US 12,430,376 B2
User-focused, ontological, automatic text summarization using a bi-directional neural network for selecting answers based on their uncommon words to user queries
Po-Hsu Chen, Dublin, OH (US); Jordan L. Vasko, Columbus, OH (US); Mitch R. Gauthier, Columbus, OH (US); and Amy Leibrand, Columbus, OH (US)
Assigned to Battelle Memorial Institute, Columbus, OH (US)
Filed by Battelle Memorial Institute, Columbus, OH (US)
Filed on Oct. 29, 2021, as Appl. No. 17/514,179.
Claims priority of provisional application 63/108,214, filed on Oct. 30, 2020.
Prior Publication US 2022/0138241 A1, May 5, 2022
Int. Cl. G06F 16/34 (2025.01); G06F 16/28 (2019.01); G06F 16/3329 (2025.01); G06F 16/3332 (2025.01); G06F 16/36 (2019.01); G06F 16/907 (2019.01); G06F 40/295 (2020.01)
CPC G06F 16/345 (2019.01) [G06F 16/3329 (2019.01); G06F 16/3338 (2019.01); G06F 16/367 (2019.01); G06F 40/295 (2020.01)] 20 Claims
OG exemplary drawing
 
1. An ontology-based text summarization system, comprising:
input interface circuitry;
output interface circuitry;
non-transitory storage circuitry to store one or more machine-readable instruction sets;
neural network circuitry configured to perform Name Entity Recognition and Semantic Relation;
a long short-term memory circuitry; and
processor circuitry communicatively coupled to the input interface circuitry, the output interface circuitry, and to the non-transitory storage circuitry, the processor circuitry to:
search the Internet;
identify, in response to the search of the Internet, a plurality of information sources on the Internet that are related to a topic;
identify at least one of the plurality of information sources on the Internet as being related to the topic using an ontology with specific concepts and relations between the concepts;
receive a plurality of queries provided by a user via the input interface circuitry;
process sentences in a bi-directional manner to ensure a complete view of content of the sources using the long short-term memory circuitry, wherein the bi-directional manner is attributed to a bi-directional long short-term memory included in the neural network circuitry and is tasked to use a conditional random field model to perform name entity recognition and semantic relation detection;
tokenize each sentence identified as relevant;
generate a score that includes a sum of term-frequency values for uncommon words included in each sentence identified as relevant;
extract a plurality of the concepts and a plurality of relationships from the at least one of the plurality of information sources related to the topic based on the plurality of queries using the neural network circuitry; and
determine a relevancy of the source by comparing the plurality of the concepts and the plurality of relationships included in the plurality of information sources related to the topic, wherein a sentence with a highest score is selected as an answer to each user query.