A twitterbot that will complete lyrics it hears on twitter
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
src
.gitignore
.sample.env
Gemfile
Gemfile.lock
Procfile
Readme.md
common_words.txt
hamilton_corpus.txt
sabaton_corpus.txt

Readme.md

Lyric Bot

Given a corpus of lyrics (eg, all the lyrics to Hamilton) in a text file, this bot will detect partial lyrics in public tweets and post the next few lines.

Specifically, if the last few words of a tweet match the last few words of a line from the corpus, it's considered a match.

Additionally, tweets in reply to this bot will be checked for lyrics, as will tweets at-mentioning this bot.

Setup

You'll need a twitter account and a twitter app (see https://apps.twitter.com)

Run bundle install to install dependencies.

Run the bot with ruby bot.rb

You'll need to provide some configuration as environment variables. You can provide these a few ways:

  • On the command line: CONSUMER_KEY=abc123 CONSUMER_SECRET=def789 [...] ruby bot.rb
  • Via a .env file (see .sample.env in this repo)
  • As environment variables in your hosting environment. Eg if you're using heroku, heroku config:set CONSUMER_KEY=abc123 will set the environment variable such that the bot can access it.

The variables

CONSUMER_KEY=abc123
CONSUMER_SECRET=def789
ACCESS_TOKEN=ghi101
ACCESS_SECRET=jkl112
MATCH_SCORE_THRESHOLD=14
SEND_TWEETS=true

The keys/secrets/tokens are from your twitter app. If you're not familiar with those, you'll need to do some reading in the twitter api guide.

The MATCH_SCORE_THRESHOLD variable determines how sensitive the lyric detection is. Higher numbers will require more words to match, resulting in better but less-frequent matches. Lower numbers will require fewer words to match, resulting in more, worse matches. See "Matching logic" below for details. 14 seems to work pretty well for Hamilton (yielding around 1 match every couple hours). You may need to adjust this for your specific corpus, though.

SEND_TWEETS is used for testing. If set to true, tweets are actually sent. Otherwise they're just logged to stdout.

Matching logic

The exact number of words needed to trigger a match vary: words are weighted by how common they are ("the" and "me" are weighted much lower than "weighted" and "subjective"). Words are worth 2 to 9 points. A candidate tweet and a candidate lyric line need to have enough matching words such that the sum of the weighted score of those words exceeds a specific (configurable) threshold.

TODOs

  • Handle song-boundaries. Right now, someone tweeting the last line of a song might get replied to with the first lines of the next song.
  • Generally clean up code. It was a Sunday afternoon hack project.
  • Write some simple tests around the matching logic