# US 2025/0053562 A1 · Machine learning recollection as part of question answering using a corpus

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2025/0053562 A1
- Document ID: US20250053562A1
- Status: Published application
- Kind: A1
- Application number: 18/533,620
- Filed: 2023-12-08
- Published: 2025-02-13
- Earliest claimed: 2023-08-07
- Assignees: Adobe Inc.
- Applicants: Adobe Inc.
- Inventors: Joseph D. Barrow; Darrell Jan Dykstra; Varun Manjunatha; Nedim Lipka
- Web page: https://patentgrep.com/patent/US20250053562A1

## CPC

- G06F 16/24522: {Translation of natural language queries to structured queries}
- G06F 16/248: Presentation of query results
- G06N 3/045: Combinations of networks
- G06N 3/0455: Auto-encoder networks; Encoder-decoder networks
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

Machine learning recollection techniques are described as part of question answering using a corpus. Inputs are received identifying a search query and a corpus of search data that is to be searched based on the search query. The search query is decomposed to form a plurality of decomposed queries and retrieval search results are generated by searching the corpus of search data using one or more additional terms based on the decomposed queries. A search result is synthesized based on the retrieval search results using a text generation machine-learning model. The search result is presented for display in a user interface.

## Claims

### Claim 1

A method comprising:
receiving, by a processing device, inputs identifying a search query and a corpus of search data that is to be searched based on the search query;
decomposing, by the processing device, the search query to form a plurality of decomposed queries;
generating, by the processing device, retrieval search results by searching the corpus of search data using one or more additional terms based on the decomposed queries;
synthesizing, by the processing device, a search result based on the retrieval search results using a text generation machine-learning model to generate readable text supported by cites to portions of the corpus of search data used as sources for the readable text; and
presenting, by the processing device, the readable text of the search result for display in a first panel of a user interface nonmodally with a second panel of the user interface that displays a portion of the corpus of search data used as a source for a portion of the readable text identified from a user selection in the first panel.

### Claim 2 (depends on 1)

The method as described in claim 1, wherein the decomposing the search query to form the plurality of decomposed queries is performed using a question decomposition technique implemented using natural language processing by a machine-learning model.

### Claim 3 (depends on 1)

The method as described in claim 1, further comprising generating hypothesis results specifying the additional terms to be searched based on the decomposed queries using a query expansion technique as applied to the decomposed queries.

### Claim 4 (depends on 1)

The method as described in claim 1, further comprising generating hypothesis results specifying the additional terms to be searched based on the decomposed queries using a blind relevance feedback technique as applied to the decomposed queries.

### Claim 5 (depends on 1)

The method as described in claim 1, further comprising generating hypothesis results specifying the additional terms to be searched based on the decomposed queries using a reformulation technique as applied to the decomposed queries.

### Claim 6 (depends on 1)

The method as described in claim 1, wherein the generating the retrieval search results by searching the corpus of search data is performed using embeddings as part of a vector-based search utilizing an embedding space.

### Claim 7 (depends on 1)

The method as described in claim 1, further comprising generating hypothesis results specifying the additional terms to be searched based on the decomposed queries and wherein the generating the retrieval search results by searching the corpus of search data is performed using text of the hypothesis results as part of a text-based search, the retrieval search results including respective portions taken from the corpus of search data.

### Claim 8 (depends on 1)

The method as described in claim 1, further comprising generating hypothesis results specifying the additional terms to be searched based on the decomposed queries and generating decomposed query search results based on the decomposed queries, the generating of the hypothesis results specifying additional terms to be searched is based at least in part on the decomposed query search results.

### Claim 9 (depends on 8)

The method as described in claim 8, wherein the generating of the decomposed query search results is based on a search performed of the corpus of search data using the decomposed queries.

### Claim 10 (depends on 1)

The method as described in claim 1, further comprising generating document summaries of digital documents included as part of the corpus of search data and wherein the synthesizing of the search result is also based on the document summaries.

### Claim 11 (depends on 1)

The method as described in claim 1, wherein the text generation machine-learning model is configured using a generative pretrained transform architecture that is trained using text and configured to predict a series of tokens representing individual pieces of text.

### Claim 12 (depends on 1)

The method as described in claim 1, wherein the presenting includes displaying representations in the user interface of:
digital documents included in the corpus of search data;
the search query;
the search result;
the portions of the corpus of search data used as the sources of the search result; and
citations to the portions.

### Claim 13

(canceled)

### Claim 14

A system comprising:
a search input module implemented by a processing device to receive inputs identifying a search query and a corpus of search data that is to be searched based on the search query;
a query decomposition module implemented by the processing device to decompose the search query to form a plurality of decomposed queries;
a hypothesis search module implemented by the processing device to generate hypothesis results specifying additional terms to be searched based on the decomposed queries;
a retrieval module implemented by the processing device to generate retrieval search results by searching the corpus of search data using the hypothesis results;
a synthesis module implemented by the processing device to synthesize a search result based on the retrieval search results using a text generation machine-learning model to generate readable text that cites to portions of the corpus of search data used as sources for the readable text, and
the search input module is further implemented by the processing device to present the readable text of the search result for display in a first panel of a user interface nonmodally with a second panel of the user interface that displays a portion of the corpus of search data used as a source for a portion of the readable text identified from a user selection in the first panel.

### Claim 15 (depends on 14)

The system as described in claim 14, wherein the retrieval search results include respective portions taken from the corpus of search data.

### Claim 16 (depends on 14)

The system as described in claim 14, further comprising a retrieval module implemented by the processing device to generate decomposed query search results based on the decomposed queries and wherein the generating the hypothesis results by the hypothesis search module is based at least in part on the decomposed query search results.

### Claim 17 (depends on 16)

The system as described in claim 16, wherein the generating of the decomposed query search results is based on a search performed of the corpus of search data using the decomposed queries.

### Claim 18 (depends on 14)

The system as described in claim 14, further comprising a document summarization module implemented by the processing device to generate document summaries of digital documents included as part of the corpus of search data and wherein synthesis of the search result is also based on the document summaries.

### Claim 19

One-or-more computer-readable storage media storing instructions that, responsive to execution by a processing device, causes the processing device to perform operations including:
receiving inputs defining a search query using text and one or more digital documents for inclusion as part of a corpus of search data;
decomposing the search query to form a plurality of decomposed queries;
generating hypothesis results specifying additional terms to be searched based on the decomposed queries;
generating retrieval search results by searching the corpus of search data using the hypothesis results;
synthesizing a search result based on the retrieval search results using a text generation machine-learning model to generate readable text supported by cites to portions of the corpus of search data used as sources for the readable text; and
presenting the readable text of the search result for display in a first panel of a user interface nonmodally with a second panel of the user interface that displays a portion of the corpus of search data used as a source for a portion of the readable text identified from a user selection in the first panel.

### Claim 20 (depends on 19)

The one-or-more computer-readable storage media as described in claim 19, wherein the text generation machine-learning model is configured using a generative pretrained transform architecture that is trained using text and configured to predict a series of tokens representing individual pieces of text.

### Claim 21 (depends on 19)

The one-or-more computer-readable storage media as described in claim 19, wherein the presenting includes displaying the portion of the corpus of search data used as the source for the portion of the readable text in the second panel in response to detecting the user selection in the first panel.
