Sentiment Analysis (SA) is a field of computational study that focuses on analyzing and understanding people’s opinions, attitudes, and emotions towards an entity. An entity could be an individual, an event, a topic, a product etc., which is most likely to be covered by reviews and such reviews can be found in abundance on social media platforms. The increase in the number of social media users and the growing amount of user-generated code-mixed content such as reviews, comments, posts etc., on social media have resulted in a rising demand for efficient tools capable of effectively analyzing such content to detect the sentiments. However, SA of social media text is challenging due to the complex nature of the code-mixed text. To tackle this issue, in this paper, we team MUCS, describe learning models submitted to “Sentiment Analysis in Tamil and Tulu” -DravidianLangTech@Recent Advances In Natural Language Processing (RANLP) 2023. Using fastText embeddings to train the Machine Learning (ML) models to perform SA in code-mixed Tamil and Tulu texts, the proposed methodology exhibited F1 scores of 0.14 and 0.204 securing 13th and 15th rank for Tamil and Tulu texts respectively.

The task of Language Identification (LI) in text processing refers to automatically identifying the languages used in a text document. LI task is usually been studied at the document level and often in high-resource languages while giving less importance to low-resource languages. However, with the recent advance- ment in technologies, in a multilingual country like India, many low-resource language users post their comments using English and one or more language(s) in the form of code-mixed texts. Combination of Kannada and English is one such code-mixed text of mixing Kannada and English languages at various levels. To address the word level LI in code-mixed text, in CoLI-Kanglish shared task, we have focused on open-sourcing a Kannada-English code-mixed dataset for word level LI of Kannada, English and mixed-language words written in Roman script. The task includes classifying each word in the given text into one of six predefined categories, namely: Kannada (kn), English (en), Kannada-English (kn-en), Name (name), Lo-cation (location), and Other (other). Among the models submitted by all the participants, the best performing model obtained averaged-weighted and averaged-macro F1 scores of 0.86 and 0.62 respectively.

Substantial amount of text data which is increasingly being generated and shared on the internet and social media every second affect the society positively or negatively almost in any aspect of online world and also business and industries. Sentiments/opinions/reviews’ of users posted on social media are the valuable information that have motivated researchers to analyze them to get better insight and feedbacks about any product such as a video in Instagram, a movie in Netflix, or even new brand car introduced by BMW. Sentiments are usually written using a combination of languages such as English which is resource rich and regional languages such as Tamil, Kannada, Malayalam, etc. which are resource poor. However, due to technical constraints, many users prefer to pen their opinions in Roman script. These kinds of texts written in two or more languages using a common language script or different language scripts are called code-mixing texts. Code-mixed texts are increasing day-by-day with the increase in the number of users depending on various online platforms. Analyzing such texts pose a real challenge for the researchers. In view of the challenges posed by the code-mixed texts, this paper describes three proposed models namely, SACo-Ensemble, SACo-Keras, and SACo-ULMFiT using Machine Learning (ML), Deep Learning (DL), and Transfer Learning (TL) approaches respectively for the task of Sentiments Analysis in Tamil-English and Malayalam-English code-mixed texts.

This paper describes the models submitted by the team MUCS for Offensive Language Identification in Dravidian Languages-EACL 2021 shared task that aims at identifying and classifying code-mixed texts of three language pairs namely, Kannada-English (Kn-En), Malayalam-English (Ma-En), and Tamil-English (Ta-En) into six predefined categories (5 categories in Ma-En language pair). Two models, namely, COOLI-Ensemble and COOLI-Keras are trained with the char sequences extracted from the sentences combined with words as features. Out of the two proposed models, COOLI-Ensemble model (best among our models) obtained first rank for Ma-En language pair with 0.97 weighted F1-score and fourth and sixth ranks with 0.75 and 0.69 weighted F1-score for Ta-En and Kn-En language pairs respectively.

This paper describes the models submitted by the team MUCS for “Hope Speech Detection for Equality, Diversity, and Inclusion-EACL 2021” shared task that aims at classifying a comment / post in English and code-mixed texts in two language pairs, namely, Tamil-English (Ta-En) and Malayalam-English (Ma-En) into one of the three predefined categories, namely, “Hope_speech”, “Non_hope_speech”, and “other_languages”. Three models namely, CoHope-ML, CoHope-NN, and CoHope-TL based on Ensemble of classifiers, Keras Neural Network (NN) and BiLSTM with Conv1d model respectively are proposed for the shared task. CoHope-ML, CoHope-NN models are trained on a feature set comprised of char sequences extracted from sentences combined with words for Ma-En and Ta-En code-mixed texts and a combination of word and char ngrams along with syntactic word ngrams for English text. CoHope-TL model consists of three major parts: training tokenizer, BERT Language Model (LM) training and then using pre-trained BERT LM as weights in BiLSTM-Conv1d model. Out of three proposed models, CoHope-ML model (best among our models) obtained 1st, 2nd, and 3rd ranks with weighted F1-scores of 0.85, 0.92, and 0.59 for Ma-En, English and Ta-En texts respectively.

The increase in domain specific text processing applications are demanding tools and techniques for domain specific Text Classification (TC) which may be helpful in many downstream applications like Machine Translation, Summarization, Question Answering etc. Further, many TC algorithms are applied on globally recognized languages like English giving less importance for local languages particularly Indian languages. To boost the research for technical domains and text processing activities in Indian languages, a shared task named ”TechDOfication2020” is organized by ICON’20. The objective of this shared task is to automatically identify the technical domain of a given text which provides information about coarse grained technical domains and fine grained subdomains in eight languages. To tackle this challenge we, team MUCS have proposed three models, namely, DL-FineTuned model applied for all subtasks, and VC-FineTuned and VC-ngrams models applied only for some subtasks. n-grams and word embedding with a step of fine-tuning are used as features and machine learning and deep learning algorithms are used as classifiers in the proposed models. The proposed models outperformed in most of subtasks and also obtained first rank in subTask1b (Bangla) and subTask1e (Malayalam) with f1 score of 0.8353 and 0.3851 respectively using DL-FineTuned model for both the subtasks.

Machine Translation (MT) is the task of automatically converting the text in source language to text in target language by preserving the meaning. MT usually require large corpus for training the translation models. Due to scarcity of resources very less attention is given to translating into low resource languages and in particular into Indic languages. In this direction, a shared task called “Adap-MT 2020: Low Resource Domain Adaptation for Indic Machine Translation” is organized to illustrate the capability of general domain MT when translating into Indic languages and low resource domain adaptation of MT systems. In this paper, we, team MUCS, describe a simple word extraction based domain adaptation approach applied to English-Hindi MT only. MT in the proposed model is carried out using Open-NMT - a popular Neural Machine Translation tool. A general domain corpus is built effectively combining the available English-Hindi corpora and removing the duplicate sentences. Further, domain specific corpus is updated by extracting the sentences from generic corpus that contains the words given in the domain specific corpus. The proposed model exhibited satisfactory results for small domain specific AI and CHE corpora provided by the organizers in terms of BLEU score with 1.25 and 2.72 respectively. Further, this methodology is quite generic and can easily be extended to other low resource language pairs as well.

This paper describes our first experiment on Neural Machine Translation (NMT) based query translation for Amharic-Arabic Cross-Language Information Retrieval (CLIR) task to retrieve relevant documents from Amharic and Arabic text collections in response to a query expressed in the Amharic language. We used a pre-trained NMT model to map a query in the source language into an equivalent query in the target language. The relevant documents are then retrieved using a Language Modeling (LM) based retrieval algorithm. Experiments are conducted on four conventional IR models, namely Uni-gram and Bi-gram LM, Probabilistic model, and Vector Space Model (VSM). The results obtained illustrate that the proposed Uni-gram LM outperforms all other models for both Amharic and Arabic language document collections.