Method and apparatus for retrieving document, and device and program product
Patent Information
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- BEIJING ZITIAO NETWORK TECH CO LTD
- Filing Date
- 2024-11-29
- Publication Date
- 2026-06-04
Smart Images
Figure CN2024135849_04062026_PF_FP_ABST
Abstract
Description
Methods, apparatus, devices, and procedures for retrieving documents. Technical Field
[0001] This disclosure relates to the field of computer technology, and more specifically, to a method, apparatus, computing device, computer-readable storage medium, and computer program product for retrieving documents. Background Technology
[0002] Retrieval-Augmented Generation (RAG) is a natural language processing technique that combines the capabilities of information retrieval and generative text generation. Within this framework, the model first uses information retrieval techniques to retrieve text fragments or documents relevant to the input query from a large-scale knowledge base. This retrieved information is then used as additional contextual or reference knowledge and fed into the generative model to generate more accurate, richer, and context-relevant text output. The core of the RAG method lies in its ability to effectively integrate external knowledge into the generation process, thereby improving the quality and diversity of the model's responses. This approach is particularly suitable for tasks requiring extensive knowledge and deep understanding, such as question-answering systems, dialogue systems, text summarization, and content creation.
[0003] In the field of information retrieval, open-source general-purpose models and domain-specific models are two common model types, each with different characteristics and applicable scenarios. Open-source general-purpose models are typically trained on large-scale, diverse datasets, aiming to provide broad applicability and high performance. Domain-specific models, on the other hand, are trained on datasets specific to a particular domain or task, and these models are generally able to handle domain-related tasks and data more accurately. Summary of the Invention
[0004] This disclosure provides a method, apparatus, computing device, computer-readable storage medium, and computer program product for retrieving documents. It first obtains preliminary search results based on information such as text similarity, and then sorts the preliminary search results using a general ranking model and a domain-specific ranking model to obtain the final search results. This combines the advantages of both the general model and the domain-specific model to improve the overall search accuracy.
[0005] According to a first aspect of this disclosure, a method for retrieving documents is provided, the method comprising retrieving a set of document blocks matching input text from a document library. The method further comprises ranking the document blocks in the set of document blocks using multiple ranking models, including a general ranking model and at least one domain-specific ranking model. Furthermore, the method comprises determining a subset of the set of document blocks as the retrieval results based on the ranking results of the multiple ranking models.
[0006] According to a second aspect of this disclosure, an apparatus for retrieving documents is provided, comprising: a retrieval unit configured to retrieve a set of document blocks matching an input text from a document library; a sorting unit configured to sort the document blocks in the set of document blocks using multiple sorting models, the multiple sorting models including a general sorting model and at least one domain-specific sorting model; and a result determination unit configured to determine a subset of the set of document blocks as the retrieval result based on the sorting results of the multiple sorting models.
[0007] According to a third aspect of this disclosure, a computing device is provided, comprising: at least one processing unit; at least one memory coupled to the at least one processing unit and storing instructions for execution by the at least one processing unit, the instructions, when executed by the at least one processing unit, causing the computing device to perform the method as described in the first aspect of this disclosure.
[0008] According to a fourth aspect of this disclosure, a non-transient computer storage medium is provided, including machine-executable instructions that, when executed by a device, cause the device to perform the method as described in the first aspect of this disclosure.
[0009] According to a fifth aspect of this disclosure, a computer program product is provided, including machine-executable instructions that, when executed by a device, cause the device to perform the method as described in the first aspect of this disclosure.
[0010] It should be understood that the summary section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description
[0011] The above and other objects, features, and advantages of embodiments of the present disclosure will become more readily understood from the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of the present disclosure will be described by way of example and non-limitation, wherein:
[0012] Figure 1 shows a schematic diagram of an example environment in which embodiments of the present disclosure may be implemented;
[0013] Figure 2 shows a schematic flowchart of a method for retrieving documents according to an embodiment of the present disclosure;
[0014] Figure 3 shows a schematic diagram of the structure of a multi-model document retrieval system according to an embodiment of the present disclosure;
[0015] Figure 4 shows a block diagram of an apparatus for retrieving documents according to an embodiment of the present disclosure; and
[0016] Figure 5 shows a block diagram of an electronic device according to an embodiment of the present disclosure.
[0017] In all the accompanying figures, the same or similar reference numerals denote the same or similar elements. Detailed Implementation
[0018] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.
[0019] In the description of embodiments of this disclosure, the term "comprising" and similar terms should be understood as open-ended inclusion, i.e., "including but not limited to". The term "based on" should be understood as "at least partially based on". The term "one embodiment" or "the embodiment" should be understood as "at least one embodiment". The terms "first", "second", etc., may refer to different or the same objects unless explicitly stated. Other explicit and implicit definitions may also be included below.
[0020] As mentioned above, open-source general-purpose models and domain-specific models each have their advantages in the retrieval domain. However, currently, the industry often uses only a single model for retrieval. Open-source general-purpose models perform well on general datasets but poorly on domain-specific datasets; conversely, models trained on domain-specific data perform well on domain-specific datasets but poorly on general datasets, resulting in poor overall retrieval recall and hit rates in practical applications. Therefore, combining the advantages of both open-source general-purpose models and domain-specific models to improve retrieval accuracy presents a challenge for those skilled in the art.
[0021] To address or mitigate the aforementioned problems and / or other potential issues, embodiments of this disclosure propose a method for document retrieval. This method first obtains preliminary search results based on information such as text similarity, and then sorts these preliminary search results using both a general ranking model and a domain-specific ranking model to obtain the final search results. This approach combines the advantages of both general and domain-specific models, thereby improving overall search accuracy.
[0022] The basic principles and implementation of this disclosure are illustrated below with reference to the accompanying drawings. It should be understood that the exemplary embodiments given are merely intended to enable those skilled in the art to better understand and implement the embodiments of this disclosure, and are not intended to limit the scope of this disclosure in any way.
[0023] Figure 1 illustrates a schematic diagram of an environment 100 capable of implementing various embodiments of the present disclosure. As shown in Figure 1, the environment 100 may include input text 101, a multi-model document retrieval system 102, a document library 103, and search results 104. The input text 101 may be a question description text manually entered by a user through an input device (e.g., a mobile phone, tablet, personal computer, etc.). Alternatively, the input text 101 may also be read directly from a storage device (e.g., a database, cloud storage, etc.) or received from other devices via a communication connection.
[0024] As shown in Figure 1, input text 101 can be provided to a multi-model document retrieval engine 102 to retrieve search results 104 from a document library 103. The multi-model document retrieval engine 102 can consist of multiple pre-trained open-source general-purpose models and domain-specific models. The multi-model document retrieval engine 102 can be deployed on a standalone physical network server, a network server cluster or distributed system composed of multiple physical network servers, or a cloud network server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms.
[0025] Document library 103 can contain a large amount of knowledge information related to input text 101 and store it in the form of document blocks. Document library 103 can support users to search using keywords, titles, and other methods. Optionally, document library 103 can also provide a corresponding vector database to support users to search by comparing the similarity of text vectors. Search results 104 can be multiple document blocks retrieved from document library 103 that are highly relevant to input text 101. Search results 104 can be used as reference knowledge and provided to the generative model along with input text 101 to generate an answer to input text 101.
[0026] Figure 2 illustrates a schematic flowchart of a document retrieval method 200 according to an embodiment of the present disclosure. In some embodiments, method 200 may be implemented by, for example, a multi-model document retrieval device 102 as shown in Figure 1. It should be understood that method 200 may also include additional actions not shown and / or actions shown may be omitted, and the scope of the present disclosure is not limited in this respect.
[0027] As shown in Figure 2, in box 210, method 200 may include: retrieving a set of document blocks matching the input text from a document library. The document library 103 may contain a vast amount of knowledge related to the input text 101 and store it in the form of document blocks. The input text 101 may be manually entered by the user or read directly from a storage device.
[0028] Text vector models are methods for converting text data into numerical vectors for processing and analysis in computers. These models are able to capture semantic relationships and similarities between texts, which is crucial for natural language processing (NLP) tasks. In some embodiments, the multi-model document retrieval system 102 may first generate multiple vector representations of the input text 101 using multiple text vector models, which may include a general text vector model and at least one domain-specific text vector model.
[0029] General-purpose text vector models are typically trained on large-scale, diverse datasets, thus enabling them to handle various task types and possess broad cross-domain applicability. Optionally, general-purpose text vector models can be Bag of Words (BOW), Word Embeddings, Sentence Embeddings, etc., and this disclosure makes no limitation thereto. Domain-specific text vector models, trained on datasets specific to a particular domain or task, are generally more accurate in handling industry-specific terminology and jargon, avoiding misunderstandings and ambiguities. The similarity between different words differs between the two models. For example, if the abbreviation for "test environment" within a company is "TE," a domain-specific text vector model would consider "TE" and "test environment" to have a relatively close vector similarity, while a general-purpose text vector model would not.
[0030] Then, the multi-model document retrieval unit 102 can retrieve document blocks similar to multiple vector representations in the vector database of the document library 103 based on vector similarity, and select the top N document blocks with the highest similarity. Optionally, cosine similarity can be used to evaluate their similarity by calculating the cosine value of the angle between these vectors. Thus, multiple document blocks with high textual similarity to the input text 101 can be obtained.
[0031] Elasticsearch (ES) is a search engine that provides full-text search, structured search, analytics, and combinations of these three capabilities. The core principle of ES search is to use an inverted index to quickly locate documents containing query terms. An inverted index is a data structure that records which documents each term appears in and where it appears. When a user enters query terms, ES search uses the inverted index to quickly find documents containing those terms, sorts these documents according to a specific algorithm, and ultimately displays the documents that best match the user's needs.
[0032] Term Frequency-Inverse Document Frequency (TF-IDF) is a commonly used weighting technique in information retrieval and data mining. It assesses the importance of a word to a document within a document set or corpus. A higher TF-IDF value indicates greater importance of the word in the document and stronger discriminative power. In Elasticsearch (ES) search, TF-IDF scores are used to measure the relevance between documents and query terms. Specifically, when a user enters a query term, ES search calculates the TF-IDF value of the query term in each document and ranks the documents based on these values. Documents with higher TF-IDF values are more relevant to the query term and therefore rank higher in the search results. In some embodiments, the multi-model document retriever 102 can also retrieve document blocks matching the input text 101 based on term frequency using elastic search. This leverages ES search's lack of semantic understanding to obtain multiple document blocks that consider only the term frequency of core keywords while excluding invalid sentence components.
[0033] In some embodiments, the multi-model document retrieval system 102 can first perform a merging and deduplication operation on the document blocks retrieved based on vector similarity and word frequency, and then determine the result of the merging and deduplication operation as a set of document blocks. Thus, the multi-model document retrieval system 102 completes the initial screening step. In this step, since the model only needs to match based on text similarity or word frequency, it has the advantages of small model size and fast processing speed.
[0034] In box 220, method 200 may include: ranking document blocks in a document block set using multiple ranking models, including a general ranking model and at least one domain-specific ranking model. Unlike text vector models, ranking models are able to understand the semantic information of the input text 101 and the document blocks in the document library 103. A confidence score is used in tasks such as Natural Language Processing (NLP) and information retrieval to quantify the reliability of a prediction or answer. It is typically a value between 0 and 1, representing how confident the system is in its given answer or prediction. Thus, the ranking model can be trained to "determine whether a document block can answer the user's question and output a confidence score." The general ranking model and the domain-specific ranking model score the answer differently due to their different training data. Therefore, combining the ranking results of the general ranking model and the domain-specific ranking model can integrate their respective advantages.
[0035] Based on the training method of the ranking model described above, in some embodiments, the multi-model document retrieval machine 102 can first determine the semantic confidence of each document block in the document block set using a general ranking model and at least one domain-specific ranking model, respectively. Then, the multi-model document retrieval machine 102 can rank the document blocks in the document block set based on the semantic confidence. Thus, the multi-model document retrieval machine 102 completes the fine-grained ranking step. In this step, the ranking model can understand some semantics and has the characteristics of moderate model size, moderate processing speed, and high accuracy.
[0036] In box 230, method 200 may include: determining a subset of the document block set as the retrieval result based on the ranking results of multiple ranking models. The multi-model document retrieval system 102 may judge the ranking results obtained from different ranking models to determine the final retrieval result. In some embodiments, the multi-model document retrieval system 102 may, in response to the consistency of the ranking results of multiple ranking models, determine a subset of the document block set (e.g., the top K document blocks of the ranking results) as the retrieval result based on the ranking results.
[0037] In some embodiments, the multi-model document retrieval system 102, in response to inconsistencies in the ranking results of multiple ranking models, uses a generative model combined with domain-specific explanatory terms corresponding to the input text to determine a subset of the document block set as the retrieval results. The generative model has higher semantic understanding capabilities than the ranking model, and by incorporating domain-specific explanatory terms, it achieves higher retrieval accuracy. Thus, the multi-model document retrieval system 102 completes the final filtering step. In this step, the generative model is slower but has the highest accuracy compared to other steps.
[0038] In some embodiments, the multi-model document retrieval system 102 can achieve a certain degree of freedom in its arrangement based on a directed acyclic graph (DAG), allowing users to adjust the configuration by editing JSON configuration files based on actual scenarios. For example, when there are many documents and the required documents are not found, the topN parameter of the coarse screening step can be increased to increase the number of document blocks entering the fine ranking step. As another example, when the fine ranking model performs poorly, the topK parameter of the fine ranking step can be increased to allow more document blocks to enter the generative model with stronger semantic understanding for final selection.
[0039] Figure 3 shows a schematic diagram of the structure of a multi-model document retrieval system 300 according to an embodiment of the present disclosure. The multi-model document retrieval system 300 can be considered as a specific implementation of the multi-model document retrieval system 102 shown in Figure 1. As shown in Figure 3, in step 301, the multi-model document retrieval system 300 can obtain input text 301. The input text 301 can be manually entered by the user or directly read from a storage device. In steps 302 and 303, the multi-model document retrieval system 300 can respectively train a text vector model 302-1 and a general text vector model 303-1 using domain data to obtain document blocks with high text similarity to the input text 301.
[0040] In step 302-2, the multi-model document retrieval machine 300 can retrieve text-similar document blocks in a specific domain from the text vector library corresponding to the document library. Then, in step 302-3, the multi-model document retrieval machine 300 can select the top N most similar document blocks based on cosine similarity. In step 303-2, the multi-model document retrieval machine 300 can retrieve text-similar document blocks in a general domain from the text vector library corresponding to the document library. Then, in step 302-3, the multi-model document retrieval machine 300 can also select the top N most similar document blocks based on cosine similarity.
[0041] In step 304, the multi-model document retrieval machine 300 can also obtain multiple document blocks that match the input text based on word frequency. In step 304-1, the multi-model document retrieval machine 300 can obtain multiple document blocks through Elasticsearch (ES) retrieval, considering only the frequency of core keywords. Then, in step 304-2, the top N most matching document blocks are selected based on the TF-IDF score. It is worth noting that N is a positive integer and can be adjusted by the user based on the actual scenario.
[0042] In step 305, the multi-model document retrieval system 300 can perform a merging and deduplication operation on the multiple document blocks obtained in steps 302, 303, and 304 to obtain the initial screening result. Then, the multi-model document retrieval system 300 can perform the next stage of fine ranking operation based on the initial screening result.
[0043] In steps 306 and 307, the multi-model document retrieval system 300 can also utilize different types of models to obtain ranking results. Specifically, in step 306-2, the multi-model document retrieval system 300 can use domain-specific knowledge trained on the ranking model 306-1 with domain data to rank the initial screening results, and select the top K most matching document blocks based on the ranking score. In step 307-2, the multi-model document retrieval system 300 can use the general knowledge of the general ranking model 307-1 to rank the initial screening results, and similarly select the top K most matching document blocks based on the ranking score. It is worth noting that K is a positive integer and can be adjusted by the user based on the actual scenario.
[0044] Then, in step 308, the multi-model document retrieval machine 300 can compare and judge the ranking results obtained in steps 306 and 307 respectively. If the results are consistent, it directly proceeds to step 311 to output the retrieval results. If they are inconsistent, it proceeds to step 309 to input the different ranking results into a generative model with strong semantic understanding, and combines them with domain terms 310 for final filtering. The filtering results of the generative model can be directly used as the final retrieval results.
[0045] Exemplary embodiments of the present disclosure have been described above with reference to Figures 1 to 3. Compared with existing document retrieval methods, the document retrieval scheme of the present disclosure first obtains preliminary retrieval results based on information such as text similarity, and then sorts the preliminary retrieval results using a general ranking model and a domain-specific ranking model respectively to obtain the final retrieval results. This allows the scheme to combine the advantages of both the general model and the domain-specific model, thereby improving the overall retrieval accuracy.
[0046] Figure 4 shows a schematic block diagram of a document retrieval apparatus 400 according to an embodiment of the present disclosure. As shown in Figure 4, the apparatus 400 includes a retrieval unit 402, a sorting unit 404, and a result determination unit 406.
[0047] In some embodiments, the retrieval unit 402 is configured to retrieve a set of document blocks that match the input text in a document library; the sorting unit 404 is configured to sort the document blocks in the document block set using multiple sorting models, including a general sorting model and at least one domain-specific sorting model; and the result determination unit 406 is configured to determine a subset of the document block set as the retrieval result based on the sorting results of the multiple sorting models.
[0048] In some embodiments, the retrieval unit 402 is further configured to generate multiple vector representations of the input text through multiple text vector models, the multiple text vector models including a general text vector model and at least one domain-specific text vector model; and to retrieve document blocks similar to the multiple vector representations based on vector similarity in the vector database of the document library.
[0049] In some embodiments, the retrieval unit 402 is further configured to retrieve document blocks that match the input text based on word frequency using elastic search.
[0050] In some embodiments, the retrieval unit 402 is further configured to perform a merging and deduplication operation on document blocks retrieved based on vector similarity and word frequency; and to determine the result of the merging and deduplication operation as the document block set.
[0051] In some embodiments, the sorting unit 404 is further configured to determine the semantic confidence of each document block in the document block set using the general sorting model and the at least one domain-specific sorting model, respectively; and to sort the document blocks in the document block set based on the semantic confidence.
[0052] In some embodiments, the result determination unit 406 is further configured to determine a subset of the document block set as the retrieval result based on the ranking result in response to the consistency of the ranking results of the plurality of ranking models.
[0053] In some embodiments, the result determination unit 406 is further configured to, in response to inconsistencies in the ranking results of the plurality of ranking models, determine a subset of the document block set as the retrieval result by using a generative model and combining domain explanatory terms corresponding to the input text.
[0054] It should be noted that further actions or steps shown in Figures 1 to 3 can be implemented using the device 400 shown in Figure 4. For example, device 400 may include more modules or units to implement the actions or steps described above, or some of the units or modules shown in Figure 4 may be further configured to implement the actions or steps described above. This will not be repeated here.
[0055] Figure 5 shows a schematic block diagram of an example device 500 that can be used to implement embodiments of the present disclosure. As shown, device 500 includes a computing unit 501, which can perform various appropriate actions and processes according to computer program instructions stored in read-only memory (ROM) 502 or loaded from storage unit 506 into random access memory (RAM) 503. Various programs and data required for the operation of device 500 may also be stored in RAM 503. The computing unit 501, ROM 502, and RAM 503 are interconnected via bus 504. Input / output (I / O) interface 505 is also connected to bus 504.
[0056] Multiple components in device 500 are connected to I / O interface 505, including: input unit 506, such as keyboard, mouse, etc.; output unit 507, such as various types of monitors, speakers, etc.; storage unit 508, such as disk, optical disk, etc.; and communication unit 509, such as network card, modem, wireless transceiver, etc. Communication unit 509 allows device 500 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.
[0057] The computing unit 501 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 501 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 501 performs the various methods and processes described above, such as method 200. For example, in some embodiments, method 200 may be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 508. In some embodiments, part or all of the computer program may be loaded and / or installed on device 500 via ROM 502 and / or communication unit 509. When the computer program is loaded into RAM 503 and executed by the computing unit 501, one or more steps of method 200 described above may be performed. Alternatively, in other embodiments, the computing unit 501 may be configured to perform method 200 by any other suitable means (e.g., by means of firmware).
[0058] In some embodiments, the methods and processes described above can be implemented as a computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of this disclosure.
[0059] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination thereof. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.
[0060] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper cables, fiber optic cables, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to computer-readable storage media within the respective computing / processing device.
[0061] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages and conventional procedural programming languages. The computer-readable program instructions may execute entirely on a user's computer, partially on a user's computer, as a standalone software package, partially on a user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.
[0062] These computer-readable program instructions can be provided to a processing unit of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the processing unit of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner. Thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing aspects of the functions / actions specified in one or more blocks of the flowchart and / or block diagram.
[0063] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.
[0064] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of devices, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those shown in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.
[0065] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.
Claims
1. A method for retrieving documents, comprising: Retrieve a set of document blocks in the document library that match the input text; The document blocks in the document block set are sorted using multiple sorting models, including a general sorting model and at least one domain-specific sorting model; and Based on the ranking results of the multiple ranking models, a subset of the document block set is determined as the retrieval result.
2. The method of claim 1, wherein retrieving a set of document blocks matching the input text from the document library comprises: Multiple vector representations of the input text are generated through multiple text vector models, including a general text vector model and at least one domain-specific text vector model; as well as Based on vector similarity, document blocks that are similar to the multiple vector representations are retrieved from the vector database of the document library.
3. The method according to claim 2, wherein retrieving a set of document blocks matching the input text from the document library further comprises: By using elastic search, document blocks that match the input text are retrieved based on word frequency.
4. The method according to claim 3, wherein obtaining the document block set comprises: Merge and deduplicate document blocks retrieved based on vector similarity and word frequency; as well as The result of the merging and deduplication operation is determined as the document block set.
5. The method according to claim 1, wherein sorting the document blocks in the document block set using multiple sorting models comprises: The semantic confidence of each document block in the document block set is determined by the general ranking model and the at least one domain-specific ranking model, respectively. as well as Based on the semantic confidence, the document blocks in the document block set are sorted.
6. The method according to claim 1, wherein determining a subset of the document block set as the retrieval result based on the ranking results of the plurality of ranking models includes: In response to the consistency of the ranking results of the multiple ranking models, a subset of the document block set is determined as the retrieval result based on the ranking results.
7. The method according to claim 1, wherein determining a subset of the document block set as the retrieval result based on the ranking results of the plurality of ranking models includes: In response to the inconsistency in the ranking results of the multiple ranking models, a subset of the document block set is determined as the retrieval result by using a generative model and combining it with domain-specific explanatory terms corresponding to the input text.
8. A multi-model recall system, comprising: The retrieval unit is configured to retrieve a set of document blocks in the document library that match the input text; A sorting unit is configured to sort document blocks in the document block set using multiple sorting models, including a general sorting model and at least one domain-specific sorting model; and The result determination unit is configured to determine a subset of the document block set as the retrieval result based on the sorting results of the multiple sorting models.
9. A computing device, comprising: At least one processing unit; At least one memory coupled to the at least one processing unit and storing instructions for execution by the at least one processing unit, the instructions, when executed by the at least one processing unit, causing the computing device to perform the method as described in any one of claims 1 to 7.
10. A computer storage medium comprising machine-executable instructions that, when executed by a device, cause the device to perform the method as claimed in any one of claims 1 to 7.
11. A computer program product comprising machine-executable instructions that, when executed by a device, cause the device to perform the method as described in any one of claims 1 to 7.