CN116340502A - Information retrieval method and device based on semantic understanding - Google Patents

Abstract

An information retrieval method and device based on semantic understanding are disclosed. The information retrieval method includes: acquiring first text information and a plurality of second text information; determining the semantic similarity between the first text information and each second text information; according to the semantic similarity between the first text information and each second text information Similarity, selecting at least one text information to be retrieved from a plurality of second text information; extracting third text information semantically related to the first text information from at least one text information to be retrieved to form a third text information set; Obtain multi-text summaries corresponding to at least two third text information in the third text information set; and determine a retrieval result corresponding to the first text information based on the multi-text summaries. The information retrieval method according to some embodiments of the present application can efficiently and accurately complete advanced information retrieval tasks such as intelligent question answering through data processing operations such as information screening, extraction, and summarization.

Description

Information retrieval method and device based on semantic understanding

technical field

The present application relates to the field of natural language processing, in particular to an information retrieval method and device based on semantic understanding, a computing device, a computer-readable storage medium and a computer program product.

Background technique

With the rapid development of the Internet, more and more information can be retrieved from the Internet through search engines, and the search results show the characteristics of massive data, diversified forms, and comprehensive coverage. On the one hand, this increases the possibility for users to search for results. On the other hand, users will feel at a loss in the face of massive search results and cannot obtain accurate answers in a short time. For example, traditional search engines based on keyword matching and single-document abstraction are limited to returning a list of web pages or documents related to the user's retrieval question, but cannot give an accurate answer to the question (users need to combine information such as title and abstract from related web pages or documents) search or get answers to questions), which cannot meet the needs and expectations of users to obtain information quickly.

As users expect more and more from search engines, the form of information retrieval begins to change from a basic form such as recalling a list of relevant web pages or documents to an advanced form such as intelligent question-and-answer retrieval. The purpose of intelligent question and answer retrieval is to answer users' questions with concise and accurate natural language, and its emergence is dedicated to providing more effective information acquisition tools. In order to realize advanced information retrieval forms such as intelligent question answering, intelligent information retrieval technology based on semantic understanding or machine reading comprehension came into being. However, the information retrieval methods based on semantic understanding or machine reading comprehension in related technologies have the following problems: First, the retrieval methods based on keywords or character string matching in related technologies can only retrieve articles with the same text, and cannot retrieve articles with different texts but different texts. Information with the same semantics will easily cause the loss of important information resources highly related to the retrieval problem, and the retrieval results will be limited in breadth and accuracy. Setting rules will lead to a large amount of calculation and work, which has low efficiency, and keywords cannot completely and accurately reflect the characteristics of the entire retrieval problem, resulting in low accuracy of retrieval results.

Contents of the invention

In view of this, the present application provides an information retrieval method and device based on semantic understanding, a computing device, a computer-readable storage medium, and a computer program product, hoping to alleviate or overcome some or all of the above-mentioned defects and other possible defects.

According to the first aspect of the present application, an information retrieval method based on semantic understanding is provided, including: acquiring first text information indicating a retrieval target and a plurality of second text information indicating candidate retrieval objects; determining the first text Semantic similarity between the information and each second text information in the plurality of second text information; according to the semantic similarity between the first text information and each second text information in the plurality of second text information, Selecting at least one text information to be retrieved from the plurality of second text information; respectively extracting third text information semantically related to the first text information from the at least one text information to be retrieved to form a third text information set ; Obtain a multi-text abstract corresponding to at least two third text information in the third text information set; determine a search result corresponding to the first text information based on the multi-text abstract.

In the information retrieval method according to some embodiments of the present application, the multi-text summarization includes a generative multi-text summarization.

In the information retrieval method according to some embodiments of the present application, determining the semantic similarity between the first text information and each second text information in the plurality of second text information includes: acquiring the first text information The corresponding first semantic feature vector and the plurality of second semantic feature vectors respectively corresponding to the plurality of second text information; calculating the first semantic feature vector and each second semantic feature in the plurality of second semantic feature vectors Vector similarity; according to the similarity between the first semantic feature vector and each second semantic feature vector, determine the semantic similarity between the first text information and each second text information in the plurality of second text information .

In the information retrieval method according to some embodiments of the present application, calculating the similarity between the first semantic feature vector and each of the plurality of second semantic feature vectors includes: based on the plurality of second semantic The distance between the feature vector and the first semantic feature vector, calculating the first similarity between the first semantic feature vector and each second semantic feature vector in the plurality of second semantic feature vectors; based on the plurality of second semantic features The cosine of the angle between the vector and the first semantic feature vector, calculating the second similarity between the first semantic feature vector and each second semantic feature vector in the plurality of second semantic feature vectors; based on the first similarity and at least one of the second similarities, determining the similarity between the first semantic feature vector and each second semantic feature vector in the plurality of second semantic feature vectors.

In the information retrieval method according to some embodiments of the present application, obtaining the first semantic feature vector corresponding to the first text information and the plurality of second semantic feature vectors respectively corresponding to the plurality of second text information includes: using semantic The understanding model determines the first semantic feature vector corresponding to the first text information; the plurality of second semantic feature vectors respectively corresponding to the plurality of second text information are obtained from the preset semantic feature vector index library, and the preset The plurality of second semantic feature vectors determined by using the semantic understanding model are stored in the semantic feature vector index library.

In the information retrieval method according to some embodiments of the present application, the third text information semantically related to the first text information is respectively extracted from the at least one text information to be retrieved to form a third text information set, including: for all For each text information to be retrieved in the at least one text information to be retrieved, the fourth text information indicating the candidate retrieval result corresponding to the first text information is determined from the text information to be retrieved by using a reading comprehension model; from each Extracting third text information including fourth text information from the text information to be retrieved; constructing a third text information set based on the third text information extracted from each text information to be retrieved.

In the information retrieval method according to some embodiments of the present application, extracting the third text information including the fourth text information from each text information to be retrieved includes one of the following steps: extracting the first text information from each text information to be retrieved The sentence where the four text information is located is used as the third text information; the natural paragraph where the fourth text information is extracted from each text information to be retrieved is used as the third text information; the fourth text information is extracted from each text information to be retrieved , as the third text information.

In the information retrieval method according to some embodiments of the present application, for each text information to be retrieved, the reading comprehension model is used to determine from the text information to be retrieved the candidate retrieval result corresponding to the first text information The fourth text information includes: performing the following steps for each text information to be retrieved: forming the first text information to be processed by splicing the first text information and the text information to be retrieved; performing word segmentation processing on the first text information to be processed To obtain the word segmentation sequence, the word segmentation sequence includes the first word segmentation sequence corresponding to the first text information and the second word segmentation sequence corresponding to the text information to be retrieved; input the word segmentation sequence into the reading comprehension model to obtain the second word segmentation sequence The first probability and the second probability corresponding to each participle in the sequence, the first probability corresponding to each participle indicates the probability that the participle is the beginning participle of the fourth text information, and the second probability corresponding to each participle Indicates the probability that the participle is the end participle of the fourth text information; according to the first probability and the second probability corresponding to each participle in the second participle sequence, determine the beginning of the fourth text information from the second participle sequence word segmentation and ending word segmentation; according to the beginning word segmentation and ending word segmentation of the fourth text information, determine the fourth text information from the text information to be retrieved.

In the information retrieval method according to some embodiments of the present application, obtaining multi-text summaries corresponding to at least two third text information in the third text information set includes: for each third text information in the third text information set For text information, according to at least one of the first probability corresponding to the start word and the second probability corresponding to the end word of the fourth text information included in the third text information, determine the retrieval matching degree corresponding to the third text information; According to the retrieval matching degree corresponding to each third text information in the third text collection, select at least two third text information from the third text collection; Splicing is performed in descending order of matching degrees to form second text information to be processed; a text summary model is used to generate a multi-text summary corresponding to the second text information to be processed.

In the information retrieval method according to some embodiments of the present application, for each third text information in the third text information set, according to the first At least one of the probability and the second probability corresponding to the end participle, and determining the retrieval matching degree corresponding to the third text information includes: determining the retrieval matching degree corresponding to the third text information based on at least one of the following values: The arithmetic mean of the first probability corresponding to the start participle and the second probability corresponding to the end participle; the geometric mean of the first probability corresponding to the start participle and the second probability corresponding to the end participle; The maximum value of the first probability corresponding to the participle and the second probability corresponding to the end participle; the minimum value of the first probability corresponding to the start participle and the second probability corresponding to the end participle.

In the information retrieval method according to some embodiments of the present application, determining the search result corresponding to the first text information based on the multi-text abstract includes: generating a search result corresponding to the first text information based on the multi-text abstract The first search result; based on the at least two third text information corresponding to the multi-text abstract, generate a second search result corresponding to the first text information; determine the search result according to the first search result and the second search result A search result corresponding to the first text information, so that the search result includes at least one of the first search result and the second search result.

In the information retrieval method according to some embodiments of the present application, according to the semantic similarity between the first text information and each second text information in the plurality of second text information, from the plurality of second text information Selecting at least one text information to be retrieved includes: sorting the plurality of second text information according to the descending order of the semantic similarity between the first text information and each second text information; Select the first M second text information as the M text information to be retrieved, where M is a preset positive integer.

In the information retrieval method according to some embodiments of the present application, the retrieval target includes a question to be retrieved, and the retrieval result includes an answer corresponding to the question to be retrieved.

According to another aspect of the present application, an information retrieval device based on semantic understanding is proposed, including: a first acquisition module configured to acquire first text information indicating a retrieval target and a plurality of second texts indicating candidate retrieval objects information; a first determination module configured to determine the semantic similarity between the first text information and each second text information in the plurality of second text information; a selection module configured to determine according to the first text information Semantic similarity between a text information and each second text information in the plurality of second text information, at least one text information to be retrieved is selected from the plurality of second text information; an extraction module is configured to obtain Extract third text information semantically related to the first text information from the at least one text information to be retrieved to form a third text information set; a second acquisition module configured to acquire the third text information set A multi-text abstract corresponding to at least two third text information; a second determining module configured to determine a search result corresponding to the first text information based on the multi-text abstract.

According to yet another aspect of the present application, there is provided a computing device comprising a memory and a processor, wherein the memory has stored therein a computer program which, when executed by the processor, causes the processor to perform according to The steps of the information retrieval method in some embodiments of the present application.

According to yet another aspect of the present application, there is provided a computer-readable storage medium on which computer-readable instructions are stored, and when executed, the computer-readable instructions implement the information retrieval method according to some embodiments of the present application.

According to another aspect of the present application, there is provided a computer program product comprising computer instructions which, when executed by a processor, implement the information retrieval method according to some embodiments of the present application.

In the information retrieval method and device based on semantic understanding according to some embodiments of the present application, through a three-stage information processing process such as information screening based on semantic similarity, information extraction based on semantic correlation, and multi-text summarization, efficiently and accurately Completed advanced information retrieval tasks such as intelligent question answering. Specifically, firstly, according to the semantic similarity between the first text information (such as the question to be retrieved) and the second text information (such as the candidate article used to obtain the answer to the question to be retrieved) (rather than a simple literal match) Screen a relatively small amount of text information to be retrieved from a large amount of second text information, thereby significantly improving the overall work efficiency without losing important information retrieval resources that are highly relevant to the retrieval problem (that is, ensuring retrieval breadth and accuracy). , and overcome the problem of lack of important retrieval resources caused by exact keyword matching in related technologies and the low efficiency caused by complex process and huge amount of calculation; secondly, for each text information to be retrieved in the first stage of screening, based on Semantic relevance re-extracts or retrieves the third text information corresponding to the first text information (such as the question to be retrieved) (that is, the text information related to the candidate answer corresponding to the question to be retrieved), so as to reuse the inherent semantic features of the question to be retrieved The semantic features of the information to be retrieved realize the extraction of candidate answers and their related texts (that is, the third text information set), which further ensures the high correlation between the third text information set and the first text information and the high accuracy of the final retrieval results. Accuracy; Finally, generate a multi-text abstract (such as a generative text abstract) of at least two third text information in the third text information set, as the final answer to the question to be retrieved, such a multi-text abstract is due to the fusion of multiple Candidate answer-related text information with a high retrieval matching degree can further improve the quality and accuracy of retrieval results.

These and other advantages of the present application will be apparent from and elucidated with reference to the embodiments described hereinafter.

Description of drawings

Embodiments of the present application will now be described in greater detail and with reference to the accompanying drawings, in which:

FIG. 1 shows an exemplary application scenario of an information retrieval method based on semantic understanding according to some embodiments of the present application;

Fig. 2 shows an exemplary functional block diagram of an information retrieval method based on semantic understanding according to some embodiments of the present application;

FIG. 3 shows a flowchart of an information retrieval method based on semantic understanding according to some embodiments of the present application;

Fig. 4 shows an example flow chart of the semantic similarity determination step in the information retrieval method based on semantic understanding according to some embodiments of the present application;

FIG. 5 shows a schematic diagram of determining semantic similarity using a semantic understanding model according to some embodiments of the present application;

6 shows an example flow chart of the steps of extracting semantically relevant information in the semantic understanding-based information retrieval method according to some embodiments of the present application;

FIG. 7 shows a schematic diagram of extracting semantically relevant information using a reading comprehension model according to some embodiments of the present application;

Fig. 8 shows an example flow chart of the multi-text abstract acquisition steps in the information retrieval method based on semantic understanding according to some embodiments of the present application;

FIG. 9 shows a schematic diagram of generating a multi-text summary using a text summary model according to some embodiments of the present application;

Fig. 10 is a schematic diagram of a complete process of an information retrieval method based on semantic understanding according to some embodiments of the present application;

Fig. 11 shows an exemplary structural block diagram of an information retrieval device based on semantic understanding according to some embodiments of the present application;

Figure 12 schematically illustrates an example block diagram of a computing device according to some embodiments of the application.

Detailed ways

Example embodiments will now be described more fully with reference to the accompanying drawings. Example embodiments may, however, be embodied in many forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided so that this application will be thorough and complete, and will fully convey the concept of example embodiments to those skilled in the art. The same reference numerals denote the same or similar parts in the drawings, and thus their repeated descriptions will be omitted.

Furthermore, the described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are provided in order to give a thorough understanding of the embodiments of the application. However, those skilled in the art will appreciate that the technical solutions of the present application may be practiced without one or more of the specific details, or other methods, components, devices, steps, etc. may be employed. In other instances, well-known methods, apparatus, implementations, or operations have not been shown or described in detail to avoid obscuring aspects of the application.

The block diagrams shown in the drawings are merely functional entities and do not necessarily correspond to physically separate entities. That is, these functional entities may be implemented in software, or in one or more hardware modules or integrated circuits, or in different networks and/or processor devices and/or microcontroller devices entity.

The flow charts shown in the drawings are only exemplary illustrations, and do not necessarily include all contents and operations/steps, nor must they be performed in the order described. For example, some operations/steps can be decomposed, and some operations/steps can be combined or partly combined, so the actual order of execution may be changed according to the actual situation. It will be understood that although the terms first, second, third etc. may be used herein to describe various components, these components should not be limited by these terms. As used herein, the term "and/or" and similar terms include all combinations of any one, a plurality, and all of the associated listed items.

Those skilled in the art can understand that the accompanying drawings are only schematic diagrams of exemplary embodiments, and the modules or processes in the accompanying drawings are not necessarily necessary for implementing the present application, and thus cannot be used to limit the protection scope of the present application.

Before introducing the embodiments of the present application in detail, some related concepts are firstly explained for the sake of clarity.

1. Text summarization refers to the extraction of key information or key information from one or more source texts (such as articles or article collections) through various techniques to summarize and display the main content or effective information of the source texts. In this paper, according to different input types, text summarization can be divided into single-text summarization (that is, the summation of a single text) and multi-text summarization (that is, the summarization of a text collection composed of multiple texts); according to different implementation technologies, text summarization It can be divided into extractive text summarization (for example, a summary composed of one or more fragments extracted directly from the source text) and generative text summarization (for example, a summary generated by understanding, summarizing, and reasoning about the source text).

2. Semantic understanding model, in this paper refers to the neural network model used for semantic understanding or semantic encoding of natural language text information, its input can be the serialized representation of the text to be processed (such as word segmentation sequence), and the output can be is the semantic feature vector corresponding to the text information; in this application, the semantic understanding model can use a pre-trained natural language processing model, such as BERT model, Roberta model, Albert model, etc.

3. Reading comprehension model, also known as machine reading comprehension model or question answering model, refers to a neural network model used to semantically understand natural language articles or corpus and answer related questions in the text, and its input can be the text of the question to be processed And the serialized representation of the corresponding text to be retrieved (such as an article), the output can be the answer or candidate answer corresponding to the text to be processed, or the probability of each word in the article as the beginning word and end word of the candidate answer. In this application, the reading comprehension model can also use a pre-trained natural language processing model, such as BERT model, Roberta model, Albert model, etc.

4. A text summarization model refers to a neural network model used to generate content summaries of input texts. Its input can be one or more texts, and its output can be corresponding text summaries. In this paper, the text summarization model can adopt a pre-trained encoder-decoder structure, such as the framework structure of (source text) sequence to (summary text) sequence based on deep neural network, where the encoder (such as semantic encoder) will The source text sequence is transformed into a corresponding sequence of semantic vectors, and a decoder (e.g. recurrent decoder) generates a summary text sequence based on the sequence of semantic vectors (e.g. via attention mechanism and recurrent decoding).

Aiming at the problems of complex retrieval process, huge amount of calculation, low efficiency, low retrieval breadth and low precision in related retrieval technology, this application proposes an information retrieval method based on semantic understanding. According to the information retrieval method of the present application, the analysis and comparison of the intrinsic semantic features of questions (i.e. the first text information, retrieval object) and articles (i.e. the second text information, retrieval object) have been fully utilized (replacing the external literals adopted by related technologies) or keyword matching), such as semantic similarity (semantic understanding model), semantic relevance (question-and-answer reading comprehension model), to screen and extract multiple candidate search results related text information from a large amount of original data to be retrieved (such as articles) (i.e. the third text information set, candidate answer-related information set); at the same time (for example, through the text summarization model) use the multi-text summarization automatic extraction technology to generate multi-text summaries based on at least a part of the relevant information of multiple candidate retrieval results, as the final retrieval result at least part of (such as an answer to a question).

First, perform a preliminary semantic screening of candidate retrieval objects (i.e., a large amount of original data or articles) through, for example, a semantic understanding model, that is, to screen out at least one piece of information to be retrieved that has a high semantic similarity with the query to be retrieved (i.e., a small amount of Articles with a high degree of relevance to the question), and because the articles with a high degree of relevance are retained at the semantic level, the breadth of retrieval resources is ensured and a retrieval basis is provided for retrieval accuracy. Secondly, extract or extract candidate retrieval results corresponding to the question to be retrieved (such as a candidate answer to the question) at the semantic level (that is, on the basis of semantic understanding) of at least one information to be retrieved from the preliminary screening by, for example, the reading comprehension model, thereby Forming the related text set of the candidate retrieval result (ie, the third text information set formed by the text containing the candidate retrieval result) further improves the accuracy of the retrieval result. Finally, perform multi-text abstraction extraction on at least a part of the texts in the relevant text set of the candidate retrieval results (for example, the text corresponding to the candidate retrieval results with high search matching degree) to obtain at least a part of the final retrieval results (such as the final answer to the question to be retrieved) , this kind of multi-text summarization can integrate the semantic features of multiple specific and high-matching candidate retrieval results or answers, which once again improves the robustness and accuracy of the retrieval results.

Fig. 1 shows an exemplary application scenario 100 of an information retrieval method based on semantic understanding according to some embodiments of the present application. As shown in FIG. 1 , the application scenario 100 may include a server 110 and may optionally include an external database 120 , a network 130 and a terminal device 140 , wherein the terminal device 140 may be controlled by a user 150 .

The information retrieval method according to some embodiments of the present application may be deployed on and implemented by the server 110 . The server 110 may be configured to: firstly, acquire first text information indicating a retrieval target and a plurality of second text information indicating candidate retrieval objects; secondly, determine each of the first text information and the plurality of second text information Semantic similarity of a second text information; Again, according to the semantic similarity between the first text information and each second text information in a plurality of second text information, select at least one text to be retrieved from a plurality of second text information information; then, extract third text information semantically related to the first text information from at least one text information to be retrieved to form a third text information set; subsequently, acquire at least two third text information in the third text information set The corresponding multi-text abstract; finally, based on the multi-text abstract, the retrieval result corresponding to the first text information is determined.

Exemplarily, server 110 may store and execute instructions that may perform various methods described herein. The server 110 can be a single server or a server cluster, or can provide cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and large Cloud servers or cloud server clusters for basic cloud computing services such as data and artificial intelligence platforms. It should be understood that the servers referred to herein may typically be server computers with substantial memory and processor resources, although other embodiments are possible. In addition, the server 110 is only shown as an example, in fact, other devices or combinations of devices having computing capabilities and storage capabilities may be used instead or additionally to provide corresponding services.

As shown in FIG. 1 , optionally, the application scenario 100 may further include an external database 120 and a network 130 . The server 110 can be connected to the external database 120 through the network 130, so that for example, the text to be processed can be obtained from the database 120, including the first text information indicating the retrieval target (such as the question to be retrieved) and the first text information indicating the candidate retrieval object (such as corresponding to the question to be retrieved). multiple second text information of candidate articles), and, for example, store the obtained retrieval results or question answers in the database 120 and so on. Exemplarily, the database 120 may be an independent data storage device or device group, or may also be a back-end data storage device or device group related to other online services (such as online services providing functions such as intelligent customer service and voice assistant).

As shown in FIG. 1 , optionally, the application scenario 100 may further include a terminal device 140 which may be connected to the server 110 through the network 130 . As shown in FIG. 1 , a user 150 of a terminal device 140 can access the server 110 via the network 130 through the terminal device 140 , so as to obtain services provided by the server 110 . For example, the user 150 can input instructions through the user interface provided by the terminal device 140, for example, through a physical input device (such as a keyboard and/or mouse, etc.) or a virtual button (such as a touch screen), through a voice or gesture command, etc., so as to start the deployment on the An information retrieval scheme on the server 110, sending first text information indicating a retrieval target or question (and/or a plurality of second text information indicating candidate retrieval objects), receiving obtained retrieval results or answers to questions, etc.

Illustratively, examples of network 130 include a local area network (LAN), a wide area network (WAN), a personal area network (PAN), and/or a combination of communication networks such as the Internet. Each of the server 110 , the database 120 and the terminal device 140 may include at least one communication interface (not shown) capable of communicating through the network 130 . Such communication interfaces may be one or more of the following: any type of network interface (e.g., a network interface card (NIC)), wired or wireless (such as IEEE 802.11 wireless LAN (WLAN)) wireless interface, global microwave Access Interoperability (Wi-MAX) interface, Ethernet interface, Universal Serial Bus (USB) interface, cellular network interface, Bluetooth interface, Near Field Communication (NFC) interface, etc.

As shown in FIG. 1 , terminal device 140 may be any type of mobile computing device, including mobile computers (e.g., personal digital assistants (PDAs), laptop computers, notebook computers, tablet computers, netbooks, etc.), mobile phones (e.g., , cellular phones, smartphones, etc.), wearable computing devices (such as smart watches, head-mounted devices, including smart glasses, etc.), or other types of mobile devices. In some embodiments, the terminal device 140 may also be a stationary computing device, such as a desktop computer, a game console, a smart TV, and the like. In addition, when the application scenario 100 includes multiple terminal devices 140, the multiple terminal devices 140 may be computing devices of the same or different types.

Exemplarily, the terminal device 140 may include a display screen and a terminal application that can interact with a terminal user via the display screen. The terminal application may be a local application program, a web page (Web) application program, or a small program as a lightweight application. In the case that the terminal application is a local application program that needs to be installed, the terminal application can be installed in the user terminal. In the case that the terminal application is a web application, the terminal application can be accessed through a browser. In the case that the terminal application is a small program, you can search for relevant information of the terminal application (such as the name of the terminal application, etc.), scan the graphic code of the terminal application (such as a barcode, two-dimensional code, etc.) Open the terminal application directly without installing the terminal application.

It should be understood that although the server 110, the database 120 and the terminal device 140 are shown and described as separate structures, they may also be different components of the same computing device. Among them, for example, the server 110 can provide background computing functions, the database 120 can provide data exchange, storage, and acquisition functions, and the terminal device 140 can provide foreground functions for interacting with users, such as receiving user input and providing output to users through terminal applications. Optionally, the information retrieval method based on semantic understanding according to some embodiments of the present application is not limited to being implemented on the server side shown in FIG. achieve together.

Fig. 2 shows an exemplary functional block diagram of an information retrieval method based on semantic understanding according to some embodiments of the present application. It should be noted that the rounded corner boxes in Figure 2 represent various data information, such as the first text information to be processed and multiple second text information, the text information to be retrieved, the third text information set, multi-text abstracts and retrieval results ; Rectangular boxes with square corners represent processing operations for various data information, including, for example, "information screening based on semantic similarity", "information extraction based on semantic correlation" and "acquisition of multi-text summaries" as shown in Figure 2. As shown in Figure 2, the information retrieval method according to some embodiments of the present application can utilize "three-stage" information processing operations (namely, "information screening based on semantic similarity", "information extraction based on semantic relevance", and "multi-text Abstract acquisition") to determine the exact retrieval result corresponding to the first text information (ie, the retrieval target or retrieval question).

As shown in FIG. 2 , in the information retrieval method based on semantic understanding according to some embodiments of the present application, firstly, in the first stage of information processing operation, the first text information 210 and multiple second text information 220 can be based on Semantic similarity among them (for example, carry out semantic similarity comparison by means of a semantic understanding model), carry out information screening on a plurality of second text information 220 (that is, a large amount of original candidate retrieval data), to select, for example, the same as the first text information 210 At least one text information 230 to be retrieved with high semantic similarity is used for the subsequent retrieval process; secondly, in the second stage of information processing operation, from each text information 230 to be retrieved (for example, by means of a reading comprehension model) or Extract the third text information related to the semantics of the first text information 210 (for example, the set of text information related to candidate retrieval results) to form the third set of text information 240; again, in the third information processing operation, for the third set of text information 240 (for example, by means of a text summarization model) to obtain a multi-text abstract 250 , and finally obtain a final retrieval result 260 according to the multi-text abstract 250 . It should be noted that the three-stage operations shown in Figure 2, including information screening, information extraction, and abstract acquisition, can all be implemented using corresponding neural network models. For example, the three-stage operations can use semantic understanding models, reading comprehension models, and text abstract models, respectively. However, in this application, this is not limiting. In other words, the above three-stage information processing operations can also be implemented in various ways other than the neural network model.

Fig. 3 is a flowchart of an information retrieval method based on semantic understanding according to some embodiments of the present application. The information retrieval method based on semantic understanding shown in FIG. 3 is implemented in the application scenario shown in FIG. 1 , and its execution subject may be the server 110 shown in FIG. 1 . Optionally, the information retrieval method according to some embodiments of the present application may also be executed on the terminal device 140 shown in FIG. 1 , or may also be jointly implemented by the server 110 and the target device 140 .

As shown in FIG. 3, the information retrieval method based on semantic understanding according to some embodiments of the present application may include the following steps:

S310, a step of acquiring text information;

S320, a step of determining semantic similarity;

S330, a step of selecting information to be retrieved;

S340, a semantic related information extraction step;

S350, a multi-text summary acquisition step;

S360, a step of determining the retrieval result.

The execution process of the above steps S310-S360 will be described in detail below in conjunction with FIG. 2 .

In step S310 (text information acquisition step), first text information indicating a retrieval target and a plurality of second text information indicating candidate retrieval objects are acquired.

In some embodiments, the retrieval goal refers to obtaining answers to questions to be retrieved (or to be processed) and/or information related to keywords to be retrieved through the information retrieval process. Therefore, the first text information may include (for example, input by the user) the information to be retrieved Text information corresponding to questions and/or search keywords. The second text information may include candidate search objects corresponding to the first text information (for example, articles corresponding to the question to be searched), that is, candidate search objects that may contain the answer to the question to be searched and/or information related to search keywords. The acquisition of text format data such as the first text information and the second text information is conducive to the smooth development of the subsequent data processing process, because various natural language processing models used in subsequent information screening, extraction, abstract acquisition, etc. The format of the input/output data is mostly text format, even serialized text format.

In some embodiments, the step of obtaining text information shown in S310 may include, for example, receiving from the terminal device the retrieval target or the question to be retrieved in text form input by the user as the first text information, and obtaining multiple Candidate retrieval data or articles are used as a plurality of second text information. Optionally, the reduction target or the question to be retrieved and the candidate retrieval data input by the user may also be other non-text form data (such as voice data, video data, table data, etc.), at this time, the text information acquisition step is receiving or obtaining these The non-text information needs to be converted (for example, using a suitable format conversion tool) into text information afterwards, so as to obtain the corresponding first text information and a plurality of second text information. In this paper, the questions to be retrieved can be various questions to be solved or processed, including but not limited to: (simple and intuitive) common sense questions, (complex and abstract) logical reasoning questions, numerical reasoning questions, etc.

In step S320 (semantic similarity determining step), the semantic similarity between the first text information and each second text information in the plurality of second text information is determined.

According to the idea of this application, in order to significantly reduce the amount of retrieval and improve retrieval efficiency while ensuring the breadth and accuracy of retrieval, a large amount of second text information ( That is, a relatively small amount of text information to be retrieved is screened out from the source data or candidate retrieval objects), so that the text information to be retrieved can be used as the retrieval object, and the retrieval result corresponding to the first text information (such as the question to be retrieved) can be retrieved from it.

In some embodiments, the semantic similarity of different text information can be determined by calculating the similarity of corresponding semantic features. In other words, the similarity of semantic features can be used to represent the intrinsic semantic similarity of different text information. Generally, for the needs of quantification, the semantic features of text information can be expressed in the form of vectors, that is, semantic feature vectors. The semantic feature vector of text information may refer to a vector extracted from the text information to represent its overall semantics. The semantic feature vector of text information can be obtained, for example, by using a semantic understanding model. In this way, the determination of the semantic similarity between the first text information and each second text information in multiple second text information can be converted into the relationship between the semantic feature vector of the first text information and the semantic feature vector of each second text information The calculation of the similarity. Optionally, other ways can also be used to calculate the semantic similarity, for example, the semantic (vector) sequence corresponding to the serialized representation of the first text information (such as a word segmentation sequence) and the serialized representation of each second text information ( For example, the comparison of the semantic (vector) sequence corresponding to the word segmentation sequence) to calculate the similarity between the two.

Since, in the process of determining the semantic similarity in step S320, the semantic features (vectors) of the first text information have covered the semantic information of each part (such as word segmentation or words) in the first text information and their interrelated information, The semantic features (vectors) of the second text information cover the semantic information and associated information of various parts (such as sentences, words, etc.) in the second text information, so the first text information and the second text information determined based on the semantic features Semantic similarity can indicate the similarity or closeness between the two at the intrinsic and essential semantic level, rather than the consistency or matching of simple, superficial, and partial text (such as keywords, partial paragraphs, or sentences in the overall text) Therefore, it can be significantly improved in terms of retrieval breadth and precision.

Taking Chinese as an example, since there are a large number of synonyms, even if there are no words in an article that are exactly the same as at least some of the words (such as keywords) in the retrieval question, the article may still be semantically different from the retrieval question. Similar or close, such a simple reliance on keyword matching will lose a large number of articles with similar semantics to the retrieval question, resulting in the lack of important retrieval resources and limited retrieval breadth; on the other hand, due to the existence of polysemy in Chinese, even if the literal The meaning of the same word may vary widely in different contexts, so keyword matching may not necessarily ensure that the required information matching the search question is found or retrieved, making it difficult to guarantee the accuracy of the search. In this application, information retrieval is realized by semantic comparison (that is, the calculation of semantic similarity) based on the first text information and the second text information. The defect of low precision is not high, and the preliminary screening of multiple second text information is realized efficiently.

In step S330 (the step of selecting information to be retrieved), at least one text to be retrieved is selected from the plurality of second text information according to the semantic similarity between the first text information and each second text information in the plurality of second text information information.

According to the concept of the present application, after determining the semantic similarity between the first text information and each second text information in the plurality of second text information, the actual retrieval can be screened from the plurality of second text information based on the semantic similarity Object, such as at least one text message to be retrieved. Generally, one or more second text information with higher semantic similarity with the first text information can be selected as the text information to be retrieved among multiple second text information, because the higher the semantic similarity, the corresponding second text information The closer the text information is to the first text information at the intrinsic semantic level. In other words, the second text information with higher semantic similarity with the first text information is more likely to be selected as the text information to be retrieved, wherein the number of selected text information to be retrieved can be preset. Compared with the original candidate retrieval objects or data (that is, multiple second text information), the number of text information to be retrieved (as the basis for subsequent retrieval) obtained through the screening operation based on semantic similarity is greatly reduced, thereby significantly reducing the Retrieval volume, improve retrieval efficiency.

In some embodiments, step S320 (the step of selecting information to be retrieved) may include: according to the descending order of semantic similarity between the first text information and each second text information, sorting the plurality of second text information Sorting; selecting the first M second text information from the sorting as M text information to be retrieved, where M is a preset positive integer. As mentioned above, in the screening process of information to be retrieved, the top M second text information with the highest similarity with the first text information can be selected as the retrieval basis or the actual retrieval object, where M can be preset according to the actual application scenario. For example, for a relatively complex or abstract problem to be retrieved (that is, the first text information), M can be set to a relatively large value to ensure the breadth and richness of the data to be retrieved, which is conducive to retrieval of more accurate Answer. On the other hand, in order to ensure retrieval efficiency, M cannot be set too large, for example, it can be set to a minimum value suitable for specific application scenarios.

In step S340 (semantic related information extraction step), third text information semantically related to the first text information is extracted from at least one text information to be retrieved to form a third text information set.

According to the concept of the present application, after the information preliminary screening of the first stage (that is, at least one text information to be retrieved as the basis of retrieval is selected from a plurality of second text information, such as articles to be retrieved), the second stage can be entered The process of information extraction, that is, to extract the third text information semantically related to the first text information from each information to be retrieved (that is, the candidate retrieval result related information corresponding to the first text information, such as the corresponding information extracted from the articles to be retrieved) Information related to candidate answers to questions to be processed), so that each of the third text information corresponding to each of the information to be retrieved forms a third text information set, which is used in the subsequent multi-text abstract acquisition process in the third stage, as a multi-text summary The basis for text summarization.

In some embodiments, the semantic related information extraction step S340 may be to extract a third text information semantically related to the first text information for each text information to be retrieved, so as to form a third text information set, and each third text information in the set The three pieces of text information can be in one-to-one correspondence with each text information to be retrieved. In this paper, the semantic correlation between the third text information and the first text information can be understood as that the third text information extracted from each text information to be retrieved is the candidate retrieval result related information corresponding to the first text information. For example, when the retrieval target indicated by the first text information is a question to be retrieved, the third text information may be related information of candidate answers matching the question extracted from the article to be retrieved (ie, the text information to be retrieved). In some embodiments, the third text information may be at least a part of the information extracted from the corresponding text information to be retrieved that contains the reduction result corresponding to the first text information, for example, the third text information may be a sentence in the article to be retrieved words, a paragraph, etc. Regarding the extraction method of the third text information, a (machine) reading comprehension model (such as a BERT-based neural network model, etc.) or a question-answering model can be used to implement. For example, first, input the first text information (such as a question) and the information to be retrieved (such as text) into the model, so as to obtain the answer to the question from the article; then, extract or extract the information containing the answer to the question (such as the answer to the question) from the article The sentence, natural paragraph or question answer itself), as the third text information. For the specific process, refer to FIG. 6 and FIG. 7 and their related descriptions.

As the information extraction step of the second stage, S340 is mainly used to further simplify the amount of information on the basis of the preliminary screening of information in the first stage, and use, for example, a reading comprehension model to screen out (that is, extract) relevant information of candidate search results from the information to be retrieved , as the basis of the subsequent third stage of multi-text summarization, so as to further reduce the amount of information processing in the retrieval process and improve the efficiency of data processing on the basis of ensuring the breadth and accuracy of the retrieval data.

In step S350 (multi-text abstract acquisition step), the multi-text abstract corresponding to at least two third text information in the third text information set is acquired.

According to the idea of the present application, after the first stage of information screening based on semantic similarity (that is, screening out at least one text information to be retrieved as the basis of retrieval from a plurality of second text information, such as articles to be retrieved) and the second After the semantic-related information extraction in the stage (that is, extracting the third text information from the information to be retrieved to form the third text information set), you can enter the third stage of the multi-text summary acquisition process, that is: first, from the third text information set Select at least two third text information (such as those with a high degree of relevance or retrieval matching with the first text information) from among them, and then use a text summarization model for natural language (such as a deep neural network-based encoder-decoder model) extract from at least two third text information or generate a multi-text abstract based on these third text information, as at least a part of the final retrieval result.

In some embodiments, the multi-text summarization involved in the information retrieval method according to the present application (ie, the text summarization obtained from at least two third texts in the third text collection) may be a generative text summarization. This kind of generative text summarization (such as generated by the text summarization model) can automatically and efficiently provide appropriate and accurate answers to more abstract and complex questions due to the process of semantic feature understanding, summarization, and reasoning, which significantly expands the The applicable scenarios and scope of the retrieval method improve the robustness and accuracy of the retrieval results and greatly enhance the user experience. Optionally, the multi-text summarization involved in step S350 may also be an extractive text summarization.

In some embodiments, a sequence-to-sequence language processing model can be used to obtain a multi-text summary, such as a model based on a deep neural network (such as RNN and/or CNN) encoder-decoder structure, wherein the input of the encoder part is In the third text information set, at least two third text information (with the first text information) have higher matching degrees or higher correlation degrees, and the output of the code part of the decoder is a multi-text summary result. The model running step may include: splicing at least two third text information according to the order of matching degree, and obtaining the input word segmentation sequence through word segmentation and inputting it into the encoder to obtain the semantic vector sequence corresponding to the input text information; however, using the attention mechanism Combining the input corresponding to each time step of the decoder to obtain the sequence of context vectors; finally, using the decoder to decode to obtain the decoded text sequence corresponding to each time step (or cyclic decoding moment), that is, a multi-text summary.

In step S360 (retrieval result determination step), the search result corresponding to the first text information is determined based on the multi-text abstract.

According to the concept of the present application, after obtaining at least two (highly relevant) multi-text abstracts of the third text information, the multi-text abstracts may be directly used as the final retrieval result or at least a part thereof. Since the basis of the multi-text summary is a plurality of third text information selected from the third text collection (that is, the relevant information collection of candidate retrieval results corresponding to the first text information) (with a high degree of correlation with the first text information), Therefore, the multi-text summarization integrates the content or intrinsic semantic features of the relevant information of multiple candidate retrieval results with high matching degree and the semantic correlation between them. Based on the related information of the result, the summary result is obtained by taking the essence of the overall information again, thereby further improving the matching accuracy of the final retrieval result (such as the answer to the question) and the first text information (such as the question to be processed).

In some embodiments, the step of determining the retrieval result may include: for example, presenting a multi-text abstract as a retrieval result corresponding to the first text information, or presenting as an information retrieval result. In some application scenarios, when the retrieval target includes a question to be retrieved, the retrieval result may include answer information corresponding to the question to be retrieved. In other words, the multi-text summarization can be directly used as the final answer information of the question to be processed. Due to the fusion of the semantic features and interrelated features of multiple candidate retrieval results (that is, multiple candidate answers corresponding to the question to be processed), such a multi-text summary can give direct and accurate answer information for the question information to be retrieved, which can Improve user experience and accuracy of search results.

In some embodiments, the retrieval result determination step S360 may also include: generating the first retrieval result corresponding to the first text information based on the multi-text abstract; generating the first text based on at least two third text information corresponding to the multi-text abstract The second search result corresponding to the information: according to the first search result and the second search result, determine the search result corresponding to the first text information, so that the search result includes at least one of the first search result and the second search result. For example, the search results may not only include the first search result based on the multi-text abstract as the (exact) main search result, but may also include at least two third text information based on the multi-text abstract (or the third text information set) generated second retrieval results, for example, a list of at least two third text information with high correlation with the first text information in the third text information set obtained in the information extraction extraction process of the second stage, as (optional ) auxiliary search results. Such a dual search result form enriches the diversity and selectivity of search results, and can meet the individual needs of different users. For example, when the user is dissatisfied with the precise retrieval result based on the multi-text abstract, he can refer to the optional auxiliary candidate retrieval result list to quickly query the corresponding information.

In the information retrieval method based on semantic understanding according to some embodiments of the present application, advanced information retrieval tasks such as intelligent question answering are efficiently and accurately completed through a three-stage information processing process, that is, the first stage: using a method based on (compared to The first text information (such as the question to be retrieved)) screening of the second text information (that is, the candidate retrieval object, such as the candidate article used to obtain the answer to the question to be retrieved) of the semantic similarity; the second stage: based on semantic correlation Extraction of a third set of text information (such as candidate answer-related texts); third stage: generation of multi-text summaries based on the third set of text information.

Specifically, firstly, according to the semantic similarity between the first text information and the second text information (rather than pure literal matching), a relatively small amount of text information to be retrieved is screened out from a large amount of second text information, so as to ensure that no In the case of losing important information retrieval resources that are highly relevant to the retrieval problem (that is, to ensure the breadth and accuracy of retrieval), the overall work efficiency is significantly improved, and the problem of missing important retrieval resources caused by exact keyword matching in related technologies and due to The problem of inefficiency caused by complex process and huge amount of calculation; secondly, for each text information to be retrieved that is screened in the first stage, based on semantic correlation, it is extracted or retrieved again corresponding to the first text information (such as the question to be retrieved). The third text information (that is, the relevant text information of the candidate answer corresponding to the question to be retrieved), so as to realize the candidate answer and its related text (ie, the first The extraction of three text information sets) further ensures the higher relevance of the third text information set and the first text information and the higher accuracy of the final retrieval result; finally, at least two of the third text information sets are generated in the third text information set. Multi-text summarization of three-text information (such as generative text summarization), that is, from multiple high-matching candidate answer-related texts through summarization and reasoning to generate multi-text summaries as the final answer to the question to be retrieved (such as the retrieval result Part of it), such a multi-text summary can further improve the quality and accuracy of the retrieval results due to the fusion of the semantic features and associated features of multiple candidate answer-related text information (ie, the third text information) with a high retrieval matching degree.

Fig. 4 shows an example flowchart of the semantic similarity determination steps of the information retrieval method according to some embodiments of the present application. Fig. 5 shows a schematic diagram of determining semantic similarity using a semantic understanding model according to some embodiments of the present application.

As shown in Fig. 4, the semantic similarity determining step S320 may include: S320a-S320c. Each step shown in FIG. 4 will be described in detail below with reference to FIG. 5 .

In step S320a, a first semantic feature vector corresponding to the first text information and a plurality of second semantic feature vectors corresponding to the plurality of second text information are acquired.

In order to quantify the semantic features of text information, semantic feature vectors can be used to represent the semantic features of text information, so that the similarity between two vectors (such as the distance between vectors and/or the cosine of the angle) can be used to describe each The semantic similarity between different text information corresponding to the vector. In this way, the determination of the semantic similarity between the first text information and each second text information in multiple second text information can be converted into the relationship between the semantic feature vector of the first text information and the semantic feature vector of each second text information The calculation of the similarity.

The acquisition of the semantic feature vector of text information can be realized by using the semantic understanding model for natural language processing. The semantic understanding model can also be called a semantic encoder, which can be a variety of pre-trained models for semantic understanding, such as BERT neural network, Roberta neural network, Albert neural network, etc., which are used to convert input text sequences (such as the first The word segmentation sequence corresponding to the text information) and/or the word vector sequence are converted into semantic feature vectors representing the overall semantics of the input text.

As shown in Figure 5, in the semantic feature vector acquisition stage (i.e., the semantic encoding stage), for the first text information 510 and each second text information 520, first, they are respectively input to the text preprocessor 530 for serialization processing (such as word segmentation processing) to obtain the first word segmentation sequence ([CLS], q(1), ..., q(k)) and the second word segmentation sequence corresponding to the first text information 510 and the second text information 520 respectively Sequence ([CLS], p(1), ..., p(m)), where [CLS] represents the text start tag character or global character, q(1)-q(k) represents the first text information ( For example, a total of k participle of the question to be processed), p(1)-p(m) represents a total of m participles of the second text information (such as an article), k and m can be positive integers greater than or equal to 1 (generally , m>k); Subsequently, the preprocessed first word sequence ([CLS], q(1),...,q(k)) and the second word sequence ([CLS], p(1) , ..., p(m)) are respectively input to the semantic understanding model 540, and through the process of semantic understanding or semantic encoding, the first semantic feature vector 550 corresponding to the first word segmentation sequence and the first semantic feature vector 550 corresponding to the second word segmentation sequence are respectively obtained. Two semantic feature vectors 560 .

As shown in Figure 5, the first word segmentation sequence ([CLS], q(1), ..., q(k)) is processed by each network layer in the neural network of the semantic understanding model 540, and the global character [CLS] corresponds to The output vector obtained by processing the neural network unit of the first text information 510 can represent the overall semantic feature of the first text information 510, and the output vector is the first word segmentation sequence or the first semantic feature vector 550 corresponding to the first text information 510; similarly, the second The word segmentation sequence ([CLS], p(1), ..., p(m)) is processed by each network layer of the semantic understanding model 540, and the output vector obtained by processing the neural network unit corresponding to the global character [CLS] can represent The overall semantic feature of the second text information 520 , the vector is the second word segmentation sequence or the second semantic feature vector 560 corresponding to the second text information 520 .

Exemplarily, the text preprocessor 530 shown in FIG. 5 may include a word segmentation tool to segment the text to be processed into one or more words, word segmentation sequences or text sequences. The word segmentation tool may be, for example, Jieba word segmentation, LTP, THULAC, NLPIR, etc. In addition to using word segmentation tools, optionally, the serialization process or word segmentation process of the text to be processed can also be achieved by manual labeling, random segmentation, complete segmentation into multiple single characters, and other methods. Generally, the models 540 for semantic understanding of the first text information 510 and the second text information 520 may be the same; 520 performs semantic understanding or semantic coding.

Optionally, if the input format of the semantic understanding model 540 is a vector sequence, the text preprocessor 530 needs to perform vectorization after performing word segmentation processing on the first and second text information to obtain the corresponding first and second word sequence processing to obtain the first and second word vector sequences; then the input of the first and second word vector sequences are respectively input to the semantic understanding model for semantic encoding. Exemplarily, the text preprocessor 530 may further include a word vector tool for converting text into word vectors, so as to implement vectorization processing for the word-segmented text sequence. Word embedding tools may include, for example, one-hot, word2vec, Glove, etc.

In some embodiments, step S320a may include: using the semantic understanding model to determine the first semantic feature vector corresponding to the first text information; The second semantic feature vector, wherein a plurality of second semantic feature vectors determined by using a semantic understanding model are stored in a preset semantic feature vector index library.

Since the retrieval objects such as articles (that is, the second text information) are generally relatively fixed known data stored in advance, such as in a server database, and the data volume of retrieval objects is often very large, therefore, in order to improve data processing Efficiency, it can be considered that the semantic understanding operation (that is, the acquisition or calculation of semantic feature vector or semantic code) of a large amount of second text information (that is, retrieval objects such as articles) is performed offline or offline in advance, and the obtained multiple A second semantic feature vector is stored in a preset semantic feature vector index library for online semantic similarity calculation (with the first text information such as the question to be retrieved) when information retrieval occurs. On the other hand, the semantic understanding or semantic coding process of the first text information can be carried out online after starting the information retrieval, so that in the information retrieval process, it is only necessary to directly call the corresponding ( Offline pre-determined) second semantic feature vector, without the need to use the semantic understanding model to perform semantic coding operations on a huge amount of second text information online in real time, thus (especially in the case of large data volumes of objects or articles to scenario) greatly improves data processing efficiency and significantly optimizes network resource allocation and scheduling.

In step S320b, the similarity between the first semantic feature vector and each of the plurality of second semantic feature vectors is calculated.

As mentioned above, the semantic similarity between different text information can be characterized by the similarity between the corresponding semantic feature vectors. Therefore, as shown in FIG. 5, after obtaining the first semantic feature vector 550 corresponding to the first text information 510 and the second semantic feature vector 560 corresponding to each second text information 520, it is necessary to calculate the vector similarity between the two degree in order to determine the semantic similarity 570 between the first text information 510 and the second text information 520 .

In some embodiments, the similarity between two vectors can be reflected in the similarity in direction between the two vectors, such as cosine similarity; it can also be reflected in the closeness of the distance between the vectors, that is, similarity based on distance. Specifically, step S320b may include: calculating the first semantic feature vector and each second semantic feature in the multiple second semantic feature vectors based on the distance between the multiple second semantic feature vectors and the first semantic feature vector The first similarity of the vector; based on the cosine of the angle between the multiple second semantic feature vectors and the first semantic feature vector, calculate the first semantic feature vector and each second semantic feature in the multiple second semantic feature vectors The second similarity of the vectors: based on at least one of the first similarity and the second similarity, determine the similarity between the first semantic feature vector and each second semantic feature vector.

For example, when calculating the distance between two semantic feature vectors, the smaller the distance, the more similar the two semantic feature vectors are, that is, the greater the vector similarity; conversely, the larger the distance, the less similar the two semantic feature vectors are , that is, the smaller the vector similarity is. For example, when calculating the cosine of the angle between two semantic feature vectors, the smaller the angle, the more similar the two semantic feature vectors are, that is, the greater the vector similarity; conversely, the larger the angle, the more semantic The less similar the eigenvectors are, the smaller the vector similarity is. Therefore, the similarity between the first semantic feature vector and the second semantic feature vector can be calculated solely based on the vector distance or the included angle cosine alone, or can be based on both the vector distance and the included angle cosine (that is, the first similarity and the second The weighted sum of the two similarities) calculates the similarity between the first semantic feature vector and the second semantic feature vector. Whether it is to calculate the distance between multiple second semantic feature vectors and the first semantic feature vector or to calculate the cosine of the angle between multiple second semantic feature vectors and the first semantic feature vector, since these two calculation methods are relatively simple Moreover, the semantic feature vector can accurately represent the semantics, so it not only meets the needs of precise retrieval, but also improves retrieval efficiency and saves retrieval time. The two characteristics complement each other, which is very beneficial for scenarios where there is a large amount of second text information.

In some embodiments, the distance between the first semantic feature vector and the second semantic feature vector may include Euclidean distance, Manhattan distance, Chebyshev distance and the like. Optionally, besides cosine similarity and distance-based similarity, other methods may also be used to calculate the similarity between the first semantic feature vector and the second semantic feature vector, such as correlation coefficient or similarity coefficient.

In step S320c, according to the similarity between the first semantic feature vector and each second semantic feature vector, the semantic similarity between the first text information and each second text information in the plurality of second text information is determined.

Exemplarily, after obtaining the similarity between the first semantic feature vector and each second semantic feature vector, the similarity between the first semantic feature vector and each second semantic feature vector can be directly used as the semantic Similarity; Optionally, appropriate data processing can also be performed on the similarity between semantic feature vectors (such as increasing or reducing the degree of discrimination of the similarity between semantic feature vectors, performing normalization operations, etc.), The processed data is then used as semantic similarity.

Fig. 6 shows an example flowchart of the steps of extracting semantically related information in the information retrieval method according to some embodiments of the present application. Fig. 7 shows a schematic diagram of extracting semantic related information by using a reading comprehension model according to some embodiments of the present application.

As shown in FIG. 6 , the semantic related information extraction step S340 may include: S340a-S340c. Each step shown in FIG. 6 will be described in detail below with reference to FIG. 7 .

In step S340a, for each text information to be retrieved, a reading comprehension model is used to determine fourth text information indicating a candidate retrieval result corresponding to the first text information.

According to the idea of the present application, in the second stage of the semantic related information extraction step S340, it is necessary to extract the relevant information of the candidate retrieval results from each text information to be retrieved, that is, the third text information, so as to form the third text information set, as The basis for subsequent multi-text summarization. Exemplarily, each third text information in the set of third text information (that is, information related to candidate retrieval results) may be limited to be extracted or extracted from the corresponding information to be retrieved, including, for example, the candidate retrieval results corresponding to the first text information. The text information of the result (that is, the fourth text information). For example, when the first text information represents a question to be processed, the fourth text information may represent a candidate answer corresponding to the question to be processed, and the third text information may represent an article fragment related to the candidate answer. Therefore, in order to obtain the third text information set, it is necessary to first determine the corresponding fourth text information (i.e., candidate retrieval results) from each information to be retrieved, and then extract the third text information (i.e., candidate retrieval results) that contains the candidate retrieval results. Related Information).

In some embodiments, the reading comprehension model refers to a neural network model used for semantically understanding natural language articles or corpus and answering related questions. Its input can be the first text information (such as the text of the question to be processed) and the The serialized representation of text information (such as the corresponding article to be retrieved), the output can be candidate retrieval results (such as the answer or candidate answer corresponding to the text to be processed), or the first word corresponding to each word in the word segmentation sequence corresponding to the text information to be retrieved Probability and second probability, wherein the first probability represents the probability that the participle is the beginning participle of the fourth text information indicating the candidate retrieval result, and the second probability corresponding to each participle represents that the participle is the fourth text information The probability of ending participle of

In some embodiments, step S340a may include performing the following steps for each text information to be retrieved:

forming first text information to be processed by splicing the first text information and the text information to be retrieved;

performing word segmentation processing on the first text information to be processed to obtain a word segmentation sequence, the word segmentation sequence including a first word segmentation sequence corresponding to the first text information and a second word segmentation sequence corresponding to the text information to be retrieved;

Inputting the word segmentation sequence into the reading comprehension model to obtain the first probability and the second probability corresponding to each word segmentation in the second word segmentation sequence, the first probability corresponding to each word segmentation indicates that the word segmentation is the fourth text information The probability of starting the word segmentation, and the second probability corresponding to each word segmentation indicates the probability that the word segmentation is the ending word segmentation of the fourth text information;

According to the first probability and the second probability corresponding to each word segment in the second word segment sequence, determine the start word segment and the end word segment of the fourth text information from the second word segment sequence;

The fourth text information is determined from the text information to be retrieved according to the start participle and the end participle of the fourth text information.

As shown in FIG. 7 , in the process of extracting the fourth text information (ie, candidate retrieval results) from each information to be retrieved, first, the first text information 710 (such as a question to be retrieved) and the corresponding information to be retrieved 720 ( For example, an article) is input into the text preprocessor 730, so as to perform preprocessing operations such as splicing and serialization on it, and transform it into the input format required by the reading comprehension model 750, that is, the text or word segmentation sequence 740. As shown in FIG. 7 , the word segmentation sequence 740 sequentially includes: the start marker "[CLS]" of the text information to be processed, the first word segmentation sequence 740a corresponding to the first text information 710 (that is, q(1), ..., q(k)), text delimiter "[SEP]", second word segmentation sequence 740b (ie, p(1), . . . , p(m)). Subsequently, the word segmentation sequence 740 is input into the reading comprehension model 750, and after the processing and calculation of the multi-layer neural network in the model, the corresponding values of each word segmentation p(1), ..., p(m) in the second word segmentation sequence 740b can be output The first probability and the second probability, wherein the first probability represents that the participle is the probability that the participle is the beginning participle of the fourth text information (such as a candidate answer to the question) indicating the candidate retrieval result, and the second probability represents that the participle is the participle of the fourth text information Probability of ending participle. Thereafter, according to the first probability and the second probability, the start participle and the end question are determined, for example, the respective first probabilities and If the largest first probability and the largest second probability are found in the second probability, then their respective corresponding participle can be regarded as the start participle and the end participle respectively. As shown in Figure 7, for example, among the first probability and the second probability corresponding to p(1)-p(m), the first probability corresponding to p(m ₁ ) is the largest, and the second probability corresponding to p(m ₂ ) The probability is the largest, and m ₁ <m ₂ , then it can be determined that the m ₁ th word segment p(m ₁ ) in the text information to be retrieved 720 is the beginning word segment of the fourth text information, and the m ₂ th word segment p in the text information 720 to be retrieved is (m ₂ ) is the end participle of the fourth text information. Finally, the fourth text information can be determined based on the initial word segmentation p(m ₁ ) and p(m ₂ ), that is, the word segmentation between the m ₁ th word segment and the m ₂ th word segment (including p(m ₁ ) and p(m ₂ )) is determined as the fourth text information, which is used as the output text of the reading comprehension model.

Optionally, when the serial number m ₁ of the participle p(m ₁ ) corresponding to the largest first probability is greater than or equal to the serial number m ₂ of the participle p(m ₂ ) corresponding to the largest second probability, that is m ₁ >= m ₂ , the following two methods can be adopted to determine the start participle and end participle of the fourth text information:

The first type: firstly, the participle p(m ₁ ) corresponding to the largest first probability in the second participle sequence is used as the fourth text information to start participle, and then in each participle after the participle p(m ₁ ) in the second participle sequence Find the participle p(m ₃ ) with the highest second probability (that is, m ₃ >m ₁ ), as the end participle of the fourth text information;

The second type: firstly, the participle p(m ₂ ) corresponding to the largest second probability in the second participle sequence is used as the end participle of the fourth text information, and then it is ranked in each participle before the participle p(m ₂ ) in the second participle sequence The word segment p(m ₄ ) with the highest first probability (that is, m ₄ <m ₂ ) is searched for as the start word segment of the fourth text information.

Optionally, the start and end word segmentations of the fourth text information can also be determined in the following manner: First, among all word segmentation pairs in the second word segmentation sequence 740b, determine the first probability corresponding to the preceding word segmentation and the following word segmentation The sum of the corresponding second probabilities, and then select the participle pair with the largest sum of probabilities, wherein the preceding participle is used as the starting participle, and the post participle is used as the ending participle. In addition, the start participle and the end participle of the fourth text information may also be determined in other ways.

In step S340b, third text information including fourth text information is extracted from each text information to be retrieved.

Based on the idea of this application, after determining the fourth text information (i.e. candidate retrieval results) from each information to be retrieved, the third text information containing the fourth text information (i.e. candidate retrieval results) can be extracted from the information to be retrieved. Result-related information), and then the third text information extracted from each of the information to be retrieved can form a third text information set.

As shown in FIG. 7 , after obtaining the fourth text information 760 , third text information 770 including the fourth text information 760 may be directly extracted from the information to be retrieved 720 . For example, the third text information 770 may be expressed as text information corresponding to a word sequence (..., p(m ₁ ),..., p(m ₂ ),...), that is, it contains (m ₁ ) and the fourth text information 760 defined by the ending participle p(m ₂ ).

In some embodiments, step S340b may include one of the following steps: extract the sentence where the fourth text information is located from each text information to be retrieved as the third text information; extract the fourth text information from each text information to be retrieved The natural paragraph where the text information is located is used as the third text information; the fourth text information is extracted from each text information to be retrieved as the third text information. In other words, the third text information may be a natural sentence, a natural paragraph, or the fourth text information itself in the corresponding information to be retrieved. Optionally, the fourth text information may also be other forms of text information that include the third text information in the corresponding information to be retrieved.

In step S340c, based on the third text information extracted from each text information to be retrieved, a third text information set is constructed.

After the third text information including the fourth text information is extracted from each information to be retrieved, each third text information extracted from each information to be retrieved can be used as an element to form a third text information set. Optionally, a part of the third text information may also be selected from each third text information to form the third text information set. For example, it is possible to select a plurality of third text information with a higher matching degree or correlation with the selected retrieval of the first text information to form a set, where the retrieval matching degree or correlation degree can be determined by the fourth text information corresponding to the third text information. At least one of the first probability and the second probability corresponding to the start participle and the end participle is represented. By selecting part of the third text information according to the retrieval matching degree to form the third text information set, the amount of data processing can be further reduced while the retrieval accuracy is ensured, and the retrieval efficiency can be improved.

In the embodiment shown in Figure 6 and Figure 7, by using the reading comprehension model to analyze the text information to be retrieved one by one and extract the semantically related information (that is, the third text information), it is possible to find the information for the first text information from each text information to be retrieved. Candidate retrieval results (that is, fourth text information, such as candidate answers) corresponding to a text information (such as a question to be retrieved), and extract the third text information including the candidate retrieval results. This process is also based on the semantic understanding and analysis of the first text information and the text information to be retrieved, so it can be ensured that the retrieved candidate retrieval results and information related to the candidate retrieval results (that is, the third text information) are consistent with the first text information at the intrinsic semantic level. A higher degree of relevance and retrieval matching of text information. In addition, the reading comprehension model gives each participle the first probability and the second probability of being the start participle and the end participle, which provides a flexible extraction method for determining the third text information, thus being able to adapt to different information retrieval scenarios.

Fig. 8 shows an example flow chart of the steps of obtaining multi-text abstracts in the information retrieval method according to some embodiments of the present application. Fig. 9 shows a principle diagram of generating a multi-text summary by using a text summary model according to some embodiments of the present application.

As shown in FIG. 8 , the step S350 of obtaining a multi-text summary may include steps S350a-S350d, and the above steps will be described in detail below with reference to FIG. 9 .

In step S350a, for each third text information in the third text information set, at least One, determining the retrieval matching degree corresponding to the third text information.

According to the idea of this application, in the multi-text abstract acquisition operation in the third stage of the information retrieval process, the basic information of the multi-text abstract (that is, the object to be abstracted) can be obtained from the third text information set (that is, the relevant information of the candidate retrieval results) select. In order to properly reduce the amount of data and calculation while ensuring the accuracy of retrieval, several third text information with a high degree of correlation with the first text information or a high degree of retrieval matching can be selected from the third text information set as the third text information. The basis for the summary fetching of the stage. In this paper, retrieval matching degree can refer to the semantic dimension of the third text information (i.e. information related to candidate retrieval results) or the corresponding fourth text information (i.e. candidate retrieval results, e.g. question answers) with the first text information (e.g. The degree of matching between retrieval questions). A higher search matching degree indicates that the candidate search result corresponding to the third text information matches the search target or the question to be searched corresponding to the first text information, or the candidate search result is more accurate.

In some embodiments, the search matching degree of each third text information in the third text information set can be based on the first probability corresponding to the start word and the end word corresponding to the fourth text information included in the third text information. at least one of the two probabilities, because the size of the first probability and the second probability reflects the determined fourth text information (that is, the candidate retrieval result) or the third text information and the first text information (that is, the retrieval target) to a certain extent. ) degree of matching or correlation.

Specifically, step S3 50a (for each third text information in the third text information set, according to the first probability corresponding to the start participle and the second probability corresponding to the end participle of the fourth text information included in the third text information At least one, determining the retrieval matching degree corresponding to the third text information) may include determining the retrieval matching degree corresponding to the third text information based on at least one of the following values:

(1) the arithmetic mean of the first probability corresponding to the start participle and the second probability corresponding to the end participle;

(2) the geometric mean of the first probability corresponding to the start participle and the second probability corresponding to the end participle;

(3) the maximum value in the first probability corresponding to the start participle and the second probability corresponding to the end participle;

(4) The minimum value of the first probability corresponding to the start participle and the second probability corresponding to the end participle.

Any one or more of the above four values may be used to determine the retrieval matching degree of the third text information or the fourth text information, so as to adapt to different situations. In addition, the four determination methods provided all have a small amount of calculation, can save retrieval time, and can also effectively determine an accurate retrieval matching degree.

Optionally, the manner of determining the retrieval matching degree of the third text information is not limited to the above four manners, but other manners may also be used, for example, based on the semantic similarity between the third text information and the first text information.

In step S350b, at least two pieces of third text information are selected from the third text set according to the retrieval matching degree corresponding to each third text information in the third text set.

After the retrieval matching degree corresponding to each third text information in the third text set is determined, based on the retrieval matching degree, a plurality of third text information with higher retrieval matching degree can be selected from the third text set as the multi-text summary object. In some embodiments, S3 50b may include: according to the descending order of the retrieval matching degree between the first text information and each third text information, perform a search on each third text information in the third text information set Sorting: selecting the first N third text messages from the sorting, where N is a preset positive integer greater than 2. In this way, the top N third text information with the highest matching degree with the first text information retrieval can be selected as the basis or object of the multi-text abstract, where N can be preset according to the actual application scenario. For example, for relatively complex or abstract problems to be retrieved (ie, the first text information), N can be set to a relatively large value to ensure the breadth and richness of the summarized data in the process of multi-text summarization, which is beneficial to More accurate retrieval results or answers to questions can be retrieved; on the other hand, in order to ensure the efficiency of the summarization process, N cannot be set too large. For example, it can be set to be suitable for specific application scenarios while ensuring retrieval accuracy minimum value.

In step S3 50c, at least two pieces of third text information are spliced according to their corresponding retrieval matching degrees from high to low, so as to form second text information to be processed.

In some embodiments, the acquisition of multi-text summaries can be accomplished by using a text summarization model, so before inputting into the text summarization model, the data of summarization objects (that is, a plurality of third text information) needs to be preprocessed to fit the model input format . Generally, the data preprocessing process of the abstract object may include a splicing process of multiple third text information, for example, splicing at least two third text information together in order of retrieval matching degree to obtain the second text information to be processed . Optionally, the data preprocessing process of the summary object can also include the serialization process of the second text information to be processed to match the input format of the text summary model (assuming that the text summary model is a text sequence or word segmentation sequence), for example, the second to be processed Process the text information for word segmentation to obtain the corresponding word sequence.

In step S350d, a text summary model is used to generate a multi-text summary corresponding to the second text information to be processed.

In some embodiments, the text summarization model can adopt a pre-trained encoder-decoder structure, such as a framework structure from (source text) sequence to (summary text) sequence based on a deep neural network. As shown in Figure 9, the text summarization model can include an encoder 910 and a decoder 920, wherein the encoder 910 (such as a semantic encoder) converts the source text sequence into a corresponding semantic vector sequence, and the decoder 920 (such as a cyclic decoder) A summary text sequence is then generated based on the sequence of semantic vectors (e.g. via attention mechanisms and recurrent decoding). Specifically, the deep neural network adopted by the text summarization model may include a convolutional neural network (CNN) and a recurrent neural network (RNN), wherein the encoder 910 and the decoder 920 may adopt the same or different one or more neural network structures , as long as the corresponding input and output functions can be realized.

Exemplarily, the process of using the text summarization model to generate a multi-text summarization of the second text information to be processed is described below with reference to FIG. 9 . As shown in Figure 9, the first is the preprocessing stage, that is, the second text information to be processed 930 is input to the text preprocessor 940 so as to serialize it, so as to obtain the word segmentation sequence or text corresponding to the second text information to be processed 930 Sequence (T ₁ , T ₂ , T ₃ ,..., T _n ), where T _i (i=1, 2,..., n) represents the i-th text in the text sequence, and n represents the text sequence in Chinese total number of copies.

Subsequently, it enters the encoding stage, that is, the preprocessed text sequence (T ₁ , T ₂ , T3, ..., T _n ) is input to the encoder 910 of the text summarization model for semantic understanding, and the text Sequence (T ₁ , T ₂ , T ₃ ,..., T _n ) one-to-one correspondence semantic vector sequence (H ₁ , H ₂ , H ₃ ,..., H _n ), where H _i (i=1 , 2,...,n) represents the i-th semantic vector in the sequence of semantic vectors, which corresponds to the i-th text or word segment T _i in the text sequence, and n represents the total number of semantic vectors in the sequence of semantic vectors. Here, the semantic vector sequence (H ₁ , H ₂ , H ₃ , . . . , H _n ) can also be called the hidden state vector sequence of the encoder 910 . As shown in Figure 9, the semantic coding stage can use a bidirectional cyclic neural network to realize the semantic coding of text sequences, which can make semantic coding more integrated between different texts or semantic associations between word segments (such as forward association and reverse Association) features, thereby more accurately reflecting the overall semantic features of the second text information to be processed.

Next, the multi-text summarization process enters the decoding stage. In some embodiments of the present application, the decoding stage adopts a cyclic decoding process based on multiple time steps, wherein the number of decoding time steps or decoding moments can be predetermined according to specific application scenarios and the like. The entire decoding process of the decoder 920 can be summarized as follows: First, using the attention mechanism, the semantic vector sequence (H ₁ , H ₂ , H ₃ , ..., H _n ) output by the encoder 910 is transformed into a content vector sequence ( C ₁ , C ₂ , C ₃ ,..., C _r ), wherein r represents the number of decoding moments corresponding to the decoder 920; subsequently, the decoder 920 is used (for example, a recurrent neural network structure) to perform cyclic decoding at each decoding moment To obtain the decoded text sequence (A ₁ , A ₂ , A ₃ , . . . , A _r ) as a sequence of multi-text summaries, and finally obtain the final multi-text summaries by splicing.

FIG. 9 shows the decoding process of the t-th decoding moment or decoding time step in the decoding stage of the multi-text summary. As shown in Figure 9, at the current decoding moment of the decoding stage of the multi-text abstract, that is, the t-th decoding moment, first, the output decoding The hidden state vector S _t-1 of the device and the semantic vector sequence (H ₁ , H ₂ , H ₃ , ..., H _n ) determine the attention distribution 950 corresponding to the current decoding moment, which can also be called the attention weight sequence. As shown in FIG. 9 , each black column in the attention distribution 950 can represent the attention weight corresponding to each semantic vector in the semantic vector sequence (H ₁ , H ₂ , H ₃ , . . . , H _n ), For example, the longer the bar, the greater the corresponding attention weight. In this way, subsequently, as shown in FIG. 9 , the weighted sum can be calculated for each semantic vector of the semantic vector sequence (H ₁ , H ₂ , H ₃ , . . . , H _n ) based on the attention weight sequence or attention respectively 950, To obtain the content vector C _t of the current decoding moment (that is, the tth decoding moment). Afterwards, in the decoder 920, according to the content vector C _t at the current decoding moment, the hidden state vector S _t-1 of the decoder at the previous decoding moment (that is, the t-1th decoding moment), and the output of the decoder 920 at the previous decoding moment The decoded text A _t-1 of the decoder determines the hidden state vector S _t of the decoder at the current decoding moment; then in the decoder 920, according to the content vector C _t of the current decoding moment (that is, the t-th decoding moment), the current decoding moment The decoder hidden state vector S _t and the decoded text A _t-1 output by the decoder 920 at the previous decoding time determine the (final) decoded text A _t corresponding to the current decoding time.

As shown in Figure 9, in the process of determining the (final) decoded text A _t corresponding to the current decoding moment, the decoder 920 (such as classification prediction function) can be used to predict the vocabulary probability distribution 960 corresponding to the current decoding moment, that is, all possible The output text or vocabulary becomes the probability distribution of the final decoded text _At . The vocabulary probability distribution 960 shown in FIG. 9 is shown in the form of a plurality of histograms, each of which represents the probability that a vocabulary becomes the decoded text at the current moment; obviously, the higher the histogram, the corresponding vocabulary becomes the corresponding one at the current decoding moment. The more likely it is to decode the text. Therefore, as shown in FIG. 9 , the vocabulary or text corresponding to the highest histogram in the vocabulary probability distribution 960 may be selected as the decoded text A _t corresponding to the current decoding moment (ie, the t-th decoding moment).

In the decoding phase of the multi-text summary acquisition process according to some embodiments of the present application, the attention mechanism can give weights to each semantic vector in the semantic vector sequence, so as to distinguish the content vector conversion and subsequent The degree of criticality in the decoding process, thereby improving the efficiency and accuracy of the neural network. For example, in the decoding stage of the text summarization model, each vector in the input semantic vector sequence corresponds to the word segmentation corresponding to the second text information to be processed, but the degree of association between these word segmentation and the problem to be processed is different, so through the attention weight (or attention distribution) to reflect the importance of different semantic vectors in the decoding process at the current decoding moment, thereby improving the efficiency and accuracy of the model prediction reasoning process.

In addition, in the decoding stage of the multi-text summary acquisition process according to some embodiments of the present application, during the calculation process of the decoded text at the current decoding moment (that is, the t-th decoding moment), the input of the decoder 920 not only includes the current decoding moment The content vector C _{t of} and the decoding vector S _t-1 outputted at the previous decoding moment (that is, the t-1th decoding moment) and including the decoded text A _t-1 of the previous decoding moment, so that the text summarization model The input of the decoder 920 incorporates more different types of relevant features (such as the context features corresponding to A _t-1 ), which is beneficial to the final decoded text A _t at the current decoding moment to more accurately reflect the overall semantics of the second text information to be removed Features, especially contextual semantic relationship features or association features. This enables the final output of decoded text sequences and/or multi-text summaries to more precisely match the problem at hand.

Fig. 10 is a schematic diagram of the entire process of the information retrieval method according to some embodiments of the present application. In the embodiment shown in FIG. 10 , the first text information is a question to be retrieved, the second text information is an article, the information to be retrieved is an article to be retrieved, the third text information is an article fragment, and the third text information is a candidate answer. As shown in FIG. 10 , the upper part of the dotted line indicates the online operation process, and the lower part of the dotted line indicates the offline operation process.

As shown in Figure 10, in the offline situation, use the semantic understanding model to determine candidate retrieval objects, that is, the semantic feature vectors of multiple articles, and then build a semantic feature vector index library locally based on these semantic feature vectors, or use these semantic feature vectors Stored in the preset semantic feature vector index library, it can be called directly during online retrieval. In this way, in the scenario of a large amount of articles to be retrieved, the data processing efficiency is greatly improved, and the network resource scheduling is significantly optimized.

Then, as shown in Figure 10, after the question to be retrieved is obtained, the online information retrieval starts, and the online information retrieval process can be divided into three stages: article screening, fragment extraction and text summarization. As shown in Figure 10, in the article screening stage, firstly, the query to be retrieved is input into the semantic understanding model to determine the first semantic feature vector corresponding to the query to be retrieved; Call the second semantic feature vector corresponding to each article and calculate the similarity between the first semantic feature vector and each second semantic feature vector; then, sort the articles according to their corresponding semantic similarity, and select the top M articles As an article to be retrieved, M is a positive integer. As shown in Figure 10, in the fragment extraction stage, the reading comprehension model is used to extract article fragments containing candidate answers that match the questions to be retrieved from the M articles screened in the previous stage. In the text summarization stage, as shown in Figure 10, the text summarization model is used to obtain or generate multi-text summaries corresponding to multiple article fragments selected from M article fragments (for example, the first N article fragments with high matching degree), and Present multi-text summaries to users as final search results or answers to questions.

In the embodiment shown in FIG. 10 , the entire information retrieval process is divided into two parts, online and offline, and the two complement each other to achieve accurate and efficient retrieval of answers to questions. On the one hand, since the retrieval objects such as articles (that is, the second text information) are generally relatively fixed known data pre-stored in, for example, a server database, and the data volume of retrieval objects is often very large, therefore, for To improve the efficiency of data processing, the semantic understanding operation (that is, the acquisition or calculation of semantic feature vectors or semantic codes) of a large amount of second text information (that is, retrieval objects such as articles) is performed offline or offline in advance, and the obtained A plurality of second semantic feature vectors are stored in a preset semantic feature vector index library for online semantic similarity calculation (with the first text information such as the question to be retrieved) when information retrieval occurs. This avoids using the semantic understanding model to perform semantic coding operations on a huge amount of second text information online in real time, thereby (especially in the scenario of large data volumes of objects or articles to be retrieved) greatly improving data processing efficiency and significantly Optimize network resource configuration and scheduling. On the other hand, regarding the online or online operation part, three natural language processing models based on deep neural networks (namely, semantic understanding model, reading comprehension model, and text summarization model) are used to realize article screening, fragment extraction, and multi-text summarization Data processing such as generation effectively guarantees the robustness and accuracy of retrieval results or retrieval answers.

Fig. 11 is an exemplary structural block diagram of an information retrieval device 1100 based on semantic understanding according to some embodiments of the present application. As shown in FIG. 11 , the information retrieval apparatus 1100 may include: a first acquisition module 1110 , a first determination module 1120 , a selection module 1130 , an extraction module 1140 , a second acquisition module 1150 and a second determination module 1160 .

The first acquiring module 1110 may be configured to acquire first text information indicating a retrieval target and a plurality of second text information indicating candidate retrieval objects. The first determination module 1120 may be configured to determine the semantic similarity between the first text information and each second text information in the plurality of second text information. The selecting module 1130 may be configured to select at least one text information to be retrieved from the plurality of second text information according to the semantic similarity between the first text information and each second text information in the plurality of second text information. The extracting module 1140 may be configured to respectively extract third text information semantically related to the first text information from at least one text information to be retrieved to form a third text information set. The second obtaining module 1150 may be configured to obtain multi-text abstracts corresponding to at least two third text information in the third text information set. The second determining module 1160 may be configured to determine the retrieval result corresponding to the first text information based on the multi-text abstract.

It should be noted that the various modules described above may be implemented in software or hardware or a combination of both. Several different modules can be realized in the same software or hardware structure, or one module can be realized by several different software or hardware structures.

In the information retrieval device according to some embodiments of the present application, firstly, according to the semantic similarity between the first text information and the second text information (rather than pure literal matching), a relatively small amount of second text information is screened out from a large amount of text information. The text information to be retrieved, thus significantly improving the overall work efficiency without losing important information retrieval resources that are highly relevant to the retrieval problem (that is, ensuring the breadth and accuracy of retrieval), and overcoming the problem of exact keyword matching in related technologies The problem of lack of important retrieval resources and the low efficiency caused by the complex process and huge amount of calculation; secondly, for each text information to be retrieved that is screened in the first stage, based on the semantic correlation, it is extracted or retrieved again. The third text information corresponding to the text information (such as the question to be retrieved) (that is, the text information related to the candidate answer corresponding to the question to be retrieved), so as to realize the candidate answer and The extraction of its relevant text (that is, the third text information set) further ensures the higher relevance of the third text information set and the first text information and the higher accuracy of the final retrieval result; finally, the third text information set is generated A multi-text summary (such as a generative text summary) of at least two third text information in the , that is, from a plurality of high-matching candidate answer-related texts, for example, a multi-text summary is generated by summarizing and reasoning as the final answer to the question to be retrieved (For example, as a part of the retrieval result), such a multi-text summary can further improve the quality and accuracy of the retrieval result because it incorporates multiple text information (ie, the third text information) related to the candidate answer with a high retrieval matching degree.

FIG. 12 schematically illustrates an example block diagram of a computing device 1200 according to some embodiments of the application. The computing device 1200 may represent a device to implement various means or modules described herein and/or perform various methods described herein. Computing device 1200 may be, for example, a server, desktop computer, laptop computer, tablet, smart phone, smart watch, wearable device, or any other suitable computing device or computing system, which may include resources from a full range of storage and processing resources Various classes of devices from resource devices to low resource devices with limited storage and/or processing resources. In some embodiments, the information retrieval apparatus 1100 based on semantic understanding described above with respect to FIG. 11 may be implemented in one or more computing devices 1200 .

As shown in FIG. 12 , example computing device 1200 includes a processing system 1201 , one or more computer-readable media 1202 , and one or more I/O interfaces 1203 communicatively coupled to each other. Although not shown, computing device 1200 may also include a system bus or other data and command transfer system that couples the various components to each other. A system bus can include any one or combination of different bus structures, such as a memory bus or memory controller, a peripheral bus, a universal serial bus, and/or a processing bus utilizing any of a variety of bus architectures device or local bus. Alternatively, such as control and data lines may also be included.

The processing system 1201 represents functionality to perform one or more operations using hardware. Accordingly, processing system 1201 is illustrated as including hardware elements 1204 that may be configured as processors, functional blocks, or the like. This may include implementing in hardware as an application specific integrated circuit or other logic device formed using one or more semiconductors. Hardware element 1204 is not limited by the material from which it is formed or the processing mechanism employed therein. For example, a processor may be comprised of semiconductor(s) and/or transistors (eg, electronic integrated circuits (ICs)). In this context, processor-executable instructions may be electronically-executable instructions.

Computer readable medium 1202 is illustrated as including memory/storage 1205 . Memory/storage 1205 represents memory/storage associated with one or more computer-readable media. Memory/storage 1205 may include volatile media (such as random access memory (RAM)) and/or non-volatile media (such as read only memory (ROM), flash memory, optical disks, magnetic disks, etc.). Memory/storage 1205 may include fixed media (eg, RAM, ROM, fixed hard drives, etc.) as well as removable media (eg, flash memory, removable hard drives, optical disks, etc.). Exemplarily, the memory/storage device 1205 may be used to store data such as first text information, second text information, third text information, and search results mentioned in the above embodiments. Computer-readable medium 1202 may be configured in various other ways as described further below.

One or more I/O (input/output) interfaces 1203 represent interfaces that allow a user to enter commands and information into computing device 1200 and also allow information to be displayed to the user and/or sent to other components or devices using various input/output devices. Function. Examples of input devices include keyboards, cursor control devices (e.g., mice), microphones (e.g., for voice input), scanners, touch capabilities (e.g., capacitive or other sensors configured to detect physical touch), cameras (e.g., For example, visible or invisible wavelengths (such as infrared frequencies to detect motion not involving touch as gestures), network cards, receivers, etc. may be employed. Examples of output devices include display devices, speakers, printers, tactile response devices, network cards, transmitters, and the like.

Computing device 1200 also includes information retrieval policies 1206 . Information retrieval strategy 1206 may be stored as computer program instructions in memory/storage device 1205, or may be hardware or firmware. The information retrieval strategy 1206 can implement all the functions of the various modules of the information retrieval device 1100 based on semantic understanding described in relation to FIG. 11 together with the processing system 1201 .

Various techniques may be described herein in the general context of software, hardware, components or program modules. Generally, these modules include routines, programs, objects, elements, components, data structures, etc. that perform particular tasks or implement particular abstract data types. As used herein, the terms "module", "function" and the like generally represent software, firmware, hardware or a combination thereof. The features of the techniques described herein are platform-independent, meaning that the techniques can be implemented on a variety of computing platforms with a variety of processors.

An implementation of the described modules and techniques may be stored on or transmitted across some form of computer readable media. Computer-readable media may include a variety of media that may be accessed by computing device 1200 . By way of example, and not limitation, computer readable media may include "computer readable storage media" and "computer readable signal media."

A "computer-readable storage medium" refers to a medium and/or device capable of persistently storing information, and/or a tangible storage device, as opposed to a mere signal transmission, carrier wave, or signal itself. Thus, computer-readable storage media refers to non-signal bearing media. Computer-readable storage media include media such as volatile and nonvolatile, removable and non-removable media and/or media suitable for storing information such as computer-readable instructions, data structures, program modules, logic elements/circuits, or ) method or technology to implement hardware such as storage devices. Examples of computer readable storage media may include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, Digital Versatile Disk (DVD) or other optical storage device, hard disk, cassette tape, magnetic tape, disk storage device or other magnetic storage device, or other storage device, tangible medium, or article of manufacture suitable for storing desired information and accessible by a computer.

"Computer-readable signal medium" refers to a signal-bearing medium configured as hardware to transmit instructions to computing device 1200 , such as via a network. Signal media typically embody computer readable instructions, data structures, program modules or other data in a modulated data signal such as a carrier wave, data signal or other transport mechanism. Signal media also includes any information delivery media. By way of example, and not limitation, signal media include wired media such as a wired network or direct-wire, and wireless media such as acoustic, RF, infrared and other wireless media.

As previously described, hardware elements 1204 and computer-readable media 1202 represent instructions, modules, programmable device logic, and/or fixed device logic implemented in hardware that, in some embodiments, may be used to implement the techniques described herein. At least some aspects. Hardware elements may include integrated circuits or systems on a chip, application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), complex programmable logic devices (CPLDs), and other implementations in silicon or components of other hardware devices. In this context, a hardware element may serve as a processing device that performs program tasks defined by the instructions, modules, and/or logic embodied by the hardware element, as well as a hardware device for storing instructions for execution, e.g., as previously described computer readable storage media.

Combinations of the foregoing can also be used to implement the various techniques and modules described herein. Thus, software, hardware or program modules and other program modules may be implemented as one or more instructions and/or logic on some form of computer readable storage medium and/or embodied by one or more hardware elements 1204 . The computing device 1200 may be configured to implement specific instructions and/or functions corresponding to software and/or hardware modules. Thus, a module may be implemented at least in part in hardware as a module executable by the computing device 1200 as software, eg, by using the computer-readable storage medium of the processing system and/or the hardware element 1204 . Instructions and/or functions may be executed/operable by, for example, one or more computing devices 1200 and/or processing system 1201 to implement the techniques, modules, and examples described herein.

The techniques described herein may be supported by these various configurations of computing device 1200 and are not limited to the specific examples of the techniques described herein.

In particular, according to the embodiments of the present application, the processes described above with reference to the flowcharts can be implemented as computer programs. For example, embodiments of the present application provide a computer program product, which includes a computer program carried on a computer-readable medium, where the computer program includes program code for executing at least one step in the method embodiments of the present application.

In some embodiments of the present application, one or more computer-readable storage media are provided, on which are stored computer-readable instructions that, when executed, implement semantic-based Understand information retrieval methods. Each step of the information retrieval method based on semantic understanding according to some embodiments of the present application can be converted into computer-readable instructions through programming, and stored in a computer-readable storage medium. When such a computer-readable storage medium is read or accessed by a computing device or a computer, the computer-readable instructions therein are executed by a processor on the computing device or computer to implement the method according to some embodiments of the present application.

In the description of this specification, descriptions of the terms "one embodiment", "some embodiments", "example", "specific examples", or "some examples" mean specific features described in conjunction with the embodiment or examples, A structure, material or characteristic is included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms are not necessarily directed to the same embodiment or example. Furthermore, the described specific features, structures, materials or characteristics may be combined in any suitable manner in any one or more embodiments or examples. In addition, those skilled in the art can combine and combine different embodiments or examples and features of different embodiments or examples described in this specification without conflicting with each other.

Any process or method description in a flow diagram or otherwise described herein may be understood to represent a module, segment or portion of code comprising one or more executable instructions for implementing the steps of a custom logical function or process, and that the scope of preferred embodiments of the present application includes additional implementations in which functions may be performed out of the order shown or discussed, including substantially concurrently or in the reverse order depending on the functions involved, which should be considered Those skilled in the art to which the embodiments of the present application belong can understand.

The logic and/or steps represented in the flowcharts or otherwise described herein, for example, can be considered as a sequenced listing of executable instructions for implementing logical functions, can be embodied in any computer-readable medium, For use with instruction execution systems, devices, or devices (such as computer-based systems, systems including processors, or other systems that can fetch instructions from instruction execution systems, devices, or devices and execute instructions), or in conjunction with these instruction execution systems, devices or equipment used. For the purposes of this specification, a "computer-readable medium" may be any device that can contain, store, communicate, propagate or transmit a program for use in or in conjunction with an instruction execution system, device or device.

It should be understood that each part of the present application may be realized by hardware, software, firmware or a combination thereof. In the embodiments described above, various steps or methods may be implemented by software or firmware stored in memory and executed by a suitable instruction execution system. For example, if it is implemented in hardware, it can be implemented by any one or combination of the following technologies known in the art: discrete logic circuits with logic gates for implementing logic functions on data signals; Application-specific integrated circuits of logic gate circuits, programmable gate arrays (Programmable Gate Array), field programmable gate arrays (Field Programmable Gate Array), etc.

Those of ordinary skill in the art can understand that all or part of the steps of the methods in the above embodiments can be completed by hardware related to program instructions, and the program can be stored in a computer-readable storage medium. When the program is executed, it includes: One or a combination of the steps of the method embodiment.

In addition, each functional unit in each embodiment of the present application may be integrated into one processing module, each unit may exist separately physically, or two or more units may be integrated into one module. The above-mentioned integrated modules can be implemented in the form of hardware or in the form of software function modules. If the integrated modules are implemented in the form of software function modules and sold or used as independent products, they can also be stored in a computer-readable storage medium.

Claims (17)

1. An information retrieval method based on semantic understanding, comprising:

acquiring first text information indicating a search target and a plurality of second text information indicating candidate search objects;

determining semantic similarity of the first text information and each of the plurality of second text information;

selecting at least one text message to be searched from the plurality of second text messages according to the semantic similarity between the first text message and each of the plurality of second text messages;

respectively extracting third text information semantically related to the first text information from the at least one text information to be searched to form a third text information set;

acquiring multiple text summaries corresponding to at least two pieces of third text information in the third text information set;

and determining a search result corresponding to the first text information based on the multi-text abstract.

2. The method of claim 1, wherein the multi-text summary comprises generating a formula multi-text summary.

3. The method of claim 1 or 2, wherein the determining the semantic similarity of the first text information to each of the plurality of second text information comprises:

Acquiring a first semantic feature vector corresponding to the first text information and a plurality of second semantic feature vectors corresponding to the plurality of second text information respectively;

calculating the similarity between the first semantic feature vector and each of the plurality of second semantic feature vectors;

and determining the semantic similarity of the first text information and each of the plurality of second text information according to the similarity of the first semantic feature vector and each of the second semantic feature vectors.

4. The method of claim 3, wherein the computing the similarity of the first semantic feature vector to each of the plurality of second semantic feature vectors comprises:

calculating a first similarity of the first semantic feature vector and each of the plurality of second semantic feature vectors based on the distances of the plurality of second semantic feature vectors from the first semantic feature vector;

calculating a second similarity of the first semantic feature vector and each of the plurality of second semantic feature vectors based on cosine of an included angle between the plurality of second semantic feature vectors and the first semantic feature vector;

And determining a similarity of the first semantic feature vector to each of the plurality of second semantic feature vectors based on at least one of the first similarity and the second similarity.

5. The method of claim 3, wherein the obtaining the first semantic feature vector corresponding to the first text information and the plurality of second semantic feature vectors corresponding to the plurality of second text information respectively comprises:

determining a first semantic feature vector corresponding to the first text information by using a semantic understanding model;

and acquiring a plurality of second semantic feature vectors corresponding to the plurality of second text information respectively from a preset semantic feature vector index library, wherein the plurality of second semantic feature vectors determined by the semantic understanding model are stored in the preset semantic feature vector index library.

6. The method according to claim 1 or 2, wherein the extracting third text information semantically related to the first text information from the at least one text information to be retrieved, respectively, to form a third set of text information, comprises:

for each piece of text information to be searched in the at least one piece of text information to be searched, determining fourth text information indicating a candidate search result corresponding to the first text information from the piece of text information to be searched by utilizing a reading understanding model;

Extracting third text information containing fourth text information from each text information to be retrieved;

a third set of textual information is constructed based on third textual information extracted from each of the textual information to be retrieved.

7. The method of claim 6, wherein the extracting third text information including fourth text information from each text information to be retrieved comprises one of:

extracting sentences in which fourth text information is located from each text information to be searched as third text information;

extracting a natural paragraph in which the fourth text information is located from each text information to be searched as third text information;

and extracting fourth text information from each text information to be retrieved as third text information.

8. The method of claim 6, wherein the determining, for each text message to be retrieved, fourth text information indicating a candidate retrieval result corresponding to the first text information from the text messages to be retrieved using the reading understanding model comprises: the following steps are performed for each text message to be retrieved:

forming first text information to be processed by splicing the first text information and the text information to be retrieved;

Performing word segmentation on the first text information to be processed to obtain a word segmentation sequence, wherein the word segmentation sequence comprises a first word segmentation sequence corresponding to the first text information and a second word segmentation sequence corresponding to the text information to be retrieved;

inputting the word segmentation sequence into a reading understanding model to obtain a first probability and a second probability corresponding to each word segmentation in the second word segmentation sequence, wherein the first probability corresponding to each word segmentation represents the probability that the word segmentation is the beginning word segmentation of the fourth text information, and the second probability corresponding to each word segmentation represents the probability that the word segmentation is the ending word segmentation of the fourth text information;

determining beginning word segmentation and ending word segmentation of the fourth text information from the second word segmentation sequence according to the first probability and the second probability corresponding to each word segmentation in the second word segmentation sequence;

and determining fourth text information from the text information to be retrieved according to the beginning segmentation and ending segmentation of the fourth text information.

9. The method of claim 8, wherein the obtaining the multiple text summaries corresponding to at least two third text information in the third set of text information comprises:

for each third text message in the third text message set, determining a retrieval matching degree corresponding to the third text message according to at least one of a first probability corresponding to a start word segmentation and a second probability corresponding to an end word segmentation of fourth text message contained in the third text message;

Selecting at least two third text messages from the third text sets according to the retrieval matching degree corresponding to each third text message in the third text sets;

splicing the at least two third text messages according to the sequence from high to low of the corresponding retrieval matching degree so as to form second text messages to be processed;

and generating a multi-text abstract corresponding to the second text information to be processed by using a text abstract model.

10. The method of claim 9, wherein the determining, for each third text message in the third text message set, the search matching degree corresponding to the third text message according to at least one of a first probability corresponding to a start word segmentation and a second probability corresponding to an end word segmentation of fourth text message included in the third text message, includes: determining the retrieval matching degree corresponding to the third text information based on at least one of the following values:

an arithmetic average of the first probability corresponding to the beginning word segmentation and the second probability corresponding to the ending word segmentation; a geometric average of the first probability corresponding to the beginning segmentation and the second probability corresponding to the ending segmentation; maximum values in the first probability corresponding to the beginning word segmentation and the second probability corresponding to the ending word segmentation; and the minimum value of the first probability corresponding to the beginning word segmentation and the second probability corresponding to the ending word segmentation.

11. The method according to claim 1 or 2, wherein the determining, based on the multi-text summary, a search result corresponding to the first text information includes:

generating a first retrieval result corresponding to the first text information based on the multi-text abstract;

generating a second search result corresponding to the first text information based on the at least two third text information corresponding to the multi-text abstract;

and determining a search result corresponding to the first text information according to the first search result and the second search result, so that the search result comprises at least one of the first search result and the second search result.

12. The method according to claim 1 or 2, wherein the selecting at least one text information to be retrieved from the plurality of second text information according to the semantic similarity of the first text information to each of the plurality of second text information comprises:

sorting the plurality of second text messages according to the sequence from the big semantic similarity of the first text message to each second text message;

and selecting the first M pieces of second text information from the ordering as M pieces of text information to be searched, wherein M is a preset positive integer.

13. The method according to claim 1 or 2, wherein the search target includes a question to be searched, and the search result includes an answer corresponding to the question to be searched.

14. An information retrieval apparatus based on semantic understanding, comprising:

a first acquisition module configured to acquire first text information indicating a search target and a plurality of second text information indicating candidate search objects;

a first determination module configured to determine a semantic similarity of the first text information to each of the plurality of second text information;

a selection module configured to select at least one text message to be retrieved from the plurality of second text messages according to semantic similarity of the first text message to each of the plurality of second text messages;

an extraction module configured to extract third text information semantically related to the first text information from the at least one text information to be retrieved, respectively, to form a third text information set;

the second acquisition module is configured to acquire multiple text summaries corresponding to at least two third text messages in the third text message set;

And a second determining module configured to determine a search result corresponding to the first text information based on the multi-text summary.

15. A computing device, comprising:

a memory and a processor, wherein the memory is configured to store,

wherein the memory has stored therein a computer program which, when executed by the processor, causes the processor to perform the method of any of claims 1-13.

16. A computer readable storage medium having stored thereon computer readable instructions which, when executed, implement the method of any of claims 1-13.

17. A computer program product comprising a computer program which, when executed by a processor, implements the steps of the method according to any of claims 1-13.

Images