Watson Workspace will no longer be available after February 28, 2019. Learn more.

Ideas

Moments algorithm - Hashtags

Comments

9 comments

  • Avatar
    Jonathan Brunn

    Thanks for the feedback! We have multiple experiments internally running on improvements for moments, and hashtags are among some of the features we have been looking at.

    By the way, this algorithm field is to capture the general pattern of a clustering algorithm we use within moments. There are other aspects of the moment summarization which also evolve in parallel, including the focus message identification and summary phrase / title generation. Within message clustering, not every change we push will result in a change to this algorithm string as well.

    If you've used moments for a while, you may have noticed a significant improvement in the summary phrase generation recently. This did not affect the clustering, so we did not update the algorithm feature to indicate a new major revision.

  • Avatar
    Arnd Layer

    I don't know if it's technically feasible, but wouldn't it be nice, if customers could implement their own algorithms for moments?

  • Avatar
    Michael Würdemann (Edited )

    @arnd - good idea. At least a customer should be able to choose which algorithm they prefer...

  • Avatar
    Ami Dewar

    By "adjust the algorithm" do you mean making adjustments behind the scenes that effect an entire organization's output  or end user facing adjustments where an individual could tell the system what is important to them/what's not important and thus tune what is highlighted?

  • Avatar
    Jonathan Brunn

    Thanks for the feedback. We are also looking at various models which allow further customization and extension to moments. If you have specific use cases that call for this, I'd be happy to discuss, either here or in a DM in Workspace.

  • Avatar
    Daniel Lieber

    Having extensions to algorithms would be helpful. There should also be a relatively easy way to distinguish responses (at some level) between core Watson and extensions so we can narrow down responses as necessary, especially as it relates to supporting environments.

  • Avatar
    Arnd Layer

    Am I correct assuming the in the future multiple algorithms will work in parallel?

    What I was thinking was customer providing their own moments algorithms as micro services. 

  • Avatar
    Jonathan Brunn

    To Dan's comment, I might have assumed people saw this, but it's relatively new. We do have one extension already in beta for moments. Apps can add their own "message focuses": https://developer.watsonwork.ibm.com/docs#add-a-message-focus. The app signals that something is interesting in a conversation.

    You can optionally connect your Conversation workspace to your app, then Watson Work will create the focus annotations on behalf of the app, based on the language model you trained in Watson Conversation. Conversation works well for conditions based purely on NLP of the message. The API allows apps to use any condition they want to extend the key or focus messages included in the moment.

    Focuses created by different apps, including our built in processing, can be easily distinguished in the API. So apps do have one way to extend the moments experience already, though we aren't planning to stop here.

    "Algorithm" is a pretty broad term which could mean a lot of different things with respect to moments. Really interested in hearing more details on your use cases. This will influence which types of extensions to moments we prioritize first, and what aspects of the "algorithm" we allow to be externalized.

    Ami's question above is an important one too as it will influence how we expose these concepts in our API and Workspace experience and what we prioritize.

  • Avatar
    Arnd Layer

    I agree that message-focus annotations are a good way to emphasize messages within a moment although it is indirect. I imagine that not each message-focus annotation should have the same weight in moments.

    My primary expectation for moments is that they - in the future - will break up the timeline and pull together messages that belong to the same conversation. This is based on the assumption and experience that multiple conversations will happen at the same time within one space. Topics may reduce the issue but will not mae it disappear. So based on this expectation, "time-gap" can only be the first step in improving moments.

    This is also the basis for my expectation that multiple algorithms will be active at the same time in order to define moments. As I don't immediately have an ideat how this could work, I think this is not a trivial thing to implement. So I'll be patient as long as it's developing.

    Having more than one algorithm active at the same time makes it easy to imagine thos algorithms could be 'pluggable' 3rd party extensions.

Please sign in to leave a comment.