Challenge Description
Sentiment analysis is a rapidly growing field, as large databases become increasingly available, and companies have an interest in automatically extracting sentiment from internet users.
With the increasing amount of content and debate on social media platforms such as Twitter, there is an interest in automatically extracting meaning and sentiment from users' posts, to be able to evaluate the aggregate opinion of a large number of users. Capturing sentiment in language is important in these times where decisions and reactions are created and updated in seconds.
For this challenge, students are provided with tweets from
[Figure Eight's Data for Everyone platform](https://appen.com/open-source-datasets/)
. Your task will be to classify tweets with positive, neutral, or negative labels, reflecting the sentiment of the user writing the tweet.
For example:
"I love this ice-cream, it is so good ! :-)"
-> positive.
"Last week I visited San Francisco"
-> neutral.
"Disappointed by the last Avengers movie..."
-> negative.
You are provided with an example notebook which covers the basic elements of NLP models for sentiment analysis, along with a baseline. The evaluation criterion is based on your model's ability to predict the correct labels for the testing data.
At the end of the challenge, if you wish to do so, there is a bonus task which consists of detecting the relevant words inside the tweet which are most responsible for the attributed sentiment. For example in the phrase
"I really like this song"
, the relevant words are:
"really like"
.
This is not evaluated with the kaggle platform however; the testing labels are provided in the test.csv file (under
selected_text
), for you to compute the performance of your own model.