SentimentAnalyzer.py

from nltk.util import pr
from numpy.lib.function_base import average
import pandas as pd
import codecs
from nltk.tokenize import word_tokenize
from sklearn.metrics import accuracy_score
from sklearn.metrics import f1_score
import re


data = pd.read_csv("HindiSentiWordnet.txt", delimiter=' ')
# print((data))

# print(data.index)
fields = ['POS_TAG', 'ID', 'POS', 'NEG', 'LIST_OF_WORDS']

# making a dictionary for every word in senti word net and tagging it with it's respective pos tag and +ve and -ve score
words_dict = {}
for i in data.index:
    # print (data[fields[0]][i], data[fields[1]][i], data[fields[2]][i], data[fields[3]][i], data[fields[4]][i])

    words = data[fields[4]][i].split(',')
    for word in words:
        words_dict[word] = (data[fields[0]][i], data[fields[2]][i], data[fields[3]][i])
    # print(words_dict)

# This function determines Senti of text.

# print(Senti("कल गुलाम राम नहीं जीता"))
def Senti(text):
   
    
    # words = text.split(" ")


    # not_stop_words = [word for word in words if word not in set(STOP_WORDS_HI) ]

    # # print(not_stop_words)
    # words = word_tokenize(not_stop_words)



    words = word_tokenize(text)

    votes = []
    pos_polarity = 0
    neg_polarity = 0
    neu_polarity = 0
    seen = 0
    #adverbs, nouns, adjective, verb are only used
    allowed_words = ['a','v','r','n']
    for word in words:
        if word in words_dict:
            #if word in dictionary, it picks up the positive and negative score of the word
            pos_tag, pos, neg = words_dict[word]
            # print(pos_tag, pos, neg)
            if pos_tag in allowed_words:
                if pos > neg:
                    pos_polarity += pos
                    # print(pos_polarity)
                    votes.append(1)
                elif neg > pos:
                    neg_polarity += neg
                    # print(neg_polarity)
                    votes.append(-1)
        
    #calculating the no. of positive and negative words in total in a review to give class labels
    pos_votes = votes.count(1)
    neg_votes = votes.count(-1)
    # print(votes.count())
    if pos_votes > neg_votes:
        seen = 1
    elif neg_votes > pos_votes:
        seen = -1
    else:
        if pos_polarity < neg_polarity:
            seen = -1
        elif pos_polarity > neg_polarity:
            seen = 1
        elif pos_polarity == neg_polarity:
            seen = 0

    

    return seen


def condition(text, seen):
    words = word_tokenize(text)
    dummy = words
    dumm1 = dummy
    
    if "नहीं" in words:
        # words = words.replace("नही", 'omit')
        # i = 0
        # message = ""
        # msg = ""
        # for word in words:
        #     if word != "नही":
        #         message += word
        #         i += 1
                
        #     else:
        #         i+=1
        #         for i in range(len(words)):
        #             msg += words[i]
        #         break
        if (seen==1):
            seen = -1
        elif (seen==0):
            seen = -1
        else:
            seen = 1
    

    elif "लेकिन" in dummy:
        # dummy = dummy.replace("पर", 'omit')
        message = ""
        msg = ""
        i = 0
        for word in dummy:
            if word != "लेकिन":
                message += word
                i+=1
            else:
                i+=1
                for i in range(len(words)):
                    msg += words[i]
                break
        if (Senti(message)== -1 and Senti(msg)== 1):
            seen = 1
        elif (Senti(message)== -1 and Senti(msg)==-1):
            seen = -1
        elif (Senti(message)== -1 and Senti(msg)== 0):
            seen = -1
        elif (Senti(message)== +1 and Senti(msg)== -1):
            seen = -1
        elif (Senti(message)== +1 and Senti(msg)== +1):
            seen = 1
        elif (Senti(message)== +1 and Senti(msg)== 0):
            seen = 1
        elif (Senti(message)== 0 and Senti(msg)== +1):
            seen = 1
        else:
            seen = 0
    elif "पर​" in dumm1:
        # dummy = dummy.replace("पर", 'omit')
        message = ""
        msg = ""
        i = 0
        for word in dumm1:
            if word != "पर":
                message += word
                i+=1
            else:
                i+=1
                for i in range(len(words)):
                    msg += words[i]
                break
        if (Senti(message)== -1 and Senti(msg)== 1):
            seen = -1
        elif (Senti(message)== -1 and Senti(msg)==-1):
            seen = -1
        elif (Senti(message)== -1 and Senti(msg)== 0):
            seen = -1
        elif (Senti(message)== +1 and Senti(msg)== -1):
            seen = 1
        elif (Senti(message)== +1 and Senti(msg)== +1):
            seen = 1
        elif (Senti(message)== +1 and Senti(msg)== 0):
            seen = 1
        elif (Senti(message)== 0 and Senti(msg)== +1):
            seen = 1
        else:
            seen = -1

    
    return seen

# to calculate accuracy and F1_score

## accuracy score of 56.46%

pred_y = []
actual_y = []
pos_reviews = codecs.open("pos_hindi.txt", "r", encoding='utf-8', errors='ignore').read()
for line in pos_reviews.split('#'):
    data = line.strip('\n')
    if data:
        pred_y.append(condition(data, Senti(data)))
        # print(pred_y)
        actual_y.append(1)
        # print(actual_y)
# print(len(actual_y))        # number of sentences taken till this point
neg_reviews = codecs.open("neg_hindi.txt", "r", encoding='utf-8', errors='ignore').read()
for line in neg_reviews.split('#'):
    data=line.strip('\n')
    if data:
        pred_y.append(condition(data, Senti(data)))
        actual_y.append(-1)
# print(len(actual_y))        # number of sentences taken till this point
neu_reviews = codecs.open("neu_hindi.txt", "r", encoding='utf-8', errors='ignore').read()
for line in neu_reviews.split('#'):
    data = line.strip('\n')
    if data:
        pred_y.append(condition(data, Senti(data)))
        # print(pred_y)
        actual_y.append(0)
        # print(actual_y)
# print(len(actual_y))        # number of sentences taken till this point

print('Accuracy-score -->  ',accuracy_score(actual_y, pred_y, normalize=True, sample_weight=None) * 100)
print('F-measure -->  ',f1_score(actual_y,pred_y, average='micro'))



#accuracy score of 59%
pred_y = []
actual_y = []
pos_reviews = codecs.open("pos_hindi.txt", "r", encoding='utf-8', errors='ignore').read()
for line in pos_reviews.split('#'):
    data = line.strip('\n')
    if data:
        pred_y.append(Senti(data))
        # print(pred_y)
        actual_y.append(1)
        # print(actual_y)
# print(len(actual_y))        # number of sentences taken till this point
neg_reviews = codecs.open("neg_hindi.txt", "r", encoding='utf-8', errors='ignore').read()
for line in neg_reviews.split('#'):
    data=line.strip('\n')
    if data:
        pred_y.append(Senti(data))
        actual_y.append(-1)
# print(len(actual_y))        # number of sentences taken till this point
neu_reviews = codecs.open("neu_hindi.txt", "r", encoding='utf-8', errors='ignore').read()
for line in neu_reviews.split('#'):
    data = line.strip('\n')
    if data:
        pred_y.append(Senti(data))
        # print(pred_y)
        actual_y.append(0)
        # print(actual_y)
# print(len(actual_y))        # number of sentences taken till this point

print('Accuracy-score -->  ',accuracy_score(actual_y, pred_y, normalize=True, sample_weight=None) * 100)
print('F-measure -->  ',f1_score(actual_y,pred_y, average='micro'))


# //////////////////////////////////////////////////
if __name__ == '__main__':
    data1 = "वो बहुत खुश था"                                        # positive
    data2 = "राम की मृत्यु हो गयी"                                     # negative
    data3 = "एक दिन चुन्नू हिरण उस जंगल में रहने के लिए आया।"            # neutral
    print(Senti(data1))
    print(condition(data1, Senti(data1)))
    print(Senti(data2))
    print(condition(data2, Senti(data2)))
    print(Senti(data3))
    print(condition(data3, Senti(data3)))
    
    


## experimentations: 

# if "भाग्यवान" in words_dict:
#     pos_tag, pos, neg = words_dict["भाग्यवान"]
#     print(pos_tag, pos, neg)
# else:
#     print("Not reading")






#  testing

# neg_reviews = codecs.open("neg_hindi.txt", "r", encoding='utf-8', errors='ignore').read()
# for line in neg_reviews.split('#'):
#     data = line.strip('\n')
#     if data:
#         print(Senti(data))



# analysis during program makings

# added neutral thingy --> decrease in accuracy by 0.1%
# senti word net --> line 2687 changes by yash for -ve words --> after 2nd letter
# senti word net --> line 2006 changes by yash for +ve words after 7th word
# senti word net --> line 1971 changes by prayush +ve words after 7th word
# add nahi in senti word net at some point --> lines 21-23,68,73,74,99,129,143,145,165,173,180,193,199,203,205,225,259 are 1 cuz "nahi" is not in wordnet
# negation --> 78,98,141,224,247
# we are not considering idioms --> eg: "hawa nikal gyi" in line 106, चारों खाने चित्त --> 166
# stemming of words was infact decreasing the accuracy score of the model.Removing iy also increased the speed of the program

# only changes in neg senti word net
# 200 sentences -->  42.64
# 250 sentences -->  43.34
# 271 sentences -->  43.84
# 350 sentences -->  45.33
# 475 sentences -->  45.93
# changes in both +ve and -ve word net
# all sentences without the "nahi" condition -->  59.177
# all sentences without the "nahi" condition and without stemming -->  59.277