Urdu Sentiment Classification

Sentiment Analysis is a classic task in NLP. There is a lot of research done for different languages but not in the Urdu language. Although there are some papers available for sentiment analysis on Urdu, but not data or source code is provided to reproduce the results. This is my first attempt to run logic regression on the Urdu dataset.

Let's start coding for logistic regression. I'm using Urdu Corpus V1 for this tutorial. Here is what data looks.

# load data and take a quick look
import re
import pandas as pd
import seaborn as sns
import matplotlib.pyplot as plt
from sklearn.model_selection import train_test_split
from sklearn.feature_extraction.text import TfidfVectorizer

raw_data = pd.read_csv('../data/sentiment_urdu.csv')
raw_data.head(5)

Original datasets have 3 classes but one class has only 20 records. I've removed it because it will cause some issues for classification. Here it is what the bar chat looks like with three classes.

# check the size of the data and its class distribution
sentences = raw_data['sentence'].tolist()
sentiments = raw_data['sentiment'].tolist()

sns.countplot(x='sentiment', data=raw_data)
plt.show()

Sentiment Count for each class

I've removed the "O" class form dataset with the following code.

indexNames = raw_data[raw_data['sentiment'] == "O"].index
raw_data.drop(indexNames, inplace=True)
sns.countplot(x='sentiment', data=raw_data)
plt.show()

Sentiment Count for each Class

Now, perform pre-processing on the data for better results. Currently, I'm only removing English text, symbols and numbers.




# text cleaning and pre-processing:
def delete_urdu_english_symbols(sentences):
    cleaned = []
    for sentence in sentences:
        text = re.sub(r"\d+", " ", sentence)
        # English punctuations
        text = re.sub(r"""[!"#$%&'()*+,-./:;<=>?@[\]^_`{|}~]+""", " ", text)
        # Urdu punctuations
        text = re.sub(r"[:؛؟’‘٭ء،۔]+", " ", text)
        # Arabic numbers
        text = re.sub(r"[٠‎١‎٢‎٣‎٤‎٥‎٦‎٧‎٨‎٩]+", " ", text)
        text = re.sub(r"[^\w\s]", " ", text)
        # Remove English characters and numbers.
        text = re.sub(r"[a-zA-z0-9]+", " ", text)
        # remove multiple spaces.
        text = re.sub(r" +", " ", text)
        text = text.split(" ")
        # some stupid empty tokens should be removed.
        text = [t.strip() for t in text if t.strip()]
        cleaned.append(" ".join(text))
    return cleaned

X = delete_urdu_english_symbols(sentences)
Y = sentiments

Now the important part is to split the data for training and testing. I'm using scikit function train_test_split for ease.

# Feel free to use different ratios to split the data.
train_text, test_text, train_labels, test_labels = train_test_split(X, Y, test_size=0.20, random_state=42)

Here is the classifier for configuration. 5000 max features is a general guess. You can use a different size for your convenience. Also, I'm using TFIDF feature for LogisticRegression. There are three class and LogisticRegression can handle multi-class data so let's see how it performs on the Urdu dataset.

# training: tf-idf + logistic regression
max_feature_num = 5000
train_vectorizer = TfidfVectorizer(max_features=max_feature_num)
train_vecs = train_vectorizer.fit_transform(train_text)
test_vecs = TfidfVectorizer(max_features=max_feature_num,vocabulary=train_vectorizer.vocabulary_).fit_transform(test_text)

# train model
from sklearn.linear_model import LogisticRegression
clf = LogisticRegression().fit(train_vecs, train_labels)

# test model
test_pred = clf.predict(test_vecs)
from sklearn.metrics import precision_recall_fscore_support,accuracy_score
acc = accuracy_score(test_labels, test_pred)
pre, rec, f1, _ = precision_recall_fscore_support(test_labels, test_pred, average='macro')
print('acc', acc)
print('precision', pre)
print('rec', rec)
print('f1', f1)

The first attempt is not bad. It gives us nearly 84% accuracy for LogisticRegression. You can use SVM to get better classification results. I will use different approaches in the future for this purpose.

Now save the model for later use. I'm saving the model as well as tfidf as well.

import pickle

# save model and other necessary modules
all_info_want_to_save = {
    'model': clf,
    'vectorizer': TfidfVectorizer(max_features=max_feature_num,vocabulary=train_vectorizer.vocabulary_)
}
save_path = open("sentiment_urdu_logistic_regression.pickle","wb")
pickle.dump(all_info_want_to_save, save_path)

And that's it. If you want to test the model, use the following code for it.

import pickle

# reload your model and use it to make predictions for test text
# you should adjust the code so as to load to your saved model/components
def test_trained_model(model_path, test_text):
    saved_model_dic = pickle.load(open(model_path,"rb"))
    saved_clf = saved_model_dic['model']
    saved_vectorizer = saved_model_dic['vectorizer']
    print(len(saved_vectorizer.vocabulary))
    new_test_vecs = saved_vectorizer.fit_transform(test_text)
    return saved_clf.predict(new_test_vecs)


# load sample test data
import pandas as pd
test_data = pd.read_csv('coursework1_train.csv')
test_text = test_data['text'].tolist()[-5000:]
test_labels = test_data['sentiment'].tolist()[-5000:]

print('test data size', len(test_labels))

# test model
from sklearn.metrics import precision_recall_fscore_support,accuracy_score
new_test_pred = test_trained_model("sentiment_urdu_logistic_regression.pickle", test_text)
acc = accuracy_score(test_labels, new_test_pred)
pre, rec, f1, _ = precision_recall_fscore_support(test_labels, new_test_pred, average='macro')
print('acc', acc)
print('precision', pre)
print('rec', rec)
print('f1', f1)