US 12,450,474 B2
Systems and methods for numerical reasoning by a partially supervised numeric reasoning module network
Amrita Saha, Singapore (SG); Shafiq Rayhan Joty, Singapore (SG); and Chu Hong Hoi, Singapore (SG)
Assigned to Salesforce, Inc., San Francisco, CA (US)
Filed by Salesforce, Inc., San Francisco, CA (US)
Filed on Jan. 29, 2021, as Appl. No. 17/162,289.
Claims priority of provisional application 63/086,851, filed on Oct. 2, 2020.
Prior Publication US 2022/0108169 A1, Apr. 7, 2022
Int. Cl. G06N 3/08 (2023.01); G06N 3/04 (2023.01)
CPC G06N 3/08 (2013.01) [G06N 3/04 (2013.01)] 20 Claims
OG exemplary drawing
 
1. A method of an artificial intelligence (AI) agent deployed on a computing device to conduct a dialogue with a user, the method comprising:
receiving, at a question-answering neural network comprising an entity-specific cross-attention model and a sampling network implemented on one or more processors, an input comprising a pair of a query and a passage, wherein the query includes a set of words in an original word order;
generating a dependency parse tree of the query including at least a root and immediate children of the root;
generating a simplified parse tree including:
a plurality of nodes, wherein each node of the plurality of nodes includes all words from a respective subtree rooted at each of the immediate children of the root of the dependency parse tree in the original word order of the query, and
a plurality of edges between a first node of the plurality of nodes and each other node of the plurality of nodes;
generating a program form of the query with at least a step of the program form for each of the plurality of edges of the simplified parse tree;
identifying a first set of numerical entities along with a first set of respective mention locations from the passage as separate from other entities in the passage;
ranking, via the entity-specific cross-attention model, the first set of numerical entities depending on a respective query relevance associated with each numerical entity from the first set of numerical entities based on an interaction between the passage and a respective step of the program form;
sampling, by the sampling network, a subset of numerical entities as discrete arguments from the ranked first set of numerical entities;
executing one or more discrete operations corresponding to the discrete arguments to provide an output answer to the query; and
generating, via a user interface of the AI agent, an output based on the output answer to the query by executing one or more discrete operations corresponding to the discrete arguments.