/ml-engineering-intern-task

Dive Logo Dive

Octernship info Timelines and Stipend
Assignment Deadline 19th June 2023
Octernship Duration 3rd July 2023 - 3rd October 2023
Monthly Stipend $500 USD

Assignment

Identify action items from meeting transcripts

Task instructions

There are two objectives:

  1. Training data generation / gathering
  2. Develop NLP model for Transcript to Action items pipeline

1. Training Data

Any successful ML implementation depends on data on which models can be trained and tested against

  • Generate or gather publicly available meeting transcript data to be used in the next phase of model development
  • Use the following CSV format to generate data:
start_time,end_time,speaker,text
“HH:MM:SS”,“HH:MM:SS”,“Alice”,“Hello everyone!”
"01:00:00","01:01:20","Bob","Today we are going to discuss about overall product metrics"
"12:01:34","12:01:50","Tay","Awesome, thanks for informing about that!"
  • Document the entire process of generating the training and tradeoffs taken in APPROACH.md

2. Develop NLP Model

From meeting transcripts, identify action items (tasks to be done) that were identified during the course of the meeting.

Input: transcript text as generated in 1st objective

“10:00:00”, “10:01:50”, “Bob“, “... Alice, can you take the UX bug? ...”
“12:25:00”, “12:25:30”, “Alice”, ”... We need to plan for offsite next month ...”

Output: action items

  • format: {"text": ..., "assignee": "Name or UNKNOWN"}
  • for the above example:
{"text": "UX bug", "assignee": "Alice"}
{"text": "plan for offsite next month", "assignee": "UNKNOWN"}

(optional) Stretch goal: Keep the timestamp of when the action item was detected in the transcript

  • format: "ts": "HH:MM:SS"
  • for the above example:
{"text": "UX bug", "assignee": "Alice", "ts": "10:00:00"}
{"text": "plan for offsite next month", "assignee": "UNKNOWN", "ts": "12:25:00"}

Task Expectations

  • Evaluation criteria: 50% of Approach Document, 50% of functional code and demo video
  • Creativity in sourcing training data
  • Ability to develop end-to-end PoC
  • Using a pre-trained SOTA is acceptable, along with appropriate citation
  • Code Quality - remove any unnecessary code, avoid large functions
  • Good commit history - we won’t accept a repo with a single giant commit 🙅‍♀️

Task submission

Please use the GitHub Flow for assignment submission

  1. Creating and working on a new branch.
  2. Raising a Pull Request for submission.
  3. Using GitHub Discussions to ask any relevant questions regarding the project.
  4. Final submission Checklist:
  • APPROACH.md Document
    • note all the approaches considered
    • document any assumptions made and why
    • what did you pick first and why (it’s okay to say - because I am familiar with the process)
    • what did you stop considering and why e.g. (didn’t use LSTM due to low accuracy)
    • citation: links to any code, article or paper referred to
    • what are the limitations of the current approach
  • SUBMISSION.md file in the repo, with steps to
    • setup the codebase (including installing dependencies)
    • run the pipeline, which generates a transcript and prints the action items (one per line)
  • loom.com video 📹 recording of the demo, where you:
    • generate a new transcript file
    • feed it to the model to extract a list of action items