CN116340502A

CN116340502A - Information retrieval method and device based on semantic understanding

Info

Publication number: CN116340502A
Application number: CN202310331474.3A
Authority: CN
Inventors: 王寒; 石智中; 梁霄; 雷涛
Original assignee: China International Financial Ltd By Share Ltd
Current assignee: China International Financial Ltd By Share Ltd
Priority date: 2023-03-30
Filing date: 2023-03-30
Publication date: 2023-06-27

Abstract

An information retrieval method and device based on semantic understanding are disclosed. The information retrieval method comprises the following steps: acquiring first text information and a plurality of second text information; determining semantic similarity of the first text information and each piece 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 second text message; respectively extracting third text information semantically related to the first text information from 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. According to the information retrieval method, advanced information retrieval tasks such as intelligent question and answer can be efficiently and accurately completed through data processing operations such as information screening, extraction, abstract and the like.

Description

Information retrieval method and device based on semantic understanding

Technical Field

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

Background

With the rapid development of the Internet, more and more information can be retrieved from the Internet through a search engine, and the search result has the characteristics of data sea quantity, diversified forms, comprehensive coverage and the like. The method improves the possibility that the user searches the results, and the user can be out of favor in the face of massive search results, so that accurate answers can not be obtained in a short time. For example, conventional search engines based on keyword matching and single document abstracts are limited to returning a list of web pages or documents related to a user's search question, but cannot give an accurate answer to the question (the user needs to find or obtain an answer to the question from the related web pages or documents in combination with information such as a title and an abstract), and cannot meet the needs and expectations of quickly acquiring information.

As users' expectations for search engines increase, the modality of information retrieval begins to transition from a primary modality, such as recall of a basic related web page or document list, to a superior modality, such as intelligent question-and-answer retrieval. The purpose of intelligent question-and-answer retrieval is to answer the user's questions in a compact, accurate natural language, the advent of which has been focused on providing more efficient information acquisition tools. To implement advanced information retrieval modalities such as intelligent question-answering, intelligent information retrieval techniques based on semantic understanding or machine reading understanding have been developed. However, the related art information retrieval method based on semantic understanding or machine reading understanding has the following problems: firstly, the search mode based on keyword or character string matching in the related technology can only search articles with the same characters, can not search information with different characters but the same semantics, is easy to cause the deletion of important information resources highly related to the search problem, and has limited search result breadth and low accuracy; secondly, the related art retrieval mode based on keyword extraction and comparison causes larger calculation amount and workload due to the use of more complex preset rules, has lower efficiency, and the keyword can not completely and accurately reflect the characteristics of the whole retrieval problem, so that the accuracy of the retrieval result is not high.

Disclosure of Invention

In view of this, the present application provides a semantic understanding based information retrieval method and apparatus, computing device, computer readable storage medium and computer program product, which desirably mitigate or overcome some or all of the above-mentioned drawbacks and other possible drawbacks.

According to a first aspect of the present application, there is provided an information retrieval method based on semantic understanding, including: 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 of 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.

In information retrieval methods according to some embodiments of the present application, the multi-text summary includes generating a multi-text summary.

In an information retrieval method according to some embodiments of the present application, determining a semantic similarity of the first text information and each of the plurality of second text information includes: 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 second text information in the plurality of second text information according to the similarity of the first semantic feature vector and each second semantic feature vector.

In an information retrieval method according to some embodiments of the present application, calculating a similarity of a first semantic feature vector to each of the plurality of second semantic feature vectors includes: 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.

In the information retrieval method according to some embodiments of the present application, obtaining 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 includes: 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.

In an information retrieval method according to some embodiments of the present application, 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 text information set, including: 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.

In an information retrieval method according to some embodiments of the present application, extracting third text information including fourth text information from each text information to be retrieved includes one of the following steps: 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.

In an information retrieval method according to some embodiments of the present application, for each text information to be retrieved, determining fourth text information indicating a candidate retrieval result corresponding to the first text information from the text information to be retrieved using the reading understanding model includes: 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.

In the information retrieval method according to some embodiments of the present application, obtaining multiple text summaries corresponding to at least two third text information in the third text information set includes: 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.

In the information retrieval method according to some embodiments of the present application, for each third text information in the third text information set, determining, according to at least one of a first probability corresponding to a start word of the fourth text information and a second probability corresponding to an end word of the fourth text information included in the third text information, a 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: 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.

In the information retrieval method according to some embodiments of the present application, determining, based on the multi-text abstract, a retrieval 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.

In the information retrieval method according to some embodiments of the present application, selecting 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 of the plurality of second text information, including: 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.

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, there is provided an information retrieval apparatus based on semantic understanding, including: 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 and 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.

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 the steps of an information retrieval method according to some embodiments of the present application.

According to yet another aspect of the present application, there is provided a computer readable storage medium having stored thereon computer readable instructions which, when executed, implement an 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 an 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, advanced information retrieval tasks such as intelligent question-answering are efficiently and accurately completed through three-stage information processing procedures such as semantic similarity-based information screening, semantic relevance-based information extraction, multi-text summarization and the like. Specifically, firstly, a relatively small amount of text information to be searched is screened from a large amount of second text information according to the semantic similarity (instead of mere literal matching) of the first text information (such as a problem to be searched) and the second text information (such as a candidate article for acquiring an answer to the problem to be searched), so that the overall working efficiency is remarkably improved under the condition of ensuring that important information search resources with high relevance to the search problem are not lost (namely, the search breadth and accuracy are ensured), and the problems of important search resource deletion caused by accurate keyword matching and inefficiency caused by complex flow and huge calculation amount of the related technology are overcome; secondly, extracting or searching third text information (namely candidate answer related text information corresponding to the to-be-searched question) corresponding to the first text information (such as the to-be-searched question) again based on semantic relevance aiming at each to-be-searched text information obtained by screening in the first stage, so that the extraction of the candidate answer and related text (namely third text information set) thereof is realized by utilizing the semantic features in the to-be-searched question and the semantic features of the to-be-searched information again, and the higher relevance of the third text information set and the first text information and the higher accuracy of a final search result are further ensured; and finally, generating multiple text summaries (such as generated text summaries) of at least two third text information in the third text information set as final answers of the questions to be searched, wherein the multiple text summaries can further improve the quality and accuracy of the search results due to the fact that a plurality of candidate answer related text information with higher search matching degree are fused.

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

Drawings

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

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

FIG. 2 illustrates an exemplary functional block diagram of a semantic understanding based information retrieval method according to some embodiments of the present application;

FIG. 3 illustrates a flow chart of a semantic understanding based information retrieval method according to some embodiments of the present application;

FIG. 4 illustrates an example flowchart of semantic similarity determination steps in a semantic understanding based information retrieval method according to some embodiments of the present application;

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

FIG. 6 illustrates an example flowchart of semantic related information extraction steps in a semantic understanding based information retrieval method according to some embodiments of the present application;

FIG. 7 illustrates a schematic diagram of extracting semantically related information using a reading understanding model according to some embodiments of the present application;

FIG. 8 illustrates an example flowchart of multiple text excerpt retrieval steps in a semantic understanding based information retrieval method according to some embodiments of the present application;

FIG. 9 illustrates a schematic diagram of generating multiple text summaries using a text summary model in accordance with some embodiments of the present application;

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

FIG. 11 illustrates an exemplary block diagram of a semantic understanding based information retrieval apparatus according to some embodiments of the present application;

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

Detailed Description

Example embodiments will now be described more fully with reference to the accompanying drawings. However, the exemplary embodiments can 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 disclosure will be thorough and complete, and will fully convey the concept of the example embodiments to those skilled in the art. The same reference numerals in the drawings denote the same or similar parts, and thus a repetitive description thereof 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 to give a thorough understanding of embodiments of the present application. One skilled in the relevant art will recognize, however, that the aspects of the application can be practiced without one or more of the specific details, or with other methods, components, devices, steps, etc. In other instances, well-known methods, devices, implementations, or operations are not shown or described in detail to avoid obscuring aspects of the application.

The block diagrams shown in the figures are merely functional entities and do not necessarily correspond to physically separate entities. That is, the 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.

The flow diagrams depicted in the figures are exemplary only, and do not necessarily include all of the elements and operations/steps, nor must they be performed in the order described. For example, some operations/steps may be decomposed, and some operations/steps may be combined or partially combined, so that the order of actual execution may be changed according to actual situations. 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, many, and all of the associated listed items.

Those skilled in the art will appreciate that the drawings are schematic representations of example embodiments, and that the modules or flows in the drawings are not necessarily required to practice the present application, and therefore, should not be taken to limit the scope of the present application.

Before describing embodiments of the present application in detail, some related concepts will be explained first for clarity.

1. Text summarization refers to extracting key information or gist information from one or more source texts (e.g., articles or article sets) by various techniques to summarize and present the main content or valid information of the source texts. Herein, text summaries can be divided into single text summaries (i.e., summaries of a single text) and multiple text summaries (i.e., summaries of a set of texts made up of multiple texts) according to the type of input; depending on implementation techniques, the text summaries may be divided into a decimated text summary (e.g., a summary of one or more segments extracted directly from the source text) and a generated text summary (e.g., a summary generated via understanding, summarizing, reasoning about the source text).

2. A semantic understanding model, herein referred to as a neural network model for performing semantic understanding or semantic encoding on text information in natural language, may be input as a serialized representation (e.g., word segmentation sequence) of text to be processed, and may output as a semantic feature vector corresponding to the text information; in this application, the semantic understanding model may employ a pre-trained natural language processing model, such as a BERT model, a robeta model, an Albert model, and the like.

3. The reading understanding model, also called machine reading understanding model or question and answer model, refers to a neural network model used for carrying out semantic understanding on articles or corpora in natural language and answering related questions, and the input of the neural network model can be a serialization representation of a text of a to-be-processed question and a corresponding text to be searched (for example, an article), and the output can be an answer or a candidate answer corresponding to the text to be processed, or the probability that each word in the article starts words and ends words as the candidate answer. In this application, the reading understanding model may also employ a pre-trained natural language processing model, such as a BERT model, a robeta model, an Albert model, and the like.

4. The text summary model refers to a neural network model for generating a content summary of input text, where the input may be one or more texts and the output may be a corresponding text summary. In this context, the text summarization model may employ a pre-trained encoder-decoder structure, such as a deep neural network based (source text) sequence to (summary text) sequence framework, where an encoder (e.g., a semantic encoder) converts the source text sequence into a corresponding semantic vector sequence and a decoder (e.g., a circular decoder) generates the summary text sequence based on the semantic vector sequence (e.g., through a attention mechanism and circular decoding).

Aiming at the problems of complex retrieval process, huge calculation amount, low efficiency, low retrieval breadth and precision and the like in the related retrieval technology, the application provides an information retrieval method based on semantic understanding. According to the information retrieval method, problems (namely, first text information and retrieval targets) and internal semantic features of articles (namely, second text information and retrieval targets) are fully utilized for analysis and comparison (external word or keyword matching adopted in place of related technologies), such as semantic similarity (semantic understanding model) and semantic relevance (question-answer reading understanding model), and a plurality of candidate retrieval result related text information (namely, third text information set and candidate answer related information set) are screened and extracted from a large amount of original data (such as articles) to be retrieved; at the same time (e.g., via a text summarization model) a multiple text summarization is generated as at least a portion of the final search result (e.g., an answer to a question) using multiple text summarization automatic extraction techniques based on at least a portion of the multiple candidate search result related information.

Firstly, a candidate retrieval object (i.e. a large amount of original data or articles) is subjected to semantic-level preliminary screening through a semantic understanding model, namely at least one piece of information to be retrieved (i.e. a small amount of articles with high correlation with problems) with high semantic similarity with the problems to be retrieved is screened out, meanwhile, the articles with high correlation are reserved at the semantic level, so that the breadth of retrieval resources is ensured, and a retrieval basis is provided for retrieval precision. And secondly, at least one piece of information to be searched obtained through preliminary screening is respectively extracted or extracted to obtain candidate search results (such as candidate answers of questions) corresponding to the questions to be searched on a semantic level (namely on the basis of semantic understanding) through a reading understanding model, so that a text set related to the candidate search results (namely a third text information set formed by texts containing the candidate search results) is formed, and the accuracy of the search results is further improved. And finally, extracting multiple text summaries from at least one part of texts (such as texts corresponding to the candidate search results with higher search matching degree) in the candidate search result related text set so as to obtain at least one part of the final search result (such as the final answer of the question to be searched), wherein the multiple text summaries can be used for fusing semantic features of a plurality of candidate search results or answers with higher search matching degree, and the robustness and the accuracy of the search results are improved again.

Fig. 1 illustrates an exemplary application scenario 100 of a semantic understanding based information retrieval method according to some embodiments of the present application. As shown in fig. 1, the application scenario 100 may include a server 110, and optionally may 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.

Information retrieval methods according to some embodiments of the present application may be deployed at server 110 and implemented by server 110. The server 110 may be configured to: firstly, acquiring first text information indicating a search target and a plurality of second text information indicating candidate search objects; secondly, determining semantic similarity of the first text information and each of the plurality of second text information; thirdly, selecting at least one text message to be searched from the plurality of second text messages according to the semantic similarity of the first text message and each of the plurality of second text messages; then, respectively extracting third text information semantically related to the first text information from at least one text information to be searched to form a third text information set; then, multi-text abstracts corresponding to at least two third text messages in the third text message set are obtained; and finally, determining a retrieval result corresponding to the first text information based on the multi-text abstract.

For example, the server 110 may store and execute instructions that may perform the various methods described herein. The server 110 may be a single server or a cluster of servers, or may be a cloud server or a cluster of cloud servers capable of providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, web services, cloud communications, middleware services, domain name services, security services, CDNs, and big data and artificial intelligence platforms. It should be understood that the servers referred to herein may typically be server computers having a significant amount of memory and processor resources, but other embodiments are also possible. Moreover, server 110 is shown by way of example only, and in fact, other devices or combinations of devices having computing power and storage capabilities may alternatively or additionally be used to provide corresponding services.

As shown in fig. 1, the application scenario 100 may optionally further include an external database 120 and a network 130. The server 110 may be connected to the external database 120 through the network 130, for example, to acquire a text to be processed from the database 120, including first text information indicating a search target (e.g., a question to be searched) and a plurality of second text information indicating candidate search objects (e.g., candidate articles corresponding to the question to be searched), and for example, to store the obtained search result or answer to the question to the database 120, or the like. The database 120 may be a stand-alone data storage device or group of devices, or may be a back-end data storage device or group of devices associated with other online services, such as online services providing intelligent customer service, voice assistant, etc., as examples.

As shown in fig. 1, the application scenario 100 may optionally 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 may access a server 110 via a network 130 through the terminal device 140 in order to obtain a service provided by the server 110. For example, the user 150 may input instructions through a user interface provided by the terminal device 140, such as through a physical input device (e.g., keyboard and/or mouse, etc.) or virtual keys (e.g., touch screen), through voice or gesture instructions, etc., to initiate an information retrieval scheme deployed on the server 110, to send first text information indicating a retrieval target or question (and/or a plurality of second text information indicating candidate retrieval objects), to receive resulting retrieval results or answers to questions, etc.

Examples of the network 130 include, illustratively, 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 over 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)), a wired or wireless (such as IEEE 802.11 Wireless LAN (WLAN)) wireless interface, a worldwide interoperability for microwave access (Wi-MAX) interface, an ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a bluetooth interface, a Near Field Communication (NFC) interface, etc.

As shown in fig. 1, the terminal device 140 may be any type of mobile computing device, including a mobile computer (e.g., personal Digital Assistant (PDA), laptop computer, notebook computer, tablet computer, netbook, etc.), mobile telephone (e.g., cellular telephone, smart phone, etc.), wearable computing device (e.g., smart watch, head mounted device, including smart glasses, etc.), or other type of mobile device. In some embodiments, terminal device 140 may also be a stationary computing device, such as a desktop computer, a gaming machine, a smart television, or the like. Further, in the case where the application scenario 100 includes a plurality of terminal devices 140, the plurality of terminal devices 140 may be the same or different types of computing devices.

By way of example, the terminal device 140 may include a display screen and a terminal application that may interact with the terminal user via the display screen. The terminal application may be a local application, a Web page (Web) application, or an applet as a lightweight application. In the case where the terminal application is a local application program that needs to be installed, the terminal application may be installed in the user terminal. In the case where the terminal application is a Web application, the terminal application may be accessed through a browser. In the case that the terminal application is an applet, the terminal application may be directly opened on the terminal device 140 by searching for related information of the terminal application (e.g., name of the terminal application, etc.), scanning a graphic code of the terminal application (e.g., bar code, two-dimensional code, etc.), etc., without installing the terminal application.

It should be appreciated that although server 110, database 120, and terminal device 140 are shown and described herein as separate structures, they may be different components of the same computing device. For example, the server 110 may provide background computing functionality, the database 120 may provide data exchange, storage, retrieval functionality, and the terminal device 140 may provide foreground functionality for interaction with a user, such as receiving user input and providing output to the user through a terminal application. Alternatively, the information retrieval method based on semantic understanding according to some embodiments of the present application is not limited to be implemented on the server side shown in fig. 1, but may also be implemented on the terminal device side, or may be implemented jointly on the terminal device side and the server side.

FIG. 2 illustrates an exemplary functional block diagram of a semantic understanding based information retrieval method according to some embodiments of the present application. It should be noted that, rounded boxes in fig. 2 represent various data information, such as a first text information to be processed and a plurality of second text information, a text information to be retrieved, a third text information set, a multi-text summary, and a retrieval result; square-corner rectangular boxes represent processing operations for various data information, including, for example, "semantic similarity-based information screening", "semantic relevance-based information extraction", and "retrieval of multiple text summaries" as shown in fig. 2. As shown in fig. 2, an information retrieval method according to some embodiments of the present application may utilize "three-stage" information processing operations (i.e., "semantic similarity-based information screening", "semantic relevance-based information extraction", and "retrieval of multiple text summaries") to determine an accurate retrieval result corresponding to a first text information (i.e., a 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, first, in a first stage information processing operation, at least one text information 230 to be retrieved, for example, having a higher semantic similarity with the first text information 210, may be selected from a plurality of second text information 220 (i.e., a large amount of original candidate retrieval data) by performing information screening based on the semantic similarity between the first text information 210 and the plurality of second text information 220 (for example, by performing semantic similarity comparison with the semantic understanding model), for use in a subsequent retrieval process; next, in a second stage information processing operation, third text information (e.g., candidate search result related text information sets) semantically related to the first text information 210 is retrieved or extracted from each text information 230 to be retrieved (e.g., by means of a reading understanding model) to form a third text information set 240; again, in a third information processing operation, a multiple text excerpt 250 is obtained for the third set of textual information 240 (e.g., by means of a text excerpt model), and finally a final search result 260 is obtained from the multiple text excerpt 250. It should be noted that the three-phase operations including information filtering, information extraction and abstract acquisition shown in fig. 2 may be implemented using corresponding neural network models, for example, the three-phase operations may be implemented using a semantic understanding model, a reading understanding model and a text abstract model, respectively, but this is not limitative in the present application, in other words, the three-phase information processing operations described above may be implemented in various ways other than the neural network model.

Fig. 3 is a flow chart of a semantic understanding based information retrieval method 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 the execution subject thereof may be the server 110 shown in fig. 1. Alternatively, the information retrieval method according to some embodiments of the present application may also be performed on the terminal device 140 shown in fig. 1, or may also be implemented jointly by the server 110 and the target device 140.

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

s310, a text information acquisition step;

s320, determining semantic similarity;

s330, selecting information to be retrieved;

s340, extracting semantic related information;

s350, a multi-text abstract obtaining step;

s360, a search result determining step.

The execution of the above-described respective steps S310 to S360 is described in detail below with reference to fig. 2.

In step S310 (text information obtaining step), first text information indicating a search target and a plurality of second text information indicating candidate search objects are obtained.

In some embodiments, the retrieval target refers to obtaining an answer to a question to be retrieved (or to be processed) and/or information related to a retrieval keyword through an information retrieval process, and thus, the first text information may include text information (e.g., input by a user) corresponding to the question to be retrieved and/or the retrieval keyword. The second text information may include candidate search objects corresponding to the first text information (e.g., articles corresponding to the questions to be searched), i.e., candidate search objects in which answers to the questions to be searched and/or search keyword related information may be implied. The acquisition of text format data such as the first text information and the second text information is advantageous for the smooth development of the subsequent data processing process, because the formats of input/output data of various natural language processing models used for the subsequent operations of information screening, extraction, abstract acquisition, and the like are mostly text formats, even serialized text formats.

In some embodiments, the text information obtaining step shown in S310 may include, for example, receiving a retrieval target or a question to be retrieved in the form of text input by a user from a terminal device as first text information, and obtaining a plurality of candidate retrieval data or articles in a text format from, for example, an external database as a plurality of second text information. Alternatively, the user-entered reduction target or question to be retrieved and the candidate retrieval data may be other non-text form data (e.g. speech data, video data, form data, etc.), in which case the text information obtaining step needs to convert (e.g. with a suitable format conversion tool) these non-text information after receiving or obtaining them into text information, thereby obtaining the corresponding first text information and the plurality of second text information. In this context, the problem to be retrieved may be a variety of problems to be solved or processed, including but not limited to: common sense problems (simple and intuitive), logical reasoning problems (complex abstractions), numerical reasoning problems, etc.

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

According to the concept of the application, in order to significantly reduce the search amount and improve the search efficiency while guaranteeing the search breadth and accuracy, a relatively small amount of text information to be searched can be screened from a large amount of second text information (namely source data or candidate search objects) based on the semantic similarity degree of the first text information and a plurality of second text information, so that the text information to be searched can be used as a search object, and a search result corresponding to the first text information (such as a problem to be searched) can be searched from the text information to be searched.

In some embodiments, the semantic similarity of the different text information may be determined by calculating the similarity of the respective corresponding semantic features, in other words, the similarity of the semantic features may be utilized to represent the inherent semantic similarity of the different text information. In general, for quantization purposes, semantic features of text information may be represented in vector form, i.e., semantic feature vectors. The semantic feature vector of the text information may refer to a vector extracted from the text information that characterizes its overall semantics. Semantic feature vectors of textual information may be obtained, for example, using a semantic understanding model. In this way, the determination of the semantic similarity of the first text information to each of the plurality of second text information may be converted into a calculation of the similarity between the semantic feature vector of the first text information and the semantic feature vector of each of the second text information. Alternatively, the semantic similarity may be calculated in other ways, for example, the similarity between the semantic (vector) sequences corresponding to the serialized representation (e.g., the word segmentation sequence) of the first text information and the semantic (vector) sequences corresponding to the serialized representation (e.g., the word segmentation sequence) of each of the second text information may be calculated by comparing the two.

Since the semantic features (vectors) of the first text information cover the semantic information of each part (e.g., word or phrase) in the first text information and the mutual association information thereof and the semantic features (vectors) of the second text information cover the semantic information of each part (e.g., sentence, phrase, etc.) in the second text information and the association information thereof in the semantic similarity determining process of step S320, the semantic similarity of the first text information and the second text information determined based on the semantic features can represent the similarity or proximity of the intrinsic semantic level of the first text information and the second text information, rather than the consistency or the matching of simple, surface, and local words (e.g., keywords, partial paragraphs or sentences in the whole text), thereby significantly improving both the search breadth and the accuracy.

Taking Chinese as an example, because a large number of synonyms exist, even if a certain article does not have words which are completely the same as at least part of words (such as keywords) in a retrieval problem, the article is still likely to be similar or close to the retrieval problem semantically, so that a large number of articles which are similar to the retrieval problem semantically can be lost simply depending on keyword matching, important retrieval resources are lost, and the retrieval breadth is limited; on the other hand, because of the presence of ambiguities in Chinese, even if words with the same word-finding meaning may vary widely in different contexts, keyword matching does not necessarily ensure that the required information matching the search problem is found or retrieved, resulting in difficulty in ensuring the search accuracy. In the method, the information retrieval is realized through semantic comparison (namely, the calculation of semantic similarity) based on the first text information and the second text information, the method is not limited to whether the literal information is consistent, the defect of low retrieval breadth and precision of the related technology is fundamentally eliminated, and the preliminary screening of a plurality of second text information is efficiently realized.

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

According to the concept of the application, after determining the semantic similarity of the first text information and each of the plurality of second text information, the actual retrieval object, such as at least one text information to be retrieved, may be selected from the plurality of second text information based on the semantic similarity. In general, one or more second text messages with higher semantic similarity with the first text message can be selected from the plurality of second text messages as the text messages to be searched, because the higher the semantic similarity is, the closer the corresponding second text message is to the first text message at the intrinsic semantic level. In other words, the second text information having 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 may be set in advance. Compared with the original candidate retrieval objects or data (namely a plurality of second text information), the number of the text information to be retrieved (serving as a subsequent retrieval basis) obtained through the screening operation based on the semantic similarity is greatly reduced, so that the retrieval amount is remarkably reduced, and the retrieval efficiency is improved.

In some embodiments, step S320 (information to be retrieved selecting step) may include: 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. As described above, in the information to be retrieved screening process, the first M pieces of second text information having the highest similarity with the first text information may be selected as the retrieval base or the actual retrieval object, where M may be preset according to the actual application scenario. For example, for a relatively complex or abstract question to be retrieved (i.e., the first text information), M may be set to a relatively large value to ensure the breadth and richness of the data to be retrieved, thereby facilitating retrieval of a more accurate answer. On the other hand, M cannot be set too large in order to ensure retrieval efficiency, and for example, can be set to a minimum value suitable for a specific application scenario.

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, respectively, to form a third text information set.

According to the concept of the present application, after the information is initially screened in the first stage (i.e., at least one text information to be searched, such as an article to be searched, which is used as a search basis, is screened out from a plurality of second text information), an information extraction process in the second stage may be performed, that is, third text information related to the first text information semantically (i.e., candidate search result related information corresponding to the first text information, such as candidate answer related information corresponding to a problem to be processed, which is extracted from the article to be searched) is extracted from each piece of information to be searched, so that the third text information corresponding to each piece of information to be searched forms a third text information set for use in a multi-text abstract acquisition process in a subsequent third stage, which is used as a basis for multi-text abstract acquisition.

In some embodiments, the extracting step S340 of the related semantic information may extract, for each text message to be retrieved, a third text message related to the first text message semantically, so as to form a third text message set, where each third text message in the set may correspond to each text message to be retrieved one by one. In this context, semantic correlation between the third text information and the first text information may be understood as that the third text information extracted at each text information to be retrieved is 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 candidate answer-related information that is extracted from an article to be retrieved (i.e., text information to be retrieved) and matches the question. In some embodiments, the third text information may be at least a portion of information extracted from the corresponding text information to be retrieved that includes a corresponding reduction of the first text information, e.g., the third text information may be a sentence, a paragraph, etc. in the article to be retrieved. As for the extraction manner of the third text information, it may be implemented using a (machine) reading understanding model (e.g., BERT-based neural network model, etc.) or a question-answering model. For example, first, a first text message (e.g., question) and a message to be retrieved (e.g., text) are entered into a model, thereby obtaining a question answer determined from an article; subsequently, information containing the answer to the question (for example, a sentence in which the answer to the question is located, a natural paragraph, or the answer to the question itself) is extracted or extracted from the article as third text information. See fig. 6 and 7 and their associated description for specific procedures.

As the second-stage information extraction step, S340 is mainly used for further reducing the information amount based on the first-stage information primary screening, and screening (i.e. extracting) candidate retrieval result related information from the information to be retrieved by using, for example, a reading understanding model, as the basis of the multi-text abstract in the subsequent third stage, thereby further reducing the information processing amount in the retrieval process and improving the data processing efficiency on the basis of guaranteeing the breadth and the accuracy of the retrieval data.

In step S350 (multiple text excerpt obtaining step), multiple text excerpts corresponding to at least two pieces of third text information in the third text information set are obtained.

According to the concept of the present application, after the first stage of semantic similarity-based information preliminary screening (i.e. screening at least one text information to be searched for, for example, an article to be searched for) and the second stage of semantic related information extraction (i.e. extracting third text information from the information to be searched for to form a third text information set), which are used as a search basis, a multi-text abstract acquisition process of a third stage may be entered, i.e.: firstly, at least two third text messages (for example, with higher correlation degree or search matching degree with the first text message) are selected from the third text message set, and secondly, a text abstract model (for example, encoder-decoder model based on deep neural network) for natural language is utilized to extract or generate multiple text abstracts based on the third text messages from the at least two third text messages as at least part of the final search result.

In some embodiments, the multiple text summaries (i.e., text summaries obtained from at least two third texts in the third set of texts) involved in the information retrieval method according to the present application may be generated text summaries. The generated text abstract (generated by the text abstract model) can automatically and efficiently provide proper and accurate answers for more abstract and complex problems due to the processes of semantic feature understanding, summarizing, reasoning and the like, the application scene and the application range of the retrieval method are remarkably enlarged, the robustness and the accuracy of the retrieval result are improved, and the user experience is greatly improved. Alternatively, the multiple text summaries involved in step S350 may also be decimated text summaries.

In some embodiments, a sequence-to-sequence language processing model may be employed to obtain a multiple text summary, such as a model based on a deep neural network (e.g., RNN and/or CNN) encoder-decoder structure, where the input of the encoder portion is at least two third textual information of a third set of textual information (of higher degree of matching or relevance to the first textual information), and the output of the decoder code portion is a multiple text summary result. The model running step may include: splicing at least two third text messages according to the high-low order of the matching degree, obtaining an input word segmentation sequence through word segmentation processing, and inputting the input word segmentation sequence into an encoder to obtain a semantic vector sequence corresponding to the input text messages; however, the context vector sequence is obtained by using an attention mechanism in combination with the input corresponding to each time step at the decoder side; and finally, decoding by using a decoder to obtain a decoded text sequence corresponding to each time step (or cyclic decoding moment), namely a multi-text abstract.

In step S360 (search result determination step), a search result corresponding to the first text information is determined based on the multiple text digests.

According to the concept of the present application, after the multiple text excerpts of at least two (more relevant) third text information are acquired, the multiple text excerpts may be directly used as a final search result or at least a part thereof. Because the basis of the multi-text abstract is a plurality of third text information with higher relativity (with the first text information) selected from a third text set (namely, a candidate search result related information set corresponding to the first text information), the multi-text abstract fuses the content or the inherent semantic features of the candidate search result related information with higher retrieval matching degree and the semantic association among the candidate search result related information, so that the multi-text abstract is an abstract result obtained by taking the essence of the whole information again on the basis of the candidate search result related information with higher retrieval precision, and the matching precision of a final search result (such as a question answer) and the first text information (such as a question to be processed) is further improved.

In some embodiments, the retrieving result determining step may include: for example, the multi-text abstract is presented as a search result corresponding to the first text information, and is presented as an information search result. In some application scenarios, when the search target includes a question to be searched, the search result may include answer information corresponding to the question to be searched. In other words, the multi-text abstract can be directly used as the final answer information of the questions to be processed. Because the semantic features and the interrelated features of a plurality of candidate retrieval results (namely a plurality of candidate answers corresponding to the questions to be processed) are fused, the multi-text abstract can give direct and accurate answer information aiming at the information of the questions to be retrieved, and user experience and accuracy of the retrieval results can be improved.

In some embodiments, the search result determining step S360 may also include: generating a first search 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 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. For example, the search result may include not only the first search result based on the multi-text summary as a (precise) main search result, but also a second search result generated based on at least two third text information (or third text information sets) based on the multi-text summary, for example, a list of at least two third text information sets having a higher degree of correlation with the first text information among the third text information sets obtained by the information extraction and extraction process of the second stage, as an (optional) auxiliary search result. The double search result form enriches the diversity and the selectivity of the search result, and can meet the personalized requirements of different users. For example, when the user is not satisfied with the accurate search result based on the multi-text excerpt, the list of candidate search results may be referred to as an optional auxiliary search result list to quickly inquire the corresponding information therefrom.

In the semantic understanding based information retrieval method according to some embodiments of the present application, advanced information retrieval tasks such as intelligent question-and-answer are efficiently and accurately completed through a three-stage information processing process, namely, the first stage: screening with second text information (i.e., candidate retrieval objects, such as candidate articles for obtaining answers to questions to be retrieved therefrom) based on semantic similarity (with respect to the first text information (e.g., questions to be retrieved)); and a second stage: extracting a third text information (such as candidate answer related text) set based on semantic relevance; and a third stage: and generating the multi-text abstract based on the third text information set.

Specifically, firstly, a relatively small amount of text information to be searched is screened from a large amount of second text information according to the semantic similarity (rather than simple literal matching) of the first text information and the second text information, so that the overall working efficiency is remarkably improved under the condition of ensuring that important information search resources with high correlation with the search problem are not lost (namely, ensuring the search breadth and accuracy), and the problems of important search resources deficiency caused by accurate keyword matching and low efficiency caused by complex flow and huge calculation amount in the related technology are overcome; secondly, extracting or searching third text information (namely candidate answer related text information corresponding to the to-be-searched question) corresponding to the first text information (such as the to-be-searched question) again based on semantic relevance aiming at each to-be-searched text information obtained by screening in the first stage, so that extracting of candidate answers and related texts (namely third text information set) thereof is realized by utilizing semantic features in the to-be-searched question and semantic features and mutual relevance of the to-be-searched information again, and higher relevance of the third text information set and the first text information and higher accuracy of a final search result are further ensured; finally, generating multiple text summaries (e.g. generating text summaries) of at least two third text messages in the third text message set, namely generating multiple text summaries from a plurality of candidate answer related texts with high matching degree through summarization and reasoning modes as final answers (e.g. as a part of retrieval results) of the questions to be retrieved, wherein the multiple text summaries can further improve the quality and accuracy of the retrieval results due to the fact that semantic features and associated features of a plurality of candidate answer related text messages with higher retrieval matching degree (i.e. third text messages) are fused.

Fig. 4 illustrates an example flow chart of semantic similarity determination steps of an information retrieval method according to some embodiments of the present application. FIG. 5 illustrates 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. The respective steps shown in fig. 4 are 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 obtained.

In order to quantify semantic features of text information, semantic feature vectors may be used to represent semantic features of text information, so that the semantic similarity between different text information corresponding to each vector may be characterized by the similarity between the two vectors (e.g., the distance between the vectors and/or the cosine of the included angle). In this way, the determination of the semantic similarity of the first text information to each of the plurality of second text information may be converted into a calculation of the similarity between the semantic feature vector of the first text information and the semantic feature vector of each of the second text information.

Regarding the acquisition of semantic feature vectors for text information, it may be implemented using a semantic understanding model for natural language processing. The semantic understanding model may also be referred to as a semantic encoder, which may be a variety of pre-trained semantic understanding models, such as a BERT neural network, a Roberta neural network, an Albert neural network, etc., for converting an input text sequence (e.g., a word segmentation sequence corresponding to the first text information) and/or a word vector sequence into semantic feature vectors that characterize the overall semantics of the input text.

As shown in fig. 5, in the semantic feature vector acquisition stage (i.e., semantic encoding stage), for the first text information 510 and each of the second text information 520, first, they are input to a text preprocessor 530, respectively, and are subjected to serialization processing (e.g., word segmentation processing) to obtain a first word segmentation sequence ([ CLS ], q (1),., q (k)) and a second word segmentation sequence ([ CLS ], p (1),., p (m)), respectively, where [ CLS ] represents a text start marker character or a global character, q (1) -q (k) represent a total of k words of the first text information (e.g., a question to be processed), p (1) -p (m) represent a total of m words of the second text information (e.g., article), and k and m may be positive integers (generally, m > k) greater than or equal to 1, respectively; 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 the first semantic feature vector 550 corresponding to the first word sequence and the second semantic feature vector 560 corresponding to the second word sequence are respectively obtained through the semantic understanding or semantic encoding process.

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

For example, 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 sequences, or text sequences. The word segmentation tool may be, for example, a nub (Jieba) word segmentation, LTP, THULAC, NLPIR, or the like. Alternatively, the serialization processing or word segmentation process of the text to be processed may be implemented by manual labeling, random segmentation, complete segmentation into a plurality of single characters, or the like. In general, the model 540 for semantic understanding of the first text information 510 and the second text information 520 may be the same; alternatively, the first text information 510 and the second text information 520 may be semantically understood or semantically encoded using different semantic understanding models, respectively.

Optionally, if the input format of the semantic understanding model 540 is a vector sequence, the text preprocessor 530 performs a vectorization process after performing a word segmentation process on the first and second text information to obtain corresponding first and second word sequences, so as to obtain first and second word vector sequences; the first and second word vector sequence inputs are then input to the semantic understanding model for semantic encoding, respectively. Illustratively, the text preprocessor 530 may also further include a word vector tool for text-to-word vector conversion to implement vectorization processing for the segmented text sequence. Word vector tools may include, for example, one-hot, word2vec, glove, etc.

In some embodiments, step S320a may include: 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 messages respectively from a preset semantic feature vector index library, wherein the preset semantic feature vector index library stores the plurality of second semantic feature vectors determined by using a semantic understanding model.

Since the retrieval object (i.e., the second text information) such as the article is generally relatively fixed known data stored in advance in, for example, a server database, and the data amount of the retrieval object tends to be very large, in order to improve the data processing efficiency, it is possible to consider that the semantic understanding operation (i.e., the acquisition or calculation of the semantic feature vector or the semantic code) of a large amount of the second text information (i.e., the retrieval object such as the article) is performed in advance while being online or offline, and the resulting plurality of the second semantic feature vectors are stored in a preset semantic feature vector index library for semantic similarity calculation (with the first text information such as a problem to be retrieved) on-line when the information retrieval occurs. On the other hand, the semantic understanding or semantic coding process of the first text information can be performed on line after the information retrieval is started, so that only second semantic feature vectors (pre-determined off-line) corresponding to each second text information are required to be directly called from a semantic feature vector index library in the information retrieval process, and semantic coding operation is not required to be performed on a large number of second text information in real time on line by utilizing a semantic understanding model, thereby greatly improving data processing efficiency (especially in the scene of a large data amount of objects or articles to be retrieved) and remarkably optimizing network resource configuration and scheduling.

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

As described above, the semantic similarity between different text information may be characterized by the similarity between corresponding semantic feature vectors. Accordingly, as shown in fig. 5, after the first semantic feature vector 550 corresponding to the first text information 510 and the second semantic feature vector 560 corresponding to each of the second text information 520 are acquired, a vector similarity therebetween needs to be calculated in order to determine a semantic similarity 570 between the first text information 510 and the second text information 520.

In some embodiments, two vector similarities may be embodied as a degree of similarity in a direction between two vectors, such as cosine similarity; it may also be embodied as the proximity of distances between vectors, i.e. based on the similarity of distances. Specifically, step S320b may include: 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; a similarity of the first semantic feature vector to each of the second semantic feature vectors is determined based on at least one of the first similarity and the second similarity.

Illustratively, when calculating the distance between two semantic feature vectors, a smaller distance indicates that the two semantic feature vectors are more similar, i.e., the vector similarity is greater; conversely, a larger distance indicates that the two semantic feature vectors are less similar, i.e., the vector similarity is smaller. Illustratively, when calculating the cosine of the angle between two semantic feature vectors, a smaller angle indicates that the two semantic feature vectors are more similar, i.e., the vector similarity is greater; conversely, a larger included angle indicates that the two semantic feature vectors are less similar, i.e., the vector similarity is smaller. Therefore, the similarity between the first semantic feature vector and the second semantic feature vector may be calculated according to the vector distance alone or according to the cosine of the included angle alone, or the similarity between the first semantic feature vector and the second semantic feature vector may be calculated according to the combination of the vector distance and the cosine of the included angle (i.e., the weighted sum of the first similarity and the second similarity). Whether the distance between the plurality of second semantic feature vectors and the first semantic feature vector or the cosine of the included angle between the plurality of second semantic feature vectors and the first semantic feature vector is calculated, the two calculation modes are relatively simple, and the semantic feature vectors can accurately represent the semantics, so that the requirement of accurate retrieval is met, the retrieval efficiency can be improved, the retrieval time is saved, the two characteristics complement each other, and the method is very beneficial to scenes with massive second text information.

In some embodiments, the distance of the first semantic feature vector from the second semantic feature vector may include a euclidean distance, a manhattan distance, a chebyshev distance, or the like. Alternatively, other methods may be used to calculate the similarity of the first semantic feature vector and the second semantic feature vector, such as a correlation coefficient or a similarity coefficient, in addition to cosine similarity and distance-based similarity.

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

For example, after the similarity between the first semantic feature vector and each of the second semantic feature vectors is obtained, the similarity between the first semantic feature vector and each of the second semantic feature vectors may be directly taken as the semantic similarity therebetween; alternatively, the similarity between the semantic feature vectors may be subjected to appropriate data processing (e.g., discrimination of increasing or decreasing the similarity between the semantic feature vectors, normalization operation, etc.), and the processed data may be then used as the semantic similarity.

Fig. 6 illustrates an example flow chart of semantically related information extraction steps in an information retrieval method according to some embodiments of the present application. FIG. 7 illustrates a schematic diagram of extracting semantically related information using a reading understanding 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. The respective steps shown in fig. 6 are described in detail below with reference to fig. 7.

In step S340a, for each text information to be retrieved, fourth text information indicating a candidate retrieval result corresponding to the first text information is determined therefrom using the reading understanding model.

According to the concept of the present application, in the semantic related information extraction step S340 of the second stage, candidate search result related information, that is, third text information, needs to be extracted from each text information to be searched, respectively, so as to form a third text information set, which is used as a basis of a subsequent multi-text abstract. For example, each third text information in the set of third text information (i.e., candidate search result related information) may be defined as text information extracted or extracted from the corresponding information to be searched, including, for example, a candidate search result (i.e., fourth text information) corresponding to the first text information. For example, when the first text information indicates a question to be processed, the fourth text information may be indicated as a candidate answer corresponding to the question to be processed, and the third text information may be indicated as a candidate answer-related article fragment. Therefore, to obtain the third text information set, it is necessary to first determine the corresponding fourth text information (i.e., candidate search results) from each piece of information to be searched, and then extract the third text information (i.e., candidate search result related information) including the candidate search results.

In some embodiments, the reading understanding model refers to a neural network model for semantically understanding an article or corpus of natural language and answering a related question, where the input may be a serialized representation of first text information (e.g., text of the question to be processed) and text information to be retrieved (e.g., corresponding text to be retrieved), and the output may be a candidate retrieval result (e.g., answer or candidate answer to be processed), or a first probability and a second probability corresponding to each word in a word segmentation sequence corresponding to the text information to be retrieved, where the first probability represents a probability that the word is a beginning word segment of fourth text information indicating the candidate retrieval result, and the second probability corresponding to each word represents a probability that the word is an ending word segment of the fourth text information

In some embodiments, step S340a may include performing the following steps 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.

In extracting fourth text information (i.e., candidate search results) from each piece of information to be searched, first, the first text information 710 (e.g., question to be searched) and the corresponding information 720 (e.g., article) to be searched are input to a text preprocessor 730, so as to perform preprocessing operations such as stitching and serialization, etc., and are converted into an input format, i.e., text or word segmentation sequence 740, required for reading the understanding model 750. As shown in fig. 7, the word segmentation sequence 740 sequentially includes: start marker of text information to be processed "[ CLS ] ]"first word sequence 740a (i.e., q (1),..q (k)), text separator" [ SEP "), which corresponds to the first text information 710]", second word sequence 740b (i.e., p (1)", p (m)). The word sequence 740 is then input to a read understanding model 750, where the individual words p (1) in the second word sequence 740b may be output, through processing and computation of the multi-layer neural network in the model,p (m) corresponds to a first probability and a second probability, wherein the first probability represents a probability that the word is a start word of fourth text information (e.g., a question candidate answer) indicating a candidate retrieval result, and the second probability represents a probability that the word is an end word of the fourth text information. After that, the beginning word segmentation and ending problem is determined according to the first probability and the second probability, for example, the largest first probability and the largest second probability can be found in the first probability and the second probability respectively corresponding to each word in the second word segmentation sequence (p (1),... As shown in FIG. 7, for example, in the first probability and the second probability of p (1) -p (m) corresponding to each other, p (m) ₁ ) The corresponding first probability is the largest, p (m ₂ ) The corresponding second probability is the largest, and m ₁ ＜m ₂ Then it can be determined that the mth of the text information 720 to be retrieved ₁ The individual word p (m ₁ ) For beginning word segmentation of the fourth text information, the mth text information 720 is to be retrieved ₂ The individual word p (m ₂ ) End word segmentation for the fourth text information. Finally, the segmentation of the words p (m) can be based on the beginning ₁ ) And p (m) ₂ ) Determining fourth text information, namely the mth in the information to be searched ₁ The individual word is the m ₂ The word between the individual words (containing p (m) ₁ ) And p (m) ₂ ) The structured text information is determined as fourth text information as the output text of the reading understanding model.

Optionally, when the word p (m ₁ ) Number m of (2) ₁ Word segment p (m) corresponding to the second probability greater than or equal to the maximum ₂ ) Number m of (2) ₂ I.e. m ₁ >＝m ₂ When the beginning segmentation and the ending segmentation of the fourth text information can be determined in the following two ways:

first kind: firstly, word segmentation p (m) corresponding to the maximum first probability in the second word segmentation sequence ₁ ) Beginning segmentation as fourth text information, followed by arranging the segmentation p (m) in a second segmentation sequence ₁ ) Searching the word p (m) with the highest second probability from the following words ₃ ) (i.e. m ₃ ＞m ₁ ) As a junction of the fourth text informationBeam segmentation;

second kind: first, word segmentation p (m) corresponding to the maximum second probability in the second word segmentation sequence ₂ ) Ending the word segmentation as the fourth text information, followed by arranging the word segmentation p (m) in the second word segmentation sequence ₂ ) Searching for the word p (m) with the highest first probability from the previous words ₄ ) (i.e. m ₄ ＜m ₂ ) As a start word of the fourth text information.

Optionally, the beginning segmentation and ending segmentation of the fourth text information may also be determined by: first, among all word pairs of the second word sequence 740b, a sum of a first probability corresponding to a preceding word and a second probability corresponding to a following word is determined, and then a word pair having the largest probability and the largest sum is selected, wherein the preceding word is used as a start word and the following word is used as an end word. In addition, the beginning segmentation and ending segmentation 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 conception of the application, after the fourth text information (i.e., candidate search result) is determined from each piece of information to be searched, third text information (i.e., candidate search result related information) containing the fourth text information can be extracted from the information to be searched, and then the third text information respectively extracted from the respective pieces of information to be searched can form a third text information set.

As shown in fig. 7, after the fourth text information 760 is obtained, the 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 represented as a word sequence (a.p (m ₁ )，...，p(m ₂ ) ,..), i.e. it contains text information corresponding to the beginning word p (m) ₁ ) And ending the word p (m) ₂ ) Fourth text information 760 is defined.

In some embodiments, step S340b may include one of the following steps: 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. In other words, the third text information may be a natural sentence, a natural paragraph, or the fourth text information itself where the fourth text information is located in the corresponding information to be retrieved. Alternatively, the fourth text information may be other text information of the corresponding information to be retrieved, including the third text information.

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

After extracting the third text information including the fourth text information from each piece of information to be retrieved, each piece of third text information extracted from each piece of information to be retrieved may be used as an element to form a third text information set. Alternatively, a part of the third text information may be selected from the respective third text information to form the third text information set. For example, a plurality of third text information with a higher search matching degree or relevance degree can be selected to form a set, where the search matching degree or relevance degree can be represented by at least one of a first probability and a second probability corresponding to a start word and an end word of fourth text information corresponding to the third text information. And selecting part of the third text information to form a third text information set according to the retrieval matching degree, so that the data processing amount can be further reduced under the condition of ensuring the retrieval precision, and the retrieval efficiency is improved.

In the embodiments shown in fig. 6 and 7, by analyzing text information to be retrieved and extracting semantic related information (i.e., third text information) one by one using a reading understanding model, candidate retrieval results (i.e., fourth text information, e.g., candidate answers) corresponding to the first text information (e.g., questions to be retrieved) can be found from each text information to be retrieved, and the third text information including the candidate retrieval results can be extracted. This process is also based on semantic understanding and analysis of the first text information and the text information to be searched, so that a higher degree of correlation and search matching between the searched candidate search result and the candidate search result related information (i.e. the third text information) and the first text information can be ensured. In addition, the reading and understanding model gives the first probability and the second probability of each word as the beginning word segmentation and the ending word segmentation, which provides a flexible extraction method for determining the third text information, so that different information retrieval scenes can be adapted.

Fig. 8 illustrates an example flowchart of multiple text excerpt retrieval steps in an information retrieval method according to some embodiments of the present application. FIG. 9 illustrates a schematic diagram of generating multiple text summaries using a text summary model according to some embodiments of the present application.

As shown in fig. 8, the multiple text excerpt obtaining step S350 may include steps S350a-S350d, which are described in detail below with reference to fig. 9.

In step S350a, for each third text message in the third text message set, a search matching degree corresponding to the third text message is determined according to at least one of a first probability corresponding to a start word of the fourth text message and a second probability corresponding to an end word of the fourth text message included in the third text message.

In accordance with the concepts of the present application, in the multiple text excerpt retrieval operation at the third stage of the information retrieval process, the basic information of the multiple text excerpt (i.e., the abstracted object) may be selected from the third text information set (i.e., the candidate retrieval result related information). In order to properly reduce the data amount and the calculation amount while ensuring the retrieval accuracy, a plurality of third text information with higher correlation degree with the first text information or higher retrieval matching degree can be selected from the third text information set to be used as the basis of the abstract acquisition in the third stage. Herein, the search matching degree may refer to a matching degree between the third text information (i.e., candidate search result related information) or the corresponding fourth text information (i.e., candidate search result, such as question answer) and the first text information (such as question to be searched) in a semantic dimension. The higher the search matching degree is, the more the candidate search result corresponding to the third text information is matched with the search target or the problem to be searched corresponding to the first text information, or the more accurate the candidate search result is.

In some embodiments, the search matching degree of each third text information in the third text information set may be determined according to at least one of a first probability corresponding to a start word of the fourth text information and a second probability corresponding to an end word of the fourth text information included in the third text information, where the magnitudes of the first probability and the second probability reflect the matching degree or the correlation degree of the fourth text information (i.e. candidate search result) or the third text information and the first text information (i.e. search target) determined thereby.

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

(1) An arithmetic average of a first probability corresponding to the beginning of word segmentation and a second probability corresponding to the ending of word segmentation;

(2) Geometric mean of a first probability corresponding to the beginning of word segmentation and a second probability corresponding to the ending of word segmentation;

(3) Maximum value in first probability corresponding to beginning word segmentation and second probability corresponding to ending word segmentation;

(4) The minimum value of the first probability corresponding to the beginning word segmentation and the second probability corresponding to the ending word segmentation.

Any one or more of the above four values may be used to determine the search matching degree of the third text information or the fourth text information to adapt to different situations. In addition, the four provided determination modes have smaller calculation amount, can save retrieval time and can effectively determine accurate retrieval matching degree.

Alternatively, the determination of the search matching degree of the third text information is not limited to the above four methods, but may be other methods, for example, based on the semantic similarity of the third text information and the first text information, or the like.

In step S350b, at least two third text information are selected from the third text sets according to the search matching degree corresponding to each third text information in the third text sets.

After determining the retrieval matching degree corresponding to each third text information in the third text set, a plurality of third text information with higher retrieval matching degree can be selected from the third text set as the object of the multi-text abstract based on the retrieval matching degree. In some embodiments, S350b may include: sequencing the third text information in the third text information set according to the sequence of the searching matching degree of the first text information and each third text information from big to small; and selecting the first N pieces of third text information from the ordering, wherein N is a preset positive integer greater than 2. In this way, the first N pieces of third text information with the highest matching degree with the first text information retrieval can be selected as the multi-text abstract base or abstract object, wherein N can be preset according to the actual application scene. For example, for a relatively complex or abstract question to be searched (i.e., the first text information), N may be set to a relatively large value to ensure the universality and richness of the summarized data in the multi-text summarization process, thereby facilitating the search of a more accurate search result or question answer; on the other hand, N cannot be set too large in order to ensure the working efficiency of the digest process, for example, can be set to a minimum value suitable for a specific application scenario with ensuring the search accuracy.

In step S350 c, at least two third text messages are spliced according to the order of the respective corresponding search matching degree from high to low, so as to form second text messages to be processed.

In some embodiments, with respect to the retrieval of multiple text summaries, a text summary model may be employed, so that summary object (i.e., a plurality of third text information) data needs to be preprocessed to fit the model input format prior to entering the text summary model. Generally, the data preprocessing process of the abstract object may include a stitching process of a plurality of third text information, for example, stitching at least two third text information together according to a sequence of searching matching degree, so as to obtain second text information to be processed. Optionally, the data preprocessing process of the abstract object may further include a serialization process of the second text information to be processed to match an input format of the text abstract model (assuming that the text abstract model is a text sequence or a word segmentation sequence), for example, performing word segmentation processing on the second text information to be processed to obtain a corresponding word segmentation sequence.

In step S350d, a multiple text excerpt corresponding to the second text information to be processed is generated by using the text excerpt model.

In some embodiments, the text summarization model may employ a pre-trained encoder-decoder structure, such as a deep neural network based (source text) sequence to (summary text) sequence framework. As shown in fig. 9, the text summarization model may include an encoder 910 and a decoder 920, wherein the encoder 910 (e.g., a semantic encoder) converts a source text sequence into a corresponding semantic vector sequence, and the decoder 920 (e.g., a circular decoder) generates a summary text sequence based on the semantic vector sequence (e.g., through an attention mechanism and circular decoding). In particular, the deep neural network employed by the text summarization model may include a Convolutional Neural Network (CNN), a cyclic neural network (RNN), where the encoder 910 and decoder 920 may employ one or more neural network structures that are the same or different, so long as the respective input-output functions are enabled.

Illustratively, a process of generating a multiple text excerpt of the second pending text information using the text excerpt model is described below with reference to fig. 9. As shown in fig. 9, the preprocessing stage is first to input the second text information to be processed 930 into the text preprocessor 940 to perform serialization processing on the second text information to obtain a word segmentation sequence or text sequence (T ₁ ，T ₂ ，T ₃ ，...，T _n ) Wherein T is _i (i=1, 2,) n represents the i-th text in the text sequence, n represents the total number of texts in the text sequence.

Subsequently, the encoding phase is entered, i.e. the text sequence (T ₁ ，T ₂ ，T3，...，T _n ) Input to text abstract modelTo perform semantic understanding thereof to obtain a text sequence (T ₁ ，T ₂ ，T ₃ ，...，T _n ) One-to-one semantic vector sequences (H ₁ ，H ₂ ，H ₃ ，...，H _n ) Wherein H is _i (i=1, 2,., n) represents an i-th semantic vector in the sequence of semantic vectors, which corresponds to an i-th text or word T in the sequence of texts _i N represents the total number of semantic vectors in the sequence of semantic vectors. Here a sequence of semantic vectors (H ₁ ，H ₂ ，H ₃ ，...，H _n ) May also be referred to as a sequence of hidden state vectors for encoder 910. As shown in fig. 9, the semantic coding stage may implement semantic coding of the text sequence by using a neural network with bidirectional circulation, so that features of semantic association (such as forward association and reverse association) between different texts or segmentation words are fused in the semantic coding, so that overall semantic features of the second text information to be processed are reflected more accurately.

The multiple text summarization process then proceeds to the decoding stage. In some embodiments of the present application, the decoding stage employs a cyclic decoding process based on a plurality of time steps, where the number of decoding time steps or decoding instants may be predetermined according to a specific application scenario or the like. The overall decoding process of decoder 920 may be summarized as follows: first, the semantic vector sequence (H ₁ ，H ₂ ，H ₃ ，...，H _n ) Conversion into a sequence of content vectors (C ₁ ，C ₂ ，C ₃ ，...，C _r ) Where r represents the number of decoding instants corresponding to decoder 920; subsequently, the decoded text sequence (a) is obtained by performing cyclic decoding at each decoding time using (e.g., a cyclic neural network structure) decoder 920 ₁ ，A ₂ ，A ₃ ，...，A _r ) And finally, obtaining a final multi-text abstract by splicing as a multi-text abstract sequence.

Fig. 9 shows the decoding process at the t-th decoding moment or step in the decoding phase of the multi-text digest. As shown in fig. 9, at the current decoding time of the decoding stage of the multi-text digest, i.e., the tth decodingFirst, the decoder hidden state vector S can be outputted by the last decoding time of the decoder 920, i.e., the t-1 st decoding time _t-1 Semantic vector sequences (H) ₁ ，H ₂ ，H ₃ ，...，H _n ) The attention profile 950 corresponding to the current decoding moment is determined, which may also be referred to as an attention weight sequence. As shown in fig. 9, each black square column in the attention distribution 950 may represent a vector sequence (H ₁ ，H ₂ ，H ₃ ，...，H _n ) The attention weights of the semantic vectors in a one-to-one correspondence, for example, the longer the square column, the greater the corresponding attention weight. Thus, subsequently, as shown in FIG. 9, the sequence of semantic vectors (H) may be determined based on the attention weight sequence or attention 950, respectively ₁ ，H ₂ ，H ₃ ，...，H _n ) The respective semantic vectors calculate a weighted sum to obtain the content vector C at the current decoding time (i.e. the t-th decoding time) _t . Then, the content vector C according to the current decoding time is decoded in the decoder 920 _t Decoder hidden state vector S at the previous decoding moment (i.e. the t-1 decoding moment) _t-1 And the decoded text a output from the previous decoding time decoder 920 _t-1 Decoder hidden state vector S for determining current decoding time _t The method comprises the steps of carrying out a first treatment on the surface of the Then in the decoder 920, the content vector C according to the current decoding time instant (i.e., the t-th decoding time instant) _t Decoder hidden state vector S at current decoding time _t And the decoded text a output from the previous decoding time decoder 920 _t-1 Determining a (final) decoded text A corresponding to the current decoding moment _t 。

As shown in fig. 9, the text a is decoded (finally) corresponding to the current decoding time _t In (2), the decoder 920 (e.g., a class prediction function) may be utilized to predict the vocabulary probability distribution 960 corresponding to the current decoding time instant, i.e., all possible output text or vocabulary into the final decoded text a _t Is a probability distribution of (c). The vocabulary probability distribution 960 shown in FIG. 9 is shown in the form of a plurality of square columns, each square column representing the probability that a vocabulary will be the decoded text at the current time; obviously, the higher the square column is, the corresponding The greater the likelihood that the vocabulary of (a) will become the decoded text corresponding to the current decoding time. Thus, as shown in FIG. 9, the vocabulary or text corresponding to the highest histogram may be selected in the vocabulary probability distribution 960 as the decoded text A corresponding to the current decoding time instant (i.e., the t decoding time instant) _t 。

In the decoding stage of the multi-text summary acquisition process according to some embodiments of the present application, the attention mechanism may weight each semantic vector in the semantic vector sequence to distinguish the key degree of the content vector conversion corresponding to the current decoding time and the subsequent decoding process of different semantic vectors, so as to improve the efficiency and accuracy of the neural network. For example, in the decoding stage of the text abstract model, each vector in the input semantic vector sequence corresponds to a word corresponding to the second text information to be processed, but the association degree of the word and the problem to be processed is different, so that the importance degree of different semantic vectors in the decoding process of the current decoding moment is reflected through attention weight (or attention distribution), and the efficiency and the accuracy of the model prediction reasoning process are improved.

Furthermore, in the calculation of the decoded text at the current decoding time (i.e., the t decoding time) in the decoding stage of the multiple text digest acquisition process according to some embodiments of the present application, the input of decoder 920 includes not only the content vector C at the current decoding time _t And a decoding vector S output from the previous decoding time (i.e., the t-1 st decoding time) _t-1 And, also, includes the decoded text A at the last decoding time _t-1 This allows the input of decoder 920 of the text excerpt model to be fused with more different kinds of relevant features (e.g., a _t-1 Corresponding contextual characteristics), facilitating the final decoded text a at the current decoding time _t The whole semantic features, especially the context semantic relation features or the association features, of the second to-be-removed text information are reflected more accurately. This allows the final output decoded text sequence and/or multi-text summary to more accurately match the problem being processed.

Fig. 10 is a schematic diagram of a complete process of an 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 searched, the second text information is an article, the information to be searched is an article to be searched, the third text information is an article fragment, and the third text information is a candidate answer. As shown in fig. 10, the upper dotted line represents an online operation process, and the lower dotted line represents an offline operation process.

As shown in fig. 10, in an offline case, semantic understanding models are used to determine semantic feature vectors of candidate search objects, i.e., a plurality of articles, and then a semantic feature vector index library is constructed locally based on the semantic feature vectors, or the semantic feature vectors are stored in a preset semantic feature vector index library for direct call during online search. Therefore, the data processing efficiency is greatly improved and the network resource scheduling is remarkably optimized under the scene of articles to be searched with large data volume.

Then, as shown in fig. 10, after the problem to be retrieved is acquired, the online information retrieval starts, and the online information retrieval process may be divided into three stages: article screening, segment extraction, and text summarization. As shown in fig. 10, in the article screening stage, firstly, a problem to be searched is input into a semantic understanding model to determine a first semantic feature vector corresponding to the problem to be searched; then, in order to screen articles with similar semanteme, invoking second semantic feature vectors corresponding to the articles from a semantic feature vector index library and calculating the similarity between the first semantic feature vector and each second semantic feature vector; and then, ordering the articles according to the semantic similarity corresponding to each article, and selecting the first M articles as articles to be searched, wherein M is a positive integer. As shown in fig. 10, in the segment extraction stage, the reading understanding model is used to extract the segment of the article containing the candidate answer matching the question to be retrieved from the M articles screened in the previous stage. In the text summarization stage, as shown in fig. 10, a text summarization model is used to obtain or generate multiple text summaries corresponding to multiple article segments (e.g., the first N article segments with higher search matching degree) selected from the M article segments, and the multiple text summaries are presented to the user 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, on-line and off-line, which complement each other to realize accurate and efficient retrieval of answers to questions. On the one hand, since the retrieval object (i.e., the second text information) such as an article is generally relatively fixed known data stored in advance in, for example, a server database, and the data amount of the retrieval object tends to be very large, in order to improve the data processing efficiency, the semantic understanding operation (i.e., the acquisition or calculation of semantic feature vectors or semantic codes) of a large amount of the second text information (i.e., the retrieval object such as an article) is performed in advance while being online or offline, and the resulting plurality of second semantic feature vectors are stored in a preset semantic feature vector index library for semantic similarity calculation (with the first text information such as a problem to be retrieved) on-line when information retrieval occurs. Therefore, semantic coding operation by utilizing a semantic understanding model aiming at a huge amount of second text information in real time on line is avoided, so that the data processing efficiency is greatly improved (especially in the scene of a large data amount of objects or articles to be searched), and the network resource configuration and scheduling are remarkably optimized. On the other hand, with respect to the online or online operation part, three natural language processing models (namely a semantic understanding model, a reading understanding model and a text abstract model) based on a deep neural network are adopted to realize data processing processes such as article screening, fragment extraction, multi-text abstract generation and the like, so that the robustness and the accuracy of a search result or a search answer are effectively ensured.

Fig. 11 is an exemplary block diagram of a semantic understanding based information retrieval apparatus 1100 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 acquisition module 1110 may be configured to acquire first text information indicating a search target and a plurality of second text information indicating candidate search objects. The first determination module 1120 may be configured to determine a semantic similarity of the first text information to each of the plurality of second text information. The selection module 1130 may be configured to select at least one text message to be retrieved from the plurality of second text messages based on the semantic similarity of the first text message to each of the plurality of second text messages. The extraction module 1140 may be 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 set of text information. The second obtaining module 1150 may be configured to obtain multiple text summaries corresponding to at least two third text information in the third set of text information. The second determination module 1160 may be configured to determine a search result corresponding to the first text information based on the multiple text summaries.

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

In the information retrieval device according to some embodiments of the present application, firstly, a relatively small amount of text information to be retrieved is screened from a large amount of second text information according to the semantic similarity (instead of mere literal matching) of the first text information and the second text information, so that the overall working efficiency is remarkably improved under the condition of ensuring that important information retrieval resources with high correlation with retrieval problems (namely ensuring retrieval breadth and accuracy) are not lost, and the problems of important retrieval resources deficiency caused by accurate keyword matching and inefficiency caused by complex flow and huge calculation amount in the related technology are overcome; secondly, extracting or searching third text information (namely candidate answer related text information corresponding to the to-be-searched question) corresponding to the first text information (such as the to-be-searched question) again based on semantic relevance aiming at each to-be-searched text information obtained by screening in the first stage, so that the extraction of the candidate answer and related text (namely third text information set) thereof is realized by utilizing the semantic features in the to-be-searched question and the semantic features of the to-be-searched information again, and the higher relevance of the third text information set and the first text information and the higher accuracy of a final search result are further ensured; finally, generating multiple text summaries (e.g. generating text summaries) of at least two third text messages in the third text message set, namely generating multiple text summaries from a plurality of candidate answer related texts with high matching degree through summarization and reasoning modes as final answers (e.g. as a part of search results) of the questions to be searched, wherein the multiple text summaries can further improve the quality and accuracy of the search results due to the fact that a plurality of candidate answer related text messages with higher search matching degree (i.e. the third text messages) are fused.

Fig. 12 schematically illustrates an example block diagram of a computing device 1200 according to some embodiments of the application. Computing device 1200 may represent devices to implement the various apparatuses or modules described herein and/or to perform the 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 devices of various levels, from a full resource device with substantial storage and processing resources to a low resource device with limited storage and/or processing resources. In some embodiments, the semantic understanding based information retrieval apparatus 1100 described above with respect to fig. 11 may be implemented in one or more computing devices 1200, respectively.

As shown in FIG. 12, the 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 one another. A system bus may 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 processor or local bus that utilizes any of a variety of bus architectures. Alternatively, control and data lines, for example, may also be included.

The processing system 1201 is representative of functionality to perform one or more operations using hardware. Thus, the processing system 1201 is illustrated as including hardware elements 1204 that may be configured as processors, functional blocks, and the like. This may include implementation in hardware as an application specific integrated circuit or other logic device formed using one or more semiconductors. The hardware element 1204 is not limited by the materials from which it is formed or the processing mechanisms employed therein. For example, the processor may be comprised of semiconductor(s) and/or transistors (e.g., electronic Integrated Circuits (ICs)). In such a context, the processor-executable instructions may be electronically-executable instructions.

Computer readable media 1202 is illustrated as including memory/storage 1205. Memory/storage 1205 represents memory/storage associated with one or more computer-readable media. The memory/storage 1205 may include volatile media (such as Random Access Memory (RAM)) and/or nonvolatile media (such as Read Only Memory (ROM), flash memory, optical disks, magnetic disks, and so forth). The memory/storage 1205 may include fixed media (e.g., RAM, ROM, a fixed hard drive, etc.) and removable media (e.g., flash memory, a removable hard drive, an optical disk, and so forth). The memory/storage 1205 may be used to store data such as the first text information, the second text information, the third text information, the search result, and the like mentioned in the above embodiments, for example. The computer readable medium 1202 may be configured in a variety of other ways as described further below.

One or more I/O (input/output) interfaces 1203 represent functionality that allows a user to enter commands and information into computing device 1200 and that also allows information to be displayed to the user and/or sent to other components or devices using various input/output devices. Examples of input devices include a keyboard, a cursor control device (e.g., a mouse), a microphone (e.g., for voice input), a scanner, touch functionality (e.g., capacitive or other sensors configured to detect physical touches), a camera (e.g., motion that does not involve touches may be detected as gestures using visible or invisible wavelengths such as infrared frequencies), a network card, a receiver, and so forth. Examples of output devices include a display device, speakers, printer, haptic response device, network card, transmitter, and the like.

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

Various techniques may be described herein in the general context of software, hardware, elements, 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. The terms "module," "functionality," and the like as used herein generally represent software, firmware, hardware, or a combination thereof. The features of the techniques described herein are platform-independent, meaning that the techniques may be implemented on a variety of computing platforms having 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 can include a variety of media that are accessible by computing device 1200. By way of example, and not limitation, computer readable media may comprise "computer readable storage media" and "computer readable signal media".

"computer-readable storage medium" refers to a medium and/or device that can permanently store 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 hardware such as volatile and nonvolatile, removable and non-removable media and/or storage devices implemented in methods or techniques suitable for storage of information such as computer-readable instructions, data structures, program modules, logic elements/circuits or other data. Examples of a computer-readable storage medium may include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital Versatile Disks (DVD) or other optical storage, hard disk, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or other storage devices, tangible media, or articles of manufacture adapted to store the desired information and which may be accessed by a computer.

"computer-readable signal medium" refers to a signal bearing medium configured to transmit instructions to hardware of computing device 1200, such as via a network. Signal media may typically be embodied in 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 include any information delivery media. By way of example, and not limitation, signal media includes wired media such as a wired network or direct-wired connection, 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 may be used in some embodiments to implement at least some aspects of the techniques described herein. The hardware elements may include integrated circuits or components of a system on a chip, application Specific Integrated Circuits (ASICs), field Programmable Gate Arrays (FPGAs), complex Programmable Logic Devices (CPLDs), and other implementations in silicon or other hardware devices. In this context, the hardware elements may be implemented as processing devices that perform program tasks defined by instructions, modules, and/or logic embodied by the hardware elements, as well as hardware devices that store instructions for execution, such as the previously described computer-readable storage media.

Combinations of the foregoing may 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 embodied on some form of computer readable storage medium and/or by one or more hardware elements 1204. Computing device 1200 may be configured to implement particular instructions and/or functions corresponding to software and/or hardware modules. Thus, for example, by using the computer-readable storage medium of the processing system and/or the hardware elements 1204, a module may be implemented at least in part in hardware as a module executable by the computing device 1200 as software. The instructions and/or functions may be executed/operable by, for example, one or more computing devices 1200 and/or processing systems 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 techniques described herein.

In particular, according to embodiments of the present application, the processes described above with reference to flowcharts may be implemented as computer programs. For example, embodiments of the present application provide a computer program product comprising a computer program embodied on a computer readable medium, the computer program comprising program code for performing at least one step of the method embodiments of the present application.

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

In the description of the present specification, the terms "one embodiment," "some embodiments," "examples," "specific examples," or "some examples," etc. describe mean that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the present application. In this specification, schematic representations of the above terms are not necessarily directed to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples. Furthermore, the different embodiments or examples described in this specification and the features of the different embodiments or examples may be combined and combined by those skilled in the art without contradiction.

Any process or method descriptions in flow charts or otherwise described herein may be understood as representing modules, segments, or portions of code which include one or more executable instructions for implementing specific logical functions or steps of the process, and additional implementations are included within the scope of the preferred embodiment of the present application in which functions may be executed out of order from that shown or discussed, including substantially concurrently or in reverse order from that shown or discussed, depending on the functionality involved, as would be understood by those reasonably skilled in the art of the embodiments of the present application.

Logic and/or steps represented in the flowcharts or otherwise described herein, e.g., a ordered listing of executable instructions for implementing logical functions, can be embodied in any computer-readable medium for use by or in connection with an instruction execution system, apparatus, or device, such as a computer-based system, processor-containing system, or other system that can fetch the instructions from the instruction execution system, apparatus, or device and execute the instructions. For the purposes of this description, a "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device.

It is to be understood that portions of the present application may be implemented in hardware, software, firmware, or a combination thereof. In the above-described embodiments, the various steps or methods may be implemented in software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented in hardware, it may be implemented using any one or combination of the following techniques, as known in the art: discrete logic circuits having logic gates for implementing logic functions on data signals, application specific integrated circuits having suitable combinational logic gates, programmable gate arrays (Programmable Gate Array), field programmable gate arrays (Field Programmable Gate Array), and the like.

Those of ordinary skill in the art will appreciate that all or part of the steps of the methods of the above embodiments may be performed by hardware associated with program instructions, and the program may be stored in a computer readable storage medium, which when executed, includes performing one or a combination of the steps of the method embodiments.

In addition, each functional unit in each embodiment of the present application may be integrated in one processing module, or each unit may exist alone physically, or two or more units may be integrated in one module. The integrated modules may be implemented in hardware or in software functional modules. The integrated modules may also be stored in a computer readable storage medium if implemented in the form of software functional modules and sold or used as a stand-alone product.

Claims

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:

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.