Data processing method, device, electronic device and computer readable medium

Abstract

This application discloses a data processing method, apparatus, electronic device, and computer-readable medium, relating to the field of computer technology. The method includes: receiving a data processing request and determining the corresponding user identifier and the corresponding time identifier; acquiring meeting data corresponding to each user identifier, sorting the meeting data based on the time identifier, inputting the sorted meeting data into a referential search model, and extracting the corresponding antecedent entities and referential entities; generating word clusters based on the antecedent entities and referential entities, clustering each word cluster to obtain each cluster, determining the object entities corresponding to each cluster, and generating entity pairs; classifying each meeting data by region and generating region identifiers corresponding to each meeting data; extracting keywords from the corresponding meeting data based on the entity pairs and region identifiers, and generating and outputting summary data based on the keywords and region identifiers. This results in clearly defined references in the generated summary data, improving the model's resolution capability.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a data processing method, apparatus, electronic device, and computer-readable medium. Background Technology

[0002] Meeting minutes are a narrative and introductory document compiled from meeting records. They include basic information about the meeting, its main points and central content, participants and related matters, and the time of the meeting. This facilitates reporting to superiors or disseminating information to relevant personnel. When compiling and processing these minutes, they may be recorded according to the meeting agenda or presented one by one according to several key issues summarized from the meeting content.

[0003] Current methods for generating meeting minutes include rule-based machine learning, which primarily define keywords and then detect the use of these keywords or sentence structures in online conversation records. Meeting participants frequently use various pronouns to refer to previously mentioned topics or content. While rule-based methods can identify key information to some extent, they fail to address these referential references. Rule-based machine learning only processes text superficially, neglecting bidirectional characteristics and contextual relationships, resulting in ambiguous meeting summaries and low model differentiation capabilities at the single-sentence and dialogue perspectives.

[0004] In the process of developing this application, the inventors discovered at least the following problems in the prior art:

[0005] The meeting summaries obtained by rule-based machine recording methods are unclear in their references, and the model has low ability to distinguish between single sentences and dialogue perspectives. Summary of the Invention

[0006] In view of this, embodiments of this application provide a data processing method, apparatus, electronic device, and computer-readable medium that can solve the problems of unclear reference in meeting summaries obtained by existing rule-based machine recording methods and poor model discrimination ability from single sentence and dialogue perspectives.

[0007] To achieve the above objectives, according to one aspect of the embodiments of this application, a data processing method is provided, comprising:

[0008] Receive data processing requests and determine the corresponding user identifier and the timestamp corresponding to the user identifier;

[0009] Obtain meeting data corresponding to each user identifier, sort the meeting data based on time identifier, and then input the sorted meeting data into the referential search model to extract the corresponding antecedent entity and referential entity.

[0010] Word clusters are generated based on antecedent entities and pronoun entities. Each word cluster is then clustered to obtain each cluster. The object entity corresponding to each cluster is determined, and then entity pairs are generated.

[0011] The data from each meeting is categorized by region, and a corresponding region identifier is generated for each meeting's data.

[0012] The content processing model based on the dialogue corpus is invoked to extract keywords from the corresponding meeting data according to entity pairs and region identifiers. Based on the keywords and region identifiers, summary data is generated and output.

[0013] Optionally, word clusters are generated based on antecedent entities and pronoun entities, including:

[0014] Determine the coreference relationships between each antecedent entity and each pronoun entity;

[0015] Input each antecedent entity, each pronoun entity, and coreference relation into the referential search model, and output the corresponding entity pair vector;

[0016] Word clusters are generated based on entity pair vectors.

[0017] Optionally, the word clusters are clustered, including:

[0018] Determine the semantic vector corresponding to each word cluster;

[0019] The semantic vectors are randomly combined to obtain a combined vector;

[0020] Calculate the similarity between the semantic vectors in each combined vector;

[0021] The word clusters corresponding to semantic vectors with similarity greater than a first preset threshold are merged to obtain clusters.

[0022] Optionally, before extracting the keywords from the corresponding meeting data, the method further includes:

[0023] Based on entity pairs, the antecedent entity and the pronoun entity are replaced with the corresponding object entity, and then updated to the data of each conference.

[0024] Optionally, before extracting the keywords from the corresponding meeting data, the method further includes:

[0025] The updated meeting data is segmented into statements to generate initial segmentation statements.

[0026] The sentence segmentation model is invoked to obtain the corresponding intent similarity based on each initially segmented sentence;

[0027] Sentence segmentation markers are added between the initial segmentation statements corresponding to intent similarity values ​​that are less than a second preset threshold, thereby generating sentence segmentation conference data.

[0028] Optionally, keywords from the corresponding meeting data can be extracted, including:

[0029] Based on the region identifier, the data from each sentence segmentation meeting is divided into regions to obtain the sentences in each region;

[0030] The information matching model is invoked to extract keywords from the sentences in the divided regions, and the corresponding keywords are obtained.

[0031] Optionally, based on keywords and region identifiers, summary data is generated and output, including:

[0032] Based on the context of the updated meeting data, the subject of each keyword is determined, and then the subject statement of each subject in each meeting data is determined.

[0033] Perform information matching on the subject statement to generate a set of key items;

[0034] Calculate the content similarity between items in the key item set, and then remove items with content similarity greater than the third preset threshold to generate a deduplicated key item set;

[0035] Search for negative identifiers in the set of key items to be deduplicated, and remove the items corresponding to the negative identifiers to generate the target set of key items;

[0036] Based on the region identifier, the corresponding set of key target items is determined, and then the corresponding set of key target items is added to the region corresponding to the region identifier, thereby generating summary data and outputting it.

[0037] Optionally, entity pairs are generated, including:

[0038] Each antecedent entity and each pronoun entity in each cluster is combined with its corresponding object entity to generate entity pairs.

[0039] Optionally, before sorting the meeting data based on time identifiers, the method further includes:

[0040] The data from each meeting is segmented into words to obtain segmented data.

[0041] The segmented data is converted into corresponding pinyin, and then the pinyin is converted into corresponding word vectors;

[0042] The word vectors are input into the word sense disambiguation model, which outputs a similarity score. Then, the correct corpus sentences that are similar to the conference data are determined based on the similarity score.

[0043] The clustering probability model is invoked to determine the probability that the segmented data matches each corpus in the correct corpus sentence;

[0044] The corpus with a probability greater than a preset threshold in the correct corpus sentences is identified as the target replacement corpus. The word segmentation data is replaced with the target replacement corpus and then updated to the conference data.

[0045] In addition, this application also provides a data processing apparatus, including:

[0046] The receiving unit is configured to receive data processing requests and determine the corresponding user identifier and the time identifier corresponding to the user identifier.

[0047] The entity extraction unit is configured to acquire meeting data corresponding to each user identifier, sort the meeting data based on time identifier, and then input the sorted meeting data into the referential search model to extract the corresponding antecedent entity and referential entity.

[0048] The entity pair generation unit is configured to generate word clusters based on antecedent entities and pronoun entities, cluster each word cluster to obtain each cluster, determine the object entity corresponding to each cluster, and then generate entity pairs.

[0049] The region identifier generation unit is configured to classify each meeting data by region and generate a region identifier corresponding to each meeting data.

[0050] The summary data generation unit is configured to invoke a content processing model based on dialogue corpus to extract keywords from the corresponding conference data according to entity pairs and region identifiers, and generate and output summary data based on the keywords and region identifiers.

[0051] Optionally, the entity pair generation unit is further configured to:

[0052] Determine the coreference relationships between each antecedent entity and each pronoun entity;

[0053] Input each antecedent entity, each pronoun entity, and coreference relation into the referential search model, and output the corresponding entity pair vector;

[0054] Word clusters are generated based on entity pair vectors.

[0055] Optionally, the entity pair generation unit is further configured to:

[0056] Determine the semantic vector corresponding to each word cluster;

[0057] The semantic vectors are randomly combined to obtain a combined vector;

[0058] Calculate the similarity between the semantic vectors in each combined vector;

[0059] The word clusters corresponding to the semantic vectors with similarity greater than the first preset threshold are merged to obtain clusters.

[0060] Optionally, the data processing apparatus further includes a decomposition unit configured to:

[0061] Based on entity pairs, the antecedent entity and the pronoun entity are replaced with the corresponding object entity, and then updated to the data of each conference.

[0062] Optionally, the data processing device further includes a sentence segmentation unit, configured to:

[0063] The updated meeting data is segmented into statements to generate initial segmentation statements.

[0064] The sentence segmentation model is invoked to obtain the corresponding intent similarity based on each initially segmented sentence;

[0065] Sentence segmentation markers are added between the initial segmentation statements corresponding to intent similarity values ​​that are less than a second preset threshold, thereby generating sentence segmentation conference data.

[0066] Optionally, the summary data generation unit is further configured to:

[0067] Based on the region identifier, the data from each sentence segmentation meeting is divided into regions to obtain the sentences in each region;

[0068] The information matching model is invoked to extract keywords from the sentences in the divided regions, and the corresponding keywords are obtained.

[0069] Optionally, the summary data generation unit is further configured to:

[0070] Based on the context of the updated meeting data, the subject of each keyword is determined, and then the subject statement of each subject in each meeting data is determined.

[0071] Perform information matching on the subject statement to generate a set of key items;

[0072] Calculate the content similarity between items in the key item set, and then remove items with content similarity greater than the third preset threshold to generate a deduplicated key item set;

[0073] Search for negative identifiers in the set of key items to be deduplicated, and remove the items corresponding to the negative identifiers to generate the target set of key items;

[0074] Based on the region identifier, the corresponding set of key target items is determined, and then the corresponding set of key target items is added to the region corresponding to the region identifier, thereby generating summary data and outputting it.

[0075] Optionally, the entity pair generation unit is further configured to:

[0076] Each antecedent entity and each pronoun entity in each cluster is combined with its corresponding object entity to generate entity pairs.

[0077] Optionally, the data processing apparatus further includes a word sense disambiguation unit, configured to:

[0078] The data from each meeting is segmented into words to obtain segmented data.

[0079] The segmented data is converted into corresponding pinyin, and then the pinyin is converted into corresponding word vectors;

[0080] The word vectors are input into the word sense disambiguation model, which outputs a similarity score. Then, the correct corpus sentences that are similar to the conference data are determined based on the similarity score.

[0081] The clustering probability model is invoked to determine the probability that the segmented data matches each corpus in the correct corpus sentence;

[0082] The corpus with a probability greater than a preset threshold in the correct corpus sentences is identified as the target replacement corpus. The word segmentation data is replaced with the target replacement corpus and then updated to the conference data.

[0083] In addition, this application also provides a data processing electronic device, including: one or more processors; and a storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the data processing method described above.

[0084] In addition, this application also provides a computer-readable medium having a computer program stored thereon, which, when executed by a processor, implements the data processing method described above.

[0085] One embodiment of the above invention has the following advantages or beneficial effects: This application receives a data processing request, determines the corresponding user identifier and the time identifier corresponding to the user identifier; obtains the meeting data corresponding to each user identifier, sorts the meeting data based on the time identifier, and then inputs the sorted meeting data into a referential search model to extract the corresponding antecedent entities and referential entities; generates word clusters based on the antecedent entities and referential entities, clusters each word cluster to obtain each cluster, determines the object entity corresponding to each cluster, and then generates entity pairs; calls a classification model to perform regional classification on each meeting data to generate regional identifiers corresponding to each meeting data; extracts keywords of the corresponding meeting data based on the entity pairs and each regional identifier, and generates and outputs summary data based on the keywords and regional identifiers. By generating word clusters based on antecedent entities and referential entities, clustering each word cluster to obtain each cluster, determining the object entity corresponding to each cluster, and then generating entity pairs, the application extracts keywords of the corresponding meeting data based on the entity pairs and each regional identifier, and generates and outputs summary data based on the keywords and regional identifiers. This approach utilizes intelligent analysis and deep learning algorithms to automatically correct various referential errors and semantic distortions in text content, resulting in clearer referential meanings in the generated meeting summaries. It improves the model's processing capabilities from both single-sentence and dialogue perspectives based on meeting data, enhancing the accuracy and completeness of the generated meeting summary structure. It provides a solution for summarizing key information and generating summaries from multi-user, multi-turn dialogue texts in online meetings. This involves collecting online dialogue data, processing it into text, and then performing content preprocessing and deep learning-based information extraction and classification. The deep learning model effectively distinguishes information relationships between different roles and fully explores the content connections between contexts, resulting in more comprehensive recognition and higher accuracy.

[0086] The further effects of the aforementioned unconventional alternative methods will be explained below in conjunction with specific implementation methods. Attached Figure Description

[0087] The accompanying drawings are provided to better understand this application and do not constitute an undue limitation thereof. Wherein:

[0088] Figure 1 This is a schematic diagram of the main flow of the data processing method according to the first embodiment of this application;

[0089] Figure 2 This is a schematic diagram of the main flow of the data processing method according to the second embodiment of this application;

[0090] Figure 3 This is a schematic diagram illustrating an application scenario of the data processing method according to the third embodiment of this application;

[0091] Figure 4This is a schematic diagram of the word sense disambiguation process of the data processing method according to an embodiment of this application;

[0092] Figure 5 This is a schematic diagram illustrating the deconstruction process of the data processing method according to an embodiment of this application;

[0093] Figure 6 This is a schematic diagram illustrating the explanation of terms related to the data processing method according to embodiments of this application;

[0094] Figure 7 This is a schematic diagram of the content classification process of the data processing method according to an embodiment of this application;

[0095] Figure 8 This is a schematic diagram of the key information extraction process of the data processing method according to an embodiment of this application;

[0096] Figure 9 This is a schematic diagram of the deduplication matching process of the data processing method according to an embodiment of this application;

[0097] Figure 10 This is a schematic diagram of the main units of a data processing apparatus according to an embodiment of this application;

[0098] Figure 11 This is an exemplary system architecture diagram to which embodiments of this application can be applied;

[0099] Figure 12 This is a schematic diagram of the structure of a computer system suitable for implementing terminal devices or servers in the embodiments of this application. Detailed Implementation

[0100] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of this application, including various details to aid understanding. These should be considered merely exemplary. Therefore, those skilled in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description. The acquisition, storage, use, and processing of data in the technical solutions of this application all comply with relevant national laws and regulations.

[0101] Figure 1 This is a schematic diagram of the main flow of the data processing method according to the first embodiment of this application, as shown below. Figure 1 As shown, the data processing methods include:

[0102] Step S101: Receive a data processing request and determine the corresponding user identifier and the time identifier corresponding to the user identifier.

[0103] In this embodiment, the execution entity of the data processing method (e.g., a server) can receive data processing requests via wired or wireless connections. Specifically, the data processing request may be, for example, a request to generate a meeting summary from meeting data or a request to generate a course summary from course data. This embodiment does not limit the specific content of the data processing request. The user identifier may be, for example, a specific role identifier for the user; for instance, it could be A, B, and C, representing roles A, B, and C respectively. The time identifier may specifically be the speaking time of the user (specifically, a role) corresponding to each user identifier; for example, role A's speaking time is 2021.12.08 10:02, role B's speaking time is 2021.12.08 10:04, and role C's speaking time is 2021.12.08 10:05. The content of the data processing request may, for example, be as follows: Figure 3 The meeting audio or chat logs shown in this application do not specifically limit the content of the data processing request.

[0104] Step S102: Obtain the meeting data corresponding to each user identifier, sort the meeting data based on the time identifier, and then input the sorted meeting data into the referential search model to extract the corresponding antecedent entity and referential entity.

[0105] Specifically, the referential search model in this embodiment is applicable to long texts with multiple pairs of related referential relationships across sentences. The referential search model can be trained as follows: Model training prioritizes searching for entities based on their preceding sentences (preceding sentences or sentences; entities refer to nouns or phrases, such as "phone" or "want to return"), generating entity pair vectors. Then, the referential entities and their corresponding antecedent entities are treated as word clusters, for example: it-phone, this-small A, that-model a, etc. Words within each cluster represent the same object (e.g., the same object, model a small A phone). Based on pre-labeled referential entities and antecedent entities, those that are labeled receive higher scores, while unlabeled referential entities and antecedent entities receive lower scores. An attention-based policy model is trained, which can improve speed in subsequent clustering. Since an antecedent entity (e.g., "small A mobile phone") may have multiple referential entities (e.g., "it", "this", "that"), a clustering algorithm is used to cluster the pronoun entity-antecedent entity pairs. If the entity pairs within a cluster have higher scores after merging, they are merged. This way, all referential entities and antecedent entities that refer to the same object entity (e.g., "small A mobile phone") are grouped into the same cluster, ultimately obtaining the unified object entity (e.g., "mobile phone") corresponding to each pronoun entity (e.g., "it", "this", "that") and its corresponding antecedent entity at once. Entities refer to nouns or phrases, such as "mobile phone" or "want to return." The entity "mobile phone" is the antecedent of the entity "it," and the two form an entity pair.

[0106] For example, such as Figure 5 As shown, the dereference of meeting data aims to find the true subject in a sentence. Pronouns are replaced with actual key information phrases. Unlike typical dereference of short texts, in multi-person, multi-turn dialogues, the data are interwoven, both related and interrupted. For example... Figure 5 As shown, the process first involves acquiring the text of a multi-person dialogue, then sorting it according to its chronological order. This is because in multi-person dialogues, coreferences of pronouns may cross roles, meaning that information events are formed by the statements of different roles. In the process of resolving pronoun references, the priority is to discover entities, antecedents, and coreferences, such as... Figure 6As shown, word segmentation identifies entities, antecedents, coreferences, and clusters in multi-person dialogue text. For example, in the dialogue text "My sister gave me her book," the entities are: "I," "sister," "she," "book," and "I." The antecedents are "I" and "sister," which are the antecedents of "she" and "I," respectively. The coreferences are: "I"-"I" coreference and "sister"-"she" coreference. The clusters are: "I"-"I" cluster and "sister"-"she" cluster. Then, by comparing with a predefined lexicon, the properties of the corresponding words (such as verbs, pronouns, phrases, adjectives, etc.) can be obtained. This leads to the retrieval of contextual pronouns and phrases, and subsequently, the antecedent entities and their corresponding pronoun entities. Then, based on a clustering + attention model, a scoring strategy for the referential search model is used to replace the pronoun entities and their antecedent entities in the text. Specifically, the objects that the pronoun entities and their antecedent entities refer to can be used to replace each pronoun entity and its corresponding antecedent entity (this is the case where the antecedent entities corresponding to each pronoun entity are different, but each pronoun entity and its corresponding antecedent entity refer to the same object), thus resolving the referential inconsistencies. It can be understood that if the object referred to is the antecedent entity corresponding to the pronoun entity, then the antecedent entity replaces the corresponding pronoun entity.

[0107] Step S103: Generate word clusters based on antecedent entities and pronoun entities, cluster each word cluster to obtain each cluster, determine the object entity corresponding to each cluster, and then generate entity pairs.

[0108] Specifically, word clusters are generated based on antecedent entities and pronoun entities, including:

[0109] Determine the coreference relationships between each antecedent entity and each pronoun entity. For example... Figure 6 As shown in the example, in the dialogue text "My sister gave me her book," word segmentation identifies the entities, antecedents, coreferences, and clusters in the multi-person dialogue text. Among them, the coreferences are: "I"-"I" coreference and "sister"-"she" coreference.

[0110] Each antecedent entity, each pronoun entity, and coreference relation are input into the referential search model, and the corresponding entity pair vectors are output. Based on the entity pair vectors, word clusters are generated. Specifically, word clusters can be composed of entity pair vectors or entity pairs corresponding to entity pair vectors. This application does not specifically limit the word clusters.

[0111] Specifically, the word clusters are clustered, including:

[0112] Determine the semantic vector corresponding to each word cluster; randomly combine the semantic vectors to obtain a combined vector; calculate the similarity between the semantic vectors in each combined vector; merge the word clusters corresponding to the semantic vectors with similarity greater than a first preset threshold to obtain a cluster.

[0113] For example, the execution entity can use a clustering algorithm to cluster pronoun entities and their antecedent entities based on semantic vector similarity. If the similarity scores of entity pairs within two clusters are high, they are merged. This allows all pronoun entities and their antecedent entities referring to the same object (e.g., "Xiao A's phone") to be grouped into the same cluster, ultimately obtaining the unified object corresponding to each pronoun entity (e.g., "it," "this," "that") and its corresponding antecedent entity in one go.

[0114] Specifically, generating entity pairs includes combining each antecedent entity and each pronoun entity in each cluster with its corresponding object entity to generate entity pairs. It can be understood that the object entity (or object) can be an antecedent entity. When all pronoun entities point to the same antecedent entity, the corresponding object entity is that antecedent entity.

[0115] Specifically, before sorting the meeting data based on time identifiers, the method also includes:

[0116] The meeting data is segmented into words to obtain segmented data; the segmented data is converted into corresponding pinyin, and then the pinyin is converted into corresponding word vectors; the word vectors are input into a word sense disambiguation model to output similarity scores, and then the correct corpus sentences that are similar to the meeting data are determined based on the similarity scores; a clustering probability model is called to determine the probability of matching each corpus in the correct corpus sentences with the segmented data; the corpus sentences with a probability greater than a preset threshold are determined as target replacement corpus, and the segmented data is replaced with the target replacement corpus, and then updated to the meeting data.

[0117] For example, such as Figure 4 The diagram shown illustrates the word sense disambiguation process in this embodiment. Before performing word sense disambiguation on the meeting text, the executing entity needs to build a pre-trained model to ultimately obtain the word sense disambiguation model. For example... Figure 4As shown, a pre-trained model is constructed by performing text segmentation on conference text corpora (e.g., address: Mianyang) to obtain commonly used vocabulary (e.g., word segmentation yields: address|Mianyang). Then, using the occurrence of words in the same sentence as a positive incentive, word clustering is performed on the corpus words (e.g., obtaining word clusters: (address, Mianyang)). The probability of each word appearing in different word clusters is obtained. Words with higher probabilities are more likely to appear in the same sentence. This is used to train the clustering probability model in the pre-trained model. Simultaneously, to eliminate the influence of accents, the obtained word segments are phonetically converted (e.g., obtaining: dizhi|mianyang), and a word vector model is constructed. The corpus sentences are represented as word vectors (e.g., the obtained word vector representation is: 230|666). Iterative training ultimately yields the word sense disambiguation model. Then, during real-time computation, the text to be disambiguated (e.g., "address in sheep") is input into the word sense disambiguation model. This model performs pre-trained text segmentation on the text to be disambiguated (e.g., obtaining: address|in|sheep), performs pinyinization (e.g., dizhi|zai|mianyang), and calls a clustering probability model to represent the text as word vectors (e.g., 230|470|666). The word vectors corresponding to the text to be disambiguated are then matched with the corpus words in the correct corpus sentence. Based on cosine distance, similar correct corpus sentences are obtained, and the ambiguous text to be disambiguated (e.g., "address in sheep") is replaced with the correct corpus sentence (e.g., "address in Mianyang"). Specifically, this can be done by replacing ambiguous words in the ambiguous text to be disambiguated (e.g., "sheep") with the correct words in the correct corpus sentence (e.g., "Mianyang").

[0118] Word sense disambiguation addresses abnormal words caused by external environment, input errors, and user accents, thereby avoiding impacting the accuracy of subsequent content classification and key information extraction.

[0119] Step S104: Classify the data of each meeting by region and generate the region identifier corresponding to each meeting data.

[0120] For example, such as Figure 7 As shown, the conference data is categorized by region, generating a corresponding region identifier for each conference data point to facilitate the automatic generation of conference summaries later. For example... Figure 7The classification of content regions shown is to distinguish the attributes of each piece of content so that it can be filled into different summary regions, such as topics, summaries, and questions, while removing meaningless content, such as small talk. The implementation of region classification requires training a deep learning-based classification model using a pre-annotated corpus. After training, the classification model is deployed in the system backend. When the execution entity calls the classification model interface, it inputs the meeting data to obtain the corresponding region classification results and region identifiers. Specifically, the training process of the classification model is as follows: First, a pre-trained model is built. Then, a pre-annotated corpus of meeting texts is input into the pre-trained model, and the corresponding annotated region identifiers (e.g., WYY, ZT, ZJ, WT, ZP, which can correspond to meaningless sentences, topic sentences, summary sentences, question sentences, and assignment sentences, respectively) are used as the expected output to iteratively train the pre-trained model, ultimately generating a classification model that can accurately classify meeting text regions. After obtaining the classification model, the executing entity can perform real-time calculations. By inputting the text corpus to be classified (which can be conference data) into the classification model, the text type can be distinguished, and the region identifiers corresponding to each content in the text to be classified (which can be conference data) can be obtained.

[0121] Step S105: Call the content processing model based on dialogue corpus to extract keywords from the corresponding meeting data according to entity pairs and region identifiers, and generate and output summary data based on keywords and region identifiers.

[0122] After obtaining the area identifiers corresponding to the meeting data, the executing entity can call the content processing model based on the dialogue corpus to extract the keywords (or key information) of the areas corresponding to each area identifier in the meeting data.

[0123] Specifically, key information extraction, or key entity identification, aims to extract core content from different types of key sentences, such as key dates, tasks, roles, locations, and plans, thereby making the final summary more readable and accurate. For example... Figure 8 As shown, the executing entity invokes strong word rules, such as "the location is located" followed by the location. This rule has a relatively small recognition range, but its accuracy is guaranteed, complementing deep learning models. Attention-based information matching models require pre-training. Their input is a single sentence, and the intermediate hidden layer extracts sentence vectors. By labeling keywords, the model positively weights the scores of key information words in the sentence, ultimately obtaining the keywords for each sentence.

[0124] For example, such as Figure 8As shown, before calling the content processing model based on the dialogue corpus, the executing entity can construct a pre-trained model and iteratively train the pre-trained model with labeled keywords to finally obtain the content processing model based on the dialogue corpus. Figure 8 As shown, the specific training process is as follows: The executing entity acquires a conference text corpus and annotates keywords (e.g., annotating keywords with time, location, subject, role, question, etc.). It then inputs the conference text corpus into the training model, using the annotated keywords as the expected output, to iteratively train the pre-trained model, ultimately obtaining a content processing model based on the dialogue corpus. This content processing model based on the dialogue corpus can specifically be a regularization / attention-based information matching model. After obtaining the content processing model based on the dialogue corpus, the executing entity can perform real-time calculations, inputting the text corpus to be extracted into the model to call the key information extraction module within the model to extract keywords, thereby obtaining the key information in the sentences, i.e., ultimately obtaining the keywords from the text corpus to be extracted.

[0125] Deduplication matching aims to eliminate the influence of similar or opposing information. Sometimes, the time, place, people, and plans related to a key matter are separated into different sentences and spoken by different characters; therefore, deduplication matching is necessary. After extracting keywords from the meeting data, the executing entity first, since the pronoun entities have already been replaced in the previous steps, information sentences on the same topic will have the same subject. Based on this, the complete information of the topic can be reorganized, such as: Zhang San on Wednesday, Zhang San completed the report, Zhang San in location A, which can form the key matter set: Zhang San completed the report in location A before Wednesday. Then, content deduplication is required. TF-IDF-based text similarity calculation is used to remove items with similarity exceeding a predetermined threshold. Simultaneously, rules are used to retrieve negative words or phrases after each item and remove the negated items. Finally, the deduplicated items are obtained.

[0126] For example, such as Figure 9 The diagram shown is a flowchart of the deduplication matching process according to an embodiment of this application. Before deduplication matching, the executing entity can first perform information matching based on the keywords of the extracted meeting data. For example, the extracted information corpus (e.g., keywords of the extracted meeting data) can be mixed and arranged in chronological order, and then the context and main statement retrieval can be performed to perform joint information matching. For example, joint information matching can be performed on "Zhang San was on Wednesday", "Zhang San completed the report", and "Zhang San was in location A" to form a set of key matters: "Zhang San completed the report in location A before Wednesday".

[0127] The resulting set of key items is deduplicated. Specifically, based on TF-IDF content similarity calculation, items with similarity scores higher than the threshold are removed from each key item in the set. Then, negative sentences are retrieved, and the items corresponding to the negative sentences are removed, ultimately yielding the valid key items.

[0128] After obtaining valid key information, the executing entity can call the output layer of the content processing model based on the dialogue corpus to organize the finally obtained valid key information and automatically fill in the form slots: through the previous steps, the key sentences of different content areas of the summary and the key entity content in the key sentences have been obtained. By filling in the slots, the complete meeting summary can be output.

[0129] This embodiment receives a data processing request, determines the corresponding user identifier and the corresponding time identifier, acquires the meeting data corresponding to each user identifier, sorts the meeting data based on the time identifier, and then inputs the sorted meeting data into a referential search model to extract the corresponding antecedent entities and referential entities. Based on the antecedent entities and referential entities, word clusters are generated, and each word cluster is clustered to obtain each cluster. The object entities corresponding to each cluster are determined, and entity pairs are generated. A classification model is invoked to perform regional classification on each meeting data, generating region identifiers corresponding to each meeting data. Based on the entity pairs and each region identifier, keywords for the corresponding meeting data are extracted, and summary data is generated and output based on the keywords and region identifiers. The process involves generating word clusters based on antecedent entities and referential entities, clustering each word cluster to obtain each cluster, determining the object entities corresponding to each cluster, and generating entity pairs. Based on the entity pairs and each region identifier, keywords for the corresponding meeting data are extracted, and summary data is generated and output based on the keywords and region identifiers. This approach utilizes intelligent analysis and deep learning algorithms to automatically correct various referential errors and semantic distortions in text content, resulting in clearer referential meanings in the generated meeting summaries. It improves the model's processing capabilities from both single-sentence and dialogue perspectives based on meeting data, enhancing the accuracy and completeness of the generated meeting summary structure. It provides a solution for summarizing key information and generating summaries from multi-user, multi-turn dialogue texts in online meetings. This involves collecting online dialogue data, processing it into text, and then performing content preprocessing and deep learning-based information extraction and classification. The deep learning model effectively distinguishes information relationships between different roles and fully explores the content connections between contexts, resulting in more comprehensive recognition and higher accuracy.

[0130] Figure 2 This is a schematic diagram of the main flow of the data processing method according to the second embodiment of this application, as shown below. Figure 2 As shown, the data processing methods include:

[0131] Step S201: Receive a data processing request and determine the corresponding user identifier and the time identifier corresponding to the user identifier.

[0132] Step S202: Obtain the meeting data corresponding to each user identifier, sort the meeting data based on the time identifier, and then input the sorted meeting data into the referential search model to extract the corresponding antecedent entity and referential entity.

[0133] Step S203: Generate word clusters based on antecedent entities and pronoun entities, cluster each word cluster to obtain each cluster, determine the object entity corresponding to each cluster, and then generate entity pairs.

[0134] Step S204: Classify the data of each meeting by region and generate the region identifier corresponding to each meeting data.

[0135] The principles of steps S201 to S204 are similar to those of steps S101 to S104, and will not be repeated here.

[0136] Step S205: Based on entity pairs, replace the corresponding object entities of the antecedent entities and pronoun entities, and then update the data of each meeting.

[0137] Step S206: Call the content processing model based on the dialogue corpus to extract keywords from the corresponding meeting data according to entity pairs and region identifiers, and generate and output summary data based on keywords and region identifiers.

[0138] Specifically, before extracting the keywords from the corresponding meeting data, the method also includes:

[0139] The updated meeting data is segmented into sentences to generate initial segmented sentences; the sentence segmentation model is called to obtain the corresponding intent similarity based on each initial segmented sentence; sentence segmentation markers are added between the initial segmented sentences with intent similarity less than a second preset threshold to generate segmented meeting data.

[0140] Specifically, automatic sentence segmentation aims to find the true context sentences in multi-turn dialogues. Because the speeches of different characters are interrupted and interspersed during multi-person dialogues, it is necessary to reconstruct the complete expression of each character's content. First, after resolving pronoun referencing, the meeting data of each character's speech is reconnected, and the meeting data is clustered according to the audience. Then, context beginning and end detection is performed, that is, the fluency of sentences is checked by using regular expressions and part-of-speech tagging. Sentences with contextual relevance are linked. Finally, a pre-trained sentence segmentation model is used. Its main principle is to identify the intent and similarity of sentences based on the pre-trained model. Sentences with intent similarity less than a threshold (this application does not specifically limit the threshold) are used as the basis for sentence segmentation, thereby generating segmented meeting data. For example, as shown... Figure 5 As shown, the automatic sentence segmentation process is as follows: After dereference resolution, the multi-person dialogue text is clustered according to its source and user. The beginning and end of sentences in the context are then detected; that is, the fluency of the sentences is checked using regular expressions and part-of-speech tagging. For sentences with contextual relevance, context reconnection is performed. Then, the context-reconnected text is input into the sentence segmentation model for preprocessing to obtain preprocessed, segmented text corpus.

[0141] Specifically, keywords are extracted from the corresponding meeting data, including:

[0142] Based on region identifiers, the meeting data for each segment is divided into regions, resulting in sentences for each region. Examples include meaningless sentences, topic sentences, summary sentences, question sentences, and assignment sentences. An information matching model is then used to extract keywords from the segmented region sentences, yielding the corresponding keywords.

[0143] Specifically, based on keywords and region identifiers, summary data is generated and output, including: determining the subject in each keyword based on the context of the updated meeting data, and then determining the subject statement in each meeting data; performing joint information matching on the subject statements to generate a set of key items; calculating the content similarity between items in the set of key items, and then removing items with content similarity greater than a third preset threshold to generate a deduplicated set of key items; retrieving negative identifiers in the deduplicated set of key items, and removing items corresponding to negative identifiers to generate a target set of key items; determining the corresponding target set of key items according to the region identifier, and then adding the corresponding target set of key items to the region corresponding to the region identifier, thereby generating summary data and outputting it.

[0144] Specifically, the negation marker can be the marker corresponding to a negative sentence, such as the "no" or "none" markers corresponding to words like "not," "not," or "nothing." Generating the target key item set involves: calculating the content similarity of each key item in the formed key item set based on TF-IDF and removing items with similarity higher than a threshold; then retrieving negative sentences and removing the items corresponding to those negative sentences, ultimately obtaining the valid key items. These valid key items are then grouped into a set, resulting in the target key item set. Each target key item in the final target key item set is assigned to a corresponding region based on a region marker. For example, target key items corresponding to a topic are assigned to the topic region, target key items corresponding to a summary are assigned to the summary region, target key items corresponding to a question are assigned to the question region, and so on. Finally, the meeting summary data is generated and output from each region and the target key items corresponding to each region.

[0145] Figure 3 This is a schematic diagram illustrating an application scenario of the data processing method according to the third embodiment of this application. The data processing method of this embodiment can be applied to scenarios involving the extraction of meeting summaries from meeting data. Figure 3As shown, the executing entity can collect meeting content, such as meeting audio or chat logs (e.g., text-based chat logs). This embodiment of the application targets an automatic summary generation application for online meetings, requiring the accuracy of the text and disambiguation of homophones in the speech-to-text transcription. This embodiment can be applied to both audio and text-based meetings, and the meeting content can originate from audio or text. Then, the executing entity can call a content processing model based on the dialogue corpus to extract summaries from the meeting content. Specifically, the executing entity can call the acquisition layer of the content processing model based on the dialogue corpus (hereinafter referred to as the model). The acquisition layer mainly collects content input from multiple terminal users through input devices, including microphone audio acquisition and text acquisition from the background of instant messaging software, to collect meeting content. Then, the model's storage layer is called to store the content data. Since new topics and suggestions are constantly generated during the meeting, the entire summary extraction process is a non-blocking, asynchronous processing flow. Pre-storing the content and processing the final meeting content uniformly helps improve accuracy. Then, the model's conversion layer is invoked, using the speech-to-text module to convert speech from different terminals into text corpora and tag them with user labels to distinguish different speech content and user roles. Next, the model's preprocessing layer is invoked to perform word sense disambiguation, pronoun resolution, and automatic sentence segmentation on the converted meeting content (details omitted here). Then, the model's extraction layer is invoked to perform content classification, key information extraction, and deduplication matching on the meeting content text after word sense disambiguation, pronoun resolution, and automatic sentence segmentation (details omitted here). The main purpose of the extraction layer is to extract the truly key content from the preprocessed corpus and the type of each key content item, for subsequent automatic form filling. Finally, the output layer is invoked to automatically generate forms based on the content-classified, key information extracted, and deduplicated text, ultimately obtaining and outputting the meeting summary data.

[0146] This application proposes a method that cleverly couples deep learning and rule matching. This improves the efficiency of long dialogue summarization and ensures the accuracy of the results. A text preprocessing method for multi-turn dialogues is proposed, capable of fusing the statements of different roles and extracting the actual intent by integrating the subject-verb-object information of the dialogue. It includes multiple modules such as referential resolution, automatic sentence segmentation, and word meaning disambiguation. It effectively reconstructs the true event content based on contextual information. A device for content information extraction and filling is proposed. By understanding different contextual information, it effectively retrieves different types of information, extracts key content, and fills it into corresponding slots. Through text cross-referencing and similarity filtering steps, high-quality summary text is output.

[0147] This application addresses the implementation of meeting summaries manually or according to rules. It provides a multi-layered intelligent algorithm scheme for content processing. On one hand, it uses intelligent algorithms for automatic key content extraction; on the other hand, it applies intelligent analysis and deep learning algorithms to the text content to achieve automatic correction of various types of semantic disambiguation and word meaning distortion. This application also addresses the issue of high accuracy requirements for meeting content. It provides a content processing model based on dialogue corpora. The word meaning disambiguation module in this method can effectively correct translation errors caused by speech-to-text conversion, especially errors caused by accents and environmental noise. Supervised classification algorithms are used for text content to achieve accurate information identification and automatic form filling, improving the accuracy and completeness of the meeting minutes structure. This application also addresses the problem of insufficient multi-angle analysis capabilities due to the overlapping information in multi-turn dialogues. It uses a model based on dialogue corpora. Through modules such as semantic disambiguation, content matching, and multi-segmentation, it achieves accurate reconstruction of the complete content information structure. Simultaneously, based on context, it enables capabilities such as synonym clustering and antonym discovery.

[0148] Figure 10 This is a schematic diagram of the main units of a data processing apparatus according to an embodiment of this application. Figure 10 As shown, the data processing device includes a receiving unit 1001, an entity extraction unit 1002, an entity pair generation unit 1003, a region identifier generation unit 1004, and a summary data generation unit 1005.

[0149] The receiving unit 1001 is configured to receive data processing requests and determine the corresponding user identifier and the time identifier corresponding to the user identifier.

[0150] The entity extraction unit 1002 is configured to acquire meeting data corresponding to each user identifier, sort the meeting data based on time identifiers, and then input the sorted meeting data into the referential search model to extract the corresponding antecedent entities and referential entities.

[0151] The entity pair generation unit 1003 is configured to generate word clusters based on antecedent entities and pronoun entities, cluster each word cluster to obtain each cluster, determine the object entity corresponding to each cluster, and then generate entity pairs.

[0152] The region identifier generation unit 1004 is configured to classify each meeting data by region and generate a region identifier corresponding to each meeting data.

[0153] The summary data generation unit 1005 is configured to call a content processing model based on dialogue corpus to extract keywords from the corresponding conference data according to entity pairs and region identifiers, and generate and output summary data based on the keywords and region identifiers.

[0154] In some embodiments, the entity pair generation unit 1003 is further configured to: determine the coreference relationship between each antecedent entity and each pronoun entity; input each antecedent entity, each pronoun entity and the coreference relationship into the reference search model, and output the corresponding entity pair vector; and generate word clusters based on the entity pair vector.

[0155] In some embodiments, the entity pair generation unit 1003 is further configured to: determine the semantic vector corresponding to each word cluster; randomly combine each semantic vector to obtain a combined vector; calculate the similarity between each semantic vector in each combined vector; and merge the word clusters corresponding to the semantic vectors with similarity greater than a first preset threshold to obtain a cluster.

[0156] In some embodiments, the data processing apparatus further includes Figure 10 The referential resolution unit (not shown) is configured to: replace the corresponding object entities of the antecedent entity and the referential entity based on the entity pair, and then update the data of each conference.

[0157] In some embodiments, the data processing apparatus further includes Figure 10 The sentence segmentation unit (not shown) is configured to: segment the updated meeting data into sentences to generate initial segmentation sentences; call the sentence segmentation model to obtain the corresponding intent similarity based on each initial segmentation sentence; add sentence segmentation markers between the initial segmentation sentences with intent similarity less than a second preset threshold, thereby generating segmented meeting data.

[0158] In some embodiments, the summary data generation unit 1005 is further configured to: divide each segmented conference data into regions based on region identifiers to obtain sentences in each region; and call an information matching model to extract keywords from the divided sentences in each region to obtain corresponding keywords.

[0159] In some embodiments, the summary data generation unit 1005 is further configured to: determine the subject in each keyword based on the context of the updated meeting data, and then determine the subject statement in each subject of each meeting data; perform information joint matching on the subject statements to generate a set of key items; calculate the content similarity between items in the set of key items, and then remove items with content similarity greater than a third preset threshold to generate a deduplicated set of key items; retrieve negative identifiers in the deduplicated set of key items, and remove items corresponding to the negative identifiers to generate a target set of key items; determine the corresponding target set of key items according to the region identifier, and then add the corresponding target set of key items to the region corresponding to the region identifier, thereby generating summary data and outputting it.

[0160] In some embodiments, the entity pair generation unit 1003 is further configured to combine each antecedent entity and each pronoun entity in each cluster with the corresponding object entity to generate entity pairs.

[0161] In some embodiments, the data processing apparatus further includes Figure 10 The word sense disambiguation unit (not shown) is configured to: segment each conference data into words to obtain segmented data; convert the segmented data into corresponding pinyin, and then convert the pinyin into corresponding word vectors; input the word vectors into a word sense disambiguation model, output a similarity score, and then determine the correct corpus sentence that is similar to the conference data based on the similarity score; call a clustering probability model to determine the probability of matching each corpus in the correct corpus sentence with the segmented data; determine the corpus with a probability greater than a preset threshold in the correct corpus sentence as the target replacement corpus, replace the segmented data with the target replacement corpus, and then update the conference data.

[0162] It should be noted that the data processing method and data processing device in this application are related in terms of specific implementation, so repeated content will not be described again.

[0163] Figure 11 An exemplary system architecture 1100 is shown that can be applied to the data processing method or data processing apparatus of the embodiments of this application.

[0164] like Figure 11 As shown, system architecture 1100 may include terminal devices 1101, 1102, and 1103, network 1104, and server 1105. Network 1104 is used as a medium to provide communication links between terminal devices 1101, 1102, and 1103 and server 1105. Network 1104 may include various connection types, such as wired or wireless communication links or fiber optic cables, etc.

[0165] Users can use terminal devices 1101, 1102, and 1103 to interact with server 1105 via network 1104 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 1101, 1102, and 1103, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, social media platform software, etc. (for example only).

[0166] Terminal devices 1101, 1102, and 1103 can be various electronic devices with data processing screens and support web browsing, including but not limited to smartphones, tablets, laptops, and desktop computers.

[0167] Server 1105 can be a server providing various services, such as a backend management server supporting data processing requests submitted by users using terminal devices 1101, 1102, and 1103 (for example only). The backend management server can receive data processing requests, determine the corresponding user identifier and the corresponding time identifier; obtain meeting data corresponding to each user identifier, sort the meeting data based on the time identifier, and then input the sorted meeting data into a referential search model to extract the corresponding antecedent entities and referential entities; generate word clusters based on the antecedent entities and referential entities, cluster each word cluster to obtain each cluster, determine the object entities corresponding to each cluster, and then generate entity pairs; call a classification model to perform regional classification on each meeting data, generating regional identifiers corresponding to each meeting data; extract keywords for the corresponding meeting data based on the entity pairs and regional identifiers, and generate and output summary data based on the keywords and regional identifiers. The process involves generating word clusters based on antecedent entities and referential entities, clustering each word cluster to obtain each cluster, determining the object entities corresponding to each cluster, and then generating entity pairs. Based on entity pairs and region identifiers, keywords corresponding to the meeting data are extracted. Summary data is generated and output based on these keywords and region identifiers. Intelligent analysis and deep learning algorithms are applied to the text content to automatically correct various referential errors and semantic distortions, resulting in clear referential meaning in the generated meeting summaries. This improves the model's processing capabilities from both single-sentence and dialogue perspectives, enhancing the accuracy and completeness of the generated meeting summary structure. A solution for summarizing key information and generating summaries of multi-user, multi-turn dialogue text in online meetings is provided. This involves collecting online dialogue data, processing it into text, and then performing content preprocessing and deep learning-based information extraction and classification. The deep learning model effectively distinguishes information relationships between different roles and fully explores the content connections between contexts, resulting in more comprehensive recognition and higher accuracy.

[0168] It should be noted that the data processing method provided in this application embodiment is generally executed by server 1105, and correspondingly, the data processing device is generally located in server 1105.

[0169] It should be understood that Figure 11 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0170] The following is for reference. Figure 12 It shows a schematic diagram of the structure of a computer system 1200 suitable for implementing a terminal device according to the embodiments of this application. Figure 12 The terminal device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0171] like Figure 12 As shown, the computer system 1200 includes a central processing unit (CPU) 1201, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 1202 or a program loaded from storage section 1208 into random access memory (RAM) 1203. The RAM 1203 also stores various programs and data required for the operation of the computer system 1200. The CPU 1201, ROM 1202, and RAM 1203 are interconnected via a bus 1204. An input / output (I / O) interface 1205 is also connected to the bus 1204.

[0172] The following components are connected to I / O interface 1205: an input section 1206 including a keyboard, mouse, etc.; an output section 1207 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 1208 including a hard disk, etc.; and a communication section 1209 including a network interface card such as a LAN card, modem, etc. The communication section 1209 performs communication processing via a network such as the Internet. A drive 1210 is also connected to I / O interface 1205 as needed. A removable medium 1211, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 1210 as needed so that computer programs read from it can be installed into storage section 1208 as needed.

[0173] 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 communication section 1209, and / or installed from removable medium 1211. When the computer program is executed by central processing unit (CPU) 1201, it performs the functions defined above in the system of this application.

[0174] It should be noted that the computer-readable medium shown in this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. Computer-readable storage media can be, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, 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 application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media can also be any computer-readable medium other than computer-readable storage media, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0175] 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 a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may 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.

[0176] The units described in the embodiments of this application can be implemented in software or hardware. The described units can also be housed in a processor; for example, a processor may be described as including a receiving unit, an entity extraction unit, an entity pair generation unit, a region identifier generation unit, and a summary data generation unit. The names of these units do not necessarily limit the specific unit itself.

[0177] In another aspect, this application also provides a computer-readable medium, which may be included in the device described in the above embodiments; or it may exist independently and not assembled into the device. The computer-readable medium carries one or more programs that, when executed by the device, cause the device to receive a data processing request, determine the corresponding user identifier and the time identifier corresponding to the user identifier; acquire meeting data corresponding to each user identifier, sort the meeting data based on the time identifier, and then input the sorted meeting data into a referential search model to extract the corresponding antecedent entities and referential entities; generate word clusters based on the antecedent entities and referential entities, cluster each word cluster to obtain each cluster, determine the object entities corresponding to each cluster, and then generate entity pairs; call a classification model to perform regional classification on each meeting data, generating regional identifiers corresponding to each meeting data; extract keywords of the corresponding meeting data according to the entity pairs and each regional identifier, and generate and output summary data based on the keywords and regional identifiers. By generating word clusters based on antecedent and pronoun entities, and clustering these word clusters to obtain individual clusters, the corresponding object entities for each cluster are determined, and entity pairs are generated. Based on the entity pairs and region identifiers, keywords from the corresponding conference data are extracted. Summary data is then generated and output based on the keywords and region identifiers.

[0178] According to the technical solution of this application, intelligent analysis and deep learning algorithms are applied to text content to automatically correct various referential errors and semantic distortions, thereby making the generated meeting summary clear and improving the model's processing capabilities based on single sentences and dialogue perspectives of meeting data, thus enhancing the accuracy and completeness of the generated meeting summary structure. A solution can be provided for summarizing key information and generating summaries of multi-user, multi-turn dialogue text in online meetings. This is achieved by collecting online dialogue data, processing it into text, and then performing content preprocessing and deep learning-based information extraction and classification. Specifically, the deep learning model layer can effectively distinguish the information relationships between different roles and fully explore the content connections between contexts, resulting in more comprehensive recognition and higher accuracy.

[0179] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can occur depending on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A data processing method, characterized in that, include: Receive a data processing request and determine the corresponding user identifier and the time identifier corresponding to the user identifier; Obtain the meeting data corresponding to each user identifier, sort the meeting data based on the time identifier, and then input the sorted meeting data into the referential search model to extract the corresponding antecedent entity and referential entity. Word clusters are generated based on the antecedent entities and the pronoun entities. Each word cluster is clustered to obtain each cluster. The object entities corresponding to each cluster are determined, and then entity pairs are generated. The step of generating word clusters based on the antecedent entities and the pronoun entities includes: determining the coreference relationship between each of the antecedent entities and each of the pronoun entities; inputting each of the antecedent entities, each of the pronoun entities, and the coreference relationship into a pronoun search model, and outputting the corresponding entity pair vectors; and generating word clusters based on the entity pair vectors. The meeting data is classified by region, and a region identifier corresponding to each meeting data is generated. The content processing model based on the dialogue corpus is invoked to extract keywords from the corresponding meeting data according to the entity pairs and the region identifiers. Based on the keywords and the region identifiers, summary data is generated and output.

2. The method according to claim 1, characterized in that, The clustering of each of the aforementioned word clusters includes: Determine the semantic vector corresponding to each of the aforementioned word clusters; The semantic vectors are randomly combined to obtain a combined vector; Calculate the similarity between the semantic vectors in each of the combined vectors; The word clusters corresponding to semantic vectors with similarity greater than a first preset threshold are merged to obtain clusters.

3. The method according to claim 1, characterized in that, Before extracting the keywords corresponding to the meeting data, the method further includes: Based on the entity pairs, the corresponding object entities of the antecedent entities and the pronoun entities are replaced, thereby updating the meeting data.

4. The method according to claim 3, characterized in that, Before extracting the keywords corresponding to the meeting data, the method further includes: The updated meeting data is segmented into statements to generate initial segmentation statements. The sentence segmentation model is invoked to obtain the corresponding intent similarity based on each of the initially segmented sentences; Sentence segmentation markers are added between the initial segmentation statements corresponding to intent similarity values ​​that are less than a second preset threshold, thereby generating sentence segmentation conference data.

5. The method according to claim 4, characterized in that, The keywords extracted from the corresponding meeting data include: Based on the region identifier, the data from each of the sentence segmentation meetings are divided into regions to obtain sentences for each region. The information matching model is invoked to extract keywords from the segmented regional statements, thus obtaining the corresponding keywords.

6. The method according to claim 5, characterized in that, The process of generating and outputting summary data based on the keywords and the region identifier includes: Based on the context of the updated meeting data, the subject in each keyword is determined, and then the subject statement in each subject in the meeting data is determined. The subject statement is subjected to information joint matching to generate a set of key items; Calculate the content similarity between each item in the set of key items, and then remove items with content similarity greater than a third preset threshold to generate a deduplicated set of key items; Retrieve negative identifiers from the set of key items to be deduplicated, and remove the items corresponding to the negative identifiers to generate the target set of key items; Based on the region identifier, the corresponding set of target key matters is determined, and then the corresponding set of target key matters is added to the region corresponding to the region identifier, thereby generating summary data and outputting it.

7. The method according to claim 1, characterized in that, The generated entity pairs include: Each antecedent entity and each pronoun entity in each cluster is combined with its corresponding object entity to generate an entity pair.

8. The method according to claim 1, characterized in that, Before sorting the meeting data based on the time identifier, the method further includes: The meeting data described above is segmented into words to obtain segmented data; The segmented data is converted into corresponding pinyin, and then the pinyin is converted into corresponding word vectors; The word vectors are input into a word sense disambiguation model, which outputs a similarity score. Then, based on the similarity score, the correct corpus sentence that is similar to the conference data is determined. A clustering probability model is invoked to determine the probability that the segmented data matches each corpus digit in the correct corpus sentence; The corpus with a probability greater than a preset threshold in the correct corpus sentences is identified as the target replacement corpus. The word segmentation data is replaced with the target replacement corpus, and then updated to the conference data.

9. A data processing apparatus, characterized in that, include: The receiving unit is configured to receive a data processing request and determine the corresponding user identifier and the time identifier corresponding to the user identifier. The entity extraction unit is configured to acquire meeting data corresponding to each user identifier, sort the meeting data based on the time identifier, and then input the sorted meeting data into the referential search model to extract the corresponding antecedent entity and referential entity. The entity pair generation unit is configured to generate word clusters based on the antecedent entity and the pronoun entity, cluster each word cluster to obtain each cluster, determine the object entity corresponding to each cluster, and then generate entity pairs. The step of generating word clusters based on the antecedent entities and the pronoun entities includes: determining the coreference relationship between each of the antecedent entities and each of the pronoun entities; inputting each of the antecedent entities, each of the pronoun entities, and the coreference relationship into a pronoun search model, and outputting the corresponding entity pair vectors; and generating word clusters based on the entity pair vectors. The region identifier generation unit is configured to classify each of the meeting data into regions and generate a region identifier corresponding to each of the meeting data. The summary data generation unit is configured to invoke a content processing model based on dialogue corpus to extract keywords from the corresponding meeting data according to the entity pairs and each of the region identifiers, and generate and output summary data based on the keywords and the region identifiers.

10. A data processing electronic device, characterized in that, include: One or more processors; Storage device for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-8.

11. A computer-readable medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-8.

Images