A method for determining if a block of text is positive, neutral, or negative is sentiment analysis. As code-mixed material in many native languages is becoming increasingly widespread, there is also an increasing need for intense research in order to produce satisfactory results. This research paper aims to classify the sentiments from a data set of comments/posts into pre-defined classes belonging to the code-mixed text in Tamil, Malayalam, and Kannada by utilizing the Sequential Deep Learning model on the code-mixed data set. The sequential model achieved an f1-score of 0.20 for Tamil-English, 0.48 for Malayalam-English, and 0.47 for Kannada-English data sets. The results were submitted to the competition 'Shared Task on Sentiment Analysis and Homophobia detection of YouTube comments in Code-Mixed Dravidian Languages' organized by DravidianLangTech.
Sentiment analysis at the word level examines how the words or phrases in the text are oriented
and how that afects the overall mood, whereas sentiment analysis at the sentence level examines
sentences that represent one perspective and makes an efort to identify its direction. The
foundation of a lexicon-based method is a corpus or list of words with a particular diference.
Then, an algorithm searches for specific words, measures their weight or counts them, and
determines the overall duality of the text [
In a country where several diferent languages are spoken, code-mixing becomes commonplace. People who live in multilingual countries employ code-mixed discourse when interacting online and in person. Code-mixing is ”the incorporation of linguistic forms from one language, such as phrases, syllables, and morphemes into an expression of a diferent language.” There have already been many experiments that were conducted to make use of sentiment analysis in monolingual texts and they have been successful [15, 25]. But there have been far fewer studies conducted for the sentiment analysis of code-mixed languages [16]. The task given in the competition is to categorize the feelings expressed in the code-mixed language data set’s YouTube comments. The objective of our study is to classify YouTube comments into the following classes: Positive, negative, unknown_state, mixed-feelings, or if the word is not in that respective language of that particular code-mixed language.
Social media corporations have always been required to fund/contribute to sentimental analysis
research to protect social media users from cyberbullying. There have been numerous studies
that have looked into sentiment analysis models. However, just a few research papers have
looked into the use of Emoji characters on social media [
The data set used here is provided by Task A proposed [19]. It is a collection of YouTube
comments in Tamil, Malayalam, and Kannada (data for all 3 from [
Due to the code-mixing and blatant disregard for grammatical rules in the data set downloaded from YouTube. To use the data set efectively, the following processes are applied. • Initially emojis, special characters, numbers, and punctuation were all eliminated as they have no functional use to a statement. • We lowercase all the characters and replace usernames with empty characters. • We then split the comment into tokens. • Next, we made a flat list of all words from the corpus and then we computed the number of occurrences of all the words in the corpus. • Padding is applied to the corpus next. • We created a sequential model and passed the required layers to it.
• To avoid over-fitting our model, we have used Early Stopping as well.
A Sequential DNN was developed for the sentiment analysis tasks. These networks received their input from the embedding vectors. The text indexes are transformed into dense vectors with defined sizes. The input length, embedding initializer, and embedding Regularizer assigned were all of ”maximum length,” ”orthogonal,” and ”L2 Regularizer,” respectively. Next, we added the LSTM layer and wrapped the layer with Bidirectional. The bi-directionality of a Keras layer was added to the model by implementing tf.keras.layers.bidirectional to the model. Finally, we used the Dense layer to classify the data into the 5 classes and used the ‘softmax’ activation function. We compiled our model and defined the loss function, optimizer, and metrics. We pick “Categorical Cross-Entropy” as the loss function because the provided task requires multi-class categorization. For the given task, we used the default optimizer “Adam” and set the learning rate to 0.01.
The model is trained in order to fine-tune the parameters to produce the desired outputs for a particular input. This is accomplished by putting inputs into the input layer, receiving an output, computing the loss function using the output, and then fine-tuning the model parameters using back-propagation. As a result, the model’s parameters will be fit and matched to the data. The batch size for the model while fitting was 256 and the number of epochs was 2. Table 2 displays the several parameters used in the Sequential model.
All the required modules and packages like TensorFlow, pandas, NumPy, Regular Expression, Natural Language Toolkit, scikit-learn, etc. are all imported to the notebook file. The feature extraction and model training is done in Python using the scikit-learn library. The text data is transformed into TF-IDF feature vectors using the scikit-learn Tfidf Vectorizer.
The metrics precision, recall, and f1-score have been calculated for the code-mixed data sets. A precision of 0.22, a recall of 0.18, and an f1-score of 0.20 were recorded for the Tamil-English data set as shown in Table 3.
A precision of 0.51, a recall of 0.57, and an f1-score of 0.48 were recorded for the MalayalamEnglish data set as shown in Table 4.
A precision of 0.48, a recall of 0.50, and an f1-score of 0.47 were recorded for the KannadaEnglish data set as shown in Table 5. The comparison between our results and other competitors’ results was shown in Tables 3,4 and 5 respectively. As a result, we tested with the sequential DNN model for the three code-mixed data sets. Any language can be used with this approach because it is language-independent.
Comparatively, we can see that the Malayalam-English data set had the highest f1-score, precision, and recall, and the Tamil-English data set had the lowest f1-score, precision, and recall with Kannada-English data set having precision and f1-score just slightly lesser than that of Malayalam. However, when compared to the other two data sets, the Tamil-English data set appeared to perform poorly with the model. However, it should be noticed that compared to the other classes, the ’Positive’ class in the Tamil-English data set has significantly too many instances. The imbalance in the data has resulted in lower accuracy. The training and development data was substantial in comparison to the other two languages, which contributed to the poor scores compared to Malayalam-English and Kannada-English data sets. ment Analysis of YouTube Comments for Dravidian Languages. In Forum for Information Retrieval Evaluation. [14] Kumar, A., Saumya, S., & Singh, J. P. (2020). NITP-AI-NLP@ Dravidian-CodeMix-FIRE2020: A Hybrid CNN and Bi-LSTM Network for Sentiment Analysis of Dravidian Code-Mixed Social Media Posts. In FIRE (Working Notes) (pp. 582-590). [15] Vilares, D., Alonso, M. A., & Gómez-Rodríguez, C. (2015, September). Sentiment analysis on monolingual, multilingual and code-switching twitter corpora. In Proceedings of the 6th workshop on computational approaches to subjectivity, sentiment and social media analysis (pp. 2-8). [16] Patra, B. G., Das, D., & Das, A. (2018). Sentiment analysis of code-mixed indian languages:
An overview of sail_code-mixed shared task@ icon-2017. arXiv preprint arXiv:1803.06745. [17] Guibon, G., Ochs, M., & Bellot, P. (2016, June). From emojis to sentiment analysis. In
WACAI 2016. [18] Yamashita, R., Nishio, M., Do, R. K. G., & Togashi, K. (2018). Convolutional neural networks: an overview and application in radiology. Insights into imaging, 9(4), 611-629. [19] Chakravarthi, B.R., Priyadharshini, R., Ponnusamy, R., Kumaresan, P.K., Sampath, K., Thenmozhi, D., Thangasamy, S., Nallathambi, R. and McCrae, J.P., 2021. Dataset for Identification of Homophobia and Transophobia in Multilingual YouTube Comments. arXiv preprint arXiv:2109.00227. [20] Shumugavadivel, Kogilavani and Subramanian, Malliga and Kumaresan, Prasanna Kumar and Chakravarthi, Bharathi Raja and B, Bharathi and Chinnaudayar Navaneethakrishnan, Subalalitha and S.K, Lavanya and Mandl, Thomas and Ponnusamy, Rahul and Palanikumar, Vasanth and Balaji J, Manoj. Overview of the Shared Task on Sentiment Analysis and Homophobia Detection of YouTube Comments in Code-Mixed Dravidian Languages. In proceedings of dravidiancodemix-2022. [21] Shanmugavadivel, K., Sampath, S.H., Nandhakumar, P., Mahalingam, P., Subramanian, M., Kumaresan, P.K. and Priyadharshini, R., 2022. An analysis of machine learning models for sentiment analysis of Tamil code-mixed data. Computer Speech Language, p.101407. [22] Chakravarthi, B.R., Hande, A., Ponnusamy, R., Kumaresan, P.K. and Priyadharshini, R., 2022. How can we detect Homophobia and Transphobia? Experiments in a multilingual code-mixed setting for social media governance. International Journal of Information Management Data Insights, 2(2), p.100119. [23] Chakravarthi, B.R., 2022. Hope speech detection in YouTube comments. Social Network
Analysis and Mining, 12(1), pp.1-19. [24] Chakravarthi, B.R., 2022. Multilingual hope speech detection in English and Dravidian languages. International Journal of Data Science and Analytics, 14(4), pp.389-406. [25] Subramanian, M., Ponnusamy, R., Benhur, S., Shanmugavadivel, K., Ganesan, A., Ravi, D., Shanmugasundaram, G.K., Priyadharshini, R. and Chakravarthi, B.R., 2022. Ofensive language detection in Tamil YouTube comments by adapters and cross-domain knowledge transfer. Computer Speech Language, 76, p.101404.