Experience Training | Yext Hitchhikers Platform
What You’ll Learn
By the end of this unit, you will be able to:
- Define what Experience Training is
- Explain how you can use experience training for three features: Featured Snippets, NLP Filters, and Spell Checking
- Demonstrate using experience training tables to approve, reject, and change predictions
Experience Training is a great way to improve the quality of search results. It allows you to train the Search algorithm by providing it feedback on its predictions. It is a critical tool for debugging backend issues and optimizing results for all search experiences.
Rejecting or modifying a prediction overrides that query immediately. All feedback (approving, rejecting, or changing) helps Search improve the relevancy of results over time. The more feedback that you provide or corrections that you apply here, the better the algorithm can perform in the future.
Experience Training allows you to train the Search algorithm in three areas:
- Featured Snippets
- NLP Filters
- Spell Checking
To train your experience in the UI, navigate to the desired experience under Search and find the Training section. Here you’ll find a screen for each of the above three areas. You can also find it via configuration as code (CaC) (learn more in the config-as-code resource doc).
Core Experience Training Features
Each Experience Training screen shows a table of predictions made by an AI/machine learning model within Search, organized by search term. There’s a row for every prediction for each search term, so a search term may have multiple predictions. There are a few reasons why this happens:
- The results set can change (through changes in your Content or in the Search configuration)
- The algorithm gets retrained as it gets more feedback and makes a new prediction
There are a few features that will come in handy when reviewing any of the three experience training tables:
- Search bar: Use the search bar above each table to search for specific search terms you want to see training predictions for.
- New vs. Completed: Use the “Show completed” toggle above each table to see the training predictions that have already been approved, rejected, or modified. This makes it easier for users to find completed training examples, especially when debugging.
- Training Actions: In the right column of each table is a set of action buttons to approve (checkmark icon), reject (x icon), or change (wand icon) training predictions suggested by the algorithm.
On the Featured Snippets training screen, you’ll see a table of search terms for which the algorithm returned one or more potential featured snippets. Featured snippets extract a direct answer to a user query from unstructured data. Check out the unit for more detail on featured snippets.
There are a few nuances to call out about the Featured Snippets training table:
- Entity Information: See the entity (hyperlinked) and field that the featured snippet came from, which is crucial context to determine whether or not a featured snippet is correct.
- Last Searched: See the most recent search time for each search term. The table is sorted by this column by default.
With featured snippets training, there are three possible actions you can take: approve, reject, or change:
- Reject: Suppress it from the results set and never return that snippet back to the end user
- Approve: Train the algorithm. Additionally, this result can now display on search verticals with the featured snippets prediction mode set to
- Change: The modified snippet will show in future search results
APPROVE_ONLY, changing the underlying content of an approved featured snippet will mean that snippet is no longer approved. You’ll have to re-approve the featured snippet if you want it to appear with the updated content.
Change Featured Snippets Modal
When you click on the change wand button, you will see a pop-up modal that will allow you to approve, reject, or change featured snippets. If the algorithm identifies the incorrect portion of text from your entity, you can select a different portion of text to be used as the featured snippet for the query going forward. You can also change the entity the featured snippet pulls from through the change featured snippets mode (you’ll learn more about this in the next unit).
When selecting the new featured snippet to return, the UI will act a bit differently depending on whether the featured snippet uses rich text or not. The right side of the modal shows the algorithm’s answer versus selected answer.
When changing plain text featured snippets, click and drag your mouse across the text you want returned. The modal will automatically expand the highlighted section to only include full words.
When changing rich text featured snippets, enable the Rich Text toggle, and select the text you want to appear in the results by individually clicking the paragraphs or sections in sequential order.
On the NLP Filters training screen, you’ll see a table of search terms that had NLP filters applied. The Filters column shows the vertical the filter was applied to, the field the filter was applied to, and the field value to filter on. Check out the unit for more detail on NLP filters.
For example, in the screenshot below, the first search term “accepts insurance” had an NLP filter applied on the Person vertical for the field “Accepts Insurance” with the value “Blue Cross”. In this case, you may want to reject the NLP filter – since the query doesn’t mention a specific insurance provider, Search shouldn’t apply an NLP filter for one!
With NLP Filters training, there are two possible actions you can take: approve or change. Approving NLP Filters helps to train the algorithm to better understand your brand and preferences.
By choosing to change filters, you can add, remove, or change the filters applied for a specific query. Since each filter has a vertical, field, and value, those are the pieces you can change.
For example, when a user searches “jobs in New York” the algorithm may apply the filters
Entity Type = Jobs and
Location = York to the Jobs vertical. However you want the location filter to be “New York”, not “York”. You can override this result by selecting the proper vertical (i.e., Jobs) in the modal and then editing the
Location filter to be near
On the Spell Checking training screen, you can review any suggested spell checks and approve or reject the corrections. These are solely based on the search terms entered. If you reject a correction, the algorithm will no longer surface that correction when that search term is run for that experience.