Нужно будет еще переделать этот гребанный рандомайзер. Он многое гадит.

Sentence_generator

#coding:utf-8
import re
from random import uniform
from collections import defaultdict
from pymongo import MongoClient, ASCENDING

sent_end = ('.!?,;:$')
comma = ',;:'
r_alphabet = re.compile(u'[a-zA-Zа-яёА-ЯЁ0-9-]+|[.,:;?!]+')

__doc__ = """

"""

class model(object):
    def __init__(self, host='localhost', port=27017, db_name='sentences'):
        db = MongoClient(host, port)[db_name]
        self.gramms = db['gramms']
        if len(self.gramms.index_information()) <= 1:
            self.gramms.ensure_index([('1', ASCENDING), ('2', ASCENDING)], unique=True)


    def init_sub_model(self, words):
        sub_model = {}
        for word in words:
            for el in self.gramms.find({'$or': [{'1': word}, {'2': word}, {'sib.3': word}]}):
                sub_model[(el['1'], el['2'])] = el['sib']
                #sanitize

        res_sub_model = {}
        is_good_data = False
        for key, value in sub_model.iteritems():
            for sib_el in value:
                next = key[1], sib_el['3']
                if next in sub_model:
                    if key in res_sub_model:
                        res_sub_model[key].append(sib_el)
                    else:
                        res_sub_model[key] = [sib_el]

                    if sib_el['3'] in sent_end and not is_good_data:
                        is_good_data = True
                        res_sub_model[sib_el['3'], '$'] = [{'3': '$', 'w': 1}]

        if ('$', '$') not in sub_model or not is_good_data:
            raise Exception('bad train data for this words')

        return res_sub_model

    def get_seq(self, t0, t1):
        res = self.gramms.find_one({'1': t0, '2': t1})
        if res:
            return res['sib']

    def add_to_seq(self, t0, t1, token, weight):
        container = self.gramms.find_one({'1': t0, '2': t1})
        if container:
            self.gramms.update(container, {'$push': {'sib': {'3': token, 'w': weight}}}, upsert=True)
        else:
            self.gramms.save({'1': t0, '2': t1, 'sib': [{'3': token, 'w': weight}]})


def gen_lines(corpus):
    data = open(corpus)
    for line in data:
        yield line.decode('utf-8').lower()


def gen_tokens(lines):
    for line in lines:
        for token in r_alphabet.findall(line):
            yield token


def gen_trigrams(tokens):
    t0, t1 = '$', '$'
    for t2 in tokens:
        yield t0, t1, t2
        if t2 in '.!?':
            yield t1, t2, '$'
            yield t2, '$', '$'
            t0, t1 = '$', '$'
        else:
            t0, t1 = t1, t2


def train(corpus):
    lines = gen_lines(corpus)
    tokens = gen_tokens(lines)
    trigrams = gen_trigrams(tokens)

    bi, tri = defaultdict(lambda: 0.0), defaultdict(lambda: 0.0)

    for t0, t1, t2 in trigrams:
        bi[t0, t1] += 1
        tri[t0, t1, t2] += 1

    m = model()
    for (t0, t1, t2), freq in tri.iteritems():
        m.add_to_seq(t0, t1, t2, freq / bi[t0, t1])
    return m


def generate_sentence(model):
    phrase = ''
    t0, t1 = '$', '$'
    while 1: # for every iteration in t1 new word random getted from model
        if (t0, t1) not in model:
            break
        t0, t1 = t1, unirand(model[t0, t1])
        if t1 == '$': break
        if t1 in sent_end or t0 == '$':
            phrase += t1
        else:
            phrase += ' ' + t1
        print phrase
    return phrase.capitalize()


def unirand(seq):
    """
    here generating next word on sequence.
    """
    sum_, freq_ = 0, 0
    for el in seq:
        sum_ += el['w']
        rnd = uniform(0, sum_)
    for el in seq:
        freq_ += el['w']
        if rnd < freq_:
            return el['3']

if __name__ == '__main__':
##    model = train('/home/alesha/develop/seo-venv/dicts/text.txt')
#    model = model(words=[u'начать', u'имении', u'граф', u'пьер'])
#    for i in range(10):
#        print generate_sentence(model)

    #model = train('../data/data.txt')
    model = model()
    words = [u'любовь', u'деньги', u'я', u'шеф']
    sub_model = model.init_sub_model(words)
    print generate_sentence(sub_model)

При изменении рандомайзера следует учесть также связи слов во входной выборке. Дабы предложение сочленялось. А не как сейчас - генерируется предложение из одного слова. А также, прикрутить pymorphy, но с более высокими n, один хер вариантов не так много нежели слов.