1 of 21

Kapotaksha Das, Amber Wood, Lance Sloan, Kenny Moore, and Zhen Qian

Insights into Peer Reviews:

A Text Analysis Pipeline Built on Google Cloud

2 of 21

Introductions

3 of 21

Presentation Outline

Overview

Connecting the Data

Deep Learning & RoBERTa

Review to Classification Pipeline

Visualization and Usage

Wrapping up

4 of 21

What is MWrite?

Helping students critique better

Student drafts an essay

Reviewed by Peers

Peers leave comments based on criteria

Feedback is incorporated to improve work

5 of 21

Peer Review Comments

A treasure trove of valuable insights

Over 500 comments are made in a single assignment

Identifying trends in feedback against criteria

Help understand the intent within feedback

Helping students give better feedback

But this cannot be done manually

6 of 21

Classifying Comments

A system to break comments down effectively

The author has identified the issue correctly but has not presented their argument in a clear manner.

Taking a written peer review:

Problem Identified

Verification

Praise

Writing or Formatting Issues

Incorrect Content

Missing Content

Tier 1: Type of Comment

Tier 2: Type of Problem

To actionable classification:

7 of 21

The Deep Learning Architecture

Classification Results

Vertex AI Notebook Manager

PyTorch

Hugging Face

Peer Review Comments

Ro

a

Model for Tier 1

Ro a

Model for Tier 2

RoBERTa is built atop

BigQuery Database

8 of 21

Using RoBERTa for Classification

Model for Tier 1 Tri-label, single-class classifier

Identifies the type of comment

Model for Tier 2 Tri-label, multi-class classifier

Identifies the type of problem

Verification

Praise

Problem Identified

Peer Review Comments

Writing or Formatting Issues

Incorrect Content

Writing or

Formatting Issues

Missing Content

Only comments identified as ʻProblem Identifiedʼ

9 of 21

Unlocking More Potential

Using Deep Learning to gain insight

Pushed to Google Cloud Platform

Used by Tableau

MWrite Database

Prediction on Google Cloud Platform

Visualization in Tableau

10 of 21

From DB To GCP

Canvas MWrite Database

GCP BigQuery Database

Openshift container executes the transfer job

Jenkins schedules and triggers the job daily

11 of 21

Controlling Data

Ensuring Privacy and Security

Only Comment ID is sent to GCP

User provides authentication and credentials

Authenticated ID used to retrieve and show pertinent information to User

Only data user should see is allowed by Tableau

Vertex AI Notebook Manager

BigQuery Database

12 of 21

Preventing Data Leaks

Instructors & Fellows

Students

2 distinct views to separate data

Authenticate User on Login

Only serves data where student is allowed

Only serves data where instructor is allowed

13 of 21

Answering through Visualizing

Making classification accessible

Instructors & Fellows

Students

Macro View

Micro View

How is the class doing?

How am I doing?

Are the criteria encouraging better feedback?

Am I able to understand the criteria?

14 of 21

Instructor View

15 of 21

Instructor View

16 of 21

Student View

17 of 21

Key Challenges

Limiting access, maintaining privacy

Keeping databases connected and up to date

Building views specific to courses and needs

Working with specific limitations of Vertex AI for model deployment

18 of 21

Solutions Designed

Develop a daily schedule to retrieve course data requiring classification dynamically

Fine-tune PeerBERT to perform two tiers of classifications on peer reviews

Use Google Cloud to handle large data and computational demands

Work with instructors and fellows to build dashboards as per their needs

19 of 21

Thank you for joining!

We will now take any questions.

20 of 21

Instructor View

21 of 21

Student View