Extracting meeting details from natural language

Prompting
1

Hi, I’m looking to format a natural language meeting request into a format that I can later connect using API for an assistant use case.

My initial approach was to build a format for the model to rephrase the text.
The format might be:
| Date | Time | Location | Subject | Participants | Reminder |

Example:
Set a meeting with David about our new job on Monday, May 22nd, we’ll have the meeting in our garden. Remind me one hour before the meeting.
Format:
| 22/05/2021 | null | Our garden | New job | Me, David | 60 minutes |

The prompt includes the instructions, and if If the user didn’t provide the information, type in “null.”

The basics of it are working using Curie and Curie-instruct-beta, but I’m having trouble mostly with the null object. In many cases, it fails to provide “null” for information that wasn’t provided. Instead, it just does not include this box or mixing it in other boxes where information was included.
I tried adding more and more instructions and examples, but I haven’t got it to reach a solid level, so I wonder if there is another approach to make it work.

Another problem is if the date is a day (Next Sunday/Today/Tomorrow…) and that the reminder time might come in minutes, hours, or days.

I will appreciate any best practices of making this task as single-valued as possible without building a massive fine-tuning file just for it.

2

You’ll probably get better results if you use labeled lists like the following:

[blah blah blah text]
Date:
Location:
Attendees:

The reason I say this is that GPT-3 has probably read plenty of emails from the internet, including meeting invites. Also, in order to get the correct date, make sure to include the current date/time in the payload.

Extract a calendar invitation from the following email in the following format. If any information is unknown, please fill in “unknown” or “TBD”.

Date: (date)
Time: (time)
Location: (location)
Attendees: (attendees)

Email:
Timestamp: 12/1/1998 4:28PM
Subject: Lunch tomorrow?
Body: Hey Joe wanna grab lunch tomorrow? Maybe that sushi place from last time. Lets say noon?

Calendar invite:
Date: 12/2
Time: 12:00
Location: TBD
Attendees: Joe

image

1 Like
3

Thanks! I’ll give it a try.

1 Like
4

This was just a quick example. It will probably work best as few-shot. Maybe even better with fine-tuning since you can have many examples.

Oh and here’s what it looked like in playground:

image
image1157×455 16.3 KB

5

Yes, it seems to perform better, but it will probably require a bunch of examples in the fine-tuning file to fit free speech format rather than email format. Thanks!

1 Like