Back to results
US 2021/0334467 A1Published application
Domain-adapted sentiment prediction for long or unbalanced text threads
AssigneeMicrosoft Technology Licensing, LLC
InventorsZhiling Wan, Chih-Hui Wang, Haichao Wei, Lili Zhou, Huiji Gao
Life of this patent
Estimating the term…
Abstract
Embodiments of the disclosed technologies use machine learning to produce thread level classification data and case level classification data.
Claims20
A method for detecting an issue relating to a service or software, the method comprising:
- determining, in response to input of a text thread into a first machine-learned model, by the first machine-learned model, embedding data for the text thread;
- the text thread indicative of at least one communication that has occurred relating to the service or software;
- determining, in response to input of the embedding data into a second machine-learned model, by the second machine-learned model, thread level sentiment classification data;
- the thread level sentiment classification data indicative of a first score of a set of ordered scores;
- determining, in response to input of thread level classification data for a case and domain feature data into a third machine-learned model, by the third machine-learned model, case level sentiment classification data;
- the case comprising a sequence of text threads;
- the domain feature data relating to the service or software but not part of the text thread;
- the case level sentiment classification data indicative of a second score of the set of ordered scores.
5 more elements
A method comprising:
- determining, in response to input of thread level classification data for a case and domain feature data into a machine-learned model, by the machine-learned model, case level classification data;
- the case comprising a sequence of text threads relating to a product and created during a time interval;
- the domain feature data relating to the product but not part of the case;
- the machine-learned model having been trained using at least two neural networks;
- when the case level classification data satisfies a criterion, causing outputting of first data indicative of an issue with the product;
- when the case level classification data does not satisfy the criterion, causing outputting of second data indicative of the case level classification data.
3 more elements
A system comprising:
- first machine learning-based means for producing thread level classification data;
- and
- second machine learning-based means for producing case level classification data in response to input of the thread level classification data and domain feature data.
Dependencies are inferred from the claim text.