uber-archive/plato-research-dialogue-system

Plato Research Dialogue System

This is v0.3.1

The Plato Research Dialogue System is a flexible framework that can be used to create, train, and evaluate conversational AI agents in various environments. It supports interactions through speech, text, or dialogue acts and each conversational agent can interact with data, human users, or other conversational agents (in a multi-agent setting). Every component of every agent can be trained independently online or offline and Plato provides an easy way of wrapping around virtually any existing model, as long as Plato's interface is adhered to.

Publication citations:

Alexandros Papangelis, Mahdi Namazifar, Chandra Khatri, Yi-Chia Wang, Piero Molino, and Gokhan Tur, "Plato Dialogue System: A Flexible conversational AI Research Platform", ArXiv Preprint [paper]

Alexandros Papangelis, Yi-Chia Wang, Piero Molino, and Gokhan Tur, “Collaborative Multi-Agent Dialogue Model Training Via Reinforcement Learning”, SIGDIAL 2019 [paper]

Plato wrote several dialogues between characters who argue on a topic by asking questions. Many of these dialogues feature Socrates including Socrates' trial. (Socrates was acquitted in a new trial held in Athens, Greece on May 25th 2012).

News

v0.2: The main update from v0.1 is that Plato RDS is now provided as a package. This makes it easier to create and maintain new conversational AI applications and all tutorials have been updated to reflect this. Plato now also comes with an optional GUI. Enjoy!

ReadMe Contents

• How does the Plato Research Dialogue System work?
• Installation
  • Installing Plato from source code (Recommended)
  • Common Issues During Installation
• Running Plato
• Building a Conversational Agent with Plato
  • Domain
  • Controller
  • Agent
    • Rule-based modules
    • Trained modules
      • Plato internal experience
      • Parsing data with Plato
      • Training Components of Conversational Agents
        • Train NLU
        • Train DST
        • Train Dialogue Policy
          • Supervised Dialogue Policy
          • Reinforce Dialogue Policy
        • Train NLG
        • Train an end-to-end conversational agent based on Seq2Seq
    • Training Multiple Conversational Agents Simultaneously
    • User Simulators
      • Agenda-Based User Simulator
      • Dialogue Act to Language User Simulator
    • Create a new module
      • Create new States, Actions, Learning Algorithms, and more
  • Graphical User Interface
• Upcoming Tutorials
• Upcoming Features
• Acknowledgements
• Conclusion

How does the Plato Research Dialogue System work?

Conceptually, a conversational agent needs to go through various steps in order to process information it receives as input (e.g., “What’s the weather like today?”) and produce an appropriate output (“Windy but not too cold.”). The primary steps, which correspond to the main components of a standard architecture (see Figure 1), are:

• Speech recognition (transcribe speech to text)
• Language understanding (extract meaning from that text)
• State tracking (aggregate information about what has been said and done so far)
• API call (search a database, query an API, etc.)
• Dialogue policy (generate abstract meaning of agent’s response)
• Language generation (convert abstract meaning into text)
• Speech synthesis (convert text into speech)

Plato has been designed to be as modular and flexible as possible; it supports traditional as well as custom conversational AI architectures, and importantly, enables multi-party interactions where multiple agents, potentially with different roles, can interact with each other, train concurrently, and solve distributed problems.

Figures 1 and 2, below, depict example Plato conversational agent architectures when interacting with human users and with simulated users, respectively. Interacting with simulated users is a common practice used in the research community to jump-start learning (i.e., learn some basic behaviours before interacting with humans). Each individual component can be trained online or offline using any machine learning library (for instance, Ludwig, TensorFlow, PyTorch, or your own implementations) as Plato is a universal framework. Ludwig, Uber's open source deep learning toolbox, makes for a good choice, as it does not require writing code and is fully compatible with Plato.

Figure 1: Plato's modular architecture means that any component can be trained online or offline and can be replaced by custom or pre-trained models. (Grayed components in this diagram are not core Plato components.)

Figure 2: Using a simulated user rather than a human user, as in Figure 1, we can pre-train statistical models for Plato's various components. These can then be used to create a prototype conversational agent that can interact with human users to collect more natural data that can be subsequently used to train better statistical models. (Grayed components in this diagram are not Plato core components since they are either available as out of the box such as Google ASR and Amazon Lex, or domain and application specific such as custom databases/APIs.)

In addition to single-agent interactions, Plato supports multi-agent conversations where multiple Plato agents can interact with and learn from each other. Specifically, Plato will spawn the conversational agents, make sure that inputs and outputs (what each agent hears and says) are passed to each agent appropriately, and keep track of the conversation.

This setup can facilitate research in multi-agent learning, where agents need to learn how to generate language in order to perform a task, as well as research in sub-fields of multi-party interactions (dialogue state tracking, turn taking, etc.). The dialogue principles define what each agent can understand (an ontology of entities or meanings; for example: price, location, preferences, cuisine types, etc.) and what it can do (ask for more information, provide some information, call an API, etc.). The agents can communicate over speech, text, or structured information (dialogue acts) and each agent has its own configuration. Figure 3, below, depicts this architecture, outlining the communication between two agents and the various components:

Figure 3: Plato's architecture allows concurrent training of multiple agents, each with potentially different roles and objectives, and can facilitate research in fields such as multi-party interactions and multi-agent learning. (Grayed components in this diagram are not core Plato components.)

Finally, Plato supports custom architectures (e.g. splitting NLU into multiple independent components) and jointly-trained components (e.g. text-to-dialogue state, text-to-text, or any other combination) via the generic agent architecture shown in Figure 4, below. This mode moves away from the standard conversational agent architecture and supports any kind of architecture (e.g., with joint components, text-to-text or speech-to-speech components, or any other set-up) and allows loading existing or pre-trained models into Plato.

Figure 4: Plato's generic agent architecture supports a wide range of customization, including joint components, speech-to-speech components, and text-to-text components, all of which can be executed serially or in parallel.

Users can define their own architecture and/or plug their own components into Plato by simply providing a Python class name and package path to that module, as well as the model’s initialization arguments. All the user needs to do is list the modules in the order they should be executed and Plato takes care of the rest, including wrapping the input/output, chaining the modules, and handling the dialogues. Plato supports serial and parallel execution of modules.

Plato also provides support for Bayesian optimization of conversational AI architectures or individual module parameters through Bayesian Optimisation of Combinatorial Structures (BOCS).

Installation

First make sure you have python version 3.6 or higher installed on your machine. Next you need to clone the Plato repository:

git clone [email protected]:uber-research/plato-research-dialogue-system.git

Next you need to install some pre-requisites:

• TensorFlow:

  pip install tensorflow>=1.14.0
  

• Install SpeechRecognition library for audio support:

  pip install SpeechRecognition
  

• For MacOS:

  brew install portaudio
  brew install gmp 
  pip install pyaudio
  

• For Ubuntu/Debian:

  sudo apt-get install python3-pyaudio
  

• For Windows: Nothing is needed to be pre-installed

Next step is installing Plato. To install Plato you should directly install it from source code.

Installing Plato from source code (Recommended)

Installing Plato from source code allows installation in editable mode which means that if you make changes to the source code, it will directly effect execution.

1. Navigate to the directory of Plato (where you cloned the plato repository in the previous step).

2. We recommend to create a new python environment. To set up the new python environment:

   2.1 install virtualenv

   sudo pip install virtualenv
   

   2.2 create a new python environment:

   python3 -m venv </path/to/new/virtual/environment>
   

   2.3 activate the new python environment:

   source </path/to/new/virtual/environment/bin>/bin/activate
   

3. Install Plato:

   pip install -e . 
   

To support speech it is necessary to install PyAudio, which has a number of dependencies that might not exist on a developer's machine. If the steps above are unsuccessful, this post on a PyAudio installation error includes instructions on how to get these dependencies and install PyAudio.

Common Issues During Installation

CommonIssues.md file contains common issues and their resolution that a user might encounter while installation.

Running Plato

To run Plato after installation, you can simply run the plato command in the terminal. The plato command receives 4 sub-commands:

• run
• gui
• domain
• parse

Each of these sub-commands receives a value for the --config argument that points to a configuration file. We will describe these configuration files in detail later in the document but remember that plato run --config and plato gui --config receive an application configuration file (examples could be found here: example/config/application/), plato domain --config receives a domain configuration (examples could be found here: example/config/domain/), and plato parse --config receives a parser configuration file (examples could be found here: example/config/parser/).

For the value that is passed to --config Plato first checks to see if the value is the address of a file on the machine. If it is, then Plato tries to parse that file. If it is not, Plato checks to see if the value is a name of a file within the example/config/<application, domain, or parser> directory.

For some quick examples, try the following configuration files for the Cambridge Restaurants domain:

plato run --config CamRest_user_simulator.yaml

plato run --config CamRest_text.yaml

plato run --config CamRest_speech.yaml

Building a Conversational Agent with Plato

An application, i.e. conversational system, in Plato contains three major parts:

• Domain
• Controller
• Agent

These parts are declared in an application configuration file. Examples of such configuration files could be found at example/config/application/ In the rest of this section we describe each of these parts in details.

Domain

For implementing a task-oriented dialog system in Plato the user needs to specify two components that constitute the domain of the dialog system:

• Ontology
• Database

Plato provides a command for automating this process of building the ontology and database. Let's say for example that you want to build a conversational agent for a flower shop, and you have the following items in a .csv (this file could be found at example/data/flowershop.csv):

id,type,color,price,occasion
1,rose,red,cheap,any
2,rose,white,cheap,anniversary
3,rose,yellow,cheap,celebration
4,lilly,white,moderate,any
5,orchid,pink,expensive,any
6,dahlia,blue,expensive,any

To automatically generate a .db SQL file and a .json Ontology file you need to create a domain configuration file within which you should specify the path to the csv file, output paths, as well as informable, requestable, and system-requestable slots: (e.g. example/config/domain/create_flowershop_domain.yaml):

GENERAL:
  csv_file_name: example/data/flowershop.csv
  db_table_name: flowershop
  db_file_path: example/domains/flowershop-dbase.db
  ontology_file_path: example/domains/flowershop-rules.json

ONTOLOGY:  # Optional
  informable_slots: [type, price, occasion]

  requestable_slots: [price, color]

  System_requestable_slots: [type, price, occasion]

and run the command:

plato domain --config create_flowershop_domain.yaml

If all went well, you should have a flowershop.json and a flowershop.db in the example/domains/ directory.

If you receive this error:

sqlite3.IntegrityError: UNIQUE constraint failed: flowershop.id

it means that the .db file has already been created.

You can now simply run Plato's dummy components as a sanity check and talk to your flower shop agent:

plato run --config flowershop_text.yaml

Controller

Controllers are objects that orchestrate the conversation between the agents. The Controller will instantiate the agents, initialize them for each dialogue, pass input and output appropriately, and keep track of statistics.

Running the command plato run runs Plato's basic controller (plato/controller/basic_controller.py). This command receives a value for the --config argument which points to a Plato application configuration file.

To run a Plato conversational agent, the user must run the following command with the appropriate configuration file:

plato run --config <FULL PATH TO CONFIG YAML FILE>

Please refer to example/config/application/ for example configuration files which contain settings on the environment and the agent(s) to be created as well as their components. The examples in example/config/application/ could be run directly using just the name of the example YAML file:

plato run --config <NAME OF A FILE FROM example/config/application/>

Alternatively a user could write their own config file and run Plato by passing the full path to their config file to --config:

plato run --config <FULL PATH TO CONFIG YAML FILE>

For the value that is passed to --config plato first checks to see if the value is the address of a file on the machine. If it is, the Plato tries to parse that file. If it is not, Plato checks to see if the value is a name of a file within the example/config/application directory.

Agent

Each conversational AI application in Plato could have one or more agents. Each agent has a role (system, user, ...) and a set of standard dialog system components (Figure 1), namely NLU, dialogue manager, dialogue state tracker, policy, NLG, and user simulator.

An agent could have one explicit module for each one of these components. Alternatively, some of these components could be combined into one or more modules (e.g. joint / end-to-end agents) that can run sequentially or in parallel (Figure 4). Plato's components are defined in plato.agent.component and all inherit from plato.agent.component.conversational_module

Figure 5. Components of Plato agents

Note that any new implementations or custom modules should inherit from plato.agent.component.conversational_module.

Each one of these modules could be either rule-based or trained. In the following subsections we will describe how to build rule-based and trained modules for agents.

Rule-based modules

Plato provides rule-based versions of all components of a Slot-Filling conversational agent (slot_filling_nlu, slot_filling_dst, slot_filling_policy, slot_filling_nlg, and the default version of agenda_based_us). These can be used for quick prototyping, baselines, or sanity checks. Specifically, all of these components follow rules or patterns conditioned on the given ontology and sometimes on the given database and should be treated as the most basic version of what each component should do.

Trained modules

Plato supports training of agents’ components modules in an online (during the interaction) or offline (from data) manner, using any deep learning framework. Virtually any model can be loaded into Plato as long as Plato’s interface Input/Output is respected. For example, if a model is a custom NLU module it simply needs to inherit from Plato's NLU abstract class (plato.agent.component.nlu) and implement the necessary abstract methods.

Plato internal experience

To facilitate online learning, debugging, and evaluation, Plato keeps track of its internal experience in a structure called the Dialogue Episode Recorder, (plato.utilities.dialogue_episode_recorder) which contains information about previous dialogue states, actions taken, current dialogue states, utterances received and utterances produced, rewards received, and a few other structs including a custom field that can be used to track anything else that cannot be contained by the aforementioned categories.

At the end of a dialogue or at specified intervals, each conversational agent will call the train() function of each of its internal components, passing the dialogue experience as training data. Each component then picks the parts it needs for training.

To use learning algorithms that are implemented inside Plato, any external data, such as DSTC2 data, should be parsed into this Plato experience so that they may be loaded and used by the corresponding component under training.

Alternatively, users may parse the data and train their models outside of Plato and simply load the trained model when they want to use it for a Plato agent.

Parsing data with Plato

Training online is as easy as flipping the 'train' flags to 'True' in the configuration for each component users wish to train.

To train from data, users simply need to load the experience they parsed from their dataset. Plato provides example parsers for DSTC2 and MetaLWOZ datasets. As an example of how to use these parsers for offline training in Plato, we will use the DSTC2 dataset, which can be obtained from the 2nd Dialogue State Tracking Challenge website:

http://camdial.org/~mh521/dstc/downloads/dstc2_traindev.tar.gz

Once the download is complete, you need to unzip the file. A config file for parsing this dataset is provided at example/config/parser/Parse_DSTC2.yaml. You can parse the data that you downloaded by first editing the value of data_path in example/config/parser/Parse_DSTC2.yaml to point to the path to where you downloaded and unzipped the DSTC2 data. Next you can run the parse script as follows:

plato parse --config Parse_DSTC2.yaml

Alternatively you could write your own config file and pass the absolute address to that file to the command:

plato parse --config <absolute pass to parse config file>

Running this command will run the parsing script for DSTC2 (which lives under plato/utilities/parser/parse_dstc2.py) and will create the training data for dialog state tracker, NLU, and NLG for both user and system under the data directory at the root directory of this repository. Now this parsed data could be used to train models for different components of Plato.

Training Components of Conversational Agents

There are multiple ways to train each component of a Plato agent: online (as the agent interacts with other agents, simulators, or users) or offline. Moreover, you can use algorithms implemented in Plato or you can use external frameworks such as TensorFlow, PyTorch, Keras, Ludwig, etc.

Ludwig is an open source deep learning framework that allows you to train models without writing any code. You only need to parse your data into .csv files, create a ludwig config (in YAML), that describes the architecture you want, which features to use from the .csv and other parameters and then simply run a command in a terminal.

Ludwig also provides an API, that Plato is compatible with. This allows Plato to integrate with Ludwig models, i.e. load or save the models, train and query them.

Train NLU

In the previous section, the DSTC2 parser of Plato generated some .csv files that can be used to train NLU and NLG. There is one NLU .csv file for the system (data/DSTC2_NLU_sys.csv) and one for the user (data/DSTC2_NLU_usr.csv). These look like this:

For training a NLU model you need to write a configuration file that looks like this:

input_features:
    -
        name: transcript
        type: sequence
        reduce_output: null
        encoder: parallel_cnn

output_features:
    -
        name: intent
        type: set
        reduce_output: null

    -
        name: iob
        type: sequence
        decoder: tagger
        dependencies: [intent]
        reduce_output: null

training:
    epochs: 100
    early_stop: 50
    learning_rate: 0.0025
    dropout: 0.5
    batch_size: 128

An example of this config file exists in example/config/ludwig/ludwig_nlu_train.yaml. The training job could be started by running:

ludwig experiment \
       --model_definition_file example/config/ludwig/ludwig_nlu_train.yaml \
       --data_csv data/DSTC2_NLU_sys.csv \
       --output_directory models/camrest_nlu/sys/

The next step is to load the model in an application config. In example/config/application/CamRest_model_nlu.yaml we provide an application config that has a model based NLU and the other components are non-ML-based. By updating the path to the mode (model_path) to the value that you provided to the --output_directory argument when you ran ludwig, you can specify the NLU model that the agent needs to use for NLU:

...

MODULE_0:
    package: applications.cambridge_restaurants.camrest_nlu
      class: CamRestNLU
      arguments:
        model_path: <PATH_TO_YOUR_MODEL>/model

...

and test that the model works:

plato run --config CamRest_model_nlu.yaml

Train DST

The DSTC2 Data Parser generated two .csv files we can use for DST: DST_sys.csv and DST_usr.csv which look like this:

Essentially, the parser keeps track of the previous dialogue state, the input from NLU, and the resulting dialogue state. We can then feed this into Ludwig to train a Dialogue State Tracker. Here's the ludwig config which can also be found at example/config/ludwig/ludwig_dst_train.yaml:

input_features:
    -
        name: dst_prev_food
        type: category
    -
        name: dst_prev_area
        type: category
    -
        name: dst_prev_pricerange
        type: category
    -
        name: dst_intent
        type: category
    -
        name: dst_slot
        type: category
    -
        name: dst_value
        type: category

output_features:
    -
        name: dst_food
        type: category
    -
        name: dst_area
        type: category
    -
        name: dst_pricerange
        type: category
    -
        name: dst_req_slot
        type: category

training:
  epochs: 100

We now need to train our model with ludwig:

ludwig experiment \
       --model_definition_file example/config/ludwig/ludwig_dst_train.yaml \
       --data_csv data/DST_sys.csv \
       --output_directory models/camrest_dst/sys/

and run a Plato agent with the model-based DST:

plato run --config CamRest_model_dst.yaml

You can of course experiment with other architectures and training parameters.

Train Dialogue Policy

So far we have seen how to train components of Plato agents using external frameworks (i.e. Ludwig). In this section, we will see how to use Plato's internal algorithms to train a dialogue policy offline, using supervised learning, and online, using reinforcement learning.

Supervised Dialogue Policy

Apart from the .csv files, the DSTC2 parser used Plato's dialogue episode recorder to also save the parsed dialogues in Plato experience logs here: logs/DSTC2_system and logs/DSTC2_user. These logs contain information about each dialogue, for example current dialogue state, action taken, next dialogue state, reward observed, input utterance, success, etc. These logs can be directly loaded into a conversational agent and can be used to fill the experience pool.

All you need to do then is write a configuration file that loads these logs (example/config/CamRest_model_supervised_policy_train.yaml):

GENERAL:
  ...

  experience_logs:
    save: False
    load: True
    path: logs/DSTC2_system

  ...

DIALOGUE:
  # Since this configuration file trains a supervised policy from data loaded
  # from the logs, we only really need one dialogue just to trigger the train.
  num_dialogues: 1
  initiative: system
  domain: CamRest

AGENT_0:
  role: system
  max_turns: 15
  train_interval: 1
  train_epochs: 100
  save_interval: 1
  
  ...

  DM:
    package: plato.agent.component.dialogue_manager.dialogue_manager_generic
    class: DialogueManagerGeneric
    arguments:
      DST:
        package: plato.agent.component.dialogue_state_tracker.slot_filling_dst
        class: SlotFillingDST

      policy:
        package: plato.agent.component.dialogue_policy.deep_learning.supervised_policy
        class: SupervisedPolicy
        arguments:
          train: True
          learning_rate: 0.9
          exploration_rate: 0.995
          discount_factor: 0.95
          learning_decay_rate: 0.95
          exploration_decay_rate: 0.995

          policy_path: models/camrest_policy/sys/sys_supervised_data
          
  ...

Note that we only run this agent for one dialogue but train for 100 epochs, using the experience that is loaded from the logs:

plato run --config CamRest_model_supervised_policy_train.yaml

After training is complete, we can test our supervised policy:

plato run --config CamRest_model_supervised_policy_test.yaml

Reinforce Dialogue Policy

In the previous section, we saw how to train a Supervised dialogue policy. We can now see how we can train a Reinforcement Learning policy, using the Reinforce algorithm. To do this, we define the relevant class in the configuration file:

...

AGENT_0:
  role: system
  max_turns: 15
  train_interval: 500
  train_epochs: 3
  train_minibatch: 200
  save_interval: 5000

  ...

  DM:
    package: plato.agent.component.dialogue_manager.dialogue_manager_generic
    class: DialogueManagerGeneric
    arguments:
      DST:
        package: plato.agent.component.dialogue_state_tracker.slot_filling_dst
        class: SlotFillingDST

      policy:
        package: plato.agent.component.dialogue_policy.deep_learning.reinforce_policy
        class: ReinforcePolicy
        arguments:
          train: True
          learning_rate: 0.9
          exploration_rate: 0.995
          discount_factor: 0.95
          learning_decay_rate: 0.95
          exploration_decay_rate: 0.995

          policy_path: models/camrest_policy/sys/sys_reinforce
          
  ...

Note the learning parameters under AGENT_0 and the algorithmic-specific parameters under the policy's arguments. We then call plato with this configuration:

plato run --config CamRest_model_reinforce_policy_train.yaml

and test the trained policy model:

plato run --config CamRest_model_reinforce_policy_test.yaml

Note that other components can also be trained online, either using Ludwig's API or by implementing the learning algorithms in Plato.

Note also that log files can be loaded and used as experience pool for any component and learning algorithm. However, you may need to implement your own learning algorithms for some Plato components.

Train NLG

To train an NLG module you need to write a configuration file that looks like this (e.g. example/config/application/CamRest_model_nlg.yaml):

---

input_features:
    -
        name: nlg_input
        type: sequence
        encoder: rnn
        cell_type: lstm
        
output_features:
    -
        name: nlg_output
        type: sequence
        decoder: generator
        cell_type: lstm

training:
    epochs: 20
    learning_rate: 0.001
    dropout: 0.2

and train your model:

ludwig experiment \
       --model_definition_file example/config/ludwig/ludwig_nlg_train.yaml \
       --data_csv data/DSTC2_NLG_sys.csv \
       --output_directory models/camrest_nlg/sys/

The next step is to load the model in Plato. Go to the CamRest_model_nlg.yaml configuration file and update the path if necessary:

...

  NLG:
    package: applications.cambridge_restaurants.camrest_nlg
    class: CamRestNLG
    arguments:
      model_path: models/camrest_nlg/sys/experiment_run/model
...

and test that the model works:

plato run --config CamRest_model_nlg.yaml

Remember that Ludwig will create a new experiment_run_i directory each time it is called, so please make sure you keep the correct path in Plato's config up to date.

Note that Ludwig also offers a method to train your model online, so in practice you need to write very little code to build, train, and evaluate a new deep learning component in Plato.

Train an end-to-end conversational agent based on Seq2Seq

For this example, we will use the MetalWoz dataset that you can download from here.

Plato supports jointly trained models through Generic Agents. Here we will see the steps needed to create a simple seq2seq conversational agent from scratch. Using MetalWOZ as an example, we need to do the following:

1. Write a MetalWOZ data parser that reads the data and procudes CSV files

As we are only training a simple seq2seq model (text to text), we need our parser to extract user and system utterances. These will be saved in .csv files that will be used by Ludwig in step 4.

For a simple implementation of a MetalWOZ parser, see utilities/parser/Parse_MetalWOZ.py

Please note that this parser will only parse one single file (one domain). You can easily modify it, however, to fit your needs. Here is a sample of the output produced by the parser for the pizza ordering domain:

Note the first user utterance does not actually exist in the data. However, we need something to prompt the model to produce the system's greeting - we could have used an empty sentence, or any other greeting (or a combination of these).

2. Write a config for plato parse

You can then run plato parse as follows:

plato parse --config Parse_MetalWOZ.yaml

3. Train an end-to-end model

To get started we can train a very simple model using Ludwig (feel free to use your favourite deep learning framework here):

input_features:
    -
        name: user
        type: text
        level: word
        encoder: rnn
        cell_type: lstm
        reduce_output: null

output_features:
    -
        name: system
        type: text
        level: word
        decoder: generator
        cell_type: lstm
        attention: bahdanau

training:
  epochs: 100

You can modify this config to reflect the architecture of your choice and train using Ludwig:

ludwig train \
       --data_csv data/metalwoz.csv \
       --model_definition_file example/config/ludwig/metalWOZ_seq2seq_ludwig.yaml \
       --output_directory "models/joint_models/"

4. Write a class inheriting from Conversational Module that loads and queries the model

This class simply needs to handle loading of the model, querying it appropriately and formatting its output appropriately. In our case, we need to wrap the input text into a pandas dataframe, grab the predicted tokens from the output and join them in a string that will be returned. See the class here: plato.agent.component.joint_model.metal_woz_seq2seq.py

5. Write a Plato generic yaml config and run your agent!

See example/config/application/metalwoz_generic.yaml for an example generic configuration file that interacts with the seq2seq agent over text. You can try it out as follows:

plato run --config metalwoz_text.yaml

Remember to update the path to your trained model if necessary! The default path assumes you run the ludwig train command from Plato's root directory.

Training Multiple Conversational Agents Simultaneously

One of Plato's main features allows two agents to interact with each other. Each agent can have a different role (for instance, system and user), different objectives, and receive different reward signals. If the agents are cooperating, some of these can be shared (e.g., what constitutes a successful dialogue).

To run multiple Plato agents on the Cambridge Restaurants domain, we run the following commands to train the agents’ dialogue policies and test them:

• Training phase: 2 policies (1 for each agent) are trained. These policies are trained using the WoLF algorithm:

  plato run --config MultiAgent_train.yaml
  

• Testing phase: uses the policy trained in the training phase to create dialogs between two agents:

  plato run --config MultiAgent_test.yaml
  

While the basic controller currently allows two agent interaction, it is fairly straightforward to extend it to multiple agents (e.g. with a blackboard architecture, where each agent broadcasts its output to other agents). This can support scenarios such as smart homes, where every device is an agent, multi-user interactions with various roles, and more.

User Simulators

Plato provides implementations for two kinds of user simulators. One is the very well known Agenda-Based User Simulator, and the other is a simulator that attempts to mimic user behaviour observed in data. However, we encourage researchers to simply train two conversational agents with Plato (one being a 'system' and one being a 'user') instead of using simulated users, when possible.

Agenda-Based User Simulator

The Agenda-Based User Simulator was proposed by Schatzmann and is explained in detail in this paper. Conceptually, the simulator maintains an "agenda" of things to say, which is usually implemented as a stack. When the simulator receives input, it consults its policy (or its set of rules) to see what content to push into the agenda, as a response to the input. After some housekeeping (e.g. removing duplicates or content that is no longer valid), the simulator will pop one or more items off the agenda that will be used to formulate its response.

The Agenda-Based User Simulator also has an error simulation module, that can simulate speech recognition / language understanding errors. Based on some probabilities, it will distort the output dialogue acts of the simulator - the intent, slot, or value (different probability for each). Here is an example of the full list of parameters that this simulator receives:

patience: 5                         # Stop after <patience> consecutive identical inputs received
pop_distribution: [0.8, 0.15, 0.05] # pop 1 act with 0.8 probability, 2 acts with 0.15, etc.
slot_confuse_prob: 0.05             # probability by which the error model will alter the output dact slot 
op_confuse_prob: 0.01               # probability by which the error model will alter the output dact operator
value_confuse_prob: 0.01            # probability by which the error model will alter the output dact value
nlu: slot_filling                   # type of NLU the simulator will use
nlg: slot_filling                   # type of NLG the simulator will use

Dialogue Act to Language User Simulator

This simulator was designed to be a simple policy-based simulator, that can operate at the dialogue act level or at the utterance level. To demonstrate how it works, the DSTC2 parser created a policy file for this simulator: user_policy_reactive.pkl (reactive because it reacts to system dialogue acts instead of user simulator state). This is actually a simple dictionary of:

System Dial. Act 1 -->  {'dacts': {User Dial. Act 1: probability}
                                  {User Dial. Act 2: probability}
                                  ...
                                  
                         'responses': {User utterance 1: probability}
                                      {User utterance 2: probability}
                                      ...
                                      
System Dial. Act 2 --> ...

The key represents the input dialogue act (e.g. coming from the system conversational agent). The value of each key is a dictionary of two elements, representing probability distributions over dialogue acts or utterance templates that the simulator will sample from.

To see an example, you can run the following configuration:

plato run --config CamRest_dtl_simulator.yaml

Create a new module

There are two ways to create a new module depending on its function. If a module, for example, implements a new way of performing NLU or dialogue policy, then you should write a class that inherits from the corresponding abstract class.

If, however, a module does not fit one of the single agent basic components, for example, it performs Named Entity Recognition or predicts dialogue acts from text, then you must write a class that inherits from the conversational_module directly. You can then load the module via a generic agent by providing the appropriate package path, class name, and arguments in the configuration.

...
MODULE_i:
    package: my_package.my_module
    Class: MyModule
    arguments:
      model_path: models/my_module/parameters/
      ...
...

Be careful! You are responsible for guaranteeing that the I/O of this module can be processed and consumed appropriately by modules before and after, as provided in your generic configuration file.

Plato also supports (logically) parallel execution of modules. To enable that you need to have the following structure in your config:

...
MODULE_i:
    parallel_modules: 5
    
    PARALLEL_MODULE_0:
        package: my_package.my_module
        Class: MyModule
        arguments:
          model_path: models/myModule/parameters/
          ...
          
    PARALLEL_MODULE_1:
        package: my_package.my_module
        Class: MyModule
        arguments:
          model_path: models/my_module/parameters/
          ...
          
    ...
...

Be careful! Outputs from the modules executed in parallel will be packed into a list. The next module (e.g. MODULE_i+1) will need to be able to handle this kind of input. The provided Plato modules are not designed to handle this, you will need to write a custom module to process input from multiple sources.

Create new States, Actions, Learning Algorithms, and more

Plato is designed to be extensible, so feel free to create your own Dialogue States, Actions, Reward Functions, Algorithms, or any other component to suit your specific needs. You only need to inherit from the corresponding class to make sure your implementation is compatible with Plato.

Graphical User Interface

Plato uses PySimpleGUI to handle graphical user interfaces. An example GUI for Plato is implemented at plato.controller.sgui_controller and you can try it out using the following command:

plato gui --config CamRest_GUI_speech.yaml

Upcoming Tutorials

1. Using pre-trained models (BERT, GPT, etc.) in Plato
2. Multi-domain Conversational Agents
3. Multi-lingual Conversational Agents

Upcoming Features

1. Principled logging and printing
2. More dataset parsers
3. More options for multi-agent controllers
4. Bayesian Optimisation in Plato

Acknowledgements

Special thanks to Mahdi Namazifar, Chandra Khatri, Yi-Chia Wang, Piero Molino, Michael Pearce, Zack Kaden, and Gokhan Tur for their contributions and support, and to studio FF3300 for allowing us to use the Messapia font.

Conclusion

Please understand that many features are still being implemented and some use cases may not be supported yet.

Enjoy!