US 12,613,929 B2Granted

Multi-stage training of a tagging model with query supplementation and adversarial enhancement

AssigneeMicrosoft Technology Licensing, LLC

InventorsSimiao Zuo, Pengfei Tang, Xinyu Hu, Qiang Lou, Jian Jiao, Denis Xavier Charles, Eren Manavoglu

Life of this patent

Estimating the term…

Loading patent…

Abstract

A technique is described herein for training a tagging model that is able to successfully interpret queries. The technique trains the tagging model in plural stages. A first stage continues training a pre-trained language model based on a set of queries, to produce a first-stage model. A second stage performs training on the basis of a set of supplemented queries and associated weak labels, to produce a second-stage model. Each supplemented query combines a query with titles of documents that match the query. A third stage performs training on the basis of a set of supplemented queries and associated strong labels, to produce a third-stage model. The third stage also uses adversarial knowledge enhancement that has the effect of making the data presented to the third-stage model more difficult for the third-stage model to interpret. This, in turn, improves the generalization capabilities and robustness of the third-stage model.

Claims20

Dependent claims

A method for training a machine-trained tagging model that applies labels to queries, comprising:

receiving a pre-trained language model;
training the pre-trained language model to interpret missing parts of a first set of queries in a first training set, to produce a first-stage model;
training a second-stage model by performing additional training on the first-stage model based on a second training set, a second-set training example in the second training set including a second-set supplemented query that is formed by combining an original second-set query with additional content, and second-set ground-truth label information,
wherein the second-set supplemented query is formed by:
extracting one or more parts of one or more target items, the one or more target items being determined by a search system to match the original second-set query;
and appending the one or more parts to the original second-set query, and
wherein the training of the second-stage model includes mapping a sequence of tokens in the second-set supplemented query to a sequence of model-generated labels;
and
training a third-stage model by performing additional training on the second-stage model based on a third training set, a third-set training example in the third training set including a third-set query and third-set ground-truth label information,
the training of the third-stage model being based on:
first loss information that represents accuracy of the third-stage model as measured against the third-set ground-truth label information, and second loss information produced by perturbing input data fed to the third-stage model to increase a level of difficulty associated with a task of interpreting the input data,
the third-stage model representing the machine-trained tagging model.

9 more elements

A computer-readable storage medium for storing computer-readable instructions, a processing system executing the computer-readable instructions to perform operations, the operations comprising each of:

receiving an original query;
extracting one or more parts of target items, the one or more target items being determined by a search system to match the original query;
appending the one or more parts to the original query, to produce a supplemented query, and
receiving ground-truth label information;
and
training a machine-trained tagging model based on a training set that includes, as one training example, the supplemented query and the ground-truth label information,
wherein the training of the machine-trained tagging model includes mapping a sequence of tokens in the supplemented query to a sequence of model-generated labels, and
the training of the machine-trained tagging model being based on:
first loss information that represents accuracy of the machine-trained tagging model as measured against the ground-truth label information, and second loss information produced by perturbing input data fed to the machine-trained tagging model to increase a level of difficulty associated with a task of interpreting the input data.

6 more elements

Dependencies are inferred from the claim text.