Tweet Sentiment Analysis Using LSTM

alishafaghiSeptember 8, 2022
0 2 minutes read

In this project I used a labeled dataset to create model for sentiment analysis of texts given to model.

The dataset contains the text of Donald Trump’s Tweet, the ex-president of united states.

Steps Of The Text Sentiment Analysis Project

There are 3 main steps in this project:

  • Text Preprocess
  • Word Embedding
  • Training The Model

So, I will explain each step and also will show the code for each section, but if you want to check out the all of the code, I will leave the link of the notebook in Kaggle.

Dataset Link:

https://www.kaggle.com/datasets/albertsavill/all-trumps-twitter-insults-with-sentiment

Notebook Link:

https://www.kaggle.com/code/alishafaghi/trump-s-tweets-sentiment-analysis-using-lstm

Data Exploration

Before anything else, I splitted the data into test and train by ratio of 30% to 70%.
In the beginning I wanted to look at the data carefully so I used the Python Pandas library to check out the data.

First, I want to see only a few results from the top.

Code:

train= pd.read_csv("../input/trump-tweets/trump-train.csv",encoding='ISO-8859-1')


test= pd.read_csv('../input/trump-tweets/trump-test.csv')

test.head()

So this is the result.

Oouput:

Next, I check out the info of data.

So here is the code and the results of what I checked.

Code:

train.info()

Output:

Code:

test.isnull().any()

Output:

Code:

train.shape

Output:

Code:

test.shape

Output:

Next I wanted to see the count of tweets categorized in each sentiment label.

Code:

sns.set_style("whitegrid")
sns.set(rc={'figure.figsize':(11,4)})
sns.countplot(train['sentiment_label'])

Output:

Text Preprocess

The next step is to preprocess the text and remove and clean it.

In this step I removed stop words, links, extra space and punctuations. Plus I corrected the miss spellings and tokenized the text.

Frist I used the NLTK library to identify the stop words, then I created a function using REGEX or regular expression to remove the things I just mentioned.

Code:

from nltk.corpus import stopwords
# load stop words
stop_word = stopwords.words('english')
def clean(text):
    #     remove urls
    text = re.sub(r'http\S+', " ", text)
    #     remove mentions
    text = re.sub(r'@\w+',' ',text)
    #     remove hastags
    text = re.sub(r'#\w+', ' ', text)
    #     remove digits
    text = re.sub(r'\d+', ' ', text)
    #     remove html tags
    text = re.sub('r<.*?>',' ', text) 
    #     remove stop words 
    text = text.split()
    text = " ".join([word for word in text if not word in stop_word])
    
    return text

Then I applied the function to test and train data separately

train['tweet'] = train['tweet'].apply(lambda x: clean(x))
test['tweet'] = test['tweet'].apply(lambda x: clean(x))

Next, I used loc in Pandas to locate the columns I want for the test and train file to be equal to.

Code:

train = train.loc[:,['tweet','sentiment_label']]
test = test.loc[:,['tweet','sentiment_label']]

Then I used label encoder to label each sentiment.

So, negative is equal to -1, neutral equal to 0 and positive is equal to 1.

label_encoder = preprocessing.LabelEncoder()
  
# Encode labels in column 'Sentiment'.
train['sentiment_label']= label_encoder.fit_transform(train['sentiment_label'])
  
train['sentiment_label'].unique()

Output:

Alright so lets check again and see if the labels changed in dataset too.

Code:

train.head()

Output:

Great, so it works and as you can see and the sentiment labels are numeric.

Next, I used train and test function to split the data into train and test by 70% to 30% respectively.

Code:

train_text,val_text,train_label,val_label=train_test_split (train.tweet, train.sentiment_label, test_size=0.3,random_state=42)

Then I used lemmatization to get the unique form of each words and after that I tokenized it.

Code:

class Lemmatizer(object):
    def __init__(self):
        self.lemmatizer = WordNetLemmatizer()
    def __call__(self, sentence):
        sentence=re.sub('(https?:\/\/)?([\da-z\.-]+)\.([a-z\.]{2,6})([\/\w \.-]*)',' ',sentence)
        sentence=re.sub('[^0-9a-z]',' ',sentence)
        return [self.lemmatizer.lemmatize(word) for word in sentence.split() if len(word)>1]
tokenizer=CountVectorizer(max_features=5000,stop_words='english',lowercase=True,tokenizer=Lemmatizer())
train_x=tokenizer.fit_transform(train_text).toarray()

tokenizer.get_params()

Output:

Next , I managed the future names.

Code:

feature_names=tokenizer.get_feature_names()
val_x=tokenizer.transform(val_text).toarray()
test_x=test.tweet
test_label=label_encoder.transform(test['sentiment_label'])
test_x_1=tokenizer.transform(test_x).toarray()

Word Embedding

Next I started the word embedding process.

Code:

early_stop=EarlyStopping(monitor='val_accuracy',patience=3)
reduceLR=ReduceLROnPlateau(monitor='val_accuarcy',patience=2)
token=Tokenizer(num_words=5000,oov_token=Lemmatizer())
token.fit_on_texts(train_text)
train_x_2=token.texts_to_sequences(train_text)
train_x_2=pad_sequences(train_x_2,maxlen=60,padding='post',truncating='post')
val_x_2=token.texts_to_sequences(val_text)
val_x_2=pad_sequences(val_x_2,maxlen=60,padding='post',truncating='post')
embedding_dimension=32
v=len(token.word_index)
model=Sequential()
model.add(Input(shape=(60,)))
model.add(Embedding(v+1,embedding_dimension))
model.add(LSTM(64,return_sequences=True))
model.add(GlobalMaxPool1D())
model.add(Dense(64))
model.add(Dense(3,activation='softmax'))

Model Training

Finally I started to train the model.

Code:

model.compile(optimizer='adam',loss='sparse_categorical_crossentropy',metrics=['accuracy'])
r=model.fit(train_x_2,train_label,validation_data=(val_x_2,val_label), epochs=50,batch_size=64,callbacks=[reduceLR,early_stop])

Output:

Next, I plotted it

Code:

plt.plot(r.history['loss'])
plt.plot(r.history['val_loss'])
plt.title('LOSS',fontdict={'size':'22'})
plt.plot()

Output:

alishafaghiSeptember 8, 2022
0 2 minutes read

Leave a Reply

Your email address will not be published. Required fields are marked *

© Copyright 2023, All Rights Reserved
Back to top button