Process for delivering responses to queries expressed in natural language based on a dynamic document corpus

The integration of traditional search engine capabilities with LLMs in a data processing and retrieval system addresses the limitations of existing technologies by delivering accurate, up-to-date natural language responses with direct source citations, enhancing user interaction and reducing re-training costs.

WO2024228712A1PCT designated stage expired Publication Date: 2024-11-07DEVOS LUKE THOMAS +7

Patent Information

Application Number
PCT/US2023/021600
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2023-05-03
Filing Date
2023-05-09
Publication Date
2024-11-07

AI Technical Summary

Technical Problem

Current search engines and Large Language Models (LLMs) struggle to deliver direct, accurate, and up-to-date natural language responses to user queries, as they rely on fixed training data and require resource-intensive re-training, lacking direct connections to primary sources and proper attribution.

Method used

A system that combines traditional search engine capabilities with LLMs, featuring a data collection pipeline and retrieval mechanism to ingest, process, and organize data, and generate natural language responses with direct citations to primary sources, enabling real-time dynamic processing without re-training costs.

Benefits of technology

The system provides accurate, up-to-date natural language responses with clear attribution to primary sources, facilitating real-time data processing and reducing the need for costly re-training of LLMs, while maintaining a user-friendly natural language interface.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure US2023021600_07112024_PF_FP_ABST
    Figure US2023021600_07112024_PF_FP_ABST
Patent Text Reader

Abstract

The present invention relates to a system and method for answering user queries by providing natural language responses containing direct citations to primary sources. The system comprises a data collection pipeline that ingests, processes, and organizes data from multiple sources, and a retrieval mechanism that processes user queries, identifies relevant data, and employs a machine learning model, such as a Large Language Model (LLM), to generate natural language responses based on the retrieved data. The generated responses are augmented with direct references to the primary sources, ensuring accurate attribution and up-to-date information. This system combines the natural language capabilities of LLMs with the direct connections to primary sources provided by traditional search engines, delivering real-time, dynamic processing of resources without incurring high re-training costs.
Need to check novelty before this filing date? Find Prior Art

Description

PROCESS FOR DELIVERING RESPONSES TO QUERIES EXPRESSED INNATURAL LANGUAGE BASED ON A DYNAMIC DOCUMENT CORPUSFIELD OF THE INVENTION

[0001] The present invention relates to the field of information processing, specifically concerning the procurement of information rapidly from a large set of data. The invention describes a process which when implemented in software delivers responses to queries made in natural language. The system is unique in that it collects primary source information from a repository of data like a search engine but delivers the response in the format of a summary document written in the same natural language as the query, with appropriate citations of those primary sources.BRIEF SUMMARY OF THE INVENTION

[0002] Search engines form the backbone of the modern internet, allowing users to discover information of interest rapidly without resorting to manual search of the immense volume of data available. However, even the best search engines deliver their responses primarily in the form of an index of resources, instead of directly attempting to answer the query submitted by a user of the search engine. Search engines additionally may struggle with queries expressed in natural language, instead requiring the user to produce a restricted query that complies with rules specific to the search engine.

[0003] The recent creation of Large Language Models (LLMs) by technology companies and research institutions has enabled a new mode of interacting with user queries, where a software system can behave in a manner that imitates a human conversational partner. The user communicates with the system in terms of natural language explanations of what they desire, enabling contextual expression far greater than what search engines support. This type of interface has proven extremely compelling in many instances, but the systems underlying these advances present multiple deficiencies.

[0004] Specifically, LLMs as currently implemented are only able to produce responses based on a fixed corpus of training data, often data that is scraped from the web and then cleaned manually by human researchers. This data is sufficient toanswer many user queries, but the process for updating the model is extremely resource intensive.

[0005] Tools have been developed that allow addition of dynamic resources, e.g. the result of web queries, to the content of a user query for an LLM before the model produces a result. While this can help address problems with the recency of data, it does not enable the LLM to identify which resources were used to compose an answer or to ensure that the answer produced faithfully represents information derived from those resources. This problem is exacerbated when information in the response was derived from a source in the training dataset and not something added to the query by such a tool.

[0006] Therefore, the need exists for a system that has the same natural language interface as an LLM but the capacity to deliver direct connections to primary sources as facilitated by a traditional search engine. Such a system should be able to dynamically process resources in real time without the exorbitant costs associated with re-training an LLM and should be able to provide natural language responses that accurately cite all resources used in composing the response.

[0007] The present invention comprises a novel approach to the architecture of a question answering and resource gathering system that builds upon the strengths of both traditional search engines and Large Language Models. The invention consists of a pipeline that ingests, processes, and organizes data, and a retrieval mechanism that processes user queries, identifies relevant data, and then employs a LLM to compose a response containing direct references to that data and a controlled amount of auxiliary information. These components refer to the architecture and algorithmic design of the system in question, not to specific software realizations thereof.BRIEF DESCRIPTION OF THE DRAWINGS

[0008] The drawings presented describe the high-level features of the invention as well as a specific realization of it to enable retrieval and presentation of relevant news for financial professionals.

[0009] Figure 1 depicts a block-level overview of the architecture defined by the invention, labeled for a generic application.

[0010] Figure 2 is a block diagram showing some of the components typically incorporated in at least some of the computer systems and other devices on which thedisclosed machine learning platform operates according to some implementations of the disclosed technology.

[0011] Figure 3 is a system diagram illustrating an example of a computing environment in which the disclosed machine learning platform operates according to some implementations of the disclosed technology.

[0012] Figure 4 is a flow diagram depicting a process performed by the disclosed platform according to some implementations of the disclosed technology.

[0013] Figure 5 is a flow diagram depicting a process performed by the disclosed platform according to some implementations of the disclosed technology.

[0014] Figure 6 is a flow diagram depicting a process performed by the disclosed platform according to some implementations of the disclosed technology.

[0015] Figure 7 is depicts an example user interface presented by some implementations of the disclosed technology.DETAILED DESCRIPTION OF THE INVENTION

[0016] The invention consists of two primary sub-processes: the data collection pipeline that ingests, processes, and organizes data from various sources, and (2) the retrieval mechanism that processes user queries, identifies relevant data, and employs machine learning models, such as those drawn from the class of Large Language Models (LLMs) to generate a natural language response that contains citations of the sources utilized in the generation of that response. For example, the citations can be direct citations to the primary sources referred to in the generation of that response.

[0017] The data collection pipeline is responsible for collecting, processing, and organizing data from multiple sources, including but not limited to websites, databases, and APIs. The pipeline comprises several subcomponents, including the data ingestion component, the data processing component, and the data organization component, as illustrated in Figure 1.

[0018] The data ingestion component fetches data from various sources, including but not limited to web data, database entries, and API calls. This component primarily concerns the collection of data and basic extraction of relevant information, as well as normalization of data to comply with internal standards. Data ingestion can be performed at an arbitrary frequency based on the requirements of the applicationof the system. Data collected can be aggregated with no limit except that imposed by the limitations of the hardware employed, and as such the information that can be retrieved by the system can be kept up-to-date indefinitely.

[0019] Data ingestion consists of software components built to interface directly with the source documents, including parsers for specific file types, clients for APIs, database connections. Once the raw documents have been loaded into system memory in a programmatically accessible format, the components of interest like text, numerical figures, and images are each extracted into standard containers, e.g. bitmaps for images, utf-16 character buffers for text. This extraction system is customized for the format of the data in question, collecting metadata like geographic location, publication time, and original source alongside the content that will be referenced by the query answering system.

[0020] After data ingestion, the data is transformed to aid in future organization and reference of relevant content. The core goals of this step are to break large data buffers into smaller components that can be atomically referenced later, as well as to generate features from the data that will aid in organization and identification. In the case of textual data, for example, irrelevant text may be removed, named entities are identified and associated with unique IDs, sentiment scores, relevance scores, and topics are generated from the content, and all this information is additionally mapped through vector embeddings to generate compact latent representations of the data. Each task corresponds to a pretrained machine learning model, which for efficiency is based off a common language encoding model when possible. The application of multiple models to extract useful features as opposed to a single end-to-end Large Language Model comes with many benefits, including the capacity to independently evaluate model performance and update sub-components asynchronously, as well as the ability to distribute data transformation and processing across networked compute resources.

[0021] The data organization component involves taking the results of data processing and organizing them into a searchable data structure. The most commonly applied data structures in this task are graph databases, for storing discrete contextual information, and vector databases, for storing compact representation of atomic data objects. These structures will facilitate efficient retrieval of relevant information during the query processing phase. Data is also stored in an unmodified format that allowsretrieval of the raw source information and the extracted features, such that search through the semantic data stores can be used to inform retrieval of accurate primary source information that would have been difficult to identify rapidly with a naive approach. Figure 4 illustrates an example process performed by the data collection pipeline.

[0022] The second main component is the retrieval mechanism, which is responsible for processing user queries, identifying relevant data from the organized data structure, and generating a natural language response using an LLM. The retrieval mechanism comprises several subcomponents, including the query processing component, the data retrieval component, and the response generation component, as illustrated in Figure 1.

[0023] User queries are processed to identify keywords, phrases, and concepts relevant to the query. This involves application of multiple pre-trained models for natural language understanding, which may extract the names of known entities, determine the linguistic sub-components of the query, and associate the content of the query with knowledge graphs or other dynamic structures for storing linguistic information that are built by the data organization system in real time. Specific techniques for implementing these models may be context-dependent, though use of a common set of models or architectures yields performance improvements. These models overlap with those employed to process input documents in the data transformation step, though certain tasks like summarization and relevance scoring are less relevant for user queries than they are for long-form input documents. User queries are also transformed to make them suitable for the data retrieval and response generation process, potentially including augmentation with other information relevant to the user or to the query's context. This processing will include techniques common in natural language understanding, such as tokenization of the text, replacement or annotation of named entities with unique keys, and embedding of the resulting transformed queries to enable numerical alignment between the query and stored data. Figure 5 illustrates an example process performed to process the received user query.

[0024] The processed query and any relevant metadata, e.g. the identities of named entities extracted from the query or semantic labels applied to the query text, are then employed by the data retrieval system. This system identifies relevant datafrom the organized data structure and collects the components of that data that may be useful for generating a response to the query. This identification process achieved through a combination of techniques such as keyword search, semantic search, or vector embedding comparison, which will individually depend on the underlying data organization. The key approach for broadly identifying relevant information from the data store will concern application of a vector embedding model and suitable comparison metric, e.g. the dot-product distance between embedding vectors. Additional identification techniques are then applied to clean the results by removing duplicates and suppressing spurious relationships. This is accomplished through the use of contextual data stores like the knowledge graph created by the data organization step from the input pipeline to find and combine synonyms or other colliding topics.

[0025] Once the relevant data is retrieved, it is organized and formatted alongside the processed query to prepare appropriate input for the response generation procedure. Special attention must be paid to how this organization is performed to render a prompt suitable for use as input to subsequent steps. In practice the format is engineered for the specific application, e.g. a form template optimized to arrange news headlines in such a way that they are reliably parsed by the specific model to which they will be provided as input. In response generation a powerful model such as an LLM is employed to generate a natural language response. The LLM is provided with the formatted prompt, which contains all additional context necessary to generate an accurate and coherent response. The generated response is then processed to ensure accurate reproduction of primary source information and compliance with any relevant filters and augmented with direct references to the data sources used in composing the response. The post-processing is extensive and involves applying the same sort of natural language understanding techniques used to process the input data to identify relevant components of the response and generate embeddings of the response components that can be used to determine relationships between the response and the input text. Response components that contain direct quotes of the input documents or that score above an application-specific threshold are determined to be drawn from the document in question and are annotated as such. The annotations can then be transmitted to the user for each part of the response, such that individual sentences or sentence clauses are appropriately marked withcitations of source documents. This ensures that users receive accurate, up-to-date information with clear attribution to the original sources. Figure 6 illustrates an example process performed to generate a user facing response illustrated in Figure 7.

[0026] The invention's architecture and algorithmic design enable the system to leverage the natural language capabilities of LLMs while providing direct connections to primary data sources, as facilitated by traditional search engines. This allows the system to deliver real-time, dynamic processing of resources without incurring the prohibitive costs associated with re-training an LLM, while also providing accurate attribution for all resources used in composing the response.Computer System

[0027] Figure 2 is a block diagram that illustrates an example of a computer system 200 in which at least some operations described herein can be implemented. As shown, the computer system 200 can include: one or more processors 202, main memory 206, non-volatile memory 210, a network interface device 212, video display device 218, an input / output device 220, a control device 222 (e.g., keyboard and pointing device), a drive unit 224 that includes a storage medium 226, and a signal generation device 230 that are communicatively connected to a bus 216. The bus 216 represents one or more physical buses and / or point-to-point connections that are connected by appropriate bridges, adapters, or controllers. Various common components (e.g., cache memory) are omitted from Figure 2 for brevity. Instead, the computer system 200 is intended to illustrate a hardware device on which components illustrated or described relative to the examples of the figures and any other components described in this specification can be implemented.

[0028] The computer system 200 can take any suitable physical form. For example, the computer system 200 can share a similar architecture to that of a server computer, personal computer (PC), tablet computer, mobile telephone, game console, music player, wearable electronic device, network-connected (“smart”) device (e.g., a television or home assistant device), AR / VR systems (e.g., head-mounted display), or any electronic device capable of executing a set of instructions that specify action(s) to be taken by the computer system 200. In some implementation, the computer system 200 can be an embedded computer system, a system-on-chip (SOC), a singleboard computer system (SBC) or a distributed system such as a mesh of computersystems or include one or more cloud components in one or more networks. Where appropriate, one or more computer systems 200 can perform operations in real-time, near real-time, or in batch mode.

[0029] The network interface device 212 enables the computer system 200 to exchange data in a network 214 with an entity that is external to the computing system 200 through any communication protocol supported by the computer system 200 and the external entity. Examples of the network interface device 212 include a network adaptor card, a wireless network interface card, a router, an access point, a wireless router, a switch, a multilayer switch, a protocol converter, a gateway, a bridge, bridge router, a hub, a digital media receiver, and / or a repeater, as well as all wireless elements noted herein.

[0030] The memory (e.g., main memory 206, non-volatile memory 210, machine- readable medium 226) can be local, remote, or distributed. Although shown as a single medium, the machine-readable medium 226 can include multiple media (e.g., a centralized / distributed database and / or associated caches and servers) that store one or more sets of instructions 228. The machine-readable (storage) medium 226 can include any medium that is capable of storing, encoding, or carrying a set of instructions for execution by the computer system 200. The machine-readable medium 226 can be non-transitory or comprise a non-transitory device. In this context, a non- transitory storage medium can include a device that is tangible, meaning that the device has a concrete physical form, although the device can change its physical state. Thus, for example, non-transitory refers to a device remaining tangible despite this change in state.

[0031] Although implementations have been described in the context of fully functioning computing devices, the various examples are capable of being distributed as a program product in a variety of forms. Examples of machine-readable storage media, machine-readable media, or computer-readable media include recordable-type media such as volatile and non-volatile memory devices 210, removable memory, hard disk drives, optical disks, and transmission-type media such as digital and analog communication links.

[0032] In general, the routines executed to implement examples herein can be implemented as part of an operating system or a specific application, component,program, object, module, or sequence of instructions (collectively referred to as “computer programs”). The computer programs typically comprise one or more instructions (e.g., instructions 204, 208, 228) set at various times in various memory and storage devices in computing device(s). When read and executed by the processor 202, the instruction (s) cause the computer system 700 to perform operations to execute elements involving the various aspects of the disclosure.Computing Environment

[0033] Figure 3 is a system diagram illustrating an example of a computing environment in which the disclosed system operates in some implementations. In some implementations, environment 300 includes one or more client computing devices 305A-D, examples of which can host the machine learning platform 100. Client computing devices 305 operate in a networked environment using logical connections through network 330 to one or more remote computers, such as a server computing device.

[0034] In some implementations, server 310 is an edge server which receives client requests and coordinates fulfillment of those requests through other servers, such as servers 620A-C. In some implementations, server computing devices 310 and 320 comprise computing systems, such as the machine learning platform 100. Though each server computing device 310 and 320 is displayed logically as a single server, server computing devices can each be a distributed computing environment encompassing multiple computing devices located at the same or at geographically disparate physical locations. In some implementations, each server 320 corresponds to a group of servers.

[0035] Client computing devices 305 and server computing devices 310 and 320 can each act as a server or client to other server or client devices. In some implementations, servers (310, 320A-C) connect to a corresponding database (315, 325A-C). As discussed above, each server 320 can correspond to a group of servers, and each of these servers can share a database or can have its own database. Databases 315 and 325 warehouse (e.g., store) information pertinent to applications described herein, including input data, intermediate processing results, output data, and / or post-processing data. Though databases 315 and 325 are displayed logically as single units, databases 315 and 325 can each be a distributed computingenvironment encompassing multiple computing devices, can be located within their corresponding server, or can be located at the same or at geographically disparate physical locations.

[0036] Network 330 can be a local area network (LAN) or a wide area network (WAN), but can also be other wired or wireless networks. In some implementations, network 330 is the Internet or some other public or private network. Client computing devices 305 are connected to network 330 through a network interface, such as by wired or wireless communication. While the connections between server 310 and servers 320 are shown as separate connections, these connections can be any kind of local, wide area, wired, or wireless network, including network 330 or a separate public or private network.Conclusion

[0037] Unless the context clearly requires otherwise, throughout the description and the claims, the words "comprise," "comprising," and the like are to be construed in an inclusive sense, as opposed to an exclusive or exhaustive sense; that is to say, in the sense of "including, but not limited to." As used herein, the terms "connected," "coupled," or any variant thereof means any connection or coupling, either direct or indirect, between two or more elements; the coupling or connection between the elements can be physical, logical, or a combination thereof. Additionally, the words "herein," "above," "below," and words of similar import, when used in this application, refer to this application as a whole and not to any particular portions of this application. Where the context permits, words in the above Detailed Description using the singular or plural number may also include the plural or singular number respectively. The word "or," in reference to a list of two or more items, covers all of the following interpretations of the word: any of the items in the list, all of the items in the list, and any combination of the items in the list.

[0038] The above Detailed Description of examples of the technology is not intended to be exhaustive or to limit the technology to the precise form disclosed above. While specific examples for the technology are described above for illustrative purposes, various equivalent modifications are possible within the scope of the technology, as those skilled in the relevant art will recognize. For example, while processes or blocks are presented in a given order, alternative embodiments mayperform routines having steps, or employ systems having blocks, in a different order, and some processes or blocks may be deleted, moved, added, subdivided, combined, and / or modified to provide alternative or sub-combinations. Each of these processes or blocks may be implemented in a variety of different ways. Also, while processes or blocks are at times shown as being performed in series, these processes or blocks may instead be performed or implemented in parallel, or may be performed at different times. Further, any specific numbers noted herein are only examples: alternative embodiments may employ differing values or ranges.

[0039] The teachings of the technology provided herein can be applied to other systems, not necessarily the system described above. The elements and acts of the various examples described above can be combined to provide further embodiments of the technology. Some alternative embodiments of the technology may include not only additional elements to those embodiments noted above, but also may include fewer elements.

[0040] These and other changes can be made to the technology in light of the above Detailed Description. While the above description describes certain examples of the technology, and describes the best mode contemplated, no matter how detailed the above appears in text, the technology can be practiced in many ways. Details of the system may vary considerably in its specific implementation, while still being encompassed by the technology disclosed herein. As noted above, specific terminology used when describing certain features or aspects of the technology should not be taken to imply that the terminology is being redefined herein to be restricted to any specific characteristics, features, or aspects of the technology with which that terminology is associated. In general, the terms used in the following claims should not be construed to limit the technology to the specific examples disclosed in the specification, unless the above Detailed Description section explicitly defines such terms. Accordingly, the actual scope of the technology encompasses not only the disclosed examples, but also all equivalent ways of practicing or implementing the technology under the claims.

[0041] To reduce the number of claims, certain aspects of the technology are presented below in certain claim forms, but the applicant contemplates the various aspects of the technology in any number of claim forms. For example, while only one aspect of the technology is recited as a computer-readable medium claim, otheraspects may likewise be embodied as a computer-readable medium claim, or in other forms, such as being embodied in a means-plus-function claim. Any claims intended to be treated under 35 ll.S.C. § 112(f) will begin with the words "means for," but use of the term "for" in any other context is not intended to invoke treatment under 35 U.S.C. § 112(f). Accordingly, the applicant reserves the right to pursue additional claims after filing this application to pursue such additional claim forms, in either this application or in a continuing application.

[0042] Although the present invention has been described in detail with reference to the preferred embodiments, various modifications and enhancements may be made without departing from the scope and spirit of the invention. Those skilled in the art will appreciate that the invention is not limited to the specific examples and embodiments described herein, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

CLAIMSWhat is claimed is:1 . A system for answering user queries with direct citations to primary sources, comprising: a. a data collection pipeline configured to ingest, process, and organize data of multiple types from multiple sources b. a retrieval mechanism configured to process user queries, identify data relevant data, and employ a machine learning model to generate natural language responses that incorporate and cite information derived from the processed data.

2. The system of claim 1 , wherein the data collection pipeline further comprises: a. a data ingestion component for fetching data from various sources; b. at least one data processing component for transforming and adding structure to the ingested data; c. at least one data organization component for collecting processed data into a searchable data structure.

3. The system of claim 1 , wherein the retrieval mechanism comprises: a. a query processing component for transforming user queries into an appropriate internal representation for data retrieval; b. a data retrieval component for identifying and extracting relevant data from the organized data structure; c. a response generation component for employing a machine learning model to generate natural language responses based on the retrieved data; d. a citation integration component for filtering and augmenting generated responses with direct references to primary sources.

Citation Information

Patent Citations

  • Methods and system for providing real-time business intelligence using natural language queries

    US10657125B1

  • Real-time predictive knowledge pattern machine

    US10943172B1

  • System for annotating a data object by creating an interface based on a selected annotation structure

    US20060129596A1

  • Knowledge discovery agent system and method

    US20120059776A1

  • Question answer system using phyisical distance data

    US20160103884A1

Cited By

  • NL2SQL method, device and equipment based on preset industry large model and medium

    CN119917524A

  • NL2SQL Method, Apparatus, Device, and Medium Based on a Preset Industry Large Model

    CN119917524B

  • Question and answer method and device, equipment, storage medium and program product

    CN121542382A