When a user visits and logs into the front end web app, they are authorized via Cognito. Upon logging in, the user will see a display of results of the already-analyzed syllabi. For the analysis results to be displayed, syllabus files would be uploaded into an S3 bucket. To extract, analyze, and store the syllabi content, separate Glue Jobs occur. A Glue Job is also used to generate analysis on new guidelines added from the front end web app. Between the data storage and front end, the main components used to display data on the web app are the AppSync API and Lambda Resolver functions.

Refer to the Component Architecture for a detailed walkthrough of the solution’s architecture.

Within the generateNLPAnalysis, two primary processes occur: a Symmetric Semantic Search, and a Yes/No Question Answering process. In the Symmetric Semantic Search, three of the most semantically similar sentences are extracted with respect to each flexibility guideline. The Semantic Search was used to enable more accurate and relevant search results as it aims to understand the meaning and context of the guideline and documents being searched.

In the Yes/No Question Answering component, the model is trained to determine whether the three extracted semantically similar sentences answers the flexibility guideline posed in a question format. It must be noted that for the Question Answering model to operate, the flexibility guidelines must be answerable only by “Yes” or “No”.

To ensure accuracy in the analysis, a confidence score analysis is run as part of the Yes/No Question Answering process. It offers the Sponsor a simple way of knowing how accurate the model is in its yes or no answer for the flexibility guideline.

1. AWS Cognito authenticates users to the dashboard web app.
2. WAF first filters the request from the web app to protect against malicious web attacks before it reaches the Appsync GraphQL API.
3. Files are uploaded to an S3 bucket which triggers an S3 Event Notifications, which will send them to the SQS queue. This bucket also allows the front end web-app to display files.
4. The extractMetadata Glue Job programmatically receives the stored events to extract syllabi metadata. It attempts to extract the course code from the file name (e.g., CSPS_syllabus.pdf). If a valid course code is not found, it searches for the course code in the content of the syllabus file. If a valid course code is found, it matches it with the corresponding campus and faculty.
5. Files that were successfully processed will then have their S3 path transferred to the generateNLPAnalysis Glue Job. The job will use the paths to fetch the syllabi, extract the raw text and analyze it with Symmetric Semantic Search and Yes/No Question Answering. 
6. The results of the analysis and syllabi metadata is sent to the storeData Glue job, which inserts them into the database. 
7. Appsync API passes the GraphQL operations from the web app to the Lambda Resolver function to be carried out against the database.
8. The API returns the responses to the front end through the Lambda Resolver and Appsync.
9. The Lambda Resolver can also initiate a new execution of the generateNLPanalysis Glue Job to analyze new guidelines added from the front-end web app. It directly passes the newly added guideline to the generateNLPanalysis Glue Job, hence generating results for all files currently stored in the database. Previously analyzed guidelines are not reprocessed.