for Indian Native Language Identi cation. The results are encouraging and can be enhanced further. Features selected for classi cation are tri (bi)-grams which are easy to aggregate. Since the e orts were put on proof of concept of HDC based classi cation, the other aspects of feature selection and e cient implementations were not addressed in this work. We believe that by deleting or rewriting lexical terms such as pls (please), u (you), y (why), r (are), sry (sorry), fyi (for your information), etc. will improve the accuracy of the system. We also believe that punctuation marks may have a role to play in natural language identi cation. In future, we would like to include such features in HDC framework. We also would like to compare the proposed approach with state of the art methods in term of training e ort, precision, recall and accuracy.