SysSieve: Extracting Actionable Insights from Unstructured Text

Sign in to queue

Description

Understanding free-form text is hard, be it bug reports or trouble tickets written by engineers or feedback/complaints from customers. We have built SysSieve, a learning system to do automated analysis of these important unstructured (yet incredibly noisy) data sources by building upon techniques from statistical NLP, ML, and information theory.  Today, this system is in production use across Windows, Bing, Skype, Office365 and CSS, as well as being leveraged to make platform improvements in our server and network hardware vendors. This video provides an overview of the SysSieve technical details and how it is being applied across product groups.

Embed

Download

The Discussion

  • User profile image
    destinchfie​ld

    I can use sys.sieve today in my enterprise apps.  I haven't solicited feedback from users via my apps yet because I don't know the best way to process their comments besides reading each one myself, one at a time, arghhhh.  sys.sieve could help.

  • User profile image
    Matt

    This sounds pretty amazing! Is it publicly available for download yet?

  • User profile image
    deiruch

    This sounds as if you're mostly hearing the most common opinions. That's only half of the story.

     

    You should probably also listen to individual stories and bug reports. As a solution, you could mark users internally that provide useful bug reports and give more visibility to other reports from those users. This way you wouldn't drown in common, loud & noisy feedback.

  • User profile image
    Navendu Jain

    Thanks for the comments!

    @destinchfie​ld, @Matt: Thanks for your interest. Currently, SysSieve is being used internally within Microsoft product teams. We are working with our cloud team to make it available to external customers; I'll add the details when this offering becomes publicly available.

    @deiruch: Thanks for your question. SysSieve analyzes each input data (e.g., customer feedback, incident/bug report) and produces the corresponding actionable output so it does not miss any data point. One of SysSieve's design principle is to avoid making any domain specific assumption to eliminate any bias. Your suggestion of assigning relatively higher weights to feedback from a subset of users can certainly be incorporated as post processing/reporting analysis.

Add Your 2 Cents