Natural language recommendation system: to pre-classify inputs or not?

Question

Does it help to "pre-classify" natural language inputs using labeled input fields? E.g., "Who," "What," "Where," "When," "Why," "How," and "How much?" Or is a single, monolithic, free-form, long-text input field equally effective and efficient for model training purposes?

Scenario 1: Without input labels

We are three research fellows, Alice, Bob and Charlie at the University of Copenhagen. We want to understand the development of the human visual system. This knowledge will help in the prevention and treatment of certain vision problems in children. Further, the rules that guide development in the visual system can be applied to other systems within the brain. Our work, therefore, has wide application to other developmental disorders affecting the nervous system. We will conduct this research in 2019 under a budget of $15,000.

Scenario 2: With input lables

Who: We are three research fellows, Alice, Bob and Charlie.

What: We want to understand the development of the human visual system.

Where: At the University of Copenhagen.

When: During the calendar year of 2019.

Why: This knowledge will help in the prevention and treatment of certain vision problems in children.

How: Further, the rules that guide development in the visual system can be applied to other systems within the brain.

How Much: The research will cost $15,000.

Use Case:

I am building an AI/ML recommendation system. Users subscribe to the system to get recommendations of research projects they might like to participate in or fund. There will be many projects from all over the globe. Far too many for a human to sort through and filter. So AI will sort and filter automatically.

Will pre-classifying input fields using labels help the training algorithm be more efficient or effective?

Answer 1

Likely yes! When you split out the inputs like this, you are adding information. How much this helps is an open question till you build your system, get some data and start training.

Of course, it would be marvelous to have a machine do all the work straight unstructured text - but you want a functional, easy-to-use website, not a research project of your own. To that end, do everything that you can to constrain the scope of the problem, and maximize the information available to your model. For example, you might want to see if you can add researchers using Google Scholar (so you can link to their profiles, and perhaps mine some information that way).

You'll be somewhat 'data constrained' till you've got a decent number of research proposals and user interactions to learn from. Tools like our NLP architect may help you get more out of your text (there are some other really cool new-generation ML-for-NLP packages you should also evaluate).