Skip to content

Twitch Bot for generating messages based on what it learned from chat

License

Notifications You must be signed in to change notification settings

oliver-lloyd/TwitchMarkovChain

 
 

Repository files navigation

TwitchMarkovChain

Twitch Bot for generating messages based on what it learned from chat


Explanation

When the bot has started, it will start listening to chat messages in the channel listed in the settings.txt file. Any chat message not sent by a denied user will be learned from. Whenever someone then requests a message to be generated, a Markov Chain will be used with the learned data to generate a sentence. Note that the bot is unaware of the meaning of any of its inputs and outputs. This means it can use bad language if it was taught to use bad language by people in chat. You can add a list of banned words it should never learn or say. Use at your own risk.

Whenever a message is deleted from chat, it's contents will be unlearned at 5 times the rate a normal message is learned from. The bot will avoid learning from commands, or from messages containing links.


How it works

Sentence Parsing

To explain how the bot works, I will provide an example situation with two messages. The messages are:

Curly fries are the worst kind of fries
Loud people are the reason I don't go to the movies anymore

Let's start with the first sentence and parse it like the bot will. To do so, we will split up the sentence in sections of size keyLength + 1 words. As keyLength has been set to 2 in the Settings section, each section has size 3.


 Curly fries are the worst kind of fries
[Curly fries:are]
      [fries are:the]
            [are the:worst]
                [the worst:kind]
                    [worst kind:of]
                          [kind of:fries]

For each of these sections of three words, the last word is considered the output, while all other words it are considered inputs. These words are then turned into a variation of a Grammar:

"Curly fries" -> "are"
"fries are"   -> "the"
"are the"     -> "worst"
"the worst"   -> "kind"
"worst kind"  -> "of"
"kind of"     -> "fries"

This can be considered a function that, when given input "the worst", will output "kind".
In order for this to know where sentences begin, we also add the first keyLength words to a seperate Database, where a list of possible starts of sentences reside.

This exact same process is applied to the second sentence as well. After doing so, the resulting grammar (and our database) looks like:

"Curly fries" -> "are"
"fries are"   -> "the"
"are the"     -> "worst" | "reason"
"the worst"   -> "kind"
"worst kind"  -> "of"
"kind of"     -> "fries"
"Loud people" -> "are"
"people are"  -> "the"
"the reason"  -> "I"
"reason I"    -> "don't"
"I don't"     -> "go"
"don't go"    -> "to"
"go to"       -> "the"
"to the"      -> "movies"
"the movies"  -> "anymore"

and in the database for starts of sentences:

"Curly fries"
"Loud people"

Note that the | is considered to be "or". In the case of the bold text above, it could be read as: if the given input is "are the", then the output is "worst" or "reason".

In practice, more frequent phrases will have higher precedence. The more often a phrase is said, the more likely it is to be generated.

Generation

When a message is generated with !generate, a random start of a sentence is picked from the database of starts of sentences. In our example the randomly picked start is "Curly fries".

Now, in a loop:

  • the output for the input is generated via the grammar,
  • and the input is shifted to remove the first word and add the output to the input.

A more programmatic example of this would be this:

# This initial sentence is either from the database for starts of sentences, 
# or from words passed in Twitch chat
sentence = ["Curly", "fries"]
for i in range(sentence_length):
    # Generate a word using last 2 words in the partial sentence, 
    # and append it to the partial sentence
    sentence.append(generate(sentence[-2:]))

It's common for an input sequence to have multiple possible outputs, as we can see in the bold part of the previous grammar. This allows learned information from multiple messages to be merged into one message. For instance, some potential outputs from the given example are

Curly fries are the reason I don't go to the movies anymore

or

Loud people are the worst kind of fries

Commands

Chat members can generate chat-like messages using the following commands (Note that they are aliases):

!generate [words]
!g [words]

Example:

!g Curly

Result (for example):

Curly fries are the reason I don't go to the movies anymore
  • The bot will, when given this command, try to complete the start of the sentence which was given.
    • If it cannot, an appropriate error message will be sent to chat.
  • Any number of words may be given, including none at all.
  • Everyone can use it.

Streamer commands

All of these commands can be whispered to the bot account, or typed in chat.
To disable the bot from generating messages, while still learning from regular chat messages:

!disable

After disabling the bot, it can be re-enabled using:

!enable

Changing the cooldown between generations is possible with one of the following two commands:

!setcooldown <seconds>
!setcd <seconds>

Example:

!setcd 30

Which sets the cooldown between generations to 30 seconds.


Moderator commands

All of these commands must be whispered to the bot account.
Moderators (and the broadcaster) can modify the blacklist to prevent the bot learning words it shouldn't.
To add word to the blacklist, a moderator can whisper the bot:

!blacklist word

Similarly, to remove word from the blacklist, a moderator can whisper the bot:

!whitelist word

And to check whether word is already on the blacklist or not, a moderator can whisper the bot:

!check word

Settings

This bot is controlled by a settings.txt file, which looks like:

{
    "Host": "irc.chat.twitch.tv",
    "Port": 6667,
    "Channel": "#<channel>",
    "Nickname": "<name>",
    "Authentication": "oauth:<auth>",
    "DeniedUsers": [
        "StreamElements",
        "Nightbot",
        "Moobot",
        "Marbiebot"
    ],
    "Cooldown": 20,
    "KeyLength": 2
}
Parameter Meaning Example
Host The URL that will be used. Do not change. "irc.chat.twitch.tv"
Port The Port that will be used. Do not change. 6667
Channel The Channel that will be connected to. "#CubieDev"
Nickname The Username of the bot account. "CubieB0T"
Authentication The OAuth token for the bot account. "oauth:pivogip8ybletucqdz4pkhag6itbax"
DeniedUsers The list of bot account who's messages should not be learned from. The bot itself it automatically added to this. ["StreamElements", "Nightbot", "Moobot", "Marbiebot"]
Cooldown A cooldown in seconds between successful generations. If a generation fails (eg inputs it can't work with), then the cooldown is not reset and another generation can be done immediately. 20
KeyLength A technical parameter which, in my previous implementation, would affect how closely the output matches the learned inputs. In the current implementation the database structure does not allow this parameter to be changed. Do not change. 2

Note that the example OAuth token is not an actual token, but merely a generated string to give an indication what it might look like.

I got my real OAuth token from https://twitchapps.com/tmi/.

Blacklist

You may add words to a blacklist by adding them on a separate line in blacklist.txt. Each word is case insensitive. By default, this file only contains <start> and <end>, which are required for the current implementation.

Words can also be added or removed from the blacklist via whispers, as is described in the Moderator Command section.


Requirements

Among these modules is my own TwitchWebsocket wrapper, which makes making a Twitch chat bot a lot easier. This repository can be seen as an implementation using this wrapper.


Other Twitch Bots

About

Twitch Bot for generating messages based on what it learned from chat

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 100.0%