/sentiment-analysis-exercise

Codes for first assignment of data science course by U of Washington in May 2013

Primary LanguagePython

In this exercise, we will

  • access the Twitter API using Python
  • estimate the public's perception (the sentiment) of a particular term or phrase
  • analyze the relationship between location and mood based on a sample of twitter data

The Twitter Application Programming Interface

Twitter provides a very rich REST API for querying the system, accessing data, and control your account. You can read more about the Twitter API

Unicode strings

Strings in the twitter data prefixed with the letter "u" are unicode strings. For example:

u"This is a string"

Unicode is a standard for representing a much larger variety of characters beyond the roman alphabet (greek, russian, mathematical symbols, logograms from non-phonetic writing systems such as kanji, etc.)

In most circumstances, you will be able to use a unicode object just like a string.

If you encounter an error involving printing unicode, you can use the encode method to properly print the international characters, like this:

unicode_string = u"aaaà çççñññ"
encoded_string = unicode_string.encode('utf-8')
print encoded_string

Problem 1: Get Twitter Data

To access the live stream, you will need to install the oauth2 library so you can properly authenticate. The command $ pip install oauth2 should work for most environments.

The steps below will help you set up your twitter account to be able to access the live 1% stream.

  1. Create a twitter account if you do not already have one.
  2. Go to https://dev.twitter.com/apps and log in with your twitter credentials.
  3. Click "Create New App"
  4. Fill out the form and agree to the terms. Put in a dummy website if you don't have one you want to use.
  5. On the next page, click the "API Keys" tab along the top, then scroll all the way down until you see the section "Your Access Token"
  6. Click the button "Create My Access Token". You can Read more about Oauth authorization.
  7. You will now copy four values into twitterstream.py. These values are your "API Key", your "API secret", your "Access token" and your "Access token secret". All four should now be visible on the API Keys page. (You may see "API Key" referred to as "Consumer key" in some places in the code or on the web; they are synonyms.) Open twitterstream.py and set the variables corresponding to the api key, api secret, access token, and access secret. You will see code like the below:
api_key = "<Enter api key>"
api_secret = "<Enter api secret>"
access_token_key = "<Enter your access token key here>"
access_token_secret = "<Enter your access token secret here>"

  1. Run the following and make sure you see data flowing and that no errors occur. $ python twitterstream.py > output.txt This command pipes the output to a file. Stop the program with Ctrl-C, but wait at least 3 minutes for data to accumulate.

  2. If you wish, modify the file to use the twitter search API to search for specific terms. For example, to search for the term "microsoft", you can pass the following url to the twitterreq function: https://api.twitter.com/1.1/search/tweets.json?q=microsoft

Output the first 20 lines of the twitter data you downloaded from the web. You should save the first 20 lines to a file turn_in.txt by using the following command:

$ head -n 20 output.txt > turn_in.txt

Problem 2: Derive the sentiment of each tweet

For this part, you will compute the sentiment of each tweet based on the sentiment scores of the terms in the tweet. The sentiment of a tweet is equivalent to the sum of the sentiment scores for each term in the tweet.

tweet_sentiment.py accepts two arguments on the command line: a sentiment file and a tweet file like the one you generated in Problem 1.

$ python tweet_sentiment.py AFINN-111.txt output.txt The file AFINN-111.txt contains a list of pre-computed sentiment scores. Each line in the file contains a word or phrase followed by a sentiment score. Each word or phrase that is found in a tweet but not found in AFINN-111.txt should be given a sentiment score of 0. See the file AFINN-README.txt for more information.

To use the data in the AFINN-111.txt file, you may find it useful to build a dictionary. Note that the AFINN-111.txt file format is tab-delimited, meaning that the term and the score are separated by a tab character. A tab character can be identified a "\t".The following snippet may be useful:

afinnfile = open("AFINN-111.txt")
scores = {} # initialize an empty dictionary
for line in afinnfile:
  term, score  = line.split("\t")  # The file is tab-delimited. "\t" means "tab character"
  scores[term] = int(score)  # Convert the score to an integer.

print scores.items() # Print every (term, score) pair in the dictionary

The data in the tweet file you generated in Problem 1 is represented as JSON, which stands for JavaScript Object Notation. It is a simple format for representing nested structures of data --- lists of lists of dictionaries of lists of .... you get the idea.

Each line of output.txt represents a streaming message. Most, but not all, will be tweets.

It is straightforward to convert a JSON string into a Python data structure; there is a library to do so called json.

To use this library, add the following to the top of tweet_sentiment.py

import json

Then, to parse the data in output.txt, you want to apply the function json.loads to every line in the file.

This function will parse the json data and return a python data stucture; in this case, it returns a dictionary.

You can read the Twitter documentation to understand what information each tweet contains and how to access it, but it's not too difficult to deduce the structure by direct inspection.

Your script should print to stdout the sentiment of each tweet in the file, one numeric sentiment score per line. The first score should correspond to the first tweet, the second score should correspond to the second tweet, and so on. The nth line of the output file should contain only a single number that represents the score of the nth tweet in the input file!

NOTE: You must provide a score for every tweet in the sample file, even if that score is zero. You can assume the sample file will only include English tweets and no other types of streaming messages.

Problem 3: Derive the sentiment of new terms. See term_sentiment.py

In this part you will be creating a script that computes the sentiment for the terms that do not appear in the file AFINN-111.txt.

Here's how you might think about the problem: We know we can use the sentiment-carrying words in AFINN-111.txt to deduce the overall sentiment of a tweet. Once you deduce the sentiment of a tweet, you can work backwards to deduce the sentiment of the non-sentiment carrying words that do not appear in AFINN-111.txt. For example, if the word soccer always appears in proximity with positive words like great and fun, then we can deduce that the term soccer itself carries a positive sentiment.

Don't feel obligated to use it, but the following paper may be helpful for developing a sentiment metric. Look at the Opinion Estimation subsection of the Text Analysis section in particular.

O'Connor, B., Balasubramanyan, R., Routedge, B., & Smith, N. From Tweets to Polls: Linking Text Sentiment to Public Opinion Time Series. (ICWSM), May 2010.

$ python term_sentiment.py <sentiment_file> <tweet_file> Your script should print output to stdout. Each line of output should contain a term, followed by a space, followed by the sentiment. That is, each line should be in the format <term:string> <sentiment:float>

For example, if you have the pair ("foo", 103.256) in Python, it should appear in the output as:

foo 103.256 The order of your output does not matter.

Problem 4: Compute Term Frequency

Write a Python script frequency.py to compute the term frequency histogram of the livestream data you harvested from Problem 1.

The frequency of a term can be calculated as [# of occurrences of the term in all tweets]/[# of occurrences of all terms in all tweets]

Your script will be run from the command line like this:

$ python frequency.py <tweet_file> You should assume the tweet file contains data formatted the same way as the livestream data.

Your script should print output to stdout. Each line of output should contain a term, followed by a space, followed by the frequency of that term in the entire file. There should be one line per unique term in the entire file. Each line should be in the format <term:string> <frequency:float>

For example, if you have the pair (bar, 0.1245) in Python it should appear in the output as:

bar 0.1245 If you wish, you may consider a term to be a multi-word phrase, but this is not required. You may compute the frequencies of individual tokens only.

Depending on your method of parsing, you may end up computing frequencies for hashtags, links, stop words, phrases, etc. If you choose to filter out these non-words, that's ok too.

Problem 5: Which State is happiest?

Write a Python script happiest_state.py that returns the name of the happiest state as a string.

Your script happiest_state.py should take a file of tweets as input. It will be called from the command line like this:

$ python happiest_state.py <sentiment_file> <tweet_file> The file AFINN-111.txt contains a list of pre-computed sentiment score.

Assume the tweet file contains data formatted the same way as the livestream data.

It's a good idea to make use of your solution to Problem 2.

There are different ways you might assign a location to a tweet. Here are three:

  • Use the coordinates field (a part of the place object, if it exists, to geocode the tweet. This method gives the most reliable location information, but unfortunately this field is not always available and you must figure out some way of translating the coordinates into a state.

  • Use the other metadata in the place field. Much of this information is hand-entered by the twitter user and may not always be present or reliable, and may not typically contain a state name.

  • Use the user field to determine the twitter user's home city and state. This location does not necessarily correspond to the location where the tweet was posted, but it's reasonable to use it as a proxy. You are free to develop your own strategy for determining the state that each tweet originates from.

You may find it useful to use this python dictionary of state abbreviations.

You can ignore any tweets for which you cannot assign a location in the United States.

In this file, each line is a Tweet object, asdescribed in the twitter documentation.

Note: Not every tweet will have a text field --- again, real data is dirty! Be prepared to debug, and feel free to throw out tweets that your code can't handle to get something working. For example, you might choose to ignore all non-English tweets.

Your script should print the two letter state abbreviation of the state with the highest average tweet sentiment to stdout.

Note that you may need a lot of tweets in order to get enough tweets with location data. Let the live stream run for a while if you wish.

Your script will not have access to the Internet, so you cannot rely on third party services to resolve geocoded locations!

Problem 6: Top ten hash tags

Write a Python script top_ten.py that computes the ten most frequently occurring hashtags from the data you gathered in Problem 1.

Your script will be run from the command line like this:

$ python top_ten.py <tweet_file> You should assume the tweet file contains data formatted the same way as the livestream data.

In the tweet file, each line is a Tweet object, as described in the twitter documentation. To find the hashtags, you should not parse the text field; the hashtags have already been extracted by twitter.

Your script should print to stdout each hashtag-count pair, one per line, in the following format:

Your script should print output to stdout. Each line of output should contain a hashtag, followed by a space, followed by the frequency of that hashtag in the entire file. There should be one line per unique hashtag in the entire file. Each line should be in the format <hashtag:string> <frequency:float>

For example, if you have the pair (bar, 30) in Python it should appear in the output as:

bar 30