Instant messaging message mixed retrieval evaluation method and device based on natural language understanding, equipment and medium

Abstract

The application discloses a natural language understanding-based instant messaging message hybrid retrieval evaluation method and device, equipment and medium, and relates to the technical field of message retrieval. The method comprises the following steps: in response to a query request issued by a user, acquiring a query vector of the query request and acquiring an instant messaging message; determining a message vector according to message text information and message meta attribute information in the instant messaging message; determining a keyword retrieval score according to a keyword matching degree between the instant messaging message and the query request; determining a vector retrieval score according to a similarity between the message vector and the query vector; and weighting and fusing the vector retrieval score and the keyword retrieval score according to a fusion weight corresponding to the query request to obtain a hybrid retrieval score of the instant messaging message and an evaluation result corresponding to the hybrid retrieval score.

Description

Technical Field

[0001] This application relates to the field of message retrieval technology, and in particular to a method, apparatus, device and medium for evaluating hybrid retrieval of instant messaging messages based on natural language understanding. Background Technology

[0002] In the daily use of instant messaging tools, users generate a large number of chat messages every day. Therefore, it is necessary to help users quickly locate important business information.

[0003] Current practices employ a single retrieval strategy, calculating the relevance of message content based on fixed matching rules and directly returning search results according to relevance. However, this approach fails to consider the diversity of user query intent. When users use vague descriptions or synonyms, the semantics are difficult to understand effectively; conversely, when users query precise identifiers, irrelevant content is easily mixed in, resulting in unsatisfactory search results in different scenarios. Therefore, accurately retrieving the required information from massive amounts of messages has become a pressing technical problem to be solved. Summary of the Invention

[0004] The main objective of this application is to provide a method, apparatus, device, and medium for hybrid retrieval and evaluation of instant messaging messages based on natural language understanding, aiming to solve the technical problem of how to accurately retrieve the required information from massive amounts of messages.

[0005] To achieve the above objectives, this application proposes a hybrid retrieval and evaluation method for instant messaging messages based on natural language understanding. The method includes: In response to a user's query request, obtain the query vector of the query request and obtain the instant messaging message; The message vector is determined based on the message text information and message meta-attribute information in the instant messaging message; The keyword retrieval score is determined based on the degree of keyword matching between the instant messaging message and the query request; The vector retrieval score is determined based on the similarity between the message vector and the query vector. Based on the fusion weight corresponding to the query request, the vector retrieval score and the keyword retrieval score are weighted and fused to obtain the mixed retrieval score of the instant messaging message, and the evaluation result corresponding to the mixed retrieval score.

[0006] In one embodiment, the step of determining the message vector based on the message text information and message meta-attribute information in the instant messaging message includes: The instant messaging messages are preprocessed to obtain separate messages, which include text, emoticons, mention messages, quote messages, links, and image descriptions. Based on the text, the emoji, the mention message, the reference message, the link, and the image description, determine the message text information and the message meta attribute information corresponding to the message text information; The message vector is obtained by fusing the message text information and the message meta-attribute information.

[0007] In one embodiment, the step of determining message text information and message meta-attribute information corresponding to the message text information based on the text, the emoji, the mention message, the reference message, the link, and the image description includes: The message text information is obtained by segmenting the text, emojis, mentions, quotations, links, and image descriptions. The message meta-attribute information is determined based on the sender identifier, receiver identifier, sending timestamp, message session, message type, and topic tag in the message text information.

[0008] In one embodiment, the step of fusing the message text information and the message meta-attribute information to obtain the message vector includes: The message text information is mapped into a text semantic vector based on a pre-trained language model; An initial meta-feature vector is constructed based on the message meta-attribute information, and the initial meta-feature vector is subjected to dimensionality reduction processing to obtain the meta-feature vector; The message vector is obtained by fusing the text semantic vector and the meta-feature vector.

[0009] In one embodiment, the step of determining the keyword retrieval score based on the keyword matching degree between the instant messaging message and the query request includes: The query requests are categorized to obtain a set of query keywords; The keyword retrieval score is obtained by calculating the degree of keyword matching between the instant messaging message and the set of query keywords.

[0010] In one embodiment, the step of weightedly fusing the vector retrieval score and the keyword retrieval score according to the fusion weight corresponding to the query request to obtain the mixed retrieval score of the instant messaging message and the evaluation result corresponding to the mixed retrieval score includes: The query request is input into a preset intent recognition model for intent classification to determine the query classification type, which includes keyword exact query type, semantic vector retrieval type and hybrid query type. The fusion weight is obtained based on the confidence levels of the keyword precise query type, the semantic vector retrieval type, and the mixed query type; The keyword retrieval score and the vector retrieval score are weighted and fused according to the fusion weight to obtain the hybrid retrieval score of the instant messaging message; The corresponding evaluation result is determined based on the hybrid retrieval score.

[0011] In one embodiment, the step of determining the corresponding evaluation result based on the hybrid retrieval score includes: Obtain the time interval between the current time and the message sending time, and calculate the message freshness decay coefficient based on the time interval; Obtain the sender authority coefficient, and based on the hybrid retrieval score, the sender authority coefficient, and the message freshness decay coefficient, obtain the retrieval ranking score of the instant messaging message; The instant messaging messages are sorted according to the search ranking score to obtain the ranking result, and the ranking result is used as the evaluation result.

[0012] Furthermore, to achieve the above objectives, this application also proposes a hybrid retrieval and evaluation device for instant messaging messages based on natural language understanding, the device comprising: The acquisition module is used to respond to a user's query request, acquire the query vector of the query request, and acquire the instant messaging message; The processing module is used to determine the message vector based on the message text information and message element attribute information in the instant messaging message; The retrieval module is used to determine the keyword retrieval score based on the degree of keyword matching between the instant messaging message and the query request; The retrieval module is further configured to determine a vector retrieval score based on the similarity between the message vector and the query vector; The evaluation module is used to perform weighted fusion of the vector retrieval score and the keyword retrieval score according to the fusion weight corresponding to the query request, so as to obtain the mixed retrieval score of the instant messaging message and the evaluation result corresponding to the mixed retrieval score.

[0013] Furthermore, to achieve the above objectives, this application also proposes a hybrid retrieval and evaluation device for instant messaging messages based on natural language understanding. The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. The computer program is configured to implement the steps of the hybrid retrieval and evaluation method for instant messaging messages based on natural language understanding as described above.

[0014] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the hybrid retrieval and evaluation method for instant messaging messages based on natural language understanding as described above.

[0015] One or more technical solutions proposed in this application have at least the following technical effects: This embodiment proposes a hybrid retrieval and evaluation method for instant messaging messages based on natural language understanding. This application responds to user queries by obtaining the query vector and the instant messaging message. It constructs a message vector based on the message text and metadata of the instant messaging message to fully represent the message's content and attribute features. A keyword retrieval score is determined based on the keyword matching degree between the instant messaging message and the query request to preserve precise matching capabilities. A vector retrieval score is determined based on the similarity between the message vector and the query vector to enhance semantic understanding. The vector retrieval score and keyword retrieval score are weighted and fused according to the fusion weight corresponding to the query request to obtain a mixed retrieval score for the instant messaging message and the corresponding evaluation result. This approach balances precise retrieval and semantic recall in different query scenarios, better adapting to diverse user query needs and effectively improving the accuracy of message retrieval. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating an embodiment of the hybrid retrieval and evaluation method for instant messaging messages based on natural language understanding in this application. Figure 2 This is a flowchart illustrating Embodiment 2 of the hybrid retrieval and evaluation method for instant messaging messages based on natural language understanding in this application. Figure 3 This is a schematic diagram of the module structure of the real-time messaging hybrid retrieval and evaluation device based on natural language understanding, as described in an embodiment of this application. Figure 4This is a schematic diagram of the device structure of the hardware operating environment involved in the hybrid retrieval and evaluation method for instant messaging messages based on natural language understanding in the embodiments of this application.

[0019] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0020] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0021] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0022] The main solution of this application embodiment is as follows: In response to a user's query request, obtain the query vector of the query request and obtain the instant messaging message; determine the message vector based on the message text information and message meta-attribute information in the instant messaging message; determine the keyword retrieval score based on the keyword matching degree between the instant messaging message and the query request; determine the vector retrieval score based on the similarity between the message vector and the query vector; and perform weighted fusion of the vector retrieval score and the keyword retrieval score according to the fusion weight corresponding to the query request to obtain the mixed retrieval score of the instant messaging message and the evaluation result corresponding to the mixed retrieval score.

[0023] In this embodiment, for ease of description, the following description will focus on the evaluation device for identifying instant messaging messages based on natural language understanding.

[0024] Because existing technologies do not take into account the differences in users' query intent, it is difficult to effectively understand the semantics when users use vague descriptions or synonyms, and when users query precise identifiers, irrelevant content is easily mixed in, resulting in unsatisfactory search results in different scenarios.

[0025] This application provides a solution that, in response to a user's query request, obtains the query vector of the query request and the instant messaging message. It constructs a message vector based on the message text and metadata of the instant messaging message to fully represent the content and attribute features of the message. A keyword retrieval score is determined based on the keyword matching degree between the instant messaging message and the query request to retain precise matching capability. A vector retrieval score is determined based on the similarity between the message vector and the query vector to enhance semantic understanding capability. Based on the fusion weight corresponding to the query request, the vector retrieval score and the keyword retrieval score are weighted and fused to obtain a mixed retrieval score for the instant messaging message and the corresponding evaluation result. This approach balances precise retrieval and semantic recall in different query scenarios, better adapts to diverse user query needs, and effectively improves the accuracy of message retrieval.

[0026] Based on this, embodiments of this application provide a hybrid retrieval and evaluation method for instant messaging messages based on natural language understanding, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the hybrid retrieval and evaluation method for instant messaging messages based on natural language understanding in this application.

[0027] In this embodiment, the instant messaging message hybrid retrieval and evaluation method based on natural language understanding includes steps S10 to S50: Step S10: In response to the query request issued by the user, obtain the query vector of the query request and obtain the instant messaging message; It should be noted that the query request is natural language text input by the user, used to represent the retrieval intent; the query vector is a high-dimensional semantic vector obtained by encoding the query request through a pre-trained language model, used to represent the semantic features of the query request; and the instant messaging message is the historical chat message to be retrieved, which is a candidate message set consisting of message text information (text, emoticons, mentions, quotes, links, image descriptions, etc.) and message meta-attribute information (sender identifier, receiver identifier, sending timestamp, message to which the message belongs, message type, topic tags, etc.).

[0028] In a specific embodiment, when a user inputs a query request through an instant messaging client and triggers a retrieval operation, the system captures the query request in real time, uses a pre-trained language model (such as BERT or BGE-M3) to semantically encode the query request, and generates a high-dimensional query vector to represent the semantics of the query. At the same time, the system retrieves the instant messaging message to be retrieved from the instant messaging message repository or index. The instant messaging message may include text content, as well as meta-attribute information such as sender identifier, receiver identifier, sending timestamp, message session, message type, and topic tags, so that the system can understand the user's fuzzy expression from a semantic level and improve the recall and accuracy of the retrieval results.

[0029] Step S20: Determine the message vector based on the message text information and message meta attribute information in the instant messaging message; It should be noted that the message vector is a high-dimensional vector representation that integrates the semantics of the message text and the characteristics of message meta-attributes, in order to improve the discriminability of the vector in the semantic space and fully represent the complete meaning of instant messaging messages in a specific business context.

[0030] Understandably, message text information refers to the content of the message that can be understood by natural language, including the text after word segmentation, the semantic description of emoticons, @mention information, quoted message content, link text and image descriptions, etc. Message meta-attribute information is structured data that represents the source and context of the message, including sender identifier, receiver identifier, sending timestamp, message session ID, message type and topic tags, etc.

[0031] In a specific embodiment, the instant messaging messages are preprocessed to obtain separated messages. The separated messages include text, emoticons, mentions, quotes, links, and image descriptions. That is, the instant messaging messages are cleaned, standardized, and feature extracted by the message preprocessing module. For example, the special formats of IM messages (including text, emoticons, @mentions, quotes, links, image descriptions, etc.) are identified and separated to obtain separated messages. The separated messages include text, emoticons, mentions, quotes, links, and image descriptions.

[0032] The message text information is obtained by segmenting the text, emoticons, mentions, quotations, links, and image descriptions. Specifically, the message text information can be segmented using a dictionary-based forward maximum matching algorithm, combined with a Hidden Markov Model (HMM) to identify out-of-vocabulary words. For example, a dictionary can be pre-built with a maximum word length (e.g., 5 characters). During segmentation, starting from the current position in the text, the longest substring is matched against the dictionary. If a match is found, it is segmented into a word; otherwise, the length is gradually reduced. The matching continues until a match is successful or a single character is formed. This forward scanning from left to right completes the initial segmentation, resulting in a coarse segmentation result based primarily on dictionary words. Since IM messages often contain out-of-vocabulary words such as names of people, product names, and business terms, relying solely on the dictionary cannot segment them correctly. Therefore, in the coarse segmentation result, continuous single-character fragments or character sequences not covered by the dictionary are extracted and input into a Hidden Markov Model (HMM) for recognition. The HMM labels each character as the beginning of a word (B), the middle of a word (M), the end of a word (E), or a single-character word (S). The Viterbi algorithm is used to solve for the most likely labeled sequence, thereby combining these out-of-vocabulary character fragments into complete words. The word segmentation result matched by the dictionary is then merged with the out-of-vocabulary words identified by the HMM to obtain the message text information.

[0033] Based on the sender identifier, receiver identifier, sending timestamp, message session, message type, and topic tag in the message text information, the message meta-attribute information is determined. That is, at this time, the preprocessing module can extract structured fields such as sender identifier, receiver identifier, sending timestamp, message session ID, message type, and topic tag from the message text information as message meta-attribute information.

[0034] The message text information is mapped to a text semantic vector based on a pre-trained language model. An initial meta-feature vector is constructed based on the message meta-attribute information, and the initial meta-feature vector is dimensionality-reduced to obtain a meta-feature vector. The text semantic vector and the meta-feature vector are fused to obtain the message vector. In other words, the multi-dimensional vectorization engine is specifically optimized for the short text characteristics of IM messages. It uses a pre-trained language model (such as BERT, BGE-M3, etc.) as the basic encoder to map the message text into a high-dimensional text semantic vector. Unlike existing technologies that only use text content to generate vectors, this application adopts a multi-dimensional feature fusion strategy, that is, concatenating the text semantic vector with the message's meta-feature vector (sender identity, publication time, message type, etc.), and performing feature fusion through a fully connected layer to generate a message vector representation, as follows:

[0035] in, It is a text semantic vector (768 dimensions). The meta-feature vector is obtained by reducing the dimensionality of the initial meta-feature vector using PCA principal component analysis. The initial meta-feature vector is obtained from the message meta-attribute information, which includes 32 dimensions of sender identity, 16 dimensions of publication time, and 8 dimensions of message type, concatenated into a 56-dimensional vector. This represents vector concatenation. The fusion weight matrix is ​​(768×824). It is the bias vector (768 dimensions).

[0036] For special formats in IM messages, these special formats can be semantically represented. For example, for emoticons, they can be converted into corresponding semantic description text (such as converting a smiling emoticon into a friendly emoticon). For @ mentions, the identity features of the mentioned user can be extracted. For quoted messages, the citation relationship can be parsed and the key information of the original message can be integrated into the vector representation of the current message. This processing method can significantly improve the semantic distinguishability of the vector representation.

[0037] To more clearly demonstrate the differences between this application and existing technologies, please refer to the comparison description: Current technologies only use pre-trained language models (such as BERT) to map text content into semantic vectors. The input is plain text, and the output is a fixed-dimensional semantic vector (usually 768-dimensional), which cannot adequately distinguish IM messages in different contexts.

[0038] This application can concatenate and fuse text semantic vectors with message meta-feature vectors, fully considering the multi-dimensional features of IM messages and improving the semantic distinguishability of the vectors. The actual comparison can then be expressed as:

[0039] At this point, a semantic recall comparison experiment can be conducted to verify the semantic discriminative ability. It should be understood that semantic recall measures the retrieval system's ability to find relevant messages, which can clearly demonstrate the actual differences. The experimental results at this point are as follows:

[0040] The experimental conditions were as follows: on a dataset of 100,000 messages, 200 test queries were evaluated, and each query returned Top-5 results. Semantic recall was defined as the proportion of messages that were relevant to the query intent among the 5 returned results.

[0041] In one feasible implementation, step S20 may include steps A11 to A13: Step A11: Preprocess the instant messaging message to obtain a separate message, which includes text, emoticons, mentions, quotes, links, and image descriptions. It should be noted that the separated message is an intermediate result obtained by identifying the content type of instant messaging messages. It may include different components such as text, emoticons, mentions, quotes, links, and image descriptions. Each component is extracted and labeled independently for separate processing.

[0042] Step A12: Determine message text information and message meta attribute information corresponding to the message text information based on the text, the emoji, the mention message, the reference message, the link, and the image description; It is understandable that text, emoticons, mentions, quotes, links, and image descriptions are information elements in a message that carry different semantic content. Among them, text is natural language content that can be segmented, emoticons are graphical expressions of emotion, mentions indicate the user being @'d and their corresponding identity information, quotes indicate historical messages referenced by the current message and their referencing relationships, links indicate URL addresses, and image descriptions are textual descriptions of the image content.

[0043] In one feasible implementation, step A12 may include steps B11-B12: Step B11: Perform word segmentation on the text, the emoji, the mention message, the quote message, the link, and the image description to obtain the message text information; Understandably, word segmentation is a process of adopting differentiated strategies for different content types. For the text portion, a dictionary-based forward maximum matching algorithm is used, combined with a Hidden Markov Model (HMM) to identify out-of-vocabulary words, completing the initial word segmentation. For emoticons, they are converted into corresponding semantic description text through a predefined mapping table before being added to the word segmentation sequence. For mention messages, the nickname or ID of the user being @ is extracted and marked as a dedicated term. For quoted messages, the original message text being quoted is parsed, and key content is extracted for word segmentation. For links, the domain name and path keywords in the URL are extracted, or descriptive text is generated based on the webpage title before being used for word segmentation. For image descriptions, the text description generated by image recognition is treated as independent text for word segmentation. Thus, all word segmentation results are merged in order to obtain the message text information.

[0044] Step B12: Determine message meta-attribute information based on the sender identifier, receiver identifier, sending timestamp, message session, message type, and topic tag in the message text information.

[0045] Understandably, the sender identifier is used to uniquely identify the sender of the message, the receiver identifier is used to identify the receiver of the message, the sending timestamp records the precise time the message occurred, the message to which the message belongs is used to indicate the dialogue context in which the message is located, the message type is used to distinguish different categories such as ordinary text messages, image messages, system messages, @ reminder messages, etc., and the topic tag identifies the topic or discussion group associated with the message.

[0046] Step A13: Perform fusion processing on the message text information and the message meta-attribute information to obtain the message vector.

[0047] Understandably, fusion processing is the process of mapping message information from different modalities to the same vector space. The goal is to make the generated message vector contain both the semantic features of the message content and the contextual attributes of the message, thereby representing a message more accurately.

[0048] In one feasible implementation, step A13 may include steps C11 to C13: Step C11: Map the message text information into a text semantic vector based on the pre-trained language model; It should be noted that the text semantic vector is a dense numerical vector obtained by encoding the natural language content in the message text information through a pre-trained language model. It includes the contextual semantics, syntactic structure, and pragmatic features of words in the message and can represent the semantic similarity of the message text in the vector space.

[0049] Understandably, pre-trained language models are deep neural network models trained in a self-supervised manner on large-scale general corpora, such as BERT, RoBERTa, and ELECTRA. They utilize the Transformer architecture to capture bidirectional contextual information and learn rich language knowledge through tasks such as masked language modeling. After inputting message text information into the model, the average value of the last hidden state can be taken as the corresponding output as the text semantic vector of the message, thereby completing the mapping from discrete text to continuous semantic vectors.

[0050] Step C12: Construct an initial meta-feature vector based on the message meta-attribute information, and perform dimensionality reduction processing on the initial meta-feature vector to obtain the meta-feature vector; It should be noted that the meta-feature vector is a low-dimensional dense vector formed by numerically encoding the meta-attribute information of the message. That is, it is a feature representation of structured attributes such as message sender, receiver, sending time, session, message type and topic tags, which is used to supplement the context information beyond the text content.

[0051] Understandably, dimensionality reduction is used to address the issues of high dimensionality, sparse data, and redundant features in the initial meta-feature vector. Since the initial meta-feature vector is usually formed by concatenating categorical attributes through one-hot encoding or embedding encoding, its dimensionality may reach hundreds or even thousands of dimensions, and many dimensions take zero values. Direct use will directly lead to overfitting. Therefore, it can be mapped to a low-dimensional space of tens of dimensions to improve the efficiency of fusion training while retaining the main information.

[0052] Step C13: Perform fusion processing on the text semantic vector and the meta-feature vector to obtain the message vector.

[0053] It is understandable that the message vector can be used to fully depict the entirety of the message. For example, in similar message retrieval, relying solely on text vectors may make the same sentence completely similar, but by fusing meta-features, it is possible to distinguish the differences in intent among different users and in different scenarios, thereby improving the accuracy of retrieval.

[0054] Step S30: Determine the keyword retrieval score based on the keyword matching degree between the instant messaging message and the query request; It should be noted that the keyword retrieval score is a quantitative value used to measure the lexical matching degree between instant messaging messages and user queries. It represents the matching degree between the word segmentation results in the message text information and the query keywords. The higher the keyword matching degree, the more query keywords are contained in the message, and the higher the keyword retrieval score.

[0055] In a specific embodiment, the query requests are classified to obtain a set of query keywords. That is, the query types corresponding to the query requests are divided into three categories: keyword precise query type, semantic vector retrieval type, and mixed query type. Therefore, classifying the query requests input by the user yields results under different types. For example, the query keyword set under the keyword precise query type is work order number, order number, and terminology; the query keyword set under the semantic vector retrieval type is the project discussed last week; and the query keyword set under the mixed query type is the various identified keywords. It should be understood that precise queries mainly rely on keyword retrieval, fuzzy queries mainly rely on vector retrieval, and mixed queries require a combination of both.

[0056] The keyword retrieval score is calculated based on the keyword matching degree between the instant messaging message and the query keyword set. The keyword retrieval score is expressed as follows:

[0057] in, The score represents the keyword retrieval score calculated for the instant messaging message doc against the query keyword set Q. TF (Term Frequency) represents the frequency of the query keyword in the message, and IDF (Inverse Document Frequency) measures the distinguishing ability of the keyword in the entire message corpus. doc is the instant messaging message, and Q is the query keyword set, thus obtaining the degree of matching between the message and the query at the term level. A higher score indicates better keyword coverage.

[0058] In one feasible implementation, step S30 may include steps D11-D12: Step D11: Classify the query requests to obtain a set of query keywords; It should be noted that the query keyword set is a collection of keywords extracted from the user's query request, which may consist of core nouns, verbs, qualifiers and their synonyms in the query.

[0059] Understandably, classification processing involves categorizing queries into different types such as inquiry, instruction, topic, and casual conversation based on their sentence structure, intent, or domain. Different keyword extraction strategies are then employed for each category. For example, inquiry queries focus on extracting key words and entity words, topic queries focus on extracting topic-centric words, and instruction queries focus on extracting action objects and operation goals. This avoids using uniform extraction rules that could lead to inaccurate keywords or omissions of key information, resulting in a more accurate set of query keywords that better reflects the user's true intent.

[0060] Step D12: Calculate the keyword retrieval score based on the keyword matching degree between the instant messaging message and the query keyword set.

[0061] It should be noted that the confidence weight is a numerical value used to quantify the importance of each keyword in the query keyword set to the query intent, representing the difference in contribution of different keywords in the retrieval.

[0062] Step S40: Determine the vector retrieval score based on the similarity between the message vector and the query vector; It should be noted that the vector retrieval score is a numerical value used to quantify the degree of matching between instant messaging messages and user query requests in the semantic space. By encoding the message text and query text into high-dimensional vectors respectively, the similarity between the two is calculated, and the similarity is mapped to a comparable score.

[0063] Understandably, the similarity between message vectors and query vectors is a geometric measure of the relationship between the vector representations obtained after semantically encoding instant messaging messages and query requests, respectively. The higher the similarity, the closer the message is to the query intent in terms of overall semantics, and the higher the vector retrieval score. This score can effectively compensate for the shortcomings of keyword matching in dealing with synonym replacement, word order adjustment and implicit semantic expression, thereby improving the rationality of the ranking of search results.

[0064] In a specific embodiment, the vector retrieval score can be expressed as:

[0065] in, Represents a message vector. This represents the query vector, at this point, For keyword retrieval scores, This is a score indicating the similarity between the message vector and the query vector.

[0066] Essentially, it calculates the cosine similarity between the message vector and the query vector. By dividing the inner product of the vectors by the product of their magnitudes, the influence of text length and vector scale is eliminated, and only the semantic consistency in direction is preserved. The closer the score is to 1, the more consistent the direction of the message and the query in the deep semantic space, and the higher the degree of semantic matching.

[0067] Step S50: Based on the fusion weight corresponding to the query request, the vector retrieval score and the keyword retrieval score are weighted and fused to obtain the mixed retrieval score of the instant messaging message and the evaluation result corresponding to the mixed retrieval score.

[0068] In a specific embodiment, the query request is input into a preset intent recognition model for intent classification to determine the query classification type, which includes keyword precise query type, semantic vector retrieval type, and hybrid query type. A fusion weight is obtained based on the confidence levels of the keyword precise query type, the semantic vector retrieval type, and the hybrid query type. The keyword retrieval score and the vector retrieval score are weighted and fused according to the fusion weight to obtain the hybrid retrieval score of the instant messaging message. The time interval between the current time and the message sending time is obtained, and a message freshness decay coefficient is obtained based on the time interval. The sender authority coefficient is obtained, and the retrieval ranking score of the instant messaging message is obtained based on the hybrid retrieval score, the sender authority coefficient, and the message freshness decay coefficient. The instant messaging messages are ranked according to the retrieval ranking score to obtain the ranking result, and the ranking result is used as the evaluation result.

[0069] This embodiment proposes a hybrid retrieval and evaluation method for instant messaging messages based on natural language understanding. In response to a user's query request, this application obtains the query vector of the query request and the instant messaging message. A message vector is constructed based on the message text information and message meta-attribute information in the instant messaging message to fully represent the content and attribute features of the message. A keyword retrieval score is determined based on the keyword matching degree between the instant messaging message and the query request to retain precise matching capability. A vector retrieval score is determined based on the similarity between the message vector and the query vector to enhance semantic understanding capability. Based on the fusion weight corresponding to the query request, the vector retrieval score and the keyword retrieval score are weighted and fused to obtain the hybrid retrieval score of the instant messaging message and the corresponding evaluation result. This method balances precise retrieval and semantic recall in different query scenarios, better adapts to diverse user query needs, and effectively improves the accuracy of message retrieval.

[0070] Based on the first embodiment of this application, in the second embodiment of this application, the same or similar content as the first embodiment can be referred to the above description, and will not be repeated hereafter.

[0071] In this embodiment, refer to Figure 2 , Figure 2 The flowchart provided for Embodiment 2 of the hybrid retrieval and evaluation method for instant messaging messages based on natural language understanding in this application also includes steps S51 to S54: Step S51: Input the query request into a preset intent recognition model for intent classification and determine the query classification type. The query classification type includes keyword precise query type, semantic vector retrieval type and hybrid query type. It should be noted that the query classification type is a retrieval strategy category obtained after the preset intent recognition model classifies the query request. Among them, the keyword precise query type refers to the query request being mainly composed of explicit proper nouns or terms, which is suitable for keyword precise matching as the core. The semantic vector retrieval type refers to the query request being colloquial or semantically ambiguous, which is suitable for vector semantic similarity as the core. The hybrid query type requires the fusion of keyword retrieval and vector retrieval.

[0072] It is understandable that the intent recognition model is a general text classification model trained on labeled samples. It can be a general model built on deep semantic classification networks such as BERT and TextCNN. By encoding and mapping the semantic features, syntactic structure and entity information of the query request, the system can dynamically balance the ability of accurate matching and semantic generalization according to the query characteristics.

[0073] In a specific embodiment, the query request can be input into a preset intent recognition model for intent classification. The query classification type is identified based on the classification confidence level output by the intent recognition model. For example, when the intent recognition model determines that the confidence level of the keyword exact query is 1, the query classification type is keyword-oriented retrieval, i.e., keyword exact query type. When the intent recognition model determines that the confidence level of the keyword exact query is 0, the query classification type is vector-oriented retrieval, i.e., semantic vector retrieval type. When the confidence level is between 0 and 1, the query classification type is mixed retrieval, i.e., mixed query type.

[0074] Step S52: Obtain the fusion weight based on the confidence levels of the keyword precise query type, the semantic vector retrieval type, and the mixed query type; It should be noted that the fusion weight is a weighted fusion ratio that combines keyword retrieval scores and vector retrieval scores under mixed query types.

[0075] Understandably, confidence is the probability value output by the pre-defined intent recognition model during classification. It is used to measure the credibility of a query request belonging to the keyword exact query type, the semantic vector retrieval type, and the mixed query type. The higher the confidence, the more reliable the classification result, so that the retrieval strategy can adaptively respond to different query expressions, thereby ensuring exact matching while taking into account semantic relevance and improving retrieval accuracy.

[0076] In a specific embodiment, the fusion weight can be obtained based on the confidence levels of the keyword precise query type, the semantic vector retrieval type, and the mixed query type. For example, when performing a keyword exact query, set the fusion weight. When the retrieval type is semantic vector, set the fusion weight. When the query type is mixed, set the fusion weight. .

[0077] Step S53: The keyword retrieval score and the vector retrieval score are weighted and fused according to the fusion weight to obtain the mixed retrieval score of the instant messaging message; It should be noted that the hybrid retrieval score is a comprehensive score obtained by weighting and summing the keyword retrieval score and the vector retrieval score according to their respective preset fusion weights, which represents the retrieval fit of two dimensions: precise matching at the term level and approximate matching at the semantic level.

[0078] In a specific embodiment, the keyword retrieval score can be expressed as:

[0079] in, For keyword retrieval scores, For vector retrieval scores, These are the weighting coefficients. This is a score for mixed retrieval.

[0080] At this point, the hybrid retrieval score can accurately characterize the relevance between the message and the query, thereby improving the quality of the retrieval recall phase.

[0081] Step S54: Determine the corresponding evaluation result based on the mixed retrieval score.

[0082] It should be noted that the evaluation result is a conclusion on whether the instant messaging message matches the query intent, and is used to determine the output sorting position of the instant messaging message.

[0083] In a specific embodiment, the time interval between the current time and the message sending time is obtained, and a message freshness decay coefficient is obtained based on the time interval; the sender authority coefficient is obtained, and the retrieval ranking score of the instant messaging message is obtained based on the mixed retrieval score, the sender authority coefficient, and the message freshness decay coefficient; the instant messaging messages are sorted according to the retrieval ranking score to obtain a ranking result, and the ranking result is used as the evaluation result, that is, the evaluation result can be the evaluation report corresponding to the final calculated score. At this time, the final calculated score, i.e., the retrieval accuracy, is expressed as:

[0084]

[0085] in, This represents the sender's authority level (administrators / core personnel have higher values). Represents the message freshness decay coefficient. The time interval between the current time and the message sending time. This is a preset time decay hyperparameter; the larger the value, the faster the message decays.

[0086] It should be understood that retrieval accuracy measures the proportion of relevant information in the results returned by the retrieval system. This application shows a significant improvement over traditional methods. For example, the evaluation results are expressed as follows:

[0087] The experiment was conducted on a dataset of 100,000 messages, with 200 test queries evaluated. The retrieval accuracy was defined as the percentage of actual relevant messages among the five returned results.

[0088] In one feasible implementation, step S54 may include steps E11 to E13: Step E11: Obtain the time interval between the current time and the message sending time, and obtain the message freshness decay coefficient based on the time interval; It should be noted that the time interval between the current time and the message sending time is the time difference between the current moment and the message sending time, used to characterize the freshness of the message. The message freshness decay coefficient is a coefficient obtained based on this time interval to measure the impact of message freshness on search ranking. The larger the time interval, the smaller the coefficient, thereby reducing the weight of older messages in the ranking and making newer messages appear relatively higher.

[0089] Step E12: Obtain the sender authority coefficient. Based on the hybrid retrieval score, the sender authority coefficient, and the message freshness decay coefficient, obtain the retrieval ranking score of the instant messaging message. It should be noted that the sender authority coefficient is a coefficient used to adjust the priority of messages sent by different senders in the search ranking, and the search ranking score of the instant messaging message is a score calculated comprehensively to indicate the display priority of the instant messaging message under the current query.

[0090] Step E13: Sort the instant messaging messages according to the search ranking score to obtain the ranking result, and use the ranking result as the evaluation result.

[0091] It should be noted that the sorting result is a message order list formed by arranging all candidate instant messaging messages from high to low according to their search ranking scores. The message order in this list represents the system's evaluation conclusion on the query request and can be directly output.

[0092] This embodiment proposes a hybrid retrieval evaluation method for instant messaging messages based on natural language understanding. This application classifies the query request into a preset intent recognition model to determine one of three types: keyword exact match, semantic vector retrieval, or hybrid query. A fusion weight is then obtained to weight and fuse the keyword retrieval score and the vector retrieval score, resulting in a hybrid retrieval score for the instant messaging message. The evaluation result is determined based on the hybrid retrieval score, and the fusion weight is dynamically determined according to the query intent. This allows keyword exact match and semantic vector retrieval to be adaptively combined in the hybrid retrieval based on confidence levels, thus taking into account the advantages of both retrieval methods and improving the accuracy of instant messaging message retrieval results.

[0093] This application also provides a hybrid retrieval and evaluation device for instant messaging messages based on natural language understanding. Please refer to [link / reference]. Figure 3 The instant messaging hybrid retrieval and evaluation device based on natural language understanding includes: The acquisition module 10 is used to respond to a query request issued by a user, acquire the query vector of the query request, and acquire an instant messaging message; Processing module 20 is used to determine the message vector based on the message text information and message element attribute information in the instant messaging message; The retrieval module 30 is used to determine a keyword retrieval score based on the degree of keyword matching between the instant messaging message and the query request; The retrieval module 30 is further configured to determine a vector retrieval score based on the similarity between the message vector and the query vector; Evaluation module 40 is used to perform weighted fusion of the vector retrieval score and the keyword retrieval score according to the fusion weight corresponding to the query request, so as to obtain the mixed retrieval score of the instant messaging message and the evaluation result corresponding to the mixed retrieval score.

[0094] The processing module 20 is further configured to preprocess the instant messaging message to obtain a separate message, which includes text, emoticons, mentions, quotes, links, and image descriptions. Based on the text, the emoji, the mention message, the reference message, the link, and the image description, determine the message text information and the message meta attribute information corresponding to the message text information; The message vector is obtained by fusing the message text information and the message meta-attribute information.

[0095] The processing module 20 is further configured to perform word segmentation processing based on the text, the emoji, the mention message, the quote message, the link, and the image description to obtain message text information; The message meta-attribute information is determined based on the sender identifier, receiver identifier, sending timestamp, message session, message type, and topic tag in the message text information.

[0096] The processing module 20 is further configured to map the message text information into a text semantic vector based on a pre-trained language model; An initial meta-feature vector is constructed based on the message meta-attribute information, and the initial meta-feature vector is subjected to dimensionality reduction processing to obtain the meta-feature vector; The message vector is obtained by fusing the text semantic vector and the meta-feature vector.

[0097] The retrieval module 30 is also used to classify the query request to obtain a set of query keywords; The keyword retrieval score is obtained by calculating the degree of keyword matching between the instant messaging message and the set of query keywords.

[0098] The evaluation module 40 is also used to input the query request into a preset intent recognition model for intent classification and to determine the query classification type. The query classification type includes keyword precise query type, semantic vector retrieval type and hybrid query type. The fusion weight is obtained based on the confidence levels of the keyword precise query type, the semantic vector retrieval type, and the mixed query type; The keyword retrieval score and the vector retrieval score are weighted and fused according to the fusion weight to obtain the hybrid retrieval score of the instant messaging message; The corresponding evaluation result is determined based on the hybrid retrieval score.

[0099] The evaluation module 40 is also used to obtain the time interval between the current time and the message sending time, and to obtain the message freshness decay coefficient based on the time interval; Obtain the sender authority coefficient, and based on the hybrid retrieval score, the sender authority coefficient, and the message freshness decay coefficient, obtain the retrieval ranking score of the instant messaging message; The instant messaging messages are sorted according to the search ranking score to obtain the ranking result, and the ranking result is used as the evaluation result.

[0100] The instant messaging hybrid retrieval and evaluation device based on natural language understanding provided in this application, employing the instant messaging hybrid retrieval and evaluation method based on natural language understanding in the above embodiments, can solve the technical problem of how to accurately retrieve the required information from massive amounts of messages. Compared with the prior art, the beneficial effects of the instant messaging hybrid retrieval and evaluation device based on natural language understanding provided in this application are the same as those of the instant messaging hybrid retrieval and evaluation method based on natural language understanding provided in the above embodiments, and other technical features in the instant messaging hybrid retrieval and evaluation device based on natural language understanding are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0101] This application provides a hybrid retrieval and evaluation device for instant messaging messages based on natural language understanding. The device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform the hybrid retrieval and evaluation method for instant messaging messages based on natural language understanding in the first embodiment described above.

[0102] The following is for reference. Figure 4This document illustrates a structural schematic diagram of a hybrid retrieval and evaluation device for instant messaging messages based on natural language understanding, suitable for implementing embodiments of this application. The hybrid retrieval and evaluation device for instant messaging messages based on natural language understanding in embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 4 The illustrated instant messaging hybrid retrieval evaluation device based on natural language understanding is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0103] like Figure 4 As shown, the real-time messaging hybrid retrieval evaluation device based on natural language understanding may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in ROM (Read Only Memory) 1002 or a program loaded from storage device 1003 into RAM (Random Access Memory) 1004. RAM 1004 also stores various programs and data required for the operation of the real-time messaging hybrid retrieval evaluation device based on natural language understanding. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via bus 1005. Input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the natural language understanding-based instant messaging hybrid retrieval evaluation device to communicate wirelessly or wiredly with other devices to exchange data. Although the figure shows a natural language understanding-based instant messaging hybrid retrieval evaluation device with various systems, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems can be implemented alternatively.

[0104] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0105] The instant messaging hybrid retrieval and evaluation device based on natural language understanding provided in this application, employing the instant messaging hybrid retrieval and evaluation method based on natural language understanding in the above embodiments, can solve the technical problem of how to accurately retrieve the required information from massive amounts of messages. Compared with the prior art, the beneficial effects of the instant messaging hybrid retrieval and evaluation device based on natural language understanding provided in this application are the same as those of the instant messaging hybrid retrieval and evaluation method based on natural language understanding provided in the above embodiments, and other technical features in this instant messaging hybrid retrieval and evaluation device based on natural language understanding are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0106] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0107] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0108] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the instant messaging message hybrid retrieval and evaluation method based on natural language understanding in the above embodiments.

[0109] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0110] The aforementioned computer-readable storage medium may be included in the hybrid retrieval evaluation device for instant messaging based on natural language understanding; or it may exist independently and not be assembled into the hybrid retrieval evaluation device for instant messaging based on natural language understanding.

[0111] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by a natural language understanding-based instant messaging hybrid retrieval evaluation device, the device performs the following actions in response to a user's query request: obtaining a query vector for the query request and obtaining an instant messaging message; determining a message vector based on the message text information and message meta-attribute information in the instant messaging message; determining a keyword retrieval score based on the keyword matching degree between the instant messaging message and the query request; determining a vector retrieval score based on the similarity between the message vector and the query vector; and weightedly fusing the vector retrieval score and the keyword retrieval score according to the fusion weight corresponding to the query request to obtain a hybrid retrieval score for the instant messaging message and an evaluation result corresponding to the hybrid retrieval score.

[0112] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0113] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0114] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0115] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described hybrid retrieval and evaluation method for instant messaging messages based on natural language understanding. This method can solve the technical problem of accurately retrieving required information from massive amounts of messages. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the hybrid retrieval and evaluation method for instant messaging messages based on natural language understanding provided in the above embodiments, and will not be repeated here.

[0116] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A hybrid retrieval and evaluation method for instant messaging messages based on natural language understanding, characterized in that, The method includes: In response to a user's query request, obtain the query vector of the query request and obtain the instant messaging message; The message vector is determined based on the message text information and message meta-attribute information in the instant messaging message; The keyword retrieval score is determined based on the degree of keyword matching between the instant messaging message and the query request; The vector retrieval score is determined based on the similarity between the message vector and the query vector. Based on the fusion weight corresponding to the query request, the vector retrieval score and the keyword retrieval score are weighted and fused to obtain the mixed retrieval score of the instant messaging message, and the evaluation result corresponding to the mixed retrieval score.

2. The method as described in claim 1, characterized in that, The step of determining the message vector based on the message text information and message meta-attribute information in the instant messaging message includes: The instant messaging messages are preprocessed to obtain separate messages, which include text, emoticons, mention messages, quote messages, links, and image descriptions. Based on the text, the emoji, the mention message, the reference message, the link, and the image description, determine the message text information and the message meta attribute information corresponding to the message text information; The message vector is obtained by fusing the message text information and the message meta-attribute information.

3. The method as described in claim 2, characterized in that, The step of determining message text information and corresponding message meta-attribute information based on the text, the emoji, the mention message, the reference message, the link, and the image description includes: The message text information is obtained by segmenting the text, emojis, mentions, quotations, links, and image descriptions. The message meta-attribute information is determined based on the sender identifier, receiver identifier, sending timestamp, message session, message type, and topic tag in the message text information.

4. The method as described in claim 2, characterized in that, The step of fusing the message text information and the message meta-attribute information to obtain the message vector includes: The message text information is mapped into a text semantic vector based on a pre-trained language model; An initial meta-feature vector is constructed based on the message meta-attribute information, and the initial meta-feature vector is subjected to dimensionality reduction processing to obtain the meta-feature vector; The message vector is obtained by fusing the text semantic vector and the meta-feature vector.

5. The method as described in claim 1, characterized in that, The step of determining the keyword retrieval score based on the keyword matching degree between the instant messaging message and the query request includes: The query requests are categorized to obtain a set of query keywords; The keyword retrieval score is obtained by calculating the degree of keyword matching between the instant messaging message and the set of query keywords.

6. The method as described in claim 1, characterized in that, The step of weightedly fusing the vector retrieval score and the keyword retrieval score according to the fusion weight corresponding to the query request to obtain the mixed retrieval score of the instant messaging message and the evaluation result corresponding to the mixed retrieval score includes: The query request is input into a preset intent recognition model for intent classification to determine the query classification type, which includes keyword exact query type, semantic vector retrieval type and hybrid query type. The fusion weight is obtained based on the confidence levels of the keyword precise query type, the semantic vector retrieval type, and the mixed query type; The keyword retrieval score and the vector retrieval score are weighted and fused according to the fusion weight to obtain the hybrid retrieval score of the instant messaging message; The corresponding evaluation result is determined based on the hybrid retrieval score.

7. The method as described in claim 6, characterized in that, The step of determining the corresponding evaluation result based on the hybrid retrieval score includes: Obtain the time interval between the current time and the message sending time, and calculate the message freshness decay coefficient based on the time interval; Obtain the sender authority coefficient, and based on the hybrid retrieval score, the sender authority coefficient, and the message freshness decay coefficient, obtain the retrieval ranking score of the instant messaging message; The instant messaging messages are sorted according to the search ranking score to obtain the ranking result, and the ranking result is used as the evaluation result.

8. A hybrid retrieval and evaluation device for instant messaging messages based on natural language understanding, characterized in that, The device includes: The acquisition module is used to respond to a user's query request, acquire the query vector of the query request, and acquire the instant messaging message; The processing module is used to determine the message vector based on the message text information and message element attribute information in the instant messaging message; The retrieval module is used to determine the keyword retrieval score based on the degree of keyword matching between the instant messaging message and the query request; The retrieval module is further configured to determine a vector retrieval score based on the similarity between the message vector and the query vector; The evaluation module is used to perform weighted fusion of the vector retrieval score and the keyword retrieval score according to the fusion weight corresponding to the query request, so as to obtain the mixed retrieval score of the instant messaging message and the evaluation result corresponding to the mixed retrieval score.

9. A hybrid retrieval and evaluation device for instant messaging messages based on natural language understanding, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the hybrid retrieval and evaluation method for instant messaging messages based on natural language understanding as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the instant messaging message hybrid retrieval and evaluation method based on natural language understanding as described in any one of claims 1 to 7.

Images