Sentiment analysis is a method for determining the positivity, negativity, or neutrality of a text block. The purpose of sentiment analysis is to study public sentiment in a manner that promotes company growth. This study seeks to classify the feelings of a dataset of comments/posts into pre-defined classifications for the code-mixed languages Tamil, Malayalam, and Kannada. The Sequential Deep Learning model is used to the code-mixed dataset to identify sentiments. The experiment was carried out using the dataset from the Codalab 2022 competition "Shared Task on Sentiment Analysis and Homophobia Detection of YouTube Comments in Code-Mixed Dravidian Languages", which included social media comments in Tamil, Malayalam, and Kannada code-mixed languages.
For the purpose of protecting social media users from cyberbullying, social media companies
have always been required to fund/contribute to sentiment analysis research. There have been
a number of studies examining models for sentiment analysis. Different fields utilize different
methodologies and models. However, few academic studies have examined the use of Emoji
characters on social media [
Recent works on sentiment analysis of mixed-code formats have expanded in number [
Dravidian languages have been code-mixed with English in the current study, such as
"TamilEnglish", "Malayalam-English", and "Kannada-English" [
The dataset utilized in this study is provided by Task A of "Sentiment Analysis and
Homophobia detection in YouTube comments." It includes YouTube comments written in
Tamil, Malayalam, and Kannada (data for all 3 from [
In this paper, a multi-task classification model for sentiment analysis of YouTube comments written in mixed-Dravidian code is developed. Each comment in the dataset must be represented by a numerical feature vector for a supervised classifier to be trained. 4.1 Task The objective is to classify each YouTube remark into one of five classes: "Positive", "Negative", "unknown state", "Mixed feelings", and "not-Tamil/Malayalam/Kannada".
Given that the YouTube dataset is code-mixed and defies grammatical standards. To successfully utilize the dataset, the following procedures are implemented.
The texts are initially transformed to lowercase and stemming and lemmatization are performed. In the following phase, all emojis, special characters, numbers, and punctuation must be removed because they serve no use in a statement. Sentences of two letters or less were deleted since they had minimal impact on the data set. Next, the training dataset was compiled. After cleaning the text, it was tokenized and encoded into a collection of token indexes.
Finally, padding was used to verify that all texts were of equal length.
For sentiment analysis tasks, a DNN has been built. The input for these networks came from the embedding vectors. Initially, word embedding was included to the model. The embedding initializer, embedding regularizer, and embedding regularizer were all assigned as "maximum length," "orthogonal," and "L2 Regularizer" After adding the LSTM layer, we wrapped it with Bidirectional. Bidirectionality was added to a Keras layer by implementing tf.keras.layers.bidirectional within the model. We classified the data into five classes using the Dense layer and the'softmax' activation function..
Our model was compiled and the loss function, optimizer, and metrics were defined. We select "Categorical Cross-Entropy" as the loss function since the provided problem involves multiclass categorization. We applied the default optimizer Adam and a learning rate of 0.01 to the provided problem. We had previously used 'accuracy,' 'precision,"recall,' and 'auc' as measures..
Consequently, we must now train our model to fine-tune the parameters in order to provide the required outputs for a given input. This is achieved by feeding inputs into the input layer, receiving an output, calculating the loss function using the output, and then fine-tuning the model parameters via backpropagation. Consequently, the parameters of the model will be fitted and matched to the data. While fitting the model, the batch size was 256 and the number of epochs was 2..
The notebook file imports all necessary modules and packages, such as TensorFlow, pandas, NumPy, Regular Expression, Natural Language Toolkit, scikit-learn, etc. Python's scikit-learn1 library is utilized for feature extraction and model training. Using scikit-Tfidf learn's Vectorizer, the text input is turned into TF-IDF feature vectors. The Tamil, Malayalam, and Kannada training sets are utilized to train the sequential model. The accuracy of the three languages is calculated. The development set is utilized to determine the accuracy of the model. The following table shows the accuracy, precision, recall, and f1-score for all three languages.
Tamil Malayalam Kannada
The table presents the weighted "macro" averages for each statistic across all three languages. For the term 'Tamil,' a precision of 0.40, an accuracy of 0.47, a recall of 0.515, and a f1-score of 0.41 were observed. 'Malayalam' was given a precision of 0.53, an accuracy of 0.55, a recall of 0.62, and a f1-score of 0.53. 'Kannada' was given a precision of 0.50, an accuracy of 0.57, a recall of 0.53, and a f1-score of 0.51. The same model was utilized for all three languages. Comparatively, the 'Malayalam' dataset had the highest precision, recall, and f1-score, although the 'Kannada' dataset had slightly greater accuracy than the 'Malayalam' dataset. Compared to the other two datasets, the "Tamil" dataset performed poorly with the model. However, it should be noted that the 'Positive' class in the 'Tamil' dataset has a disproportionately large number of instances relative to the other classes. The disparity in the data reduced the precision. The extensive training and development data relative to the other two languages may have also contributed to the low performance on the measures.
As a result, we tested the three languages in this research using the DNN Sequential model. This method can be applied to any language because it is language-independent. McCrae, John. (2020). Overview of the track on Sentiment Analysis for Dravidian Languages in Code-Mixed Text. 21-24. 10.1145/3441501.3441515.