JP-2026514423-A - Optimizing user interaction and task selection in large-scale language models

Abstract

Data representing the text query is received. Text embeddings for the text query are generated using a machine learning-based embedding generation model. Multiple chunk embeddings, each generated by an embedding generation model for multiple document chunks of multiple documents, are accessed. Multiple documents are organized into multiple document subsets. Data representing the selected(s) document subsets is retrieved. A similarity search for text embeddings is performed only on chunk embeddings associated with the document chunks contained within the selected(s) document subsets. The similarity search identifies chunk embeddings as semantically similar to the text query. The document chunks corresponding to the chunk embeddings identified by the similarity search are provided for display within the user interface.

Inventors

• アダム・ジョシュア・ビグネル
• スティーヴン・ジョンソン
• デイル・アマンダ・マルコウィッツ
• ジョシュア・トーマス・ウッドワード
• ミゲル・デ・アンドレス－クラベラ
• スコット・ブラッドリー・ハフマン

Assignees

• グーグル エルエルシー

Dates

Publication Date

20260511

Application Date

20230331

Claims (20)

1. A computer-implemented method for semantic search of a specified subset of multiple documents, A computing system including one or more computing devices receives data indicating a text query, The computing system generates text embeddings for the text query using a machine learning-based embedding generation model, The computing system accesses multiple chunk embeddings generated by the machine learning-based embedding generation model for multiple document chunks of the multiple documents, wherein the multiple documents are organized into multiple document subsets. The computing system obtains data indicating one or more selected document subsets from the plurality of document subsets, The computing system performs a similarity search for the text embeddings with respect only to chunk embeddings associated with document chunks contained in one or more selected document subsets, wherein the similarity search identifies one or more chunk embeddings as semantically similar to the text query. The computing system provides one or more of the document chunks corresponding to one or more of the chunk embeddings identified by the similarity search for display within the user interface. A method of implementation in a computer, including the methods mentioned above.
2. The method implemented in a computer according to claim 1, wherein the user interface includes a text editing interface associated with a document processing application.
3. The user interface includes a primary text editing field that allows the user of the user interface to generate a set of text, The text query includes at least a portion of the set of text generated by the user through interaction with the primary text editing field. A method implemented in a computer according to claim 2.
4. The aforementioned user interface is A primary text editing field that allows the user of the aforementioned user interface to generate a set of text, A query field separate from the primary text editing field, which allows the user to enter the text query separately from the set of text, A method implemented in a computer according to claim 2, including the above.
5. The method of implementing the one or more selected document subsets in a computer according to any one of claims 1 to 4, as specified by the user.
6. The computer implementation method according to claim 5, wherein the user interface includes a document subset selection tool that enables the user to provide user input for selecting one or more selected document subsets from the plurality of document subsets.
7. The method implemented in a computer according to claim 6, wherein the document subset selection tool provides a graphical display of the plurality of document subsets.
8. The method implemented on a computer according to claim 6 or 7, wherein the document subset selection tool enables the user to apply a set of filter logic to the plurality of document subsets, and by applying the filter logic, one or more selected document subsets are selected from the plurality of document subsets.
9. A computer implementation method according to any one of claims 1 to 8, wherein at least some of the documents included in the one or more selected subsets of documents include documents provided by the user.
10. At least some of the documents included in the one or more selected document subsets are Books, Product manual, legal opinion, Academic papers, A method implemented in a computer according to any one of claims 1 to 9, including a proprietary data file or patent document.
11. The computing system retrieves the multiple documents, The computing system performs syntactic analysis of the multiple documents into the multiple document chunks, The computing system generates the multiple chunk embeddings using the machine learning-prepared embedding generation model, A computer-implemented method according to any one of claims 1 to 10, further comprising:
12. A computer system for semantic search of a specified subset of multiple documents, One or more processors, One or more non-temporary computer-readable media, The one or more non-temporary computer-readable media include, Multiple chunk embeddings generated by a machine learning-based embedding generation model for multiple document chunks of the aforementioned multiple documents, wherein the aforementioned multiple documents are organized into multiple document subsets, and When executed by the one or more processors, the instructions cause the computer system to perform an action, The data is stored collectively, and the operation described above is: Receiving data that indicates a text query, Using the aforementioned machine learning-based embedding generation model, generate text embeddings for the text query, To obtain data indicating one or more selected document subsets from the aforementioned plurality of document subsets, Performing a similarity search for text embeddings with respect only to chunk embeddings associated with document chunks contained in the one or more selected document subsets, wherein the similarity search identifies one or more chunk embeddings as semantically similar to the text query. To provide one or more of the multiple document chunks corresponding to one or more of the chunk embeddings identified by the similarity search for display within the user interface, including, Computer system.
13. The computer system according to claim 12, wherein the user interface includes a text editing interface associated with a document processing application.
14. The user interface includes a primary text editing field that allows the user of the user interface to generate a set of text, The text query includes at least a portion of the set of text generated by the user through interaction with the primary text editing field. The computer system according to claim 13.
15. The aforementioned user interface is A primary text editing field that allows the user of the aforementioned user interface to generate a set of text, A query field separate from the primary text editing field, which allows the user to enter the text query separately from the set of text, The computer system according to claim 13, including the computer system according to claim 13.
16. The computer system according to any one of claims 12 to 15, wherein the one or more selected document subsets are specified by the user.
17. The computer system according to claim 16, wherein the user interface includes a document subset selection tool that enables the user to provide user input for selecting one or more selected document subsets from the plurality of document subsets.
18. The computer system according to claim 17, wherein the document subset selection tool provides a graphical display of the plurality of document subsets.
19. The computer system according to claim 17 or 18, wherein the document subset selection tool enables the user to apply a set of filter logic to the plurality of document subsets, and by applying the filter logic, one or more selected document subsets are selected from the plurality of document subsets.
20. The aforementioned operation is, Obtaining the aforementioned multiple documents, The process of parsing the aforementioned multiple documents into the aforementioned multiple document chunks, Using the aforementioned machine learning-prepared embedding generation model, the plurality of chunk embeddings are generated, A computer system according to any one of claims 12 to 19, further comprising:

Description

This disclosure generally relates to optimizing task execution using large-scale language models. More specifically, this disclosure relates to optimizing the interaction between the user and the large-scale language model while selecting tasks for the large-scale language model. Large-scale language models are models trained on massive datasets. This training method gives large-scale language models the ability to perform multiple types of language tasks. For example, some language models can simplify text, generate opposing viewpoints, facilitate brainstorming, and respond to user queries conversationally. By using these tasks in combination, conversational interaction between the model and the user can be facilitated, allowing the user to receive relevant information more efficiently. However, the wide variety of tasks that large-scale language models can perform makes it difficult to choose a specific task(s) at any given time. A block diagram of an exemplary computing system that performs user interaction and task selection optimization for a large language model according to an exemplary embodiment of the present disclosure is shown.The following is a block diagram of an exemplary computing device that performs a semantic search of a specific subset of multiple documents according to an exemplary embodiment of the present disclosure.The following is a block diagram of an exemplary computing device that performs the simplification of selecting specific language tasks to enhance user interaction with a large language model, according to an exemplary embodiment of the present disclosure.A block diagram of an exemplary machine learning-trained large-scale language model according to an exemplary embodiment of the present disclosure is shown.A block diagram of an exemplary machine learning-prepared language model ensemble according to an exemplary embodiment of the present disclosure is shown.This disclosure presents exemplary user interfaces for facilitating interaction between users and large language models, based on several embodiments of this disclosure.The following illustrates user interaction with the exemplary user interface in Figure 4 for assigning documents to document subsets, according to several embodiments of this disclosure.The following illustrates user interaction with the exemplary user interface in Figure 4 for assigning documents to document subsets, according to several other embodiments of the present disclosure.The present disclosure illustrates user interaction with an exemplary user interface in Figure 4 for selecting a document subset from multiple document subsets, according to several embodiments of this disclosure.The following illustrates user interaction with the exemplary user interface shown in Figure 4 for providing queries via a query field, according to several embodiments of this disclosure.This disclosure illustrates user interaction with a large language model using exemplary user interfaces for requesting the model to perform a summarization task, as described in several embodiments of this disclosure.Some embodiments of this disclosure demonstrate additional user interaction with large language models using exemplary user interfaces for requesting the model to perform a counter-perspective task.Some embodiments of this disclosure demonstrate additional user interaction with large language models using exemplary user interfaces for requesting the model to perform brainstorming tasks.Some embodiments of this disclosure demonstrate additional user interaction with large language models using exemplary user interfaces for requesting the model to perform simplification tasks.This disclosure shows various interface layouts that can implement the interface shown in the previous figure according to some embodiments of this disclosure.A flowchart illustrating an exemplary method for performing a semantic search of a specific subset of multiple documents, according to an exemplary embodiment of the present disclosure, is shown.A flowchart illustrating an exemplary method for performing large-scale language model interactions with improved explainability, according to exemplary embodiments of the present disclosure, is shown.A flowchart illustrating an exemplary method for performing the simplification of selecting a specific language task to enhance user interaction with a large language model, according to exemplary embodiments of this disclosure, is shown.A flowchart illustrating an exemplary method for performing dynamic selection of tasks for large-scale language models, according to exemplary embodiments of this disclosure, is shown. The repeated reference numbers across multiple drawings are intended to identify the same features in various embodiments. Overview In general, this disclosure concerns the optimization of task execution using large-scale language models. More specifically, this disclosure concerns optimizing the interaction between a user and a large-scal