Neural network-based artificial intelligence system for content-based recommendations using multi-perspective learned descriptors

Patent metadata

Field	Value
Publication number	US-10909442-B1
Application number	US-201715474992-A
Country	US
Kind code	B1
Filing date	Mar 30, 2017
Priority date	Mar 30, 2017
Publication date	Feb 2, 2021
Grant date	Feb 2, 2021

Abstract

At a network-accessible artificial intelligence service for generating content-based recommendations based on multi-perspective learned descriptors, text sections associated with a plurality of description perspectives, including a single-character perspective and a multi-character perspective, are extracted from various text sources. Using the text sections as input, a machine learning model which includes respective portions corresponding to the different perspectives is trained to reconstruct the input using intermediary descriptors learned from the input. An indication that a second text source is recommended with respect to a first text source is generated using a set of the learned descriptors and transmitted.

First claim

What is claimed is:

1. A system, comprising:
one or more computing devices of a neural-network based artificial intelligence service for generating recommendations;
wherein the one or more computing devices are configured to:
identify a plurality of content description perspectives to be used to characterize a plurality of books, wherein a first book of the plurality of books includes mentions of a plurality of characters, wherein the plurality of content description perspectives includes a single-character perspective and a multi-character perspective;
obtain, from the first book, a plurality of sequences of text extracts including a first sequence and a second sequence, wherein individual ones of the text extracts of the first sequence comprise mentions of one character, and wherein individual ones of the text extracts of the second sequence comprise mentions of multiple characters;
train, using at least the plurality of sequences as input, a neural network model comprising a single-character portion and a multi-character portion, wherein the neural network model is trained to:
generate a plurality of descriptors, including (a) a single-character perspective descriptor learned using at least the single-character portion and (b) a multi-character perspective descriptor learned using at least the multi-character portion, wherein the single-character perspective descriptor corresponds to characteristics of a character mentioned in the first sequence, and wherein the multi-character perspective descriptor corresponds to a relationship between multiple characters mentioned in the second sequence; and
produce, using the plurality of descriptors, a respective reconstructed text section corresponding to one or more text extracts of the input, wherein an objective function of the neural network model comprises a respective element to (a) minimize a same-perspective error and (b) maximize a cross-perspective error, wherein the same-perspective error is based at least in part on a comparison of a reconstructed text section produced by the single-character portion with an input text extract comprising a mention of one character, and wherein the cross-perspective error is based at least in part on a comparison of another reconstructed text section produced by the single-character portion with an input text extract comprising a mention of more than one character; and

provide, via a programmatic interface, a book recommendation generated using at least one descriptor of the plurality of descriptors, wherein the book recommendation comprises a natural language representation of a relationship between contents of the first book and a recommended book, wherein the natural language representation is based at least in part on the at least one descriptors.

2. The system as recited in claim 1 , wherein the plurality of content description perspectives comprises an additional perspective comprising one or more of: (a) an event perspective or (b) a location perspective, wherein the neural network model comprises a portion corresponding to the additional perspective, wherein the neural network model is trained to:
generate, using the portion of the neural network model corresponding to the additional perspective, an additional descriptor corresponding to the additional perspective.

3. The system as recited in claim 1 , wherein the at least one descriptor on which the natural language representation is based comprise a particular descriptor corresponding to a particular point in a word vector set corresponding to a dictionary, wherein the one or more computing devices are configured to:
identify, with respect to the particular descriptor, one or more neighboring words within the dictionary, based at least in on a distance metric associated individual pairs of words of the word vector set; and
utilize the one or more neighboring words to generate the natural language representation.

4. The system as recited in claim 1 , wherein the one or more computing devices are configured to:
compare (a) a first sequence of descriptors, corresponding to a particular content description perspective, wherein the first sequence of descriptors is generated from the first book with (b) a second sequence of descriptors, corresponding to the particular content description perspective, wherein the second sequence of descriptors is generated from the recommended book; and
utilize a result of the comparison to generate the book recommendation.

5. The system as recited in claim 1 , wherein the one or more computing devices are configured to:
evaluate, prior to generating the book recommendation, the trained neural network model using one or more of: (a) a clustering algorithm, or (b) a K-nearest neighbors algorithm.

6. A method, comprising:
performing, by one or more computing devices:
obtaining, from a first book of a plurality of books, a plurality of text sections, wherein individual ones of the sections are associated with a particular perspective of a plurality of content description perspectives, wherein the plurality of content description perspectives include a single-character perspective and a multi-character perspective, wherein an individual text section associated with a first single-character perspective comprises mentions of one character, and wherein an individual text section associated with a first multi-character perspective comprise mentions of multiple characters;
training, using at least the plurality of text sections as input, a neural network model comprising at least (a) a single-character perspective portion including a first set of weights and (b) a multi-character perspective portion including a second set of weights, wherein the neural network model produces, using one or more learned descriptors, a respective reconstructed text section as output corresponding to an input text section, wherein a first learned descriptor of the one or more learned descriptors represents a point corresponding to one or more input text sections within a word vector space; and
providing a book recommendation generated using the trained neural network model, wherein the book recommendation comprises a content-based similarity indicator with respect to the first book and a recommended book, wherein the content-based similarity indicator is based at least in part on a learned descriptor of the one or more learned descriptors.

7. The method as recited in claim 6 , wherein an objective function used for training the neural network model includes a first error term, wherein the first error term based at least in part on a difference between a particular input text section and a corresponding reconstructed text section produced by a particular portion of the model, wherein the particular portion corresponds to a different perspective than the particular input text section.

8. The method as recited in claim 6 , wherein the one or more learned descriptors include (a) a single-character perspective descriptor generated by the single-character perspective portion and (b) a multi-character perspective descriptor generated by the multi-character perspective portion, wherein the single-character perspective descriptor represents characteristics of a character mentioned in a first single-character perspective text section, and wherein the multi-character perspective descriptor represents characteristics of a relationship between multiple characters mentioned in a first multi-character perspective text section.

9. The method as recited in claim 6 , wherein the plurality of content description perspectives includes an additional perspective comprising one or more of: (a) an event perspective, (b) a location perspective, wherein the neural network model comprises a portion correspondi

Assignees

• Amazon Tech Inc

Inventors

• Szarvas Gyuri
• Klementiev Alex
• Frermann Lea

Classifications

• G06N3/044

  Recurrent networks, e.g. Hopfield networks · CPC title

• G06N3/08 Primary

  Learning methods · CPC title

• G06Q30/0631 Primary

  Recommending goods or services · CPC title

• G06N3/0442

  characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU] · CPC title

• G06N3/0455

  Auto-encoder networks; Encoder-decoder networks · CPC title