7 Replies Latest reply on Mar 17, 2014 5:27 AM by aimaf

    Process for sentiment analysis

    Edward Ford

      Has anyone employed a process for sentiment analysis within your internal Jive instance? I'm working on building a process now to review content, analyze it and report out trends based on what we see. It would be helpful to hear how others have done this to get some ideas on how to create a sustainable process.

        • Re: Process for sentiment analysis

          Haven't done it, but wonder if Radian 6 or one of the other companies who do monitoring might have something that could be applied, or maybe be encouraged to provide a plug-in or app to accomplish this. What say you, Mark Weitzel?

          • Re: Process for sentiment analysis

            Sentiment analysis is tricky, mostly due to the 'modifications' of the root english language to make words that normally mean a bad thing into a good thing and vice versa.  With that being said, you can either try to make general assumptions based on your community, the age of the individuals writing content, gender, location, or you can still get an overall 'trend' based on the root use of most words.

            So who has done some sentiment analysis?  Jive did some work with Networked Insights a few years back - all the content within discussion posts would be analyzed by their service and returned with sentiment based on users and content.  Interesting integration but it required quite a bit of data to get a trend, so it was not perfect for smaller communities.


            As for something 'out of the box' - I will have to have a chat with the engineering team, I know traditionally we would not send the actual 'text' of content into the analytics database, only the subject, author, views, likes, etc.  This means that all the content you would want to run reports on is actually nonexistent in the analytics package.  Why?  Because even just capturing all the actions around content creation you get Gigabytes of data, capturing all the text within discussions, blogs, documents, etc - means a Multi-X modifier to the amount of data.  Just counting every thread title means a lot requests.  Counting every thread, every word, every reply - I would guess at least 100x the amount of data.  So... expensive in many ways.

                 Luckily we do have the same team that built 'FiltrBox / Fathom' working on how to better process more and more data in Jive - so stay tuned.  But for the moment I believe Mark's solution of an App will be the best.

            • Re: Process for sentiment analysis
              Edward Ford

              Thank you for all of your thoughts on this. It was helpful. We were quickly pulling together a review of some internal discussions, so pardon my delayed response. The information that you shared was helpful in helping me gain a better understanding of sentiment analysis. As Curtis Gross and Mark Weitzel point out, creating an app would be ideal to cull all of the data necessary to get a full picture of sentiment. We didn't go that route. Instead, we focused on a few pieces of content and conducted a manual process to assess what people were saying. Based on what the team needed to report, this worked for us in this instance.

                • Re: Process for sentiment analysis

                  Hi Edward - would you mind sharing a bit more on this process? It would nice to have a low-tech solution to test and have in place before we go spending big bucks on modifications and plugins.


                  How did you select new content?

                  Did you just analyse all the words or did you search for particular words?

                  Which tools did you use for analysis?