Build and train an RNN chatbot using TensorFlow [Tutorial]

Chatbots are increasingly used as a way to provide assistance to users. Many companies, including banks, mobile/landline companies and large e-sellers now use chatbots for customer assistance and for helping users in pre and post sales queries. They are a great tool for companies which don't need to provide additional customer service capacity for trivial questions: they really look like a win-win situation!

In today’s tutorial, we will understand how to train an automatic chatbot that will be able to answer simple and generic questions, and how to create an endpoint over HTTP for providing the answers via an API.

This article is an excerpt from a book written by Luca Massaron, Alberto Boschetti, Alexey Grigorev, Abhishek Thakur, and Rajalingappaa Shanmugamani titled TensorFlow Deep Learning Projects.

There are mainly two types of chatbot: the first is a simple one, which tries to understand the topic, always providing the same answer for all questions about the same topic. For example, on a train website, the questions Where can I find the timetable of the City_A to City_B service? and What's the next train departing from City_A? will likely get the same answer, that could read Hi! The timetable on our network is available on this page: <link>. This types of chatbots use classification algorithms to understand the topic (in the example, both questions are about the timetable topic).

Given the topic, they always provide the same answer. Usually, they have a list of N topics and N answers; also, if the probability of the classified topic is low (the question is too vague, or it's on a topic not included in the list), they usually ask the user to be more specific and repeat the question, eventually pointing out other ways to do the question (send an email or call the customer service number, for example).

The second type of chatbots is more advanced, smarter, but also more complex. For those, the answers are built using an RNN, in the same way, that machine translation is performed. Those chatbots are able to provide more personalized answers, and they may provide a more specific reply. In fact, they don't just guess the topic, but with an RNN engine, they're able to understand more about the user's questions and provide the best possible answer: in fact, it's very unlikely you'll get the same answers with two different questions using these types of chatbots.

The input corpus

Unfortunately, we haven't found any consumer-oriented dataset that is open source and freely available on the Internet. Therefore, we will train the chatbot with a more generic dataset, not really focused on customer service. Specifically, we will use the Cornell Movie Dialogs Corpus, from the Cornell University.

The corpus contains the collection of conversations extracted from raw movie scripts, therefore the chatbot will be able to answer more to fictional questions than real ones. The Cornell corpus contains more than 200,000 conversational exchanges between 10+ thousands of movie characters, extracted from 617 movies.

The dataset comes as a .zip archive file. After decompressing it, you'll find several files in it:

• README.txt contains the description of the dataset, the format of the corpora files, the details on the collection procedure and the author's contact.
• Chameleons.pdf is the original paper for which the corpus has been released. Although the goal of the paper is strictly not around chatbots, it studies the language used in dialogues, and it's a good source of information to understanding more
• movie_conversations.txt contains all the dialogues structure. For each conversation, it includes the ID of the two characters involved in the discussion, the ID of the movie and the list of sentences IDs (or utterances, to be more precise) in chronological order. For example, the first line of the file is:

u0 +++$+++ u2 +++$+++ m0 +++$+++ ['L194', 'L195', 'L196', 'L197']

That means that user u0 had a conversation with user u2 in the movie m0 and the conversation had 4 utterances: 'L194', 'L195', 'L196' and 'L197'

• movie_lines.txt contains the actual text of each utterance ID and the person who produced it. For example, the utterance L195 is listed here as:

L195 +++$+++ u2 +++$+++ m0 +++$+++ CAMERON +++$+++ Well, I thought we'd start with pronunciation, if that's okay with you.

So, the text of the utterance L195 is Well, I thought we'd start with pronunciation, if that's okay with you. And it was pronounced by the character u2 whose name is CAMERON in the movie m0.

• movie_titles_metadata.txt contains information about the movies, including the title, year, IMDB rating, the number of votes in IMDB and the genres. For example, the movie m0 here is described as:

m0 +++$+++ 10 things i hate about you +++$+++ 1999 +++$+++ 6.90 +++$+++ 62847 +++$+++ ['comedy', 'romance']

So, the title of the movie whose ID is m0 is 10 things i hate about you, it's from 1999, it's a comedy with romance and it received almost 63 thousand votes on IMDB with an average score of 6.9 (over 10.0)

• movie_characters_metadata.txt contains information about the movie characters, including the name the title of the movie where he/she appears, the gender (if known) and the position in the credits (if known). For example, the character “u2” appears in this file with this description:

u2 +++$+++ CAMERON +++$+++ m0 +++$+++ 10 things i hate about you +++$+++ m +++$+++ 3

The character u2 is named CAMERON, it appears in the movie m0 whose title is 10 things i hate about you, his gender is male and he's the third person appearing in the credits.

• raw_script_urls.txt contains the source URL where the dialogues of each movie can be retrieved. For example, for the movie m0 that's it:

m0 +++$+++ 10 things i hate about you +++$+++ http://www.dailyscript.com/scripts/10Things.html

As you will have noticed, most files use the token  +++$+++  to separate the fields. Beyond that, the format looks pretty straightforward to parse. Please take particular care while parsing the files: their format is not UTF-8 but ISO-8859-1.

Creating the training dataset

Let's now create the training set for the chatbot. We'd need all the conversations between the characters in the correct order: fortunately, the corpora contains more than what we actually need. For creating the dataset, we will start by downloading the zip archive, if it's not already on disk. We'll then decompress the archive in a temporary folder (if you're using Windows, that should be C:Temp), and we will read just the movie_lines.txt and the movie_conversations.txt files, the ones we really need to create a dataset of consecutive utterances.

Let's now go step by step, creating multiple functions, one for each step, in the file corpora_downloader.py. The first function we need is to retrieve the file from the Internet, if not available on disk.

def download_and_decompress(url, storage_path, storage_dir):
   import os.path
   directory = storage_path + "/" + storage_dir
   zip_file = directory + ".zip"
   a_file = directory + "/cornell movie-dialogs corpus/README.txt"
   if not os.path.isfile(a_file):
       import urllib.request
       import zipfile
       urllib.request.urlretrieve(url, zip_file)
       with zipfile.ZipFile(zip_file, "r") as zfh:
           zfh.extractall(directory)
   return

This function does exactly that: it checks whether the “README.txt” file is available locally; if not, it downloads the file (thanks for the urlretrieve function in the urllib.request module) and it decompresses the zip (using the zipfile module).

The next step is to read the conversation file and extract the list of utterance IDS. As a reminder, its format is: u0 +++$+++ u2 +++$+++ m0 +++$+++ ['L194', 'L195', 'L196', 'L197'], therefore what we're looking for is the fourth element of the list after we split it on the token  +++$+++ . Also, we'd need to clean up the square brackets and the apostrophes to have a clean list of IDs. For doing that, we shall import the re module, and the function will look like this.

import re
def read_conversations(storage_path, storage_dir):
   filename = storage_path + "/" + storage_dir + "/cornell movie-dialogs corpus/movie_conversations.txt"
   with open(filename, "r", encoding="ISO-8859-1") as fh:
       conversations_chunks = [line.split(" +++$+++ ") for line in fh]
   return [re.sub('[[]']', '', el[3].strip()).split(", ") for el in conversations_chunks]

As previously said, remember to read the file with the right encoding, otherwise, you'll get an error. The output of this function is a list of lists, each of them containing the sequence of utterance IDS in a conversation between characters. Next step is to read and parse the movie_lines.txt file, to extract the actual utterances texts. As a reminder, the file looks like this line:

L195 +++$+++ u2 +++$+++ m0 +++$+++ CAMERON +++$+++ Well, I thought we'd start with pronunciation, if that's okay with you.

Here, what we're looking for are the first and the last chunks.

def read_lines(storage_path, storage_dir):
   filename = storage_path + "/" + storage_dir + "/cornell movie-dialogs corpus/movie_lines.txt"
   with open(filename, "r", encoding="ISO-8859-1") as fh:
       lines_chunks = [line.split(" +++$+++ ") for line in fh]
   return {line[0]: line[-1].strip() for line in lines_chunks}

The very last bit is about tokenization and alignment. We'd like to have a set whose observations have two sequential utterances. In this way, we will train the chatbot, given the first utterance, to provide the next one. Hopefully, this will lead to a smart chatbot, able to reply to multiple questions. Here's the function:

def get_tokenized_sequencial_sentences(list_of_lines, line_text):
   for line in list_of_lines:
       for i in range(len(line) - 1):
           yield (line_text[line[i]].split(" "), line_text[line[i+1]].split(" "))

Its output is a generator containing a tuple of the two utterances (the one on the right follows temporally the one on the left). Also, utterances are tokenized on the space character.

Finally, we can wrap up everything into a function, which downloads the file and unzip it (if not cached), parse the conversations and the lines, and format the dataset as a generator. As a default, we will store the files in the /tmp directory:

def retrieve_cornell_corpora(storage_path="/tmp", storage_dir="cornell_movie_dialogs_corpus"):
   download_and_decompress("http://www.cs.cornell.edu/~cristian/data/cornell_movie_dialogs_corpus.zip",      
                     storage_path,
                           storage_dir)
   conversations = read_conversations(storage_path, storage_dir)
   lines = read_lines(storage_path, storage_dir)
   return tuple(zip(*list(get_tokenized_sequencial_sentences(conversations, lines))))

At this point, our training set looks very similar to the training set used in the translation project. We can, therefore, use some pieces of code we've developed in the machine learning translation article. For example, the corpora_tools.py file can be used here without any change (also, it requires the data_utils.py).

Given that file, we can dig more into the corpora, with a script to check the chatbot input.

To inspect the corpora, we can use the corpora_tools.py, and the file we've previously created. Let's retrieve the Cornell Movie Dialog Corpus, format the corpora and print an example and its length:

from corpora_tools import *
from corpora_downloader import retrieve_cornell_corpora
sen_l1, sen_l2 = retrieve_cornell_corpora()
print("# Two consecutive sentences in a conversation")
print("Q:", sen_l1[0])
print("A:", sen_l2[0])
print("# Corpora length (i.e. number of sentences)")
print(len(sen_l1))
assert len(sen_l1) == len(sen_l2)

This code prints an example of two tokenized consecutive utterances, and the number of examples in the dataset, that is more than 220,000:

# Two consecutive sentences in a conversation
Q: ['Can', 'we', 'make', 'this', 'quick?', '', 'Roxanne', 'Korrine', 'and', 'Andrew', 'Barrett', 'are', 'having', 'an', 'incredibly', 'horrendous', 'public', 'break-', 'up', 'on', 'the', 'quad.', '', 'Again.']
A: ['Well,', 'I', 'thought', "we'd", 'start', 'with', 'pronunciation,', 'if', "that's", 'okay', 'with', 'you.']
# Corpora length (i.e. number of sentences)
221616

Let's now clean the punctuation in the sentences, lowercase them and limits their size to 20 words maximum (that is examples where at least one of the sentences is longer than 20 words are discarded). This is needed to standardize the tokens:

clean_sen_l1 = [clean_sentence(s) for s in sen_l1]
clean_sen_l2 = [clean_sentence(s) for s in sen_l2]
filt_clean_sen_l1, filt_clean_sen_l2 = filter_sentence_length(clean_sen_l1, clean_sen_l2)
print("# Filtered Corpora length (i.e. number of sentences)")
print(len(filt_clean_sen_l1))
assert len(filt_clean_sen_l1) == len(filt_clean_sen_l2)

This leads us to almost 140,000 examples:

# Filtered Corpora length (i.e. number of sentences)
140261

Then, let's create the dictionaries for the two sets of sentences. Practically, they should look the same (since the same sentence appears once on the left side, and once in the right side) except there might be some changes introduced by the first and last sentences of a conversation (they appear only once). To make the best out of our corpora, let's build two dictionaries of words and then encode all the words in the corpora with their dictionary indexes:

dict_l1 = create_indexed_dictionary(filt_clean_sen_l1, dict_size=15000, storage_path="/tmp/l1_dict.p")
dict_l2 = create_indexed_dictionary(filt_clean_sen_l2, dict_size=15000, storage_path="/tmp/l2_dict.p")
idx_sentences_l1 = sentences_to_indexes(filt_clean_sen_l1, dict_l1)
idx_sentences_l2 = sentences_to_indexes(filt_clean_sen_l2, dict_l2)
print("# Same sentences as before, with their dictionary ID")
print("Q:", list(zip(filt_clean_sen_l1[0], idx_sentences_l1[0])))
print("A:", list(zip(filt_clean_sen_l2[0], idx_sentences_l2[0])))

That prints the following output. We also notice that a dictionary of 15 thousand entries doesn't contain all the words and more than 16 thousand (less popular) of them don't fit into it:

[sentences_to_indexes] Did not find 16823 words
[sentences_to_indexes] Did not find 16649 words
# Same sentences as before, with their dictionary ID
Q: [('well', 68), (',', 8), ('i', 9), ('thought', 141), ('we', 23), ("'", 5), ('d', 83), ('start', 370), ('with', 46), ('pronunciation', 3), (',', 8), ('if', 78), ('that', 18), ("'", 5), ('s', 12), ('okay', 92), ('with', 46), ('you', 7), ('.', 4)]
A: [('not', 31), ('the', 10), ('hacking', 7309), ('and', 23), ('gagging', 8761), ('and', 23), ('spitting', 6354), ('part', 437), ('.', 4), ('please', 145), ('.', 4)]

As the final step, let's add paddings and markings to the sentences:

data_set = prepare_sentences(idx_sentences_l1, idx_sentences_l2, max_length_l1, max_length_l2)
print("# Prepared minibatch with paddings and extra stuff")
print("Q:", data_set[0][0])
print("A:", data_set[0][1])
print("# The sentence pass from X to Y tokens")
print("Q:", len(idx_sentences_l1[0]), "->", len(data_set[0][0]))
print("A:", len(idx_sentences_l2[0]), "->", len(data_set[0][1]))

And that, as expected, prints:

# Prepared minibatch with paddings and extra stuff
Q: [0, 68, 8, 9, 141, 23, 5, 83, 370, 46, 3, 8, 78, 18, 5, 12, 92, 46, 7, 4]
A: [1, 31, 10, 7309, 23, 8761, 23, 6354, 437, 4, 145, 4, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0]
# The sentence pass from X to Y tokens
Q: 19 -> 20
A: 11 -> 22

Training the chatbot

After we're done with the corpora, it's now time to work on the model. This project requires again a sequence to sequence model, therefore we can use an RNN. Even more, we can reuse part of the code from the previous project: we'd just need to change how the dataset is built, and the parameters of the model. We can then copy the training script, and modify the build_dataset function, to use the Cornell dataset.

Mind that the dataset used in this article is bigger than the one used in the machine learning translation article, therefore you may need to limit the corpora to a few dozen thousand lines. On a 4 years old laptop with 8GB RAM, we had to select only the first 30 thousand lines, otherwise, the program ran out of memory and kept swapping. As a side effect of having fewer examples, even the dictionaries are smaller, resulting in less than 10 thousands words each.

def build_dataset(use_stored_dictionary=False):
   sen_l1, sen_l2 = retrieve_cornell_corpora()
   clean_sen_l1 = [clean_sentence(s) for s in sen_l1][:30000] ### OTHERWISE IT DOES NOT RUN ON MY LAPTOP
   clean_sen_l2 = [clean_sentence(s) for s in sen_l2][:30000] ### OTHERWISE IT DOES NOT RUN ON MY LAPTOP
   filt_clean_sen_l1, filt_clean_sen_l2 = filter_sentence_length(clean_sen_l1, clean_sen_l2, max_len=10)
   if not use_stored_dictionary:
       dict_l1 = create_indexed_dictionary(filt_clean_sen_l1, dict_size=10000, storage_path=path_l1_dict)
       dict_l2 = create_indexed_dictionary(filt_clean_sen_l2, dict_size=10000, storage_path=path_l2_dict)
   else:
       dict_l1 = pickle.load(open(path_l1_dict, "rb"))
       dict_l2 = pickle.load(open(path_l2_dict, "rb"))
   dict_l1_length = len(dict_l1)
   dict_l2_length = len(dict_l2)
   idx_sentences_l1 = sentences_to_indexes(filt_clean_sen_l1, dict_l1)
   idx_sentences_l2 = sentences_to_indexes(filt_clean_sen_l2, dict_l2)
   max_length_l1 = extract_max_length(idx_sentences_l1)
   max_length_l2 = extract_max_length(idx_sentences_l2)
   data_set = prepare_sentences(idx_sentences_l1, idx_sentences_l2, max_length_l1, max_length_l2)
   return (filt_clean_sen_l1, filt_clean_sen_l2), 
           data_set, 
           (max_length_l1, max_length_l2), 
           (dict_l1_length, dict_l2_length)

By inserting this function into the train_translator.py file and rename the file as train_chatbot.py, we can run the training of the chatbot.

After a few iterations, you can stop the program and you'll see something similar to this output:

[sentences_to_indexes] Did not find 0 words
[sentences_to_indexes] Did not find 0 words
global step 100 learning rate 1.0 step-time 7.708967611789704 perplexity 444.90090078460474
eval: perplexity 57.442316329639176
global step 200 learning rate 0.990234375 step-time 7.700247814655302 perplexity 48.8545568311572
eval: perplexity 42.190180314697045
global step 300 learning rate 0.98046875 step-time 7.69800933599472 perplexity 41.620538109894945
eval: perplexity 31.291903031786116
...
...
...
global step 2400 learning rate 0.79833984375 step-time 7.686293318271639 perplexity 3.7086356605442767
eval: perplexity 2.8348589631663046
global step 2500 learning rate 0.79052734375 step-time 7.689657487869262 perplexity 3.211876894960698
eval: perplexity 2.973809378544393
global step 2600 learning rate 0.78271484375 step-time 7.690396382808681 perplexity 2.878854805600354
eval: perplexity 2.563583924617356

Again, if you change the settings, you may end up with a different perplexity. To obtain these results, we set the RNN size to 256 and 2 layers, the batch size of 128 samples, and the learning rate to 1.0.

At this point, the chatbot is ready to be tested. Although you can test the chatbot with the same code as in the test_translator.py, here we would like to do a more elaborate solution, which allows exposing the chatbot as a service with APIs.

Chatbox API

First of all, we need a web framework to expose the API. In this project, we've chosen Bottle, a lightweight simple framework very easy to use.

import pickle
import sys
import numpy as np
import tensorflow as tf
import data_utils
from corpora_tools import clean_sentence, sentences_to_indexes, prepare_sentences
from train_chatbot import get_seq2seq_model, path_l1_dict, path_l2_dict
model_dir = "/home/abc/chat/chatbot_model"
def prepare_sentence(sentence, dict_l1, max_length):
   sents = [sentence.split(" ")]
   clean_sen_l1 = [clean_sentence(s) for s in sents]
   idx_sentences_l1 = sentences_to_indexes(clean_sen_l1, dict_l1)
   data_set = prepare_sentences(idx_sentences_l1, [[]], max_length, max_length)
   sentences = (clean_sen_l1, [[]])
   return sentences, data_set

The function prepare_sentence does the following:

• Tokenizes the input sentence
• Cleans it (lowercase and punctuation cleanup)
• Converts tokens to dictionary IDs
• Add markers and paddings to reach the default length

Next, we will need a function to convert the predicted sequence of numbers to an actual sentence composed of words. This is done by the function decode, which runs the prediction given the input sentence and with softmax predicts the most likely output. Finally, it returns the sentence without paddings and markers:

def decode(data_set):
with tf.Session() as sess:
   model = get_seq2seq_model(sess, True, dict_lengths, max_sentence_lengths, model_dir)
   model.batch_size = 1
   bucket = 0
   encoder_inputs, decoder_inputs, target_weights = model.get_batch(
     {bucket: [(data_set[0][0], [])]}, bucket)
   _, _, output_logits = model.step(sess, encoder_inputs, decoder_inputs,
                                   target_weights, bucket, True)
   outputs = [int(np.argmax(logit, axis=1)) for logit in output_logits]
   if data_utils.EOS_ID in outputs:
       outputs = outputs[1:outputs.index(data_utils.EOS_ID)]
tf.reset_default_graph()
return " ".join([tf.compat.as_str(inv_dict_l2[output]) for output in outputs])

Finally, the main function, that is, the function to run in the script:

if __name__ == "__main__":
   dict_l1 = pickle.load(open(path_l1_dict, "rb"))
   dict_l1_length = len(dict_l1)
   dict_l2 = pickle.load(open(path_l2_dict, "rb"))
   dict_l2_length = len(dict_l2)
   inv_dict_l2 = {v: k for k, v in dict_l2.items()}
   max_lengths = 10
   dict_lengths = (dict_l1_length, dict_l2_length)
   max_sentence_lengths = (max_lengths, max_lengths)
   from bottle import route, run, request
   @route('/api')
   def api():
       in_sentence = request.query.sentence
     _, data_set = prepare_sentence(in_sentence, dict_l1, max_lengths)
       resp = [{"in": in_sentence, "out": decode(data_set)}]
       return dict(data=resp)
   run(host='127.0.0.1', port=8080, reloader=True, debug=True)

Initially, it loads the dictionary and prepares the inverse dictionary. Then, it uses the Bottle API to create an HTTP GET endpoint (under the /api URL). The route decorator sets and enriches the function to run when the endpoint is contacted via HTTP GET. In this case, the api() function is run, which first reads the sentence passed as HTTP parameter, then calls the prepare_sentence function, described above, and finally runs the decoding step. What's returned is a dictionary containing both the input sentence provided by the user and the reply of the chatbot.

Finally, the webserver is turned on, on the localhost at port 8080. Isn't very easy to have a chatbot as a service with Bottle?

It's now time to run it and check the outputs. To run it, run from the command line:

$> python3 –u test_chatbot_aas.py

Then, let's start querying the chatbot with some generic questions, to do so we can use CURL, a simple command line; also all the browsers are ok, just remember that the URL should be encoded, for example, the space character should be replaced with its encoding, that is, %20.

Curl makes things easier, having a simple way to encode the URL request. Here are a couple of examples:

$> curl -X GET -G http://127.0.0.1:8080/api --data-urlencode "sentence=how are you?"
{"data": [{"out": "i ' m here with you .", "in": "where are you?"}]}
$> curl -X GET -G http://127.0.0.1:8080/api --data-urlencode "sentence=are you here?"
{"data": [{"out": "yes .", "in": "are you here?"}]}
$> curl -X GET -G http://127.0.0.1:8080/api --data-urlencode "sentence=are you a chatbot?"
{"data": [{"out": "you ' for the stuff to be right .", "in": "are you a chatbot?"}]}
$> curl -X GET -G http://127.0.0.1:8080/api --data-urlencode "sentence=what is your name ?"
{"data": [{"out": "we don ' t know .", "in": "what is your name ?"}]}
$> curl -X GET -G http://127.0.0.1:8080/api --data-urlencode "sentence=how are you?"
{"data": [{"out": "that ' s okay .", "in": "how are you?"}]}

Replies are quite funny; always remember that we trained the chatbox on movies, therefore the type of replies follow that style.

To turn off the webserver, use Ctrl + C.

To summarize, we've learned to implement a chatbot, which is able to respond to questions through an HTTP endpoint and a GET API.

To know more how to design deep learning systems for a variety of real-world scenarios using TensorFlow, do checkout this book TensorFlow Deep Learning Projects.

• Facebook’s Wit.ai: Why we need yet another chatbot development framework?
• How to build a chatbot with Microsoft Bot framework
• Top 4 chatbot development frameworks for developers