Disruptive Behavior Mitigation Framework for Games

OSS Framework for Detecting Toxic Speech in Games

Project Overview

Clean Chat is an open source solution to eliminate or mitigate toxic behavior in online multiplayer video games, with the goals of making games more enjoyable and inclusive for all. In large open multiplayer games, toxicity is rampant. Clean Chat provides the ability to proactively detect toxicity in voice and text chat, and provides scores to game moderation services or player support personnel. The project provides a globally scalable framework for training, deploying, and maintaining low-latency anti-toxicity models. . The framework comes pre-packaged with several machine learning 'base models' that allow you to train, or fine tune, the models to detect toxic speech in the vernacular of your game's players.

What Clean Chat Is Not

Clean Chat is not a specific model for detecting toxicity. It is a framework where users can train and deploy their own ML model or they can build upon our pre-trained base models.

Architecture Components:

Clean Chat’s design is divided into three main components:

• The Scoring Engine: This is the online service that hosts the toxicity detection models, and conducts aggregation of the results.
• Game Telemetry Intake: Clean Chat’s means of communicating with Game Servers and handling game telemetry data from them. Note - Clean Chat is specifically designed to work as a third-party service or sidecar to the game server, and does not directly communicate with the client itself
• ML Model Pipeline: A self-contained, ML-Ops driven framework for training, deploying, and maintaining toxicity models. The Model pipeline provides several options for base models, including both open source and Google Cloud Partner Models.

1. API Backend Service

Clean Chat contains an an API Backend service. This service accepts text chat and audio files to be passed into the framework. The backend service is deployed on Google Cloud Run and exposed as an internal service with a dedicated service account.

2. Cloud Functions

Cloud Functions are used as part of this architecture to trigger serverless functions used as part of the event-driven pipeline. These functions are used for parsing data, performing quality checks, calling ML-based APIs (such as the Speech-to-text API), or a variety of other tasks.

3. Scoring Engine

The Scoring Engine is based on Apache Beam, which enables an open source, unified model for defining both batch and streaming data processing pipelines. The objective of the Scoring Engine is to analyze, score, and flag toxic or disruptive behavior. It is based on a serverless design pattern, which allows the service to scale to meet varying data traffic patterns and demand. A unique feature of this deployment, is The Scoring Engine is a hydrid model based on a swappable ML model plus heuristic rules to bias or set business specific thresholds. Google Cloud Dataflow, which is a managed service for Apache Beam, is used to deploy and scale the Scoring Engine.

4. ML Model Pipeline

The pipeline module orchestrates training and orchestration of the toxicity model.

Reference Architecture

Dependencies

The following dependencies should be install on you client prior to deploying the Clean Chat framework. The dependencies are:

1. Terraform (Install Terraform)
2. Docker (Install Docker)
3. Google Cloud SDK (Install gcloud)

Getting Started

Follow these steps to config and deploy the Clean Chat framework:

NOTE: If you're deploying from Google Cloud Shell, you may need to run ./tools/cloud_shell_setup.sh prior to running Terraform. There is a open/known issue with Terraform not honoring OS IPv4 settings.

1. Clone the repo

   git clone https://github.com/googleforgames/clean-chat
   

2. Make a copy of the config.default

   cp config.default config
   

3. Edit the config file

   This file should be updated to reflect your GCP project ID, your naming convensions, and desired parameters. By editing the config file, you are able to customize how sensitive the scoring is, the duration of windowing (which is used to aggregate scores), and a variety of other GCP specific settings.

   vi config
   

4. Initialize Terraform for deployment

   make terraform-init
   

5. Deploy the GCP services, executing the actions proposed in a Terraform plan

   make terraform-apply
   

6. Deploy the Scoring Engine (the scoring engine runs on Google Cloud Dataflow)

   make deploy-scoring-engine
   
   # NOTE: You can also deploy the scoring engine in interactive mode, for testing/debugging purposes using hte following commmand:
   make deploy-scoring-engine-interactive
   

7. Deploy the API Backend Service

   make deploy-backend-api
   

Examples

As an example, and for testing purposes, here are a few sample scripts that can be called once the framework has been deployed:

# Option #1 
# Hello World example to ensure that the Backend API service is responding.
python3 ./examples/api-backend-get-test.py

# Option #2
# Run the Chat/Text Message example
python3 ./examples/api-backend-post-text.py

# Option #3
# Run the Audio Demo
python3 ./examples/api-backend-post-audio.py

Toxicity Model Sidecar

Clean Chat comes pre-packaged with a framework to train, test, and deploy toxicity detection models. Currently, the framework supports models that detect toxicity in voice and text chat. Clean Chat provides two model training interfaces; one basic training interface that only utilizes the training features of Keras, and one that supports an ML Ops framework.

Model Choices

Cohere.AI

Cohere.ai provides a developer API to create high-quality word embeddings and classification models. Users may “fine tune” a base cohere embedding model to receive embeddings that are specific to a Game’s dataset. Cohere then provides a toxicity classification model.

Cohere is a Google Cloud Partner and available through the GCP marketplace.

To learn more about Cohere, see their documentation.

To utilize the Cohere model, run:

python .components/model/cohere/main.py 'API KEY HERE' 'DATA PATH HERE' 'COHERE MODEL TYPE HERE'

BERT

BERT is a transformer language model developed by Google in 2018. BERT is available via Tensorflow Hub, and users can fine-tune the model to their Game’s dataset.

Model Training with TFX (ML Ops)

This module presents a packaged TFX pipeline for training and deploying your own custom toxicity model. Clean Chat currently supports the TF Hub BERT Model as it's base language model in a TFX framework.

The architecture consists of:

• The model pipeline (tfx_pipeline.py). A TFX/Kubeflow pipeline to transform the training data, train the model, and push the resulting model artifact
  • transform.py (TFX transform file)
  • trainer.py (TFX training file)
• The pipeline runner (kubeflow_dag_runner.py)

You will need to setup a kubernetes cluster with kubeflow deployed on it. This is where your built pipeline will reside. You can set up your cluster with:

make create-pipeline-cluster

Your pipeline endpoint (the address of your pipeline is now contained in the enviroment variable KUBEFLOW_ENDPOINT

Before you build your pipeline, set the enviroment variables for the number of training steps, evaluation steps, adn the path to the model's training data.

export TRAINING_DATA_PATH='path here'
export TRAINING_STEPS='training steps'
export EVAL_STEPS='eval steps'

To build your TFX pipeline on the Kubeflow cluster that you just created, run:

make tfx-create-pipeline

Additional runs of the pipeline can be conducted with:

make tfx-run

Model Deployment

To deploy your pipeline to the cloud, we need both a serving container and a serving cluster. You can create the serving container with:

make build-model-serving

Next, create a Kubernetes cluster to deploy your model on:

make create-serving-cluster

Finally, to create a deployment of the model, run:

make deploy-image

If you need to update the model that is currently being served, you may do so with:

make serve-latest-model

Training BERT Model on Vertex AI Pipelines

tfx supports Vertex AI Pipelines option (tfx pipeline create --engine=vertex).
This section helps you to deploy your BERT pipeline to Vertex AI.

Before you start, you need to set up local Python environment. For tfx Docker image compatibility, we use Python 3.7.12 or 3.7.13.

On top of that, Install Docker if needed. make commands (vertex-create-pipeline and vertex-update-pipeline) for tfx build the image with Docker.

After you set up the Python environment, install modules.

pip install -r ./components/model/bert/vertex_pipeline/requirements.txt

You have to enable Vertex AI service, Container Registry and Cloud Storage. Make sure you have the OWNER permission of your GCP Project to avoid IAM problems.

gcloud services enable \
  aiplatform.googleapis.com \
  containerregistry.googleapis.com \
  storage.googleapis.com

After enabling the service, you need to set up values Vertex AI Pipelines.

cd ~/clean-chat # or path you cloned the repository.
source config
source ./components/model/bert/vertex_pipeline/env.sh
make vertex-upload-data
make vertex-create-pipeline

vertex-upload-data uploads sample data to Google Cloud Storage. If you changed it, you need to upload again.

vertex-create-pipeline creates pipeline in Vertex AI and build a docker image with Dockerfile.

This image is used as the default image in the pipeline.

Then you can run the pipeline. vertex-run-pipeline run the pipeline.

make vertex-run-pipeline

The command shows the URL for Vertex AI pipelines to monitor the run.

If you modify files, you need to update the pipeline before your new run.

make vertex-update-pipeline

Then run your pipeline again.

make vertex-run-pipeline

Roadmap

Project is currently in alpha status, and is being actively developed. Expect things to break.

See the open issues for a list of proposed features (and known issues).

Contributing

Participation in this project comes under the Contributor Covenant Code of Conduct

Please read the contributing guide for directions on writing code and submitting Pull Requests.

Clean Chat is in active development - we would love your help in shaping its future!

License

Distributed under the Apache 2.0 License. See LICENSE for more information.

References

• Jigsaw Perspective API
• Dataset - Kaggle Toxic Comment Classification Challenge
• Dataset - Kaggle League of Legends Tribunal Chatlogs