queread — natural language interactive bot library.

const Queread = require('queread')
let bot = new Queread(dataset)
bot.guess('Get me to Madrid.')
// { label: 'search', parameters: [{ destination: 'Madrid' }] }

Workings

Compiles examples into a directed graph with edges weighted by the number of example queries walking through them, one weight for each label. This ensures that words can be missing from a query, but still be matched to an example. Word order matters to the probability of a correct guess.

Guesses the meaning of a new query by computing the probability of its label, by seeing how much of the graph this query walks through. The basic formula for probability computation is the following:

Count(query, label) = Σi (walked_link(query)[i][label])
Pr(query, label) = Count(query, label) / Σl (Count(query, l))

It is O(n^2) in space in terms of the number of important words in examples, and linear in time in terms of the size of the query.

Dataset

We receive the list of examples in the JSON format returned by bot.parse(fileContent). Each example is on a separate line of the following form:

label: Some text with /words/ and parameters [type] [name].

First, the line starts with a label to which the example corresponds. Then, the example is written in plain text, with some markup to highlight important words (like /so/), and to highlight parameters: the type comes first, the name (or names) afterwards. If only the type is specified, its name is the name of the type.

To specify a parameter that includes spaces, you can use the same markup as important words: here is a /parameter with words/ [type] [name].