Another huge milestone completed!
I have all of my intents working with my entire list of recipes (not just one hard coded one). All new code and resources uploaded to GitHub
There's still plenty to do, however. I need to add in all sorts of fuzzy searching for ingredients. For example, when I ask about how much 'milk' I need for Pancakes (which is explicitly listed in the recipe) it responds with "you need 1 cup of milk", but asking about 'whole milk' responds "You don't need whole milk for this recipe". I would like to at least have it respond "I couldn't find whole milk, but you need 1 cup of milk" if not automatically responding about milk in general.
I also have been testing random recipes from my slot list and some of them aren't working. I'll have to look into this. As ironic as it is, I couldn't get any of my chocolate chip cookie recipes to respond, since the name usually has weird variations (Anna's Chocolate Chip cookies, Nestle Chocolate Chip Cookies I, etc). I'll probably implement the same fuzzy searching that will be used for ingredients.
In the next few days I'm also going to test out the skill by having my Alexa assist me in cooking so I know what I need to fix about the user experience.
And, of course, I'll work on saving the current session so the steps and ingredients can be read in order with the users prompting and without repeating the current recipe title.
There's two general approaches to the matching problem. You can use them together. One is what you suggest: use various text information retrieval techniques to try to deal with the mess in the recipes and queries. The other is to curate—manually canonicalize—the recipes.
ReplyDeleteCuration is obviously not a great plan in some ways. To scale it, someone would need to build some kind of probably crowd-sourced infrastructure. It still doesn't solve the weird query problems. But it is something you could do with your current recipe base fairly quickly yourself, and at least put off some of the problems of hard text IR.