Bot Framework: Ways to handle longer email messages with LUIS (exceeding 500 chars)

0
votes

I am creating a bot to provide simple answers and perform some light-weight actions on commonly asked questions via the e-mail channel.

Different from the other channels, the e-mail channel tends to get longer messages fed to the bot (including user's email signature, some intro, thread history, ...). I am trying to use LUIS to guess the intent for the messages, to then execute the right action. This is working fine when the messages are very short, but it is failing when messages get larger as LUIS is limited to 500 chars.

What are best practices to get around this problem when using the e-mail channel with bot framework? Is there common ways to strip out the thread history, email signature, introduction parts of the e-mail body so that only the real question can get sent to Luis? Are there ways to split the message up in different phrases and then correlate the LUIS intents somehow?

I'm looking to see what others are doing here as the e-mail channel is a bit different from all other 'chat'-oriented channels.

Thanks a lot,

Bart

1
botframeworkazure-language-understanding

1 Answers

1
votes

I made the same kind of project a few months ago (1 year ago in fact), but in the end it did not used Bot Framework as we only have email channel and were able to process more rules by directly calling Office365 API. But I think the following comments are also valid for you:

I initially found the 500 character limit really annoying, but then I also understood that LUIS is not a good tool to detect multiple intents so it's important to restrict the content to the minimum.

Is there common ways to strip out the thread history, email signature, introduction parts of the e-mail body so that only the real question can get sent to Luis? Are there ways to split the message up in different phrases and then correlate the LUIS intents somehow?

At that time, I got the same questions and found that there was no miracle tool or solution, in particular to do the following:

  • eliminate salutations
  • eliminate forwarded / replyTo messages

I had to pass over a huge amount of messages to do my own processing, using... regex, and other rules.

Moreover, you cannot cut your message into several pieces that you would send to LUIS, as you have no way to understand which sentence should be considered in priority...

Another thing that I suggest: do no trust the email title.