Run the previous `normalize.py` on FB chat data, then run this script on its output to get sentences generated using a Markov chain.

markov-single-word.py

path = "normalized_data.json"
user = "Keri Warr"

import json, sys, re
import random
from collections import defaultdict

data = json.load(open(path, "r"))

START, END = 0, 1

# get word token matcher
PUNCTUATION = r"[`~@#$%_\\'+\-/]" # punctuation that is a part of text
STANDALONE = r"(?:[!.,;()^&\[\]{}|*=<>?]|[dDpP][:8]|:\S)" # standalone characters or emoticons that wouldn't otherwise be captured
PATTERN = STANDALONE + r"\S*|https?://\S+|(?:\w|" + PUNCTUATION + r")+" # token pattern
matcher = re.compile(PATTERN, re.IGNORECASE)

def train(data):
    # get messages typed by user
    corpus = [entry[2] for entry in data if entry[1] == user]
    messages = []
    for entry in data:
        if entry[1] == user:
            messages.append(START)
            messages += matcher.findall(entry[2])
            messages.append(END)

    # lowercase all the messages for better results
    messages = [m.lower() if isinstance(m, str) else m for m in messages]

    # find word chain counts as a dictionary mapping words to dictionaries mapping words to amount of times they appear after the first word
    i, total = 0, len(messages)
    word_chains = defaultdict(lambda: defaultdict(int))
    while True:
        if i >= total - 1: break
        current, next = messages[i], messages[i + 1]
        word_chains[current][next] += 1
        i += 1
    return word_chains

def speak(word_chains):
    # generate a sentence based on probability chains
    choices = word_chains[START]
    phrase_list = []
    while True:
        random_choice = random.random() * sum(choices.values())
        for current_choice, w in choices.items():
            random_choice -= w
            if random_choice < 0:
                new_word = current_choice
                break
        else: # couldn't find the choice somehow
            raise Exception("Bad choice!")
        if current_choice == END: break
        phrase_list.append(current_choice)
        choices = word_chains[current_choice]

    # format the sentence into a human-readable string
    close_matcher = re.compile("[!.,;)\]}?]", re.IGNORECASE)
    phrase = ""
    for i, v in enumerate(phrase_list):
        if i == 0 or close_matcher.match(v): phrase += v
        else: phrase += " " + v
    
    return phrase

#result = "\n".join([str(word) + ":\t" + str(probability) for word, probability in sorted(word_chains[START].items(), key=lambda x: -x[1])])
#print(result.encode(sys.stdout.encoding, errors="replace").decode(sys.stdout.encoding))

chains = train(data)
for i in range(5000):
    print(speak(chains).encode(sys.stdout.encoding, errors="replace").decode(sys.stdout.encoding))