nmt-chatbot

Introduction
Setup
Custom summary values (evaluation)
Standard vs BPE/WPM-like (subword) tokenization, embedded detokenizer
Rules files
Tests
More detailed information about training a model
Utils
Inference
Importing nmt-chatbot
Deploying chatbot/model
Demo chatbot
Changelog

Introduction

nmt-chatbot is the implementation of chatbot using NMT - Neural Machine Translation (seq2seq). Includes BPE/WPM-like tokenizator (own implementation). Main purpose of that project is to make an NMT chatbot, but it's fully compatible with NMT and still can be used for sentence translations between two languages.

The code is built on top of NMT but because of lack of available interfaces, some things are "hacked", and parts of the code had to be copied into that project (and will have to be maintained to follow changes in NMT).

This project forks NMT. We had to make a change in our code allowing the use of a stable TensorFlow (1.4) version. Doing so allowed us also to fix some bug before official patch as well as do couple of necessary changes.

Setup

Steps to setup project for your needs: It is highly recommended that you use Python 3.6+. Python 3.4 and 3.5 is likely to work in Linux, but you will eventually hit encoding errors with 3.5 or lower in a Windows environment.

If you want to use exactly what's in tutorial made by Sentdex, use v0.1 tag. There are multiple changes after last part of tutorial.

$ git clone --recursive https://github.com/daniel-kukiela/nmt-chatbot
(or)
$ git clone --branch v0.1 --recursive https://github.com/daniel-kukiela/nmt-chatbot.git (for a version featured in Sentdex tutorial)
$ cd nmt-chatbot
$ pip install -r requirements.txt TensorFlow-GPU is one of the requirements. You also need CUDA Toolkit 8.0 and cuDNN 6.1. (Windows tutorial: https://www.youtube.com/watch?v=r7-WPbx8VuY Linux tutorial: https://pythonprogramming.net/how-to-cuda-gpu-tensorflow-deep-learning-tutorial/)
$ cd setup
(optional) edit settings.py to your liking. These are a decent starting point for ~4GB of VRAM, you should first start by trying to raise vocab if you can.
(optional) Edit text files containing rules in the setup directory.
Place training data inside "new_data" folder (train.(from|to), tst2013.(from|to), tst2013(from|to)). We have provided some sample data for those who just want to do a quick test drive.
$ python prepare_data.py ...Run setup/prepare_data.py - a new folder called "data" will be created with prepared training data
$ cd ../
$ python train.py Begin training

Version 0.3 introduces epoch-based training including custom (epoch-based as well) decaying scheme - refer to preprocessing['epochs'] in setup/settings.py for more detailed explanation and example (enabled by default).

Custom summary values (evaluation)

It is possible to add custom values logged into model logs. TensorBoard will plot those values in a separate graphs.

To add custom values, modify custom_summary function inside setup/custom_summary.py.

Data object is a list of tuples, where tuple contains:

source phrase
target phrase
nmt phrase

Return must be a dictionary, where:

key - lowercase ascii letters only plus underscore
value - float value

Function is called on every evaluation. As a result returned values will be saved in model logs and plot in TensorBoard.

Standard vs BPE/WPM-like (subword) tokenization, embedded detokenizer

v0.1 includes only standard (own, first version) tokenizer.

Standard tokenizer is based on moses-smt one. It's highly modified own python implementation of that tokenizer. The adventage of tokenizer like that is lack of duplicates in vocab file (more on that later). Biggest disadventage - it needs bunch of regex-based rules for detokenization process and it's hard to write ones that covers all cases.

Standard tokenizer splits sentences by space, period (and other grammar chars), speparates digits, etc.

BPE/WPM-like (subword) tokenizer is based on subword-nmt one, but (like for "standard" one) it's highly modified to fit our needs and for speed. The biggest adventage is ability to fit any number of words (tokens) in much smaller vocab thanks to subwords.

BPE/WPM-like (subword) tokenizer is doing similar splits like standard one, but in addition it splits every entity by char, counts most common pairs of chars (chars next to other chars in entities), and joins that pairs to make a vocab of desired number of tokens. As a result, most common words are joined back together, when the rarest ones will stay split into multiple pieces (subwords) shared with other words. That way near any number of vocab tokens can be lowered to number as small as couple of thousands (or lower - depending on training set). Basically it's something between char model and word model (as for model not beeing a char model, vocab size should be held as big as possible to include most common words in one piece as well ass some elemental tokens like single chars). Subword-based model should produce way less (to none) <unk> special tokens in output at the exchange for sentences with higher number of tokens (possibly shorter sentences outputted by network for purposes like chatbot).

Standard vs. BPE/WPM-like (subword) tokenization examples:

> Aren ' t they streaming it for free online ... ?
> Aren ' t they streaming it for free online ... ?

> THE GREATEST FOOTBALL TEAM !
> THE GRE AT EST F OO TB ALL TEAM !

> Become a tourist , I hear lots of Kerbals go on vacation there ...
> B ec ome a tou rist , I hear lots of Ker b als go on vacation there ...

Embeded detokenizer (option for standard tokenizer, forced to 'on' for BPE/WPM-like one - you can see why above) allows to do a perfect detokenization at the exchange for higher number of tokens in vocab - duplicate-like entities in vocab (variations with and without meta symbol '▁', where that meta symbol is all is needed for detokenization). Biggest adventage of that detokenizer is lack of any rules for detokenization. Detokenization is easy as doing two replaces in resulting sentence - first remove all spaces, then replace meta symbol '▁' with space character.

Examples:

> ▁B ec ome ▁a ▁tou rist , ▁I ▁hear ▁lots ▁of ▁Ker b als ▁go ▁on ▁vacation ▁there ...
> ▁I ' d ▁w ager ▁it ' s ▁appropriate ▁for ▁all ▁kinds ▁of ▁things .

Rules files

Setup folder contains multiple "rules" files (All of them are regex-based:

answers_detokenize.txt - detokenization rules (removes unnecessary spaces, legacy tokenizer only).
answers_replace - synonyms, replaces phrase or it's part with a replacement.
answers_subsentence_score.txt - rules for answer score (list of subsentences and score modifiers that can either lower or raise score when includes certain subsentences).
protected_phrases_standard.txt - ensures that matching phrases will remain untouched when building vocab file with standard tokenizer.
protected_phrases_bpe.txt - same as above but for BPE/WPM-like tokenizer.

Tests

Every rules file has related test script. Those test scripts might be treated as some kind of unit testing. Every modification of rules files might be checked against those tests but every modification should be also followed by new test cases in those scripts.

It's important to check everything before training new model. Even slight change might break something.

Test scripts will display check status, checked sentence and eventually check result (if different than assumed).

More detailed information about training a model

setup/settings.py consist of multiple settings:

untouched file/folder paths should fit for most cases
"preprocessing" dictionary should be easy to understand
"hparams" dictionary will be passed to NMT like command line options with standard usage
"score" dictionary with settings for answer scoring

setup/prepare_data.py:

walks thru files placed in "new_data" folder - train.(from|to), tst2012.(from|to)m tst2013(from|to)
tokenizes all sentences (based on settings and internal rules)
for "train" files - builds vocabulary files, makes them unique and saves up to the number of entities set in setup/settings.py file, unused vocab entities (if any - depends on settings) will be saved to separate files

train.py - starts training process

Utils

utils/run_tensorboard.py is easy to use wrapper starting Tensorboard with model folder

utils/pairing_testing_outputs.py - joins model/output_dev file with data/tst2012.form file and prints result to a console allowing easy check if things are going ok during training. The console will consist of input phrase, inference output frame, and separator.

utils/prepare_for_deployment - copies only necessary files for inference. See more below

Inference

Whenever a model is trained, inference.py, when directly called, allows to "talk to" AI in interactive mode. It will start and setup everything needed to use trained model (using saved hparams file within the model folder and setup/settings.py for the rest of settings or lack of hparams file).

For every question will be printed up to a number of responses set in setup/settings.py. Every response will be marked with one of three colors:

green - first one with that color is a candidate to be returned. Answers to the color passed blacklist check (setup/response_blacklist.txt)
orange - still proper responses, but with lower than maximum score
red - improper response - below threshold

Steps from the question to the answers:

Pass question
Compute up to number of responses set in setup/settings.py
Detokenize answers using either embedded detokenizer or rules from setup/answers_detokenized.txt (depending on settings)
Replace responses or their parts using additional rules
Score responses
Return (show with interactive mode) responses

It is also possible to process a batch of the questions by simply using command redirection:

python inference.py < input_file

or:

python inference.py < input_file > output_file

It is possible to pass specified checkpoint as a parameter to use inference using that checkpoint, for example:

python inference.py translate.ckpt-1000

Importing nmt-chatbot

The project allows being imported for the needs of inference. Simply embed folder within your project and import (notice: change folder name in a way it will not include dash (-) character - you can't import module with that character in it's name), then use:

from nmt_chatbot.inference import inference

print(inference("Hello!"))

(where nmt_chatbot is a directory name of chatbot module)

inference() takes one parameter:

question (required)

For a single question, function will return dictionary containing:

answers - list of all answers
scores - list of scores for answers
best_index - index of best answer
best_score - score of best answer

Score:

Every response starts with score['starting_score'] value from setup/settings.py (10 by default) and is further modified by various checks. Please refer to score section of setup/settings.py for more details and settings. Final score is next used to pick "best" response.

With a list of questions, the function will return a list of dictionaries.

For every empty question, the function will return None instead of result dictionary.

Deploying chatbot/model

Whether model is trained, there might be a need to export only files necessary for interference (for example to be embeded and imported it other project).

That might be achieved automatically by using prepare_for_deployment utility:

cd utils
python prepare_for_deployment.py

Script will create _deployment folder inside project's root directory and copy all necessary files depending on your current settings.

Demo chatbot

We have trained demo model. It's deployment-ready package.

As it's a demo model, it was trained on small data-set (corpus) of one million pairs and using relatively small neural network. It is not a decent model, but good enough to try it and see how it's working.

You can download package here. It includes all but also only necessary files to run inference (doesn't include training corpus for example). All model settings can be found in setup/setting.py file.

Changelog

Master

New response scoring engine (work in progress, suggestions and code improvements are welcome)
Showing score modifiers in 'live' inference mode
Fixed 'learning bpe' stage of 'prepare_data' speed issue (it's multiple times faster now)
Improved 'prepare training set' stage of 'prepare_data' speed (should run about 1/3rd faster)
Fixed model paths - pathhs are relative now, so model can be easily moved between different paths or even machines
Fixed info about importing project as a module
Updated README
Added changelog
Added table of contents
Added passing checkpoint name as a parameter for inference
Added deployment script
Added ability to cache some prepare_data common steps for multiple script run
Added epoch-based training
Added custom decaying scheme (epoch-based)
Added ability to return own evaluation values (will be plotted in TensorBoard)
Updated NMT fork (fixed train_ppl graph, added evaluation outputs saved to a separate files, added ability to pass custom evaluation callback function)
Merged latest NMT changes into our fork
Various fixes and other small improvements

v0.2

BPE/WPM-like tokenizer
Updated NMT
Enabled NMT's SMP detokenizer with our embedded BPE/WPM-like tokenizer
Fixed issue with paths on Linux and MacOS machines
Improved pair testing utility
Fixed command for tag cloning
Various fixes and other small improvements, improved readme file

v0.1

Initial commit, code for tutorial: https://pythonprogramming.net/chatbot-deep-learning-python-tensorflow/

koliadych/bot