Representation learning for tax rule bootstrapping

A system comprising:

• a data repository storing a rule, a plurality of dummy tokens representing a plurality of entities in the rule, a dependency graph comprising a rooted tree having a plurality of nodes related by a plurality of edges tagged according to a plurality of dependency tags, a canonical graph having a plurality of node labels, a document formed from the plurality of node labels, a machine readable vector which embeds the document as a sequence of numbers representative of a structure of the rule, and a plurality of additional machine readable vectors representative of a plurality of structures of a plurality of additional rules;
• a pre-processing engine configured to pre-process the rule by replacing terms in the rule with the plurality of dummy tokens;
• a document generator configured to sort the plurality of node labels into a lexicographic order to form the document;
• and
• a machine learning model execution engine configured to execute:
• a first machine learning model which receives as input the plurality of dummy tokens and outputs the dependency graph,
• a second machine learning model which receives as input the dependency graph and outputs the canonical graph, and
• a third machine learning model which receives as input the document and outputs the machine readable vector;
• a comparator configured to:
• compare the machine readable vector to the plurality of additional machine readable vectors;
• and
• identify a set of machine readable vectors, from the plurality of additional machine readable vectors, that match the machine readable vector within a range;
• and
• a display device configured to display a set of rules that correspond to the set of machine readable vectors.

A non-transitory computer readable storage medium comprising computer readable program code, the computer readable program code for causing a computer system to:

• receive a rule comprising text;
• pre-process the rule by replacing terms in the rule with a plurality of dummy tokens denoting a plurality of entities;
• generate, using a first machine learning model which takes the plurality of dummy tokens as input, a dependency graph comprising a rooted tree having a plurality of nodes related by a plurality of edges that are tagged according to a plurality of dependency tags;
• generate, using a second machine learning model which takes the dependency graph as input, a canonical version of the dependency graph, wherein the canonical version comprises a canonical graph having a plurality of node labels;
• sort the plurality of node labels into a lexicographic order to form a document;
• and
• generate, using a third machine learning model which takes the document as input, a machine readable vector that embeds the document as a sequence of numbers representative of a structure of the rule;
• compare the machine-readable vector to a plurality of additional machine readable vectors, wherein the plurality of additional machine readable vectors corresponds to a plurality of additional rules for which a plurality of computer useable program code blocks has been generated;
• identify a set of machine readable vectors, from the plurality of machine readable vectors, that match the machine readable vector within a range, wherein the set of machine readable vectors correspond to a set of rules from the plurality of additional rules;
• and
• display the set of rules to a user.