Volume 28, Issue 3 - Special Issue: Neural Machine Translation. Issue Editors: M. Anand Kumar, K. P. Soman, S. Rajendran

Sentiment analysis refers to determining the polarity of the opinions represented by text. The paper proposes an approach to determine the sentiments of tweets in one of the Indian languages (Hindi, Bengali, and Tamil). Thirty-nine sequential models have been created using three different neural network layers [recurrent neural networks (RNNs), long short-term memory (LSTM), convolutional neural network (CNN)] with optimum parameter settings (to avoid over-fitting and error accumulation). These sequential models have been investigated for each of the three languages. The proposed sequential models are experimented to identify how the hidden layers affect the overall performance of the approach. A comparison has also been performed with existing approaches to find out if neural networks have an added advantage over traditional machine learning techniques.

Sentiment polarity detection is one of the most popular sentiment analysis tasks. Sentiment polarity detection in tweets is a more difficult task than sentiment polarity detection in review documents, because tweets are relatively short and they contain limited contextual information. Although the amount of blog posts, tweets and comments in Indian languages is rapidly increasing on the web, research on sentiment analysis in Indian languages is at the early stage. In this paper, we present an approach that classifies the sentiment polarity of Bengali tweets using deep neural networks which consist of one convolutional layer, one hidden layer and one output layer, which is a soft-max layer. Our proposed approach has been tested on the Bengali tweet dataset released for Sentiment Analysis in Indian Languages contest 2015. We have compared the performance of our proposed convolutional neural networks (CNN)-based model with a sentiment polarity detection model that uses deep belief networks (DBN). Our experiments reveal that the performance of our proposed CNN-based system is better than our implemented DBN-based system and some existing Bengali sentiment polarity detection systems.

Introduction of deep neural networks to the machine translation research ameliorated conventional machine translation systems in multiple ways, specifically in terms of translation quality. The ability of deep neural networks to learn a sensible representation of words is one of the major reasons for this improvement. Despite machine translation using deep neural architecture is showing state-of-the-art results in translating European languages, we cannot directly apply these algorithms in Indian languages mainly because of two reasons: unavailability of the good corpus and Indian languages are morphologically rich. In this paper, we propose a neural machine translation (NMT) system for four language pairs: English–Malayalam, English–Hindi, English–Tamil, and English–Punjabi. We also collected sentences from different sources and cleaned them to make four parallel corpora for each of the language pairs, and then used them to model the translation system. The encoder network in the NMT architecture was designed with long short-term memory (LSTM) networks and bi-directional recurrent neural networks (Bi-RNN). Evaluation of the obtained models was performed both automatically and manually. For automatic evaluation, the bilingual evaluation understudy (BLEU) score was used, and for manual evaluation, three metrics such as adequacy, fluency, and overall ranking were used. Analysis of the results showed the presence of lengthy sentences in English–Malayalam, and the English–Hindi corpus affected the translation. Attention mechanism was employed with a view to addressing the problem of translating lengthy sentences (sentences contain more than 50 words), and the system was able to perceive long-term contexts in the sentences.

This article addresses language identification at the word level in Indian social media corpora taken from Facebook, Twitter and WhatsApp posts that exhibit code-mixing between English-Hindi, English-Bengali, as well as a blend of both language pairs. Code-mixing is a fusion of multiple languages previously mainly associated with spoken language, but which social media users also deploy when communicating in ways that tend to be rather casual. The coarse nature of code-mixed social media text makes language identification challenging. Here, the performance of deep learning on this task is compared to feature-based learning, with two Recursive Neural Network techniques, Long Short Term Memory (LSTM) and bidirectional LSTM, being contrasted to a Conditional Random Fields (CRF) classifier. The results show the deep learners outscoring the CRF, with the bidirectional LSTM demonstrating the best language identification performance.

This paper presents a comprehensive analysis and comparison of various proposed sequential models based on different deep networks such as the convolutional neural network, long short-term memory, and recurrent neural network. The different sequential models are analyzed based on the number of layers, the number of output dimensions, order, and the combination of different deep network architectures. The proposed approach is compared to a baseline model based on traditional machine learning techniques.

The paper addresses the problem of part-of-speech (POS) tagging for Malayalam tweets. The conversational style of posts/tweets/text in social media data poses a challenge in using general POS tagset for tagging the text. For the current work, a tagset was designed that contains 17 coarse tags and 9915 tweets were tagged manually for experiment and evaluation. The tagged data were evaluated using sequential deep learning methods like recurrent neural network (RNN), gated recurrent units (GRU), long short-term memory (LSTM), and bidirectional LSTM (BLSTM). The training of the model was performed on the tagged tweets, at word level and character level. The experiments were evaluated using measures like precision, recall, f1-measure, and accuracy. During the experiment, it was found that the GRU-based deep learning sequential model at word level gave the highest f1-measure of 0.9254; at character-level, the BLSTM-based deep learning sequential model gave the highest f1-measure of 0.8739. To choose the suitable number of hidden states, we varied it as 4, 16, 32, and 64, and performed training for each. It was observed that the increase in hidden states improved the tagger model. This is an initial work to perform Malayalam Twitter data POS tagging using deep learning sequential models.

English to Indian language machine translation poses the challenge of structural and morphological divergence. This paper describes English to Indian language statistical machine translation using preordering and suffix separation. The preordering uses rules to transfer the structure of the source sentences prior to training and translation. This syntactic restructuring helps statistical machine translation to tackle the structural divergence and hence provides better translation quality. The suffix separation is used to tackle the morphological divergence between English and highly agglutinative Indian languages. We demonstrate that the use of preordering and suffix separation helps in improving the quality of English to Indian language machine translation.

Machine translation (MT) is the automatic translation of the source language to its target language by a computer system. In the current paper, we propose an approach of using recurrent neural networks (RNNs) over traditional statistical MT (SMT). We compare the performance of the phrase table of SMT to the performance of the proposed RNN and in turn improve the quality of the MT output. This work has been done as a part of the shared task problem provided by the MTIL2017. We have constructed the traditional MT model using Moses toolkit and have additionally enriched the language model using external data sets. Thereafter, we have ranked the phrase tables using an RNN encoder-decoder module created originally as a part of the GroundHog project of LISA lab.

In recent years, the multilingual content over the internet has grown exponentially together with the evolution of the internet. The usage of multilingual content is excluded from the regional language users because of the language barrier. So, machine translation between languages is the only possible solution to make these contents available for regional language users. Machine translation is the process of translating a text from one language to another. The machine translation system has been investigated well already in English and other European languages. However, it is still a nascent stage for Indian languages. This paper presents an overview of the Machine Translation in Indian Languages shared task conducted on September 7–8, 2017, at Amrita Vishwa Vidyapeetham, Coimbatore, India. This machine translation shared task in Indian languages is mainly focused on the development of English-Tamil, English-Hindi, English-Malayalam and English-Punjabi language pairs. This shared task aims at the following objectives: (a) to examine the state-of-the-art machine translation systems when translating from English to Indian languages; (b) to investigate the challenges faced in translating between English to Indian languages; (c) to create an open-source parallel corpus for Indian languages, which is lacking. Evaluating machine translation output is another challenging task especially for Indian languages. In this shared task, we have evaluated the participant’s outputs with the help of human annotators. As far as we know, this is the first shared task which depends completely on the human evaluation.

Machine Translation bridges communication barriers and eases interaction among people having different linguistic backgrounds. Machine Translation mechanisms exploit a range of techniques and linguistic resources for translation prediction. Neural machine translation (NMT), in particular, seeks optimality in translation through training of neural network, using a parallel corpus having a considerable number of instances in the form of a parallel running source and target sentences. Easy availability of parallel corpora for major Indian language forms and the ability of NMT systems to better analyze context and produce fluent translation make NMT a prominent choice for the translation of Indian languages. We have trained, tested, and analyzed NMT systems for English to Tamil, English to Hindi, and English to Punjabi translations. Predicted translations have been evaluated using Bilingual Evaluation Understudy and by human evaluators to assess the quality of translation in terms of its adequacy, fluency, and correspondence with human-predicted translation.

Machine translation (MT) from English to foreign languages is a fast developing area of research, and various techniques of translation are discussed in the literature. However, translation from English to Malayalam, a Dravidian language, is still in the rising stage, and works in this field have not flourished to a great extent, so far. The main reason of this shortcoming is the non-availability of linguistic resources and translation tools in the Malayalam language. A parallel corpus with alignment is one of such resources that are essential for a machine translator system. This paper focuses on a technique that enables automatic setting up of a verb-aligned parallel corpus by exploring the internal structure of the English and Malayalam language, which in turn facilitates the task of machine translation from English to Malayalam.

Building an automatic, high-quality, robust machine translation (MT) system is a fascinating yet an arduous task, as one of the major difficulties lies in cross-linguistic differences or divergences between languages at various levels. The existence of translation divergence precludes straightforward mapping in the MT system. An increase in the number of divergences also increases the complexity, especially in linguistically motivated transfer-based MT systems. This paper discusses the development of Telugu-Tamil transfer-based MT and how a divergence index (DI) is built to quantify the number of parametric variations between languages in order to improve the success rate of MT. The DI facilitates MT in proposing where to put efforts for the given language pair to attain better and faster results. In addition, handling strategies of different types of divergences in a transfer-based approach to MT are discussed. The paper also includes the evaluation method and how an improvization takes place with the application of DI in MT.