Tokenization: Inconsistent results for longer-tailed queries

Hi All,

I’m QAing an experience for one of my clients and am running into some issues with longer-tailed queries returning inconsistent results.

Example 1: when I query for ‘How do I get started using Mobile Wallet?’ Answers returns an FAQ titled ‘How do I get started with my Mobile Wallet’ and the Links solution. However, when I query for ‘How do I get started using Mobile Wallet’ Answers returns the ‘Mobile Wallet’ service entity in addition to the FAQs related to Mobile Wallet as expected.

Example 2: when querying for ‘Do you have a student checking account?’ Answers returns the ‘All About U Student Checking Account’ in addition to 4 other accounts. However, if I query for ‘Do you offer a student checking account?’ Answers only returns the ‘All About U Student Checking Account’ as desired.

It’s interesting how in both examples one word appears to be throwing off the score of the match. I’d imagine this has to do with how we tokenize individual words within each query, but was hoping someone could provide more insight/guidance on how to best approach this.

As an immediate work around, I know I can add keywords, but I’m trying to avoid this as much as possible as it’s not a scalable approach going forward.

Thank you!
Austin

Austin - Hey glad to help, answers below:

Example 1

Are you still seeing this issue?

Looking at this now I’m seeing identical resolves for both

How do I get started using Mobile Wallet

and How do I get started using Mobile Wallet?

Example 2

This is an interesting case as like you mentioned these two Search Terms are very similar except for one token have / offer

The reason for these differing results is that have is a stop word while offer is not meaning Do you have a student checking account? is actually a better match for those additional results you saw returned:

  • Free Business Checking
  • Business Analysis Checking
  • Business Premium Checking
  • Business Interest Checking

than Do you offer a student checking account? based on the way Answers interprets stop words.

As you mentioned, one way to resolve this is to create an NLP Filter on a field, like Keywords, that has a value of student checking which will limit your results to the ones you’d like.

Hope that helps and let me know if you have any additional questions.

Mike

Hi Mike,

Thanks for getting back to me here. I just realized that the hyperlinks for Example 1 were identical, my apologies.

How do I get started using Mobile Wallet

How do I get started with Mobile Wallet

I’m assuming this behavior is a result of the Stop Words you mentioned in your comment above, but figured I’d include the correct examples here for reference.

Thanks again for the help!
Austin