CSYE7245 - Text Summarization Pipeline

Team Members
Uthsav Shetty
Dileep Ravindranath Holla
Swarna Ananthaswamy

⚠️ The instance hosting this Web Application has been shutdown to avoid recurring AWS Charges. Follow instructions below to setup and deploy your own Streamlit Application

TS-Pipeline | WebApp

Scalable Data Pipeline for scraping/collecting articles, generating text summaries & sentiment analysis, benchmarking ROUGE scores, and deploying them on the cloud to run completely on a Serverless Infrastructure on-demand. Summarization is done using bert models based on the HuggingFace Pytorch transformers library to run extractive summarizations.

Architecture

Setup

The pipeline requires an Amazon Web Services account to deploy and run. Signup for an AWS Account here. The pipeline uses the folllowing AWS Services:

Lambda
Batch
S3
STEP Functions
DynamoDB
Comprehend
CloudWatch
CloudTrail
EC2
Simple Notification Servive (SNS)

Create a new role on the AWS IAM Console and upload the policy template found at aws_config/policy.json on this repository to allow access to all required AWS Services

Clone

Clone this repo to your local machine using https://github.com/holladileep/TS-Pipeline.git

Setup `config.ini`

All scripts make use of the configparser Python library to easily pass configuration data to running scripts/deployed packages. This allows for easy replication of code with zero modifications to Python scripts. Find configuration file can be found in config/config.ini directory on this repository. Modify the file with your environment variables and place it on your S3 bucket under the config directory like so YourS3BucketName/config/config.ini ; All packages and scripts are designed to read the configuration values from this path.

[aws]
input_dir: demo/
link_op: stage1/
stage2: stage2/
bucket: <enter your bucketname here>

[sentry]
init_params: <Enter Sentry Endpoint Here>

[slack]
webhook_url: <Enter Slack WebHook here>

[flask]
app1: <Enter endpoint for Summary-Gen-1>
app2: <Enter endpoint for Summary-Gen-1>

Deploying Lambda Functions

The pipeline extensively uses AWS Lambda Functions for Serverless Computing. All directories on this repo marked with the prefix lambda- are Lambda functions that have to be deployed on AWS. All functions follow a common deployment process.

Deploy serverless Python code in AWS Lambda

Python Lambda is toolkit to easily package and deploy serverless code to AWS Lambda. Packaging is requried since AWS Lambda functions only ship with basic Python libraries and do not contain external libraries. Any external libraries to be used will be have to be packaged into a .zip and deployed to AWS Lambda. More information about Python Lambda can be found here

Setup your `config.yaml`

All lambda- directories contain a config.yaml file with the configuration information required to deploy the Lambda package to AWS. Configure the file with your access keys, secret access keys and function name before packaging and deploying the Python code. An example is as follows

region: us-east-1

function_name: Lambda_Function_1
handler: service.handler
description: Deployed lambda Function
runtime: python3.7
role: <Enter the role name created earlier>

# if access key and secret are left blank, boto will use the credentials
# defined in the [default] section of ~/.aws/credentials.
aws_access_key_id: <Enter your Access Keys>
aws_secret_access_key: <Enter your Secret Access Keys>

timeout: 15
memory_size: 512

environment_variables:
    ip_bucket: <enter_your_S3_Bucket>

# Build options
build:
  source_directories: lib

Create a Virtual Environment

pipenv shell --python 3.7
pip3 install python-lambda

All package dependencies are available in the respective lambda- directories on this repository

Install all Python dependencies

pip3 install -r requirements.txt

Initiate Lambda Deployement

lambda init

This will create the following files: event.json, __init__.py, service.py, and config.yaml Replace the created service.py and config.yaml files with the service.py and config.yaml files in the respective lambda- directory on this repository.

Package and Deploy Lmabda function

lambda deploy

This should create a new Lambda function on your AWS Lambda Console. Follow the same steps for all lambda directories on this repository to deploy packages to AWS Lambda.

Deploying Model Summarization Flask Applications

The pipeline uses two summarizers, which are packaged as standalone Flask applications running on EC2. They can be found in the docker-summary-gen-1-distil and docker-summary-gen-2-bertlarge directories on this repo. Provision an EC2 instance on your AWS account ; Recommended EC2 instance type is t2.xlarge. The Flask applications accept requests on ports 5000 and 5001 respectively.

DockerHub

All images are pushed to DockerHub. They can be found here:
summary-gen-1
summary-gen-2

Pull Images

docker pull holladileep/summary-gen-1
docker pull holladileep/summary-gen-2

Build Docker Images manually from this repository

Install and run Docker on the EC2 instance

sudo amazon-linux-extras install docker
sudo service docker start
sudo usermod -a -G docker ec2-user

This should install Docker on the instance and start the service. Verify if the service is running by executing docker info

Build Docker Images

Copy contents of docker-summary-gen-1-distil and docker-summary-gen-2-bertlarge to the instance.

Service 1

cd docker-summary-gen-1-distil
docker build -t summary-gen-1 -f Dockerfile.service ./

Service 2

cd docker-summary-gen-2-bertlarge
docker build -t summary-gen-1 -f Dockerfile.service ./

Verify that the images are built by running docker images

Run the images and start the service

docker run --rm -it -p 5001:5001 summary-gen-2:latest -model bert-large-uncased
docker run --rm -it -p 5000:5000 summary-gen-1:latest -model distilbert-base-uncased

The Flask Application should now be running on localhost:5000 and localhost:5001 respectively.

Summarization Lambda Consumers use the created Flask URLs in the pipeline. To allow the Consumers to make an API request to the endpoints, place the API endpoints for the two running services under the keys app1 and app2 inside the config.ini file.

Deploying Step Workflow

AWS Step is used the pipeline and for job orchestration. Go to AWS Console and create a new state machine. AWS Step requires the workflow to be written in Amazon States Language.

The ASL workflow.json file is available on aws_step_workflow directory on this repository. Paste the contents of the file on State Machine Definition after creating a new State Machine on the AWS Step console.

Triggering the workflow

Click on the newly created Step workflow and click on New Execution

Auto-Trigger the workflow on placing an input file on S3 Bucket

CloudTrail and CloudWatch are used to auto-trigger the pipeline on any PUT operations on the S3 Bucket. In order to set this up - create a new CloudTrail and add the source S3 Bucket. Once the trail is created, go to the CloudWatch console to create a new Rule. Provide the rule name and select the Event Pattern option. Select Simple Storage Service as the service name and Object Level Operations as the Event Type. Select PutOperations for the type of operations and specify the Bucket name. On the target screen, choose AWS Step and choose the newly created Step Workflow as the target. Deploy the created rule.

AWS Provides a handy guide to walk through the entire process. It can be found here

Deploying Streamlit App

The pipeline uses Streamlit for allowing the user to upload a file with URLs, enter a single URL or quicky summarize text and sentiments for a given text input. The app directly interacts with the built components on AWS and provides a GUI to run the pipeline without the need for the end-user to manually login to AWS Account and trigger the pipeline.

The Python code for this app can be found at streamlit_webapp/app.py. This app is deployed on the EC2 Instance.

Install required libraries

pip3 install streamlit
pip3 install boto3
pip3 install pandas
pip3 install configparser

Run app.py

Run the WebApp by running streamlit run app.py.

Deploying the AWS Batch Docker Container

The Python script for batch scraping URLs can be found in scrape-batch/ directory. Copy contents of the directory on your machine.

Create new ECR Repository

Create a new repository on ECR with the following name ts-pipeline1

Login to the repository

aws ecr get-login-password --region us-east-1 | docker login --username AWS --password-stdin <your_account_id>/ts-pipeline1

Build the Image

docker build -t ts-pipeline1 .

Tag the image

docker tag <your_ECR_repository_URL>/ts-pipeline1:latest

Push the image to ECR docker push <your_ECR_repository_URL>/ts-pipeline1:latest

Create a new Compute Environment first followed by a Job Queue on the AWS Batch Console with the default configurations. Create a Job Definition on the console and provide the ECR container ID created earlier as the input to the container attribute. The batch process is ready for execution

DynamoDB

The pipleine requires three tables to be created on DynamoDB:

articles Store the scraped data
sentiments Store sentiment scores
summary Store generated summaries and scores

Create all tables from the DynamoDB console with url as the Primary Key. There is NO need to specify additional fields.

Slack

The pipeline delivers real-time notifications via Slack. All Python processes are designed to push notifications to Slack via the Web-Hook placed in the config.ini file. Step by step instructions to create an App and generate an incoming Web-Hook can be found here.

Once the Web-Hook is generated, place the same inside the config.ini file for the key webhook_url.

TestCases

All Test Cases have been documented here

Streamlit App can be accessed using this link: TS-Pipeline | WebApp

⚠️ The instance hosting the App has been shutdown

The pipeline can be tested with the sample demo.txt file present in the tests directory. Additonally any URL can be entered in the Streamlit app and results can be seen.

Additionally, POST request can be made to the following URLs to receive a summarized response.

⚠️ POST requests made to the below API Endpoints do not work since the Flask Application is no longer running on EC2 Instance to avoid recurring AWS Charges. Replace the endpoints with your deployed Flask Application(s) to test the summarization models

Model 1

POST http://18.234.153.64:5001/summarize?ratio=0.2
Content-type: text/plain
Body: <enter your text here>

Model 2

POST http://3.87.77.113:5000/summarize?ratio=0.2
Content-type: text/plain
Body: <enter your text here>

holladileep/TextSummarizationPipeline-AWS