Natural Language Processing with Disaster Tweets

Predict which Tweets are about real disasters and which ones are not

---

Abstract:

Twitter has become an important communication channel in times of emergency. The ubiquitousness of smartphones enables people to announce an emergency they’re observing in real-time. Because of this, more agencies are interested in programmatically monitoring Twitter (i.e., disaster relief organizations and news agencies). But, it’s not always clear whether a person’s words are actually announcing a disaster.

In this notebook, I've built a machine learning model that predicts which Tweets are about real disasters and which one’s aren’t.

*Also, I've created two notebook, one of them is commented and fully detailed and easy to read and train, And other one is functional implementation of first one and more clean.

---

Dataset:

The dataset that I've used in this notebook is available on Kaggle site. [link]

---

In this notebook I've performed some steps such as :

1) EDA: a little look at dataset with some graphs

2) Clean data in two steps:

• Remove duplicated tweets

• Find similar tweets and drop them

3) Extract some features such as:

• Length of tweets

• Count of words in tweets

• Count of numbers in tweets

• Count of sentences in tweets

• Count of hashtags in tweets

• Text of hashtags

• Count of mentions in tweets

• Text of Mentions

• Count of links in tweets

• Word per length of tweet

• Punctuation count per tweet length

• Uppercase letters count per tweet length

• MinMaxScaling for numeric columns

4) Process Tweets such as:

• Lowercase tweets

• Remove URLs

• Remove Punctuation

• Remove Short words <=2 chars

• Remove Stopwords

• Lemmatization

• GetDummy for keyword column

5) TF-IDF:

• TF-IDF on tweets

• TF-IDF on text of hashtags

• TF-IDF on text of mentions

6) Train and test models:

• GradientBoostingClassifier

• NaiveBayes

• LogisticRegression

• SVM

And I've got approx. 0.8 score on main test set on kaggle site.

---