We are exploring the field of natural language processing, which is the broad study of how computers and machines can understand human to human communication and how texts are analyzed based on contextual information by machines.
Code-Mixing is a phenomenon where speakers switch between multiple languages in a single utterance [1,2]. Code-Mixing is common in multilingual countries such as India [3,4]. There is an increasing demand for sentiment analysis on social media texts which are largely code-mixed [5,6]. Sentiment analysis is the interpretation and classification of emotions (positive, negative, and neutral) within text data using text analysis techniques. The machine learning model based on monolingual data fails on code-mixed data. As the usage of the internet growing amount of code-mix multilingual data is increasing. The mixing of the scripts in the code-mixing makes it even more complicated to use the model trained on monolingual corpora [7,8,9]. In this paper, I provide a classic Machine learning algorithms trained on code mixed multilingual data.
In this paper, we present a model which can be used to find the sentiment of a given text. We want to classify the text as 'Positive ', 'Negative ', 'Mixed feelings ', 'unknown state ', 'not-Tamil' 2. Data-set [5,6] Malayalam is one of the Dravidian languages spoken in the the southern region of India with nearly 38 million Malayalam speakers in India and other countries. Tamil, is a Dravidian language natively spoken by the Tamil people of India and Sri Lanka. Tamil is the official language of the South Indian state of Tamil Nadu, as well as two sovereign states, Sri Lanka and Singapore. For this shared task, we have been provided with a new gold standard corpus by the organizers for sentiment analysis of code-mixed text in Dravidian languages (Malayalam-English and Tamil-English). The data-set consists of YouTube comments which are then marked as one of the following. ' Positive', 'Negative', 'Mixed feelings', 'unknown state', 'not-Tamil'.
The distribution of the dataset is below.
As we can see from the data the very skewed a simpler machine learning approach will be more generalized.
This is a message-level polarity classification task [10,11]. Given a YouTube comment, systems have to classify it into positive, negative, neutral, mixed emotions, or not in the intended languages.
We experimented with broadly three kinds of classic systems -an SVM classifier, a logistic classifier, and a Perceptron. We used the sci-kit learn implementation of SVM, Logistic Regression, and Perceptron. Support Vector Machines are one of the most successful classic machine learning models used for various kinds of text classification tasks. Used logistic regression with a multi-class variable as 'ovr' for multi-class classification. Perceptron is a single layer neural network and a multi-layer perceptron is called Neural Networks. We used a grid search for finding the best parameters for SVM algorithms.
For text to vector conversion, we used sklearn CountVectorizer. The CountVectorizer provides a simple way to both tokenize a collection of text documents and build a vocabulary of known words, but also to encode new documents using that vocabulary. I trained models separately for the Tamil dataset and the Malayalam dataset.
This section presents the results of the evaluation of the three architectures. We compare the performance of the above machine learning architectures to select submissions for each language. Classification Accuracy is what we usually mean when we use the term accuracy. It is the ratio of the number of correct predictions to the total number of input samples.
SVM Logistic Reg Perceptron Accuracy Score 0.63 0.677 0.614
For further analysis, I used the confusion matrix. A confusion matrix is a table that is often used to describe the performance of a classification model (or "classifier") on a set of test data for which the true values are known. It is extremely useful for measuring Recall, Precision, Specificity, Accuracy, and most importantly AUC-ROC Curve.
In confusion matrix figure 3 and figure 4, we can see that due to an unbalanced data-set many test cases were classified as negative 1.
In this paper, we have described how I trained machine learning algorithms for classification. Simple machine learning algorithms were fast to train and set the base for further research. For, future work we can train complex deep learning algorithms but we will need a more balanced dataset for complex deep learning algorithms.