New Components for Experience Training (May '21 Release)

Experience Training allows you to help make the algorithm smarter and more customized to your brand. With this release, two new components are now available for experience training: NLP Filters and Spell Checking.

NLP Filters
When you turn NLP filters on for a field in the search configuration, you are enabling Named Entity Recognition (NER), which uses Google’s open source machine learning framework, BERT, to detect potential filters present structured information from a Knowledge Graph. But, occasionally, the algorithm gets it wrong!

You can now accept or reject the NLP filters in the Experience Training tab. If you reject one, the algorithm will no longer apply the NLP filter for that query for your experience.

Spell Checking
By default, the Answers algorithm will return spell checking corrections for a given query to help guide users who may have meant something else. Users can then click on the spell correction to re-run the query with the correction. You can learn more about spell checking here.

Now, you can accept or reject these corrections in the Experience Training tab. If a correction is accepted, it helps to train the algorithm and make it smarter for future queries. If a correction is rejected, it will no longer surface that correction for that query and experience.

To learn more about this feature visit the Answers Debugging training unit.


Hi team,

Such cool new features! I recently showed this to a client and they jumped on the spell checking training right away.

They raised an intriguing point: Is it possible to proactively provide terms to be removed from spell check?


Hey Max,

We currently just have the functionality to train the algorithm based on corrections already provided by the algorithm. Once you reject a suggestion, the algorithm will no longer apply that correction for a given search term.

Would love to get more context for when you’d want to proactively remove terms from spell check. Does your client have terms they predict will get undesired corrections?