A method for quantifying the quality of online user interaction experience by integrating large models and knowledge graphs

Abstract

The present invention provides a method for quantifying the quality of online user interaction experience that integrates a large model and a knowledge graph. The quantification method obtains the bullet screen data of the audio content of the online audio and video platform and the audio data of the drama, and constructs a knowledge graph with the bullet screen data and the audio data of the drama. The discussion similarity between the bullet screen sent by each user in each audio data and other bullet screens, as well as the discussion correlation between the sent bullet screen and the audio data are calculated in combination with the constructed knowledge graph. The user experience quality is quantified according to the discussion similarity and discussion correlation. Compared with the prior art, the discussion similarity and discussion correlation represent the discussion situation of the user through the bullet screen and the correlation between the bullet screen and the audio content, reflecting the user's subjective experience of the corresponding audio content. The user experience quality is calculated by the discussion similarity and discussion correlation, which can better quantify the user experience.

Description

Technical Field

[0001] The present invention relates to the field of online user interaction experience data processing, and more specifically, to a method for quantifying the quality of online user interaction experience by integrating a large model and a knowledge graph. Background Art

[0002] Nowadays, there are experience economies such as online barrage audio and video platforms. This type of experience economy is mainly to satisfy consumers' pleasure. Although this type of experience economy is more difficult to meet users' real needs, its profitability as an experience economy is greater. In order to improve the profitability of the experience economy, it is necessary to meet users' needs as much as possible, and users' needs can be obtained through user experience quality; user experience quality (Quality of Experience, QoE) is a subjective feeling of a service generated by users in the process of interacting with services or applications. It can directly reflect the user's recognition of the service. However, user experience quality is a subjective feeling of users and is difficult to obtain directly. Even if it can be obtained, it is difficult to form a set of standards to quantify user experience quality. Summary of the invention

[0003] The present invention aims to overcome at least one defect of the above-mentioned prior art and provide a method for quantifying the quality of online user interaction experience that integrates a large model and a knowledge graph, which is used to quantify the user experience quality of online experience economy platforms such as online barrage audio and video platforms, and provide a basis for obtaining user needs.

[0004] The technical solution adopted by the present invention is:

[0005] A first aspect of the present invention provides a method for quantifying the quality of online user interaction experience by integrating a large model and a knowledge graph, the method comprising:

[0006] Acquire original data of audio content of an online audio-visual platform, wherein the original data includes episode information and audio data of all episodes of the audio content, and barrage data and user information of barrage in the audio content; the audio data includes audio description data and audio files;

[0007] Classify and store the original data;

[0008] Preprocessing the classified and stored raw data respectively according to the categories;

[0009] Mining the character aliases in all episodes according to the preprocessed raw data to construct a character alias knowledge graph of the audio content;

[0010] Calculate the discussion similarity between each user's corresponding bullet screen in each audio data according to the preprocessed original data;

[0011] Calculate the discussion relevance between each user's bullet comment corresponding to each audio data and the audio data according to the role alias knowledge graph and the preprocessed original data;

[0012] The user experience quality of each user is calculated according to the discussion relevance and discussion similarity.

[0013] As an important means of interaction between users and audio content in online audio and video platforms, barrage can well reflect the user's real-time experience of audio content. Through the barrage data of the barrage in the audio content, the user's subjective feelings about the audio content can be effectively obtained. The higher the discussion between users through barrage and the more relevant the barrage to the audio content, the better the user experience. Therefore, by obtaining the discussion similarity and discussion relevance of the barrage of the audio of each user in each episode, the user experience quality can be calculated by quantifying the discussion similarity and discussion relevance, which can well quantify the user's subjective experience.

[0014] Furthermore, the classifying of the original data specifically includes:

[0015] The original data is divided into structured data and audio file data and then stored accordingly;

[0016] The structured data includes the user information, the episode information, the barrage data, and the audio description data in the audio data;

[0017] The audio file data includes an audio file in the audio data.

[0018] The structured data mainly includes text data with a certain structural format, and the audio file data is the audio file corresponding to the audio data. By classifying the original data, the processing and management of the corresponding data are facilitated.

[0019] Furthermore, the classified and stored raw data is preprocessed according to the categories, specifically including:

[0020] For the structured data:

[0021] Obtaining basic information of the bullet screen and corresponding bullet screen text in the bullet screen data;

[0022] Obtaining a bullet-chat embedding vector corresponding to the bullet-chat text in the bullet-chat data through a pre-trained language representation model;

[0023] The barrage embedding vector, the corresponding barrage basic information and the barrage text are stored in association;

[0024] For the audio file data:

[0025] Cutting the audio file data into a preset number of audio segments;

[0026] Obtaining audio segment information corresponding to each of the audio segments after cutting;

[0027] Converting the audio clip into corresponding audio text data through a pre-trained speech recognition model;

[0028] The audio text data and the corresponding audio segment information are stored in association.

[0029] The preprocessing of the original data mainly includes processing the structured data and audio file data in the original data, wherein the preprocessing of the structured data mainly includes extracting features from the barrage data in the structured data and converting them into corresponding barrage embedding vectors, which can more conveniently calculate the similarity of features between corresponding barrage data; for audio files, the audio files are converted into corresponding audio text data, which can facilitate the subsequent acquisition of the correlation between the key information in the text of the corresponding audio file and the barrage text through text recognition.

[0030] Furthermore, mining the character aliases in all episodes according to the preprocessed raw data to construct a character alias knowledge graph of the audio content specifically includes:

[0031] Acquire all text-related data in the preprocessed original data to construct a text-related data set;

[0032] Perform entity discovery on the text-related data set using a named entity recognition algorithm to obtain an entity set;

[0033] Unsupervised role alias semantic mining is performed on the entity set based on pattern matching of suffixes and pinyin approximation, and the role alias knowledge graph is constructed according to the mining results.

[0034] Danmu is sent by users based on their personal experience. Different users have different cognition, understanding, and habits. Different names may be used to describe different roles or names in the audio data, and different names will affect the calculation and judgment of the discussion relevance and discussion similarity of the Danmu. Therefore, before calculating the discussion relevance and discussion similarity of the Danmu, it is necessary to construct the role alias knowledge graph. The role alias knowledge graph is obtained by mining the text-related data of the original data, and can unify different role name descriptions into the same entity, thereby improving the accuracy of the calculation of the discussion relevance and discussion similarity of the Danmu.

[0035] Furthermore, the calculation of the discussion similarity between each user's bullet screen in each audio data according to the preprocessed original data specifically includes:

[0036] Obtain all bullet comments corresponding to each audio data of each user, and respectively construct a bullet comment list corresponding to each audio data of each user;

[0037] Each bullet in the bullet list is the central bullet , obtain the audio data corresponding to the center bullet screen All bullet comments whose time span does not exceed the preset time range are constructed to build a set of similar bullet comments for each bullet comment , the similar barrage collection The calculation formula is expressed as:

[0038]

[0039] In the formula, Indicates the central bullet screen The time of appearance, Indicates any bullet comment corresponding to the audio data The time of appearance, Indicates the preset time range;

[0040] Traverse each of the similar bullet comment sets , calculate the similar bullet comment set In each barrage The corresponding central bullet screen The neighbor similarity of is expressed as follows:

[0041]

[0042] In the formula, and Respectively represent the center bullet screen and barrage The barrage embedding vector of

[0043] The similar bullet screen set is calculated according to the following formula The corresponding central bullet screen Degree of discussion :

[0044]

[0045] In the formula, represents a hyperparameter of [0,1];

[0046] The average of the discussion degrees of all the bullet comments in each bullet comment list is calculated as the discussion similarity of each user corresponding to each audio data.

[0047] Furthermore, the calculation of the discussion relevance between each user's bullet comment corresponding to each audio data and the audio data based on the role alias knowledge graph and the preprocessed original data specifically includes:

[0048] Get the appearance time of each bullet comment in the bullet comment list ;

[0049] According to the occurrence time Obtain the audio text data and the drama information corresponding to each bullet screen, and construct an associated text set for each bullet screen according to the corresponding audio text data and the drama information;

[0050] Performing role entity recognition on the barrage text of each barrage and the corresponding associated text set through the role alias knowledge graph, replacing and unifying the role name;

[0051] Fuzzily match the keywords of the bullet text of each bullet with the corresponding associated text set. If the keywords are matched, calculate the text similarity between the bullet text and the corresponding associated text set through the language representation model. If the text similarity exceeds a preset similarity threshold, use the text similarity as the correlation between the bullet text and the corresponding associated text set.

[0052] Otherwise, the relevance between the bullet comment and the corresponding associated text set is calculated through the large model;

[0053] The average of the relevance of all the bullet comments in each bullet comment list is calculated as the discussion relevance of each user corresponding to each audio data.

[0054] Furthermore, the calculation of the relevance between the bullet comment and the corresponding associated text set by using the large model specifically includes:

[0055] Construct a relevance calculation prompt according to the barrage text and the corresponding associated text set;

[0056] The constructed relevance calculation prompt is input into the large model, and the output result of the large model is used as the relevance between the bullet screen and the corresponding associated text set.

[0057] For the associated text set and barrage text that exceed the similarity threshold, the relevance calculation prompt is constructed, and the relevance calculation prompt includes the corresponding associated text set and barrage text, as well as auxiliary description text. The relevance calculation prompt is identified in the large model, and the existing mature language recognition large model is used to further analyze the text similarity of the associated text set and barrage text, which can further improve the accuracy of the relevance calculation.

[0058] Furthermore, the calculating of the user experience quality of each user according to the discussion relevance and discussion similarity specifically includes:

[0059] Obtaining the discussion similarity and discussion relevance of each user corresponding to each of the audio data;

[0060] Normalizing the discussion similarity and discussion relevance;

[0061] According to the normalized discussion similarity and discussion relevance, the user experience quality of each user corresponding to each audio data is calculated by the following formula:

[0062]

[0063] In the formula, represents the user experience quality of user u corresponding to the audio data s, and They respectively represent the discussion similarity and discussion relevance of user u corresponding to the audio data s, and a and b respectively represent quality thresholds of [0,1], where a≥b.

[0064] According to the specific values ​​of the discussion similarity and discussion relevance, the user's subjective experience is quantified into different levels of user experience quality, which can intuitively obtain the user's subjective experience of different audio data content, and then better optimize and adjust the audio data to further improve the user experience.

[0065] Furthermore, the quantification method also includes:

[0066] Constructing a filtering dictionary, wherein the filtering dictionary includes a plurality of filtering keywords;

[0067] The bullet screen text is filtered according to the filtering keywords in the filtering dictionary.

[0068] In actual scenarios, there may be some interfering barrages in the barrages sent by users, which will affect the calculation of the discussion similarity and discussion relevance of the barrages. Therefore, by constructing a filtering dictionary, after preprocessing the original data, the barrages are filtered according to the filtering dictionary to reduce the interfering barrages in the barrages and improve the accuracy of the calculation of the discussion similarity and discussion relevance.

[0069] Compared with the prior art, the present invention has the following beneficial effects:

[0070] 1. The present invention obtains the bullet screen data of the audio content and the audio data of the drama on the online audio-visual platform, and calculates the discussion similarity between the bullet screen sent by each user in each audio data and other bullet screens, as well as the discussion correlation between the sent bullet screen and the audio data according to the sending situation of the bullet screen data of each user in each audio data. The discussion similarity and discussion correlation represent the discussion situation of the user through the bullet screen and the correlation between the bullet screen and the audio content, reflecting the user's subjective experience of the corresponding audio content. The user experience quality is calculated by the discussion similarity and discussion correlation, which can better quantify the user experience;

[0071] 2. The present invention constructs a knowledge graph of role aliases, which is constructed based on the barrage data of audio content. It can reflect the user's description of different character names in the drama, and can uniformly map different name descriptions of the same character thing into the same entity, so as to avoid different character name descriptions affecting the calculation of the discussion similarity and discussion relevance. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 The figure is a flowchart of the steps of the quantification method of the present invention.

[0073] Figure 2 The flowchart is a step chart of the preprocessing of structured data of the present invention.

[0074] Figure 3 The present invention is a flowchart of the steps of preprocessing audio data.

[0075] Figure 4 A flowchart of the steps for constructing the knowledge graph of the present invention.

[0076] Figure 5 The following is a flowchart of the steps of similarity calculation for discussion of the present invention.

[0077] Figure 6 The flowchart of the steps of calculating the relevance for discussion of the present invention is shown in FIG. DETAILED DESCRIPTION

[0078] The drawings of the present invention are only for illustrative purposes and should not be construed as limiting the present invention. In order to better illustrate the following embodiments, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of the actual product; it is understandable to those skilled in the art that some well-known structures and their descriptions in the drawings may be omitted.

[0079] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present application.

[0080] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0081] As a typical carrier of experience economy, online barrage audio and picture platform is an important economic type in the current economic form. The audio content in online barrage audio and picture platform is a typical experience-based consumer product. Online barrage audio and picture platform satisfies users' leisure and entertainment needs by providing them with audio content products such as audio books, audio images, audio dramas, etc. It is very difficult to obtain users' real-time experience in conventional experience economy.

[0082] However, while the online barrage audio and video platform provides users with online consumption experience services, it also meets users' interactive needs through a special interactive means - barrage. Users can send barrages in the audio content provided by the online barrage platform to express their personal experience, and interact with the providers of audio content and other users of audio content through barrages. Therefore, we can obtain users' real-time experience of audio content through the barrage data in the audio content of the online barrage audio and video platform, and then extract and quantify the user's experience quality.

[0083] This embodiment provides a method for quantifying the quality of experience of users of the online audio and video platform by utilizing the bullet screen data of the online audio and video platform. The specific implementation methods of this application are described in detail below in conjunction with the accompanying drawings.

[0084] like Figure 1 As shown, this embodiment provides a method for quantifying the quality of online user interaction experience by integrating a large model and a knowledge graph. The quantification method may specifically include:

[0085] S1: Obtain the original data of the audio content of the online audio-visual platform;

[0086] In this embodiment, the original data of the audio content of the online audio and image platform can be crawled by using crawler technology. Specifically, the address source of the online audio and image platform to be crawled is first determined, and after obtaining the address source, a crawler framework is constructed according to the page structure of the online audio and image platform to be crawled and the content of the original data, and the crawler software is constructed according to the crawler framework. The original data is crawled from the online audio and image platform through the constructed crawler software.

[0087] Preferably, the original data includes episode information and audio data of all episodes of the audio content, wherein the episode represents each segment of the audio content, and the episode information represents the situation of each segment of the audio content. For example, if the audio content is an audio novel, the episode represents each chapter of the audio novel, and the episode information is an introduction or description of each chapter; the audio data includes audio description data and an audio file, and the audio description data can be obtained by analyzing the audio description data in advance, and the audio description data is used to describe the characteristics of various sounds in the audio data and mark the specific time of their occurrence; it can be understood that the episode and the audio data are in a one-to-many relationship, that is, one episode can include a plurality of audio data;

[0088] In addition, the original data also includes the barrage data and user information of the barrage in the audio content; the barrage data includes the basic information of the barrage and the barrage text; the basic information of the barrage includes at least the time when the barrage appears, the user information sending it, etc.; the barrage text is the text content sent by the barrage.

[0089] S2: Classify and store the original data;

[0090] In this embodiment, the original data can be divided into structured data and audio file data; the structured data mainly includes the user information, the episode information, the barrage data, and the audio description data in the audio data; the audio file data includes the audio file in the audio data;

[0091] Dividing the original data into the structured data and the audio file data can facilitate the storage of the original data according to different types of data, wherein the structured data is mainly text-type data, and the text-type data is convenient for storage and management according to the text content, and the audio file data can be stored and managed according to the storage time or the name of the file.

[0092] In a preferred implementation of this embodiment, the structured data can be stored in the form of a two-dimensional table, that is, the user information, the drama information, the barrage data, and the audio description data in the audio data are respectively constructed into a user information table t_user, a drama information table t_drama, a barrage information table t_danmu, and an audio information table t_sound; each two-dimensional table is associated with each other, and an associated query between the two-dimensional tables can be realized; since the structured data is mainly text-type data, the structured data can be stored in a document database;

[0093] The audio file data may be named and stored by a name associated with the audio information table, such as t_sound.sound_id, and the corresponding audio file data may be found through the audio information table.

[0094] S3: pre-processing the classified and stored raw data according to the categories;

[0095] After the raw data are classified and stored, in order to facilitate subsequent calculation and processing, the raw data need to be preprocessed according to the specific types of the raw data. Specifically,

[0096] like Figure 2 As shown, for the structured data in the original data:

[0097] A1: Obtaining basic information of the bullet screen and corresponding bullet screen text in the bullet screen data;

[0098] A2: Obtaining a bullet-chat embedding vector corresponding to the bullet-chat text in the bullet-chat data through a pre-trained language representation model;

[0099] A3: The barrage embedding vector, the corresponding barrage basic information and the barrage text are stored in association.

[0100] The processing of the structured data is mainly aimed at the barrage data, and the basic information of the barrage and the corresponding barrage text are separated, because the basic information of the barrage mainly records the user information and time information of the barrage, and the information format in the barrage information is relatively fixed, while the barrage text is text sent by the user. Different users may use different words and sentences, and even describe the same thing differently, so the barrage text is separated, converted into the corresponding barrage embedding vector, and the content of the barrage text is normalized to facilitate feature recognition of the barrage embedding vector.

[0101] like Figure 3 As shown, for the audio file data in the original data:

[0102] B1: cutting the audio file data into a preset number of audio segments;

[0103] B2: Obtaining audio segment information corresponding to each of the audio segments after cutting;

[0104] B3: Convert the audio clip into corresponding audio text data through a pre-trained speech recognition model.

[0105] Cutting the audio file data into a preset number of audio segments can reduce the types of sounds contained in each audio segment, so that the barrage sent by the user can have a higher correlation with the sounds in the audio segment; wherein the preset number of segments is actually set according to the length of the audio file data, so as to reduce the types of sounds contained in each audio segment as much as possible, so that the user's barrage is more targeted.

[0106] Preferably, the preprocessed audio segment is converted into the corresponding audio text data, which can be stored in a document database together with the corresponding structured data, so as to facilitate the associated query of the audio text data and the corresponding structured data.

[0107] Furthermore, in a preferred implementation of this embodiment, the quantization method further includes, after preprocessing the original data:

[0108] Constructing a filtering dictionary, wherein the filtering dictionary includes a plurality of filtering keywords;

[0109] The bullet screen text is filtered according to the filtering keywords in the filtering dictionary.

[0110] In actual scenarios, there may be some interfering barrages in the barrages sent by users, which will affect the calculation of the discussion similarity and discussion relevance of the barrages. Therefore, by constructing a filtering dictionary, after preprocessing the original data, the barrages are filtered according to the filtering dictionary to reduce the interfering barrages in the barrages and improve the accuracy of the calculation of the discussion similarity and discussion relevance.

[0111] S4: mining the character aliases in all episodes according to the preprocessed raw data, and constructing a character alias knowledge graph of the audio content;

[0112] Different users may have certain deviations in their understanding of the same role object, or in their habitual expressions. As a result, different users may use different names to describe the same role object, which ultimately affects recognition judgment and quantitative calculation. Therefore, it is necessary to unify the names of role objects.

[0113] In this embodiment, a role alias knowledge graph is constructed based on the original data of the audio content, such as Figure 4 As shown, the construction of the role alias knowledge graph specifically includes:

[0114] S41: Acquire all text-related data in the preprocessed original data to construct a text-related data set;

[0115] In this embodiment, specifically, all episodes of the audio content can be traversed to obtain all text-related data of all episodes, including the episode information, audio description data, basic information of barrage, barrage text and audio text data, and the episode information, audio description data, basic information of barrage, barrage text and audio text data are constructed as the text-related data set;

[0116] S42: Perform entity discovery on the text-related data set by using a named entity recognition algorithm to obtain an entity set;

[0117] In this embodiment, the text-related data set is traversed. Since the text-related data set is obtained based on the original data of the audio content and is only valid within the audio content, it is possible to avoid relying on entity links in an external knowledge base in the entity discovery process, which interferes with the entity discovery process.

[0118] S43: Perform unsupervised role alias semantic mining on the entity set based on pattern matching of suffixes and pinyin approximation, and construct the role alias knowledge graph according to the mining results.

[0119] By performing role alias semantic mining on the entity set, the role alias knowledge graph finally obtained can effectively map the same role things described by different names into a same entity, avoiding interference from different name descriptions.

[0120] S5: Calculate the discussion similarity between each user's corresponding bullet screen in each of the audio data according to the preprocessed original data;

[0121] In this embodiment, if Figure 5 As shown, the specific steps of step S5 may include:

[0122] S51: Obtain all bullet comments corresponding to each audio data of each user, and respectively construct a bullet comment list corresponding to each audio data of each user;

[0123] In this embodiment, since the data in the original data are stored in association with each other, all the bullet comments sent by the user in each audio data can be obtained through association query and constructed into the bullet comment list. For example, if a user u sends 4 bullet comments in an audio data s, the bullet comment list corresponding to the audio data s contains 4 bullet comments of the user u.

[0124] S52: Taking each bullet screen in the bullet screen list as the central bullet screen , obtain the audio data corresponding to the center bullet screen All bullet comments whose time span does not exceed the preset time range are constructed to build a set of similar bullet comments for each bullet comment ;

[0125] The similar bullet comments collection The calculation formula is expressed as:

[0126]

[0127] In the formula, Indicates the central bullet screen The time of appearance, Indicates any bullet comment corresponding to the audio data The time of appearance, represents the preset time range, which is set according to actual needs; the similar bullet screen set It shows the discussion of the corresponding barrage within the time range;

[0128] S53: Traverse each of the similar bullet comment sets , calculate the similar bullet comment set In each barrage The corresponding central bullet screen The nearest neighbor similarity of

[0129] The calculation formula of the nearest neighbor similarity is expressed as:

[0130]

[0131] In the formula, and Respectively represent the center bullet screen and barrage The barrage embedding vector of and the center barrage The higher the similarity of the neighbors, the stronger the barrage and the center barrage The content is closer;

[0132] S54: Calculate the similar bullet comment set The corresponding central bullet screen Degree of discussion;

[0133] The calculation formula of the discussion degree is expressed as:

[0134]

[0135] In the formula, Indicates the similar bullet comments set The corresponding central bullet screen The degree of discussion, represents a hyperparameter of [0,1];

[0136] S55: Calculate the average of the discussion degrees of all the bullet comments in each bullet comment list as the discussion similarity of each user corresponding to each audio data.

[0137] Since the neighbor similarity represents the barrage and the center barrage The similarity of the content, if within a certain time range, that is, based on the similar barrage set The calculated discussion degree can reflect the discussion enthusiasm among users for the same topic. The higher the discussion degree, the better the corresponding user experience. Therefore, the average of the discussion degrees of all barrages in each barrage list is calculated as the discussion similarity of each user corresponding to each audio data, and the discussion similarity is used as a criterion for quantifying the quality of user experience.

[0138] S6: Calculate the discussion relevance between each user's bullet comment corresponding to each audio data and the audio data according to the role alias knowledge graph and the preprocessed original data;

[0139] In this embodiment, specifically, Figure 6 The specific steps of step S6 may include:

[0140] S61: Obtaining the appearance time of each bullet comment in each bullet comment list ;

[0141] In this embodiment, the appearance time of each bullet comment It can be obtained through the basic information of the barrage corresponding to the barrage;

[0142] S62: According to the occurrence time Obtain the audio text data and the drama information corresponding to each bullet screen, and construct an associated text set for each bullet screen according to the corresponding audio text data and the drama information;

[0143] In this embodiment, the audio text data is obtained by converting the cut audio segment. The cut audio segment has less and more accurate sound information. If calculated based on the entire audio data, the audio data contains more sound information. Then, the relevance of a barrage sent by a user to the entire audio data is very low. Therefore, the corresponding audio text data and the drama information are constructed as the associated text set of each barrage, which can more accurately calculate whether the barrage text in the barrage is relevant to the audio text information of the audio segment through the associated text set, thereby avoiding interference from other sound information.

[0144] S63: Performing role entity recognition on the barrage text of each barrage and the corresponding associated text set through the role alias knowledge graph, and replacing and unifying the role name;

[0145] The constructed role alias knowledge graph is used to unify different role name descriptions of the same entity object in the barrage text and the corresponding associated text set, avoiding interference from multiple role name descriptions.

[0146] S64: Fuzzily match the keywords of the bullet text of each bullet with the corresponding associated text set. If the keywords are matched, calculate the text similarity between the bullet text and the corresponding associated text set through the language representation model. If the text similarity exceeds a preset similarity threshold, use the text similarity as the relevance between the bullet and the corresponding associated text set; otherwise, calculate the relevance between the bullet and the corresponding associated text set through the big model.

[0147] In this embodiment, the calculation of the relevance between the bullet comment and the corresponding associated text set by using the large model may specifically include:

[0148] Construct a relevance calculation prompt according to the barrage text and the corresponding associated text set;

[0149] The constructed relevance calculation prompt is input into the large model, and the output result of the large model is used as the relevance between the bullet screen and the corresponding associated text set.

[0150] The large model can be set as an existing language recognition processing large model, which has a strong language recognition capability and can effectively identify the correlation between two language texts;

[0151] Exemplarily, the relevance calculation prompt can be constructed as: {You are a user of an online audio and video platform. When you read the audio content of "XXX", [the barrage text corresponding to the barrage] appears in the barrage. The relevant text of the current audio content is [the related text set]. What is the correlation between the barrage content and the related text at this time? Give a specific score between [-11] based on your subjective understanding. []}; the large model will output a score of [-11] based on the input barrage and the corresponding related text set, and use the score as the similarity between the barrage and the corresponding related text set.

[0152] S65: Calculate the average of the relevance of all the bullet comments in each bullet comment list as the discussion relevance of each user corresponding to each of the audio data.

[0153] The discussion relevance indicates the degree of relevance between the barrage and the corresponding short audio clip and drama, reflecting whether the barrage sent by the user is related to the corresponding audio clip. For example, the higher the discussion relevance, the more interested the user is in the audio clip, indicating that the user's subjective experience is better.

[0154] S7: Calculate the user experience quality of each user according to the discussion relevance and discussion similarity;

[0155] In this embodiment, the specific steps of step S7 may include:

[0156] Obtain the discussion similarity and discussion relevance of each user corresponding to each audio data calculated in step S6;

[0157] Normalizing the discussion similarity and discussion relevance;

[0158] According to the normalized discussion similarity and discussion relevance, the user experience quality of each user corresponding to each audio data is calculated by the following formula:

[0159]

[0160] In the formula, represents the user experience quality of user u corresponding to the audio data s, and They respectively represent the discussion similarity and discussion relevance of user u corresponding to the audio data s, and a and b respectively represent quality thresholds of [0,1], where a≥b.

[0161] The discussion similarity and discussion relevance reflect the user's subjective experience of the audio data from two perspectives: between users and between users and audio content. According to the specific values ​​of the discussion similarity and discussion relevance, the user's subjective experience is quantified into different levels of user experience quality. The user's subjective experience of different audio data contents can be intuitively obtained, and the audio data can be better optimized and adjusted to further improve the user experience.

[0162] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the technical solution of the present invention, and are not intended to limit the specific implementation methods of the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the claims of the present invention shall be included in the protection scope of the claims of the present invention.

Claims

1. A method for quantifying the quality of online user interaction experience by integrating a large model and a knowledge graph, characterized in that: The quantification method comprises: Acquire original data of audio content on an online audio-visual platform, the original data including episode information and audio data of all episodes of the audio content, and barrage data and user information of barrage in the audio content; Classify and store the original data; Preprocessing the classified and stored raw data respectively according to the categories; Mining the character aliases in all episodes according to the preprocessed raw data to construct a character alias knowledge graph of the audio content; Calculate the discussion similarity between each user's corresponding bullet screen in each audio data according to the preprocessed original data; Calculate the discussion relevance between each user's bullet comment corresponding to each audio data and the audio data according to the role alias knowledge graph and the preprocessed original data; Calculate the user experience quality of each user according to the discussion relevance and discussion similarity; The classified and stored raw data are preprocessed according to the categories, specifically including: For structured data: Obtaining basic information of the bullet screen and corresponding bullet screen text in the bullet screen data; Obtaining a bullet-chat embedding vector corresponding to the bullet-chat text in the bullet-chat data through a pre-trained language representation model; The barrage embedding vector, the corresponding barrage basic information and the barrage text are stored in association; For audio file data: Cutting the audio file data into a preset number of audio segments; Obtaining audio segment information corresponding to each of the audio segments after cutting; Converting the audio clip into corresponding audio text data through a pre-trained speech recognition model; The audio text data and the corresponding audio segment information are stored in association; The calculating, based on the pre-processed original data, the discussion similarity between the bullet screens of each user corresponding to each of the audio data specifically includes: Obtain all bullet comments corresponding to each audio data of each user, and respectively construct a bullet comment list corresponding to each audio data of each user; Each bullet in the bullet list is the central bullet , obtain the audio data corresponding to the center bullet screen All bullet comments whose time span does not exceed the preset time range are constructed to build a set of similar bullet comments for each bullet comment , the similar barrage collection The calculation formula is expressed as: In the formula, Indicates the central bullet screen The time of appearance, Indicates any bullet comment corresponding to the audio data The time of appearance, Indicates the preset time range; Traverse each of the similar bullet comment sets , calculate the similar bullet comment set In each barrage The corresponding central bullet screen The neighbor similarity of is expressed as follows: In the formula, and Respectively represent the center bullet screen and barrage The barrage embedding vector of The similar bullet screen set is calculated according to the following formula The corresponding central bullet screen Degree of discussion : In the formula, represents a hyperparameter of [0,1]; Calculate the average of the discussion degrees of all the bullet comments in each bullet comment list as the discussion similarity of each user corresponding to each audio data; The step of calculating the discussion relevance between each user's bullet comment corresponding to each audio data and the audio data according to the role alias knowledge graph and the preprocessed original data specifically includes: Get the appearance time of each bullet comment in the bullet comment list ; According to the occurrence time Obtain the audio text data and the drama information corresponding to each bullet screen, and construct an associated text set for each bullet screen according to the corresponding audio text data and the drama information; Performing role entity recognition on the barrage text of each barrage and the corresponding associated text set through the role alias knowledge graph, replacing and unifying the role name; Fuzzily match the keywords of the bullet text of each bullet with the corresponding associated text set. If the keywords are matched, calculate the text similarity between the bullet text and the corresponding associated text set through the language representation model. If the text similarity exceeds a preset similarity threshold, use the text similarity as the correlation between the bullet text and the corresponding associated text set. Otherwise, the relevance between the bullet comment and the corresponding associated text set is calculated through the large model; The average of the relevance of all the bullet comments in each bullet comment list is calculated as the discussion relevance of each user corresponding to each of the audio data.

2. According to claim 1, a method for quantifying the quality of online user interaction experience by integrating a large model and a knowledge graph, characterized in that: The classifying of the original data specifically includes: The original data is divided into structured data and audio file data and then stored accordingly; The structured data includes the user information, the episode information, the barrage data, and the audio description data in the audio data; The audio file data includes an audio file in the audio data.

3. According to claim 1, a method for quantifying the quality of online user interaction experience integrating a large model and a knowledge graph is characterized in that: The mining of character aliases in all episodes according to the pre-processed raw data to construct a character alias knowledge graph of the audio content specifically includes: Acquire all text-related data in the preprocessed original data to construct a text-related data set; Perform entity discovery on the text-related data set using a named entity recognition algorithm to obtain an entity set; Unsupervised role alias semantic mining is performed on the entity set based on pattern matching of suffixes and pinyin approximation, and the role alias knowledge graph is constructed according to the mining results.

4. According to claim 1, a method for quantifying the quality of online user interaction experience integrating a large model and a knowledge graph is characterized in that: The calculating the relevance between the bullet comment and the corresponding associated text set by using the large model specifically includes: Construct a relevance calculation prompt according to the barrage text and the corresponding associated text set; The constructed relevance calculation prompt is input into the large model, and the output result of the large model is used as the relevance between the bullet screen and the corresponding associated text set.

5. According to claim 1, a method for quantifying the quality of online user interaction experience integrating a large model and a knowledge graph is characterized in that: The calculating the user experience quality of each user according to the discussion relevance and discussion similarity specifically includes: Obtaining the discussion similarity and discussion relevance of each user corresponding to each of the audio data; Normalizing the discussion similarity and discussion relevance; According to the normalized discussion similarity and discussion relevance, the user experience quality of each user corresponding to each audio data is calculated by the following formula: In the formula, represents the user experience quality of user u corresponding to the audio data s, and They respectively represent the discussion similarity and discussion relevance of user u corresponding to the audio data s, and a and b respectively represent quality thresholds of [0,1], where a≥b.

6. According to claim 1, a method for quantifying the quality of online user interaction experience integrating a large model and a knowledge graph is characterized in that: The quantification method further comprises: Constructing a filtering dictionary, wherein the filtering dictionary includes a plurality of filtering keywords; The bullet screen text is filtered according to the filtering keywords in the filtering dictionary.

Images