Insurance Simplification and Intent Analysis Challenge

• Our client is an insurance company that currently provides its potential customers with forms, which the customers fill to describe the insurance options they want to include in their policy. Currently, there are a lot of redundant and unstructured options listed in the form, along with a lot of text input options. This makes it difficult for the customers to easily find and select the options. Furthermore, the insurance company also loses out on the opportunity to help the customer in easily finding standard or recommended list of options that most people might find useful. 

• In the near future, the client is looking to build a simplified insurance options (benefits) configurator, where the customer can go ahead and select options according to their needs in a hierarchical manner. Example of the intended hierarchy (example only for general idea, not exactly based on the provided dataset): 

  • Q1 - Coverage for transplants?

    • Q1.1 - Kidney Transplant coverage?

      • Yes - 100% coverage

      • Copay - 20%

      • Copay - 10%

      • No

    • Q1.2 - Hip Transplant coverage?

      • Yes - 100% coverage

      • No

• This hierarchy was created for the client in the last challenge of the series. The hierarchy was created by analyzing the anonymized past history configuration data , which contains the choices made by customers on the unstructured options discussed above. This raw dataset will also be used in this challenge.
  
  In this challenge, we will add new columns to the raw dataset based on some new goals listed below and also to compile the results of the last challenges into the columns of the raw dataset.

Results of the previous challenges

• The top two submissions of the last challenge can be found in the forum. Participation in the last challenge is not required for this challenge, and for reference the specification of the last challenge can be found here.

• In the last challenge, a hierarchical tree of topics/labels was created where each node in the tree described a label. The ordering is of the following format for the winning submission: 

  • T_1     copay

  • T_1_1   visit

  • T_1_1_1 office visit copay

• In the last challenge, the submitters also created a mapping of each row  in the provided raw excel file to the leaf of the hierarchy tree that it is the best associated with. So basically there were 2 output resources from the last submission.

• In addition to the last challenge, a few more challenges were conducted in this series, where the goal was to reduce the variation in the ‘answer’ column values in the raw dataset. The results of the previous challenge submissions are also included in the forum.
  In the last challenge, the participants had the option to use the previous submissions as their starting base, or start from scratch. The winner of the last submission chose to start from scratch, but possibly took some inspiration from the submissions of the previous challenges on how to achieve variability reduction.

 

The Three Goals of this challenge

    

• Goal 1 - Generic Answer Column

• The primary goal of this challenge is to use the result of the previous challenges, particularly last challenge, and find the generic version of the ‘answer’ value of each row in the raw dataset, particularly those row with input type = text. The raw dataset can be found in the forum.

• The main problem that the client is facing that when people fill up their form, they sometimes enter text input manually or in a manner that is not consistent with how other users enter similar or the same information. And this creates the problem that there are many rows where the freely entered text value actual mean the same thing, but are written a little differently or maybe have different numeric values, some additional words, different stop words and punctuation etc. 

• To solve the above problem, they are looking to first clean each answer by removing all the unnecessary words, spaces, numbers, punctuation etc, such that the ‘answer’ value can be represented in its most generic form. 

• The generic version of the answer should be such that if two answers mean the same thing, they should ideally have the same generic answer. It should be possible to group the rows on the basis of unique generic answer assigned to that row by your algorithm. This generic answer value should be added to a new column called ‘generic_answer’.

• The requirement of finding the generic answer is very similar (almost identical) to the requirements of the first 3 challenges of this series, but the participants are free to use a new approach or improve upon the existing results of challenge 3 as they see fit. The challenge specifications of those 3 challenges can be found here: Challenge 1 and Challenge 2 (these were hosted simultaneously) and Challenge 3. Note that the results of only Challenge 3 has been provided, which combines and improves the results of first 2 challenges.

To get an idea on how to get these generic answers for each row, the participants are strongly encouraged to refer to the results of challenge 3 (combination of the first 2 + some additional improvements), which can be found in the forum. 

Furthermore, the winning submission of the previous submission also carried out considerable steps to reduce variability during its execution, so if possible that can also be directly extended to get the generic answer.

• Important - In addition to taking inspiration from the previous submission, the hierarchical categorization/tree created in the last challenge can also be used to find this generic answer. One idea would be to first group the rows using the hierarchy created in the last challenge, and then look for common patterns in the text within the group to find the generic answer. The participants are free to use the results of the last challenge in any way they find suitable.
   

• Goal 2 - Intent Column

• Once the generic answer is found, that generic answer should further be simplified into its most basic concept, called the intent from here onward. For example, the answer value “ lifetime maximum for hospice respite is limited to 15 days inpatient/15 days outpatient” can have a simple intent like ‘hospice life maximum is limited’, or maybe even ‘hospice life maximum limited’. 

• The key difference between the intent and a generic_answer value of a row is that there can be multiple unique generic answers under one intent, but not the vice versa. Also, the generic_answer is basically the most simplified version of the answer text, but the intent value does not have to be a simplified version of the answer value. It can be worded in a different way from how the answer has been worded, but the overall goal of the intent column is that, if a marketing professional reads the value in the intent column, he should know what is the 'intention' of this option selected or written by a customer.

• In some cases, it might be possible that the intent value and the generic value are identical or very similar, if suppose your algorithm found that the generic answer and the intent are worded in a similar or exactly the same manner.
  
   

• Goal 3 - Compile the heirarchy tree into columns

• This is a relatively simple goal. Just parse the already available hierarchical tree, which was created in the last challenge, and using the mapping generated in the last challenge, add the labels of each node of an answer in a separate column, such as - Root category or Level 0, Level 1, Level 2, Level 3, and so on. For example, in the example given above, the values can be:
  Root Category column: Copay
  Level 1 column: Visit

Level 2 column: Office Visit Copay
Level 3 column: N/A
Level 4 column: N/A
… Level N column: N/A
Here N is the maximum depth of the hierarchical tree generated in the last challenge. Please check the tree and assign N accordingly.

 

Data and additional code access

 

The Data is available in the forums. In addition, the winning submissions of previous challenges attempting to solve the problem have also been included in the forums. 

Important - The participants are free to start from scratch or use the provided submissions and extend/modify them to achieve the objectives of this challenge. However, the participants are required to use the output hierarchical categorization of the last challenge. For participants who choose to implement this generic_answer and intent generation algorithm from scratch, then their code should internally invoke the the winning algorithm of the last challenge and generate the hierarchy and the mapping. It is optional to use this hierarchical information in finding the generic_answer and intent values, but it is compulsory to parse the hierarchical tree, in order to populate the hierarchy related columns. 

 

Expected technologies - The participants are free to use any technique they like to achieve the results as long as everything is implemented in Python. It can be a string parsing based implementation, or it might include advanced techniques from the field of NLP, Deep Learning or Machine Learning (ML). In the ML route is chosen, the participants are free to train models themselves, or use ready-made machine learning/deep learning models available online, as long as they are available to be used in commercial software free of charge. 

Final Submission Guidelines

What to Submit

• Command line based code written to achieve the 3 goals mentioned above. Python 3.7 should be used to implement the submission.

• A PDF/Word/Markup format based report detailing the techniques and algorithms used to achieve the goals.

• The updated CSV file, with these new columns: generic_answer, intent and hierarchical columns such as root_level, level_1, level_2, level_3, and so on.

• A README.md file detailing the deployment instructions.

Technology Stack

• Python