Twitter Scraper

Github Repository: https://github.com/Marsan-Ma/twitter_scraper

Scraping twitter content from twitter streaming API, in python3. The output is used as the training corpus for my ChatBot.

How it works

Setup

1. Copy the config.yml.default to config.yml, and fill your twitter application tokens you got from twitter developers.

2. just type python3 twitter.py, the listener will start to dump corpus in: corpus/<YYYYMMDD_HHMMSS.txt>

Output

I build this corpus for training the neural network model ChatBot, thus the corpus is arranged as consecutive dialog where even sentences are the response of odd sentences, like:

Line number	Sentences
1	happy birthday!
2	thank you :)
3	game of throne is the best drama i’ve seen.
4	I’ll say the walking dead is even better.

Where (1,2), (3,4) are two independent dialog pairs.

Content Filters

Twitter streaming API supports some filter, and the most useful ones are:

Location filter

simply place the logitude and altitude in [South,West,North,East] format, for example:

Place	filter
San Francisco	stream.filter(locations=[-122.75,36.8,-121.75,37.8])
New York City	stream.filter([-74,40,-73,41])
San Francisco or New York City	stream.filter([-122.75,36.8,-121.75,37.8, -74,40,-73,41])

Language filter

there are threads questioning about language filter is not work, actually is does, it just need to accompanied with track filter, like:

This won’t work. stream.filter(languages=["en"])

This works. stream.filter(languages=["en"], track=['machine', 'learning'])

If you want language other than English

Twitter can’t tokenize some languages like Japanese or Chinese correctly, your track parameter won’t work for these languages. For example, you might setup:

stream.filter(languages=["ja"], track=['バイト'])

While you expect you will get a lot of tweets about バイト, but you simply can’t get it, because twitter can’t tokenize Japenses and Chinese correctly.

If you just want some corpus, regardless of the topics, here are the work around: use some generic English keywords instead, like:

stream.filter(languages=["zh"], track=['I', 'you', 'http', 'www', '@', '。', '，', '！', '』', ')', '...', '-', '/'])

According to document, you could add up to 400 keywords on this list, even some emoji also works.