Data-driven causal tracing method for automobile attribute viewpoints

Abstract

The present invention provides a data-driven causal tracing method, system, storage medium and electronic device for automobile attribute opinion, which relates to the field of data processing technology. The data-driven causal tracing method for automobile attribute opinion proposed in the present invention includes two aspects: a method for constructing an automobile attribute hierarchical network and a method for tracing the causal relationship of automobile attribute opinion. It only needs to use the total text of social comments on the automobile to be analyzed, without tracing the massive text corpus, which effectively solves the problems of low coverage and strong dependence on text corpus in the existing research model; and on the basis of traditional attribute opinion mining technology, it explores the correlation between opinions in the attribute hierarchy, realizes opinion causal relationship tracing, and helps enterprises to comprehensively, quickly, accurately and efficiently discover the deep-seated reasons for user opinions, better grasp the advantages and disadvantages of automobile attributes, understand user needs, and formulate more accurate product improvement strategies.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a data-driven automobile attribute viewpoint causal tracing method, system, storage medium and electronic device. Background Art

[0002] The aggregate text of social media reviews is a crucial source for companies to understand user needs. However, due to the sparse value, fragmented content, and redundant information found in online product review data, consumers and companies struggle to quickly, comprehensively, and accurately grasp the product information provided by user feedback. Users often express themselves in a casual manner, resulting in varying dimensions and inconsistent hierarchies in online product review text. Mining the hierarchical relationships between product attribute perspectives using social media review data can help companies uncover the underlying reasons behind coarse-grained user perspectives, tracing causal relationships between perspectives and ultimately developing more precise product development and improvement strategies.

[0003] Establishing a hierarchical relationship network for product attributes is a prerequisite for constructing hierarchical relationships between opinions. Existing research has primarily used patterns such as "for example, is, of" to identify hierarchical relationships between entities using large text corpora. However, similar patterns are rarely found in user product descriptions, making it difficult to extract hierarchical relationships based solely on these patterns. Regarding opinion mining, existing research has primarily focused on sentiment extraction and classification using methods such as dependency parsing and dynamic window extraction, while neglecting the deeper underlying reasons behind opinions. In reality, there are inherently close relationships between attribute opinions. For example, in a comment like "The appearance is very impressive, especially the wide front end, which really suits my aesthetic," it can be found that the deeper reason behind the user's comment about "appearance" being "impressive" is that the "front end" is "wide." This demonstrates that opinions are correlated. Extracting correlations between opinions is a new research direction in sentiment analysis, tracing the root causes of emotions. Examining the specific underlying reasons can help companies identify the specific manifestations of attributes and optimize product design. Summary of the Invention

[0004] (1) Technical problems solved

[0005] In response to the shortcomings of the existing technology, the present invention provides a data-driven automobile attribute view causal tracing method, system, storage medium and electronic device, which solves the technical problem of ignoring the inherent close relationship between attribute viewpoints.

[0006] (2) Technical solution

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0008] A data-driven causal tracing method for vehicle attribute viewpoints, including:

[0009] S1. Obtaining car attributes and corresponding attribute opinions based on the total text of social comments on the car to be analyzed;

[0010] S2. Identify the hierarchical relationship of vehicle attributes, including:

[0011] S21. Obtain attribute pairs that may have potential hierarchical relationships based on the attributes of the vehicles.

[0012] S22. Extracting hierarchical relationship features of various automobile attributes of the attribute pairs based on the total text of the social comment;

[0013] S23, using the hierarchical relationship features of the vehicle attributes as input to the pre-trained decision tree classification model to obtain the hierarchical relationship between the two attributes in each attribute pair;

[0014] S3. Build a hierarchical network of car attributes, including:

[0015] constructing a car attribute hierarchical network based on quadruple groups, wherein the quadruple groups are composed of attribute pairs, corresponding hierarchical relationships, and relationship weights;

[0016] S4. Tracing attribute view causal relationships, including:

[0017] Based on the generated hierarchical network of automobile attributes, a hierarchical relationship between viewpoints is established to form a preliminary causal relationship tracing path of viewpoints, and the strong connection paths are screened and obtained; according to each strong connection path, the causal relationship strength between viewpoints is measured.

[0018] Preferably, in S21:

[0019] Combine any two attributes and calculate the co-occurrence frequency of the two attributes in each combination:

[0020]

[0021] Where P(AB) is the co-occurrence frequency of car attribute A and car attribute B in any attribute pair; N(S) is the total amount of social comment text S, and N(AB) is the number of comments in which car attribute A and car attribute B co-occur;

[0022] Compute the pointwise mutual information of the two attributes in each combination:

[0023]

[0024] Among them, PMI(A,B) is the point mutual information of car attribute A and car attribute B;

[0025] When both the co-occurrence frequency and the point mutual information meet corresponding thresholds, the combination is determined to be an attribute pair that may have a potential hierarchical relationship.

[0026] Preferably, the hierarchical relationship characteristics of each vehicle attribute in S22 include:

[0027] (1) Frequency characteristics:

[0028] (1.1) The frequency difference between A and B appearing alone in the total text S of social comments is expressed as f1 = P(A) 单 -P(B) 单 ;

[0029] (1.2) The frequency quotient of A and B appearing alone in S is expressed as

[0030] (1.3) The difference in conditional probabilities between A and B is expressed as f3 = P(A|B) - P(B|A);

[0031] Among them, P(A) 单 、P(B) 单 Respectively represent the frequency of A and B appearing alone in S; P(A|B) represents the conditional probability of A appearing alone in S under the condition that B appears alone in S; P(B|A) represents the conditional probability of B appearing alone in S under the condition that A appears alone in S;

[0032] (2) Location features:

[0033] (2.1) A and B are in the i-th social comment text paragraph s i The frequency difference that appears in front is expressed as f4=P(A) 前 -P(B) 前 ;

[0034] (2.2) A and B are in s i The frequency quotient of the preceding value is expressed as

[0035] Among them, P(A) 前 、P(B) 前 Respectively represent A and B in s i The frequency of the preceding occurrence in the

[0036] (3) Semantic features:

[0037] (3.1) The frequency difference between A&B and B&A is expressed as f6 = P(A&B) - P(B&A);

[0038] (3.2) The frequency quotient of A&B and B&A is expressed as

[0039] (3.3) The frequency value of A$B is expressed as f8=P(A$B);

[0040] Among them, & is the hierarchical semantic pattern {of, such as, especially, including}, P(A&B) and P(B&A) represent the frequency of the existence of hierarchical semantic patterns between A and B, and between B and A; $ is the parallel semantic template {and, or, and}, P(A$B) represents the frequency of the existence of parallel semantic patterns between A and B.

[0041] Preferably, the training process of the decision tree classification model in S23 includes:

[0042] Manually construct a hierarchical dictionary of automobile attributes, which consists of attribute pairs and corresponding hierarchical relationship labels, where the labels are: contain, belong to, and irrelevant;

[0043] Extract the hierarchical relationship features f1 to f8 of each attribute pair in the dictionary, and discretize the seven continuous data of features f1 to f7 of the training data set based on the equidistance discretization method;

[0044] According to the discretized continuous features f1 to f7 and discrete feature f8, corresponding hierarchical relationship labels are integrated and the decision tree classification model is obtained by training using the CART algorithm.

[0045] Preferably, the S3 includes:

[0046] S31. Hierarchical connection:

[0047] The car attribute without any upper-level attributes is considered a first-level attribute. Based on the two quadruple combinations of <first-level attribute, attribute B, contains, weight> or <attribute B, first-level attribute, belongs to, weight>, attribute B is identified as a second-level attribute. This process is repeated to obtain deeper car attributes until the car attribute has no lower-level attributes, stopping the growth of the hierarchical network.

[0048] S32, hierarchical relationship path pruning:

[0049] Based on the preliminary hierarchical structure diagram, overlapping and misaligned paths are pruned to obtain the final vehicle attribute hierarchical network. The hierarchical relationship path pruning rules include:

[0050] Rule 1: For overlapping pruning, the longest connection path is retained based on the longest path principle;

[0051] Rule 2: For misaligned pruning, locate the original social comment text paragraph and determine whether to retain the path based on whether it conforms to the contextual semantics.

[0052] Preferably, the causal relationship tracing of the attribute viewpoints corresponding to the attributes of the vehicle in S4 includes:

[0053] The Ochiia coefficient is used to screen out the strong connection paths between hierarchical viewpoints;

[0054]

[0055] If the Ochiia coefficient of attribute view a and attribute view b is greater than the preset threshold, it is determined that there is a strong connection path between attribute view a and attribute view b;

[0056] Causal association strength measurement: In addition to direct connections, there may be multiple indirect connections between two user views. Assume that there are r strong connection paths between attribute view a and attribute view b, and each path has z i By T i A level sequence of viewpoints, represented as in Indicates the tth i point of view, o1=a, P(o j+1 |o j ) i is o on the i-th path j+1 With o j The conditional probability of obtaining the causal relationship strength T between attribute view a and attribute view b is:

[0057]

[0058] Preferably, the S4 further includes:

[0059] The co-occurrence frequency of hierarchical opinion pairs with strong connection paths is calculated as the importance of the attribute opinion pairs. Combined with the correlation strength of the attribute opinion pairs, all attribute opinion pairs are divided into misleading opinion reason pairs, key opinion reason pairs, useless opinion reason pairs and minor opinion reason pairs.

[0060] A data-driven causal tracing system for automobile attribute viewpoints, including:

[0061] An acquisition module, configured to acquire car attributes and corresponding attribute opinions based on the total text of social comments on the car to be analyzed;

[0062] The recognition module is used to identify the relationship between car attributes, including:

[0063] A judgment submodule, configured to obtain attribute pairs that may have potential hierarchical relationships based on the attributes of each vehicle;

[0064] An extraction submodule, configured to extract hierarchical relationship features of various automobile attributes of the attribute pairs based on the total text of the social comment;

[0065] A classification submodule, configured to use the hierarchical relationship features of the vehicle attributes as input to a pre-trained decision tree classification model to obtain the hierarchical relationship between the two attributes in each attribute pair;

[0066] Building blocks for constructing a hierarchical network of car attributes, including:

[0067] constructing a car attribute hierarchical network based on quadruple groups, wherein the quadruple groups are composed of attribute pairs, corresponding hierarchical relationships, and relationship weights;

[0068] The tracing module is used to trace the causal relationship of attribute views, including:

[0069] Based on the generated hierarchical network of automobile attributes, a hierarchical relationship between viewpoints is established to form a preliminary causal relationship tracing path of viewpoints, and the strong connection paths are screened and obtained; according to each strong connection path, the causal relationship strength between viewpoints is measured.

[0070] A storage medium stores a computer program for data-driven causal tracing of automobile attribute viewpoints, wherein the computer program enables a computer to execute the automobile attribute viewpoint causal tracing method as described above.

[0071] An electronic device, comprising:

[0072] One or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, the programs including a method for executing the automobile attribute viewpoint causal tracing method as described above.

[0073] (3) Beneficial effects

[0074] The present invention provides a data-driven causal tracing method, system, storage medium, and electronic device for vehicle attribute viewpoints. Compared with existing technologies, it has the following advantages:

[0075] The data-driven causal tracing method for vehicle attribute opinion analysis proposed in this paper actually comprises two aspects: a method for constructing a hierarchical network of vehicle attributes and a method for tracing the causal relationships between vehicle attribute opinion. This method utilizes only the aggregated text of social commentary about the vehicle being analyzed, eliminating the need to trace a massive text corpus. This effectively addresses the low research coverage and strong reliance on text corpora inherent in existing approaches. Furthermore, building on traditional attribute opinion mining techniques, this method explores the correlations between opinions within the attribute hierarchy and traces the fine-grained causes of opinion expression. This helps companies more comprehensively, quickly, accurately, and efficiently trace fine-grained user opinions, better understand the strengths and weaknesses of vehicle attributes, understand user needs, and develop more precise product improvement strategies. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0077] Figure 1 This is an overall framework diagram of a data-driven automobile attribute viewpoint causal tracing method provided by an embodiment of the present invention;

[0078] Figure 2 A flowchart of a data-driven automobile attribute viewpoint causal tracing method provided by an embodiment of the present invention;

[0079] Figure 3 A flow chart of another data-driven automobile attribute viewpoint causal tracing method provided by an embodiment of the present invention

[0080] Figure 4 A schematic diagram of a strongly connected path screened by the Ochiia coefficient provided in an embodiment of the present invention;

[0081] Figure 5 A car attribute hierarchy diagram (part 1) provided in an embodiment of the present invention;

[0082] Figure 6 A car attribute hierarchy diagram (part 2) provided by an embodiment of the present invention;

[0083] Figure 7 A schematic diagram of measuring causal correlation between attribute viewpoints provided by an embodiment of the present invention;

[0084] Figure 8 A schematic diagram of the classification of attribute-viewpoint-reason pairs provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0085] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0086] The embodiments of the present application solve the technical problem of ignoring the inherent close causal relationship between attribute viewpoints by providing a data-driven automobile attribute viewpoint causal tracing method and system, and realize the visualization of the attribute hierarchy diagram and the in-depth tracing of the causal relationship of the automobile attribute viewpoint hierarchy.

[0087] The technical solution in the embodiments of the present application is to solve the above technical problems, and the overall idea is as follows:

[0088] like Figure 1 As shown, the data-driven automobile attribute view causal tracing method proposed by the present invention actually includes two aspects: a method for constructing an automobile attribute hierarchical network and a method for tracing the causal relationship of automobile attribute view.

[0089] The specific method for constructing a hierarchical network of automobile attributes is as follows: 1) extracting automobile attributes and opinions from a large amount of social comment data; 2) using point mutual information and decision tree models to accurately identify the hierarchical relationship between automobile attributes; 3) forming a hierarchical network of automobile attributes through hierarchical relationship connection and pruning.

[0090] The specific method for tracing the hierarchical relationship between automotive attribute viewpoints is as follows: 1) Based on the generated automotive attribute hierarchical network, hierarchical relationships between viewpoints are established to form preliminary viewpoint causal relationship tracing paths, and strong connection paths are screened and obtained. Based on each strong connection path, the strength of the causal relationship between viewpoints is measured. 2) Classification indicators are designed to classify pairs of viewpoints and causes, helping companies develop product optimization plans.

[0091] In order to better understand the above technical solution, the above technical solution will be described in detail below with reference to the accompanying drawings and specific implementation methods.

[0092] Example:

[0093] like Figure 2 As shown, an embodiment of the present invention provides a data-driven automobile attribute viewpoint causal tracing method, including:

[0094] S1. Obtaining car attributes and corresponding attribute opinions based on the total text of social comments on the car to be analyzed;

[0095] S2. Identify car attribute relationships, including:

[0096] S21. Obtain attribute pairs that may have potential hierarchical relationships based on the attributes of the vehicles.

[0097] S22. Extracting hierarchical relationship features of various automobile attributes of the attribute pairs based on the total text of the social comment;

[0098] S23, using the hierarchical relationship features of the vehicle attributes as input to the pre-trained decision tree classification model to obtain the hierarchical relationship between the two attributes in each attribute pair;

[0099] S3. Build a hierarchical network of car attributes, including:

[0100] constructing a car attribute hierarchical network based on quadruple groups, wherein the quadruple groups are composed of attribute pairs, corresponding hierarchical relationships, and relationship weights;

[0101] S4. Trace the hierarchical relationship of attribute viewpoints, including:

[0102] S41. Based on the generated automobile attribute hierarchical network, a hierarchical relationship between viewpoints is established to form a preliminary viewpoint causal relationship tracing path, and strong connection paths are screened to obtain them; based on each strong connection path, the causal relationship strength between viewpoints is measured.

[0103] The data-driven causal tracing method for automobile attribute opinion is proposed in an embodiment of the present invention. It actually includes two aspects: a method for constructing an automobile attribute hierarchical network and a method for tracing the causal relationships between automobile attribute opinion. This method only utilizes the total text of social comments about the automobile to be analyzed, eliminating the need to trace a massive text corpus. This effectively solves the problems of low research coverage and strong dependence on text corpora in existing solutions. Furthermore, based on traditional attribute opinion mining technology, it explores the correlation between opinions in the attribute hierarchy and realizes causal relationship tracing of opinions. This helps companies to more comprehensively, quickly, accurately, and efficiently discover the underlying reasons for user opinions, better understand the advantages and disadvantages of automobile attributes, understand user needs, and formulate more precise product improvement strategies.

[0104] Further, in order to identify the actual value of attribute view pairs, such as Figure 3 As shown, in an optional embodiment, the S4 further includes:

[0105] S42. Calculate the co-occurrence frequency of the hierarchical opinion pairs with strong connection relationship paths as the importance of the attribute opinion pairs, and combine the correlation strength of the attribute opinion pairs to divide all the attribute opinion pairs into misleading opinion reason pairs, key opinion reason pairs, useless opinion reason pairs and minor opinion reason pairs.

[0106] The following are the steps of the above technical solution:

[0107] In step S1 , the attributes of the car and the corresponding attribute opinions are obtained based on the total text of social comments on the car to be analyzed.

[0108] Use existing technologies such as crawlers to crawl the total text of the social comment from the network and define it as S, cut it into i paragraphs, then S can be expressed as S:{a1s2s3...a i}.

[0109] As automobile attribute extraction and attribute opinion mining are basic tasks in social review content analysis, there are currently very mature methods that have achieved good results in practice.

[0110] Specifically, this step draws on the existing association rule-based algorithm to extract automobile attributes and adopts the Word2Vec model to merge synonyms; it adopts an opinion extraction method that takes automobile attribute words as the center and sets a fixed window size to extract adjectives and adverbs, and constructs an opinion synonym dictionary to merge synonyms, that is, attribute opinions are in the form of adjectives plus adverbs or single adjectives.

[0111] In step S2, the car attribute relationship is identified:

[0112] First, it should be noted that a four-tuple pattern is defined in the embodiment of the present invention. The pattern includes car attribute pairs and relationships, namely <attribute one, attribute two, relationship, relationship weight>. The relationships include 'include', 'belong to', and 'unrelated'. The relationship weight is the normalized value of the co-occurrence frequency.

[0113] The S2 includes:

[0114] S21. According to the attributes of each vehicle, obtain attribute pairs that may have potential hierarchical relationships.

[0115] Combine any two attributes and calculate the co-occurrence frequency of the two attributes in each combination:

[0116]

[0117] Where P(AB) is the co-occurrence frequency of car attribute A and car attribute B in any attribute pair; N(S) is the total amount of social comment text S, and N(AB) is the number of comments in which car attribute A and car attribute B co-occur;

[0118] Compute the pointwise mutual information of the two attributes in each combination:

[0119]

[0120] Among them, PMI(A,B) is the point mutual information between car attribute A and car attribute B. When the co-occurrence frequency P(AB) is greater than the product of P(A) and P(B), car attribute A and car attribute B are not independent of each other, indicating that there is a close correlation between the two, and the point mutual information PMI(A,B) is greater than 0. The larger the value of the point mutual information, the closer the relationship between the two, and the more likely there is a hierarchical relationship. The point mutual information is used to filter the car attribute combinations that meet the conditions to form a set of potential hierarchical relationships.

[0121] When both the co-occurrence frequency and the point mutual information meet corresponding thresholds, the combination is determined to be an attribute pair that may have a potential hierarchical relationship.

[0122] S22. Extracting hierarchical relationship features of various automobile attributes of the attribute pairs based on the total text of the social comment.

[0123] For the discovered potential hierarchical relationships, further precise identification of the three specific types of "contains," "belongs to," and "unrelated" is required. This embodiment of the present invention employs eight statistical features, including frequency, location, and semantics, for identifying hierarchical relationships. For the sake of convenience in the following description, assume that car attributes A and B are a pair of car attribute combinations with a potential hierarchical relationship.

[0124] a. Frequency characteristics

[0125] Coarse-grained, higher-level car attributes are often easier for consumers to evaluate and describe than fine-grained, lower-level car attributes. Therefore, this embodiment of the present invention uses frequency information as a key basis for identifying the hierarchical relationships of car attributes and sets three features to measure the frequency difference between car attribute A and car attribute B.

[0126] b. Location features

[0127] The superior attribute of a hierarchical attribute pair is often evaluated by users first, and the subordinate attributes are evaluated later. Therefore, this paper sets a position information feature to identify the hierarchical relationship by comparing the positions of each attribute in the attribute pair in the co-occurrence sentence. For example, when attribute A appears more frequently at the front of the sentence than attribute B, attribute A is more likely to be identified as the superior attribute.

[0128] c. Semantic features

[0129] The semantic relationship between attributes is mainly manifested in fixed grammatical patterns, such as {of, for example, especially, including}, etc. When such expression patterns exist between attribute pairs, they are easily identified as hierarchical attribute pairs. Therefore, this paper sets semantic features to extract the semantic pattern information between attribute pairs, which is used to identify the hierarchical relationship between attribute pairs.

[0130] The specific description of each indicator is shown in Table 1:

[0131] Table 1 Feature description

[0132]

[0133] Among them, P(A) 单 、P(B) 单 Respectively represent the frequency of A and B appearing alone in S; P(A|B) represents the conditional probability of A appearing alone in S under the condition that B appears alone in S; P(B|A) represents the conditional probability of B appearing alone in S under the condition that A appears alone in S;

[0134] P(A) 前 、P(B) 前 Respectively represent A and B in s i The frequency of the preceding occurrence in the

[0135] & is a hierarchical semantic pattern {of, such as, especially, including}, P(A&B) and P(B&A) represent the frequency of the existence of hierarchical semantic patterns between A and B, and between B and A; $ is a parallel semantic template {and, or, and}, P(A$B) represents the frequency of the existence of parallel semantic patterns between A and B.

[0136] S23. Using the hierarchical relationship features of the automobile attributes as input to a pre-trained decision tree classification model, and obtaining the hierarchical relationship between the two attributes in each attribute pair.

[0137] The eight features described above are strongly correlated with each other. In machine learning algorithms, the decision tree classification model is more suitable for identifying hierarchical relationships between vehicle attributes than naive Bayesian and logistic regression. It does not require independence between features and can be trained with a small sample set. In the decision tree classification model, continuous features need to be discretized.

[0138] Specifically, the training process of the decision tree classification model includes:

[0139] S100. Manually construct a hierarchical dictionary of automobile attributes, which consists of attribute pairs and corresponding hierarchical relationship labels, where the labels include include, belong to, and irrelevant.

[0140] S200: Extract the various vehicle attribute hierarchical relationship features f1-f8 for each attribute pair in the dictionary. To maximize the preservation of the original data distribution, the seven continuous data features f1-f7 of the training dataset are discretized into three ordered intervals based on the equidistance discretization method, representing "much greater than," "close to," and "much less than," with corresponding discretized values of (0, 1, 2). Since feature f8 indicates whether a combination of vehicle attributes exhibits a clear parallel relationship, it is divided into two cases, with discretized values of 0 and 1, indicating the absence and presence of a parallel relationship, respectively. Combined with the manually labeled categories of the vehicle attribute hierarchical relationships, the training sample set shown in Table 2 is formed.

[0141] S300. Based on the discretized continuous features f1 to f7 and discrete feature f8, the corresponding hierarchical relationship labels are integrated and the CART algorithm commonly used in decision trees is used to train and obtain the decision tree classification model. The classification ability of the features is measured by the Gini coefficient. The smaller the Gini coefficient, the higher the category purity and the better the feature classification ability. In order to ensure that the model has good generalization and avoid overfitting, the post-pruning method is used to prune the formed decision tree.

[0142] Table 2 Example of decision tree classification model sample set

[0143]

[0144] By using the decision tree classification model obtained through the above training and inputting the eight features of the attribute pairs obtained in step S22, the hierarchical relationship between the attribute pairs can be obtained. Then, the attribute pairs, relationships and relationship weights are formed into a four-tuple pattern to prepare for the construction of the attribute hierarchy network.

[0145] In step S3, a car attribute hierarchical structure network is constructed, including constructing the car attribute hierarchical structure network according to quadruple groups, wherein the quadruple groups are composed of attribute pairs, corresponding hierarchical relationships, and relationship weights.

[0146] A quadruple structure such as <attribute A, attribute B, relationship, relationship weight> is used as input to construct the attribute hierarchy. The relationship weight is the normalized co-occurrence frequency of the two attributes in the corresponding attribute pair. For example, <appearance, front, contains, 0.87> indicates that "appearance" and "front" are in a "contains" hierarchical relationship with a correlation of 0.87. Connecting a large number of quadruple hierarchical relationships will result in a cluttered hierarchical network, so connection pruning rules are required.

[0147] The S3 specifically includes:

[0148] S31. Hierarchical connection:

[0149] 1)Hierarchical connection.

[0150] The car attribute without any upper-level attributes is considered a first-level attribute. According to the two quadruple combinations of <first-level attribute, attribute B, contains, weight> or <attribute B, first-level attribute, belongs to, weight>, attribute B is identified as a second-level attribute. This process is repeated to obtain deeper car attributes until the car attribute has no lower-level attributes. The growth of the hierarchical network is stopped. Relationship weights; S32, hierarchical relationship path pruning:

[0151] 2) Hierarchical relationship path pruning.

[0152] Based on the preliminary constructed hierarchical graph, overlapping and misaligned paths are pruned. Overlapping refers to the situation where the upper-level attributes simultaneously connect to the lower-level attributes and the lower-level attributes of the lower-level attributes, forming a cumbersome closed loop phenomenon; misalignment refers to the situation where the connection between the upper and lower-level attributes lacks common sense support.

[0153] That is, based on the preliminary hierarchical structure diagram, overlapping and misaligned paths are pruned to obtain the final vehicle attribute hierarchical network; wherein the hierarchical relationship path pruning rules include:

[0154] Rule 1: For overlapping pruning, the longest connection path is retained based on the longest path principle;

[0155] Rule 2: For misaligned pruning, locate the original social comment text paragraph and determine whether to retain the path based on whether it conforms to the contextual semantics.

[0156] In step S4, the attribute view causal relationship is traced, including:

[0157] S41. Measure the strength of opinion associations. Based on the generated hierarchical network of vehicle attributes, establish hierarchical relationships between opinions, form preliminary opinion causal relationship traceability paths, and screen and obtain strong connection paths. Based on each strong connection path, measure the strength of the causal association between opinions.

[0158] The same car attribute usually matches multiple user opinions, and user opinions at different attribute levels usually have a strong correlation, thus establishing a preliminary hierarchical relationship between opinions.

[0159] In order to screen out stronger hierarchical connections, the Ochiia coefficient is used. The larger the value, the stronger the connection between viewpoints. The calculation formula of the Ochiia coefficient is as follows:

[0160]

[0161] If the Ochiia coefficient of attribute view a and attribute view b is greater than a preset threshold, it is determined that there is a strong connection path between attribute view a and attribute view b.

[0162] Causal association strength measurement: In addition to direct connections, there may be multiple indirect connections between two user views. Assume that there are r strong connection paths between attribute view a and attribute view b, and a and b are the superior and subordinate views respectively. Each path z i By T i A level sequence of viewpoints, represented as in Indicates the tth i point of view, o1=a,o Ti =b, P(o j+1 |o j ) i is o on the i-th path j+1 With o j The conditional probability of obtaining the association strength T between attribute view a and attribute view b is:

[0163]

[0164] Figure 4 A schematic diagram of a strong connection path screened by the Ochiia coefficient is given. Figure 4 In the example, there are two connection paths between the first-level viewpoint o1 and the third-level viewpoint o3: path ① and path ②. Then the connection strength T between o1 and o3 is:

[0165] T(o1,o3)=p(o3|o1)+p(o2|o1)*p(o3|o2)

[0166] S42. Classification of opinion-reason pairs. Calculate the co-occurrence frequency of hierarchical opinion pairs with strong connection paths as the importance of the attribute opinion pair. Combined with the correlation strength of the attribute opinion pairs, all attribute opinion pairs are classified into misleading opinion-reason pairs, key opinion-reason pairs, useless opinion-reason pairs, and minor opinion-reason pairs.

[0167] After obtaining a large number of opinion-reason pairs, not all information has actual value for the company. Therefore, it is necessary to classify the opinion pairs and trace back the product opinion-reason pairs that are actually related to the user experience, so as to obtain greater commercial value while improving user satisfaction. In order to further identify the actual value of opinion-reason pairs, this patent classifies them based on two indicators: the importance of opinion pairs and the degree of correlation between opinion pairs. The importance between opinion pairs is measured by the frequency of co-occurrence of opinion pairs, which is a basic representational indicator. The degree of correlation is measured by the correlation strength T value, which is an internal indicator and plays a corrective role in the representational indicator. Specifically, it is divided into the following four types of opinion-reason pairs:

[0168] a. Misleading opinion and reason pairs. These refer to opinion pairs that are mentioned frequently but have weak correlations. Users often exaggerate the relationship between them, potentially due to misperceptions. These pairs are of little value to businesses and can easily cause disruption. Businesses can reduce decision-making uncertainty and risk by identifying and filtering these misleading reason pairs.

[0169] b. Key opinion and reason pairs. These pairs appear frequently and have strong correlations. They are the focus of user attention and key drivers of business success, requiring focused attention and improvement. By analyzing and optimizing these reason pairs, businesses can improve product quality, user satisfaction, and enhance market competitiveness.

[0170] c. Useless opinion pairs. These are opinion pairs that appear infrequently and are not particularly strong. These pairs may be incidental factors to the problem and offer little reference value to the company. Companies can screen and eliminate these opinion pairs to avoid unnecessary investment and decision-making, allowing them to focus on solving core issues.

[0171] d. Minor opinion and reason pairs. These pairs are rarely mentioned but have strong connections. They are easily overlooked by users but may contain potential innovation opportunities and points for improvement, providing important insights and directions for enterprise innovation. By analyzing and evaluating minor opinion and reason pairs, enterprises can uncover overlooked issues or potential needs, thereby creating higher user value, achieving better innovation results, and gaining a competitive advantage.

[0172] To help you better understand the advantages of the above solution, the following specific examples are provided:

[0173] This example uses the mining of hierarchical relationships among opinions about vehicle attributes for a certain brand of new energy vehicles as an example. The input is product social media data, and the output is a hierarchical network diagram of vehicle attributes and a causal relationship diagram for opinions. This case, inspired by a real-world problem in the product field, identifies causal relationships among opinions about vehicle attributes and traces the fine-grained reasons behind their expression.

[0174] This paper selects the RAV4 Rongfang, a car product with a relatively rich word-of-mouth review text on the Autohome website, as the analysis object. It obtains its online review data and cleans up invalid text and repeated character reviews. Then, the review data is segmented, part-of-speech marked, stop words removed, attributes extracted and opinions mined, and synonyms merged to obtain the product candidate attribute set and opinion set. Then, the hierarchical relationship between automobile attribute opinions is mined according to the previous steps.

[0175] The car attribute hierarchy diagram is constructed. Figures 5-6 As shown in the figure, the car product hierarchy is roughly divided into eight aspects: appearance, interior, space, handling, power, comfort, fuel consumption, and cost-effectiveness. This is consistent with the classification labels originally set by Autohome, thus verifying the effectiveness of this method. In addition, by statistically analyzing the weights between hierarchical relationship attributes, it was found that the secondary attribute that users pay most attention to under the appearance attribute of this car is the "body" attribute, followed by the "paint", "headlights", "rear", and "front" attributes. The "lines" and "wheels" of the "body", the "color" of the "paint", and the "headlights" and "taillights" of the "headlights" are the more fine-grained tertiary attributes that users pay attention to. Companies should pay special attention to the fine-grained attributes that users focus on, which will help them better grasp the market voice and obtain the direction of key optimization of car attributes.

[0176] Figure 7 The strong connection relationship path established is the conditional probability value between the two viewpoints. Combined with the calculation process of the association strength provided in the previous article, the connection strength of the viewpoint pair is measured, such as T 舒适度好-座椅柔软 =0.0690, T 舒适度好-空调好 =0.0345, it can be seen that users of this car are generally satisfied with the "comfort" level. The most likely reason for the "good" level of "comfort" is the soft "seats", followed by the good "air conditioning". Therefore, the company should continue to maintain the advantages of the original car attributes and optimize them to strengthen users' good impression of this product; T 舒适度很差-隔音一般 =0.0741, T 舒适度很差-噪音较大=0.0244. The main reason why the "comfort" of this car is very poor is that the "sound insulation effect" is "average", followed by the "noise" is "loud". Therefore, manufacturers can improve automobile sound insulation and noise-related processing technologies to address users' negative emotions about product comfort and thus improve user satisfaction.

[0177] Since the influence of opinion pairs is not necessarily the same, the data tends to be far from the center. In order to truly reflect the data distribution, the median is used as the classification reference line. Here, the correlation strength of opinion pairs is selected as 0.0300, and the co-occurrence frequency of opinion pairs is 0.0010. The results are as follows Figure 8 As shown: It can be found that useless opinion pairs and misleading opinion pairs are more numerous than the other two opinion pairs, which can easily interfere with the company's optimization plan. Screening and distinguishing them can effectively help companies avoid wasting energy and reduce unnecessary decision-making risks, which also reflects the value of classification.

[0178] An embodiment of the present invention provides a data-driven automobile attribute view causal tracing system, including:

[0179] An acquisition module, configured to acquire car attributes and corresponding attribute opinions based on the total text of social comments on the car to be analyzed;

[0180] The recognition module is used to identify the relationship between car attributes, including:

[0181] A judgment submodule, configured to obtain attribute pairs that may have potential hierarchical relationships based on the attributes of each vehicle;

[0182] An extraction submodule, configured to extract hierarchical relationship features of various automobile attributes of the attribute pairs based on the total text of the social comment;

[0183] A classification submodule, configured to use the hierarchical relationship features of the vehicle attributes as input to a pre-trained decision tree classification model to obtain the hierarchical relationship between the two attributes in each attribute pair;

[0184] Building blocks for constructing a hierarchical network of car attributes, including:

[0185] constructing a car attribute hierarchical network based on quadruple groups, wherein the quadruple groups are composed of attribute pairs, corresponding hierarchical relationships, and relationship weights;

[0186] The traceability module is used to trace the hierarchical relationship of attribute views, including:

[0187] Based on the generated hierarchical network of automobile attributes, a hierarchical relationship between viewpoints is established to form a preliminary causal relationship tracing path of viewpoints, and the strong connection paths are screened and obtained; according to each strong connection path, the causal relationship strength between viewpoints is measured.

[0188] An embodiment of the present invention provides a storage medium storing a computer program for data-driven causal tracing of automobile attribute viewpoints, wherein the computer program enables a computer to execute the automobile attribute viewpoint causal tracing method described above.

[0189] An embodiment of the present invention provides an electronic device, including:

[0190] One or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, the programs including a method for executing the automobile attribute viewpoint causal tracing method as described above.

[0191] It is understandable that the data-driven automobile attribute viewpoint causal tracing system, storage medium and electronic device provided in the embodiments of the present invention correspond to the data-driven automobile attribute viewpoint causal tracing method provided in the embodiments of the present invention. The explanations, examples and beneficial effects of the relevant contents can refer to the corresponding parts in the automobile attribute viewpoint causal tracing method, and will not be repeated here.

[0192] In summary, compared with the existing technology, the present invention has the following beneficial effects:

[0193] 1. The data-driven causal tracing method for automobile attribute opinion is comprised of two aspects: a method for constructing an automobile attribute hierarchical network and a method for tracing the hierarchical relationships of automobile attribute opinion. This method utilizes only the aggregated text of social commentary on the automobile being analyzed, eliminating the need to trace a massive text corpus. This effectively addresses the low research coverage and strong reliance on text corpora of existing approaches. Furthermore, building on traditional attribute opinion mining techniques, it explores the correlations between opinions within the attribute hierarchy and enables causal tracing of opinion relationships. This helps companies more comprehensively, quickly, accurately, and efficiently uncover the underlying causes of user opinions, better understand the strengths and weaknesses of automobile attributes, understand user needs, and develop more precise product improvement strategies.

[0194] 2. The embodiment of the present invention designs a flexible and free method for constructing a hierarchical network of automobile attributes based on social media data. The hierarchical structure contains a large number of hierarchical relationships between automobile attributes, which helps companies accurately and quickly capture the fine-grained automobile attributes that users are interested in.

[0195] 2. The embodiment of the present invention proposes a set of product viewpoint causal relationship tracing methods to help enterprises trace the deep-seated causes of viewpoints. At the same time, classification indicators are designed to measure the actual value of the viewpoint causes to the enterprise, thereby providing a basis for enterprises to formulate product optimization strategies.

[0196] 3. The method proposed in the embodiment of the present invention can be extended to a wide range of fields and is particularly applicable to product fields with rich social media data, such as electronic products.

[0197] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

[0198] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A data-driven causal tracing method for automobile attribute viewpoints, characterized by: include: S1. Obtaining car attributes and corresponding attribute opinions based on the total text of social comments on the car to be analyzed; S2. Identify the hierarchical relationship of vehicle attributes, including: S21. Obtain attribute pairs that may have potential hierarchical relationships based on the attributes of the vehicles. S22. Extracting hierarchical relationship features of various automobile attributes of the attribute pairs based on the total text of the social comment; S23, using the hierarchical relationship features of the vehicle attributes as input to the pre-trained decision tree classification model to obtain the hierarchical relationship between the two attributes in each attribute pair; S3. Build a hierarchical network of car attributes, including: constructing a car attribute hierarchical network based on quadruple groups, wherein the quadruple groups are composed of attribute pairs, corresponding hierarchical relationships, and relationship weights; S4. Tracing attribute view causal relationships, including: Based on the generated hierarchical network of automobile attributes, a hierarchical relationship between viewpoints is established to form a preliminary causal relationship tracing path of viewpoints, and strong connection paths are screened and obtained; based on each strong connection path, the causal relationship strength between viewpoints is measured; In said S21: Combine any two attributes and calculate the co-occurrence frequency of the two attributes in each combination: Where P(AB) is the co-occurrence frequency of car attribute A and car attribute B in any attribute pair; N(S) is the total amount of social comment text S, and N(AB) is the number of comments in which car attribute A and car attribute B co-occur; Compute the pointwise mutual information of the two attributes in each combination: Among them, PMI(A,B) is the point mutual information of car attribute A and car attribute B; When the co-occurrence frequency and the point mutual information both meet corresponding thresholds, the combination is determined to be an attribute pair that may have a potential hierarchical relationship; The hierarchical relationship characteristics of each vehicle attribute in S22 include: (1) Frequency characteristics: (1.1) The frequency difference between A and B appearing alone in the total text S of social comments is expressed as f1 = P(A) 单 -P(B) 单 ; (1.2) The frequency quotient of A and B appearing alone in S is expressed as (1.3) The difference in conditional probabilities between A and B is expressed as f3 = P(A|B) - P(B|A); Among them, P(A) 单 、P(B) 单 Respectively represent the frequency of A and B appearing alone in S; P(A|B) represents the conditional probability of A appearing alone in S under the condition that B appears alone in S; P(B|A) represents the conditional probability of B appearing alone in S under the condition that A appears alone in S; (2) Location features: (2.1) A and B are in the i-th social comment text paragraph s i The frequency difference that appears in front is expressed as f4=P(A) 前 -P(B) 前 ; (2.2) A and B are in s i The frequency quotient of the preceding value is expressed as Among them, P(A) 前 、P(B) 前 Respectively represent A and B in s i The frequency of the preceding occurrence in the (3) Semantic features: (3.1) The frequency difference between A&B and B&A is expressed as f6 = P(A&B) - P(B&A); (3.2) The frequency quotient of A&B and B&A is expressed as (3.3) The frequency value of A$B is expressed as f8=P(A$B); Among them, & is the hierarchical semantic pattern {of, such as, especially, including}, P(A&B) and P(B&A) represent the frequency of the existence of hierarchical semantic patterns between A and B, and between B and A; $ is the parallel semantic template {and, or, and}, P(A$B) represents the frequency of the existence of parallel semantic patterns between A and B.

2. The data-driven automobile attribute viewpoint causal tracing method according to claim 1, characterized in that: The training process of the decision tree classification model in S23 includes: Manually construct a hierarchical dictionary of automobile attributes, which consists of attribute pairs and corresponding hierarchical relationship labels, where the labels are: contain, belong to, and irrelevant; Extract the hierarchical relationship features f1 to f8 of each attribute pair in the dictionary, and discretize the seven continuous data of features f1 to f7 of the training data set based on the equidistance discretization method; According to the discretized continuous features f1 to f7 and discrete feature f8, corresponding hierarchical relationship labels are integrated and the decision tree classification model is obtained by training using the CART algorithm.

3. The data-driven automobile attribute viewpoint causal tracing method according to claim 1, characterized in that: The S3 includes: S31. Hierarchical connection: The car attribute without any upper-level attributes is considered a first-level attribute. Based on the two quadruple combinations of <first-level attribute, attribute B, contains, weight> or <attribute B, first-level attribute, belongs to, weight>, attribute B is identified as a second-level attribute. This process is repeated to obtain deeper car attributes until the car attribute has no lower-level attributes, stopping the growth of the hierarchical network. S32, hierarchical relationship path pruning: Based on the preliminary hierarchical structure diagram, overlapping and misaligned paths are pruned to obtain the final vehicle attribute hierarchical network. The hierarchical relationship path pruning rules include: Rule 1: For overlapping pruning, the longest connection path is retained based on the longest path principle; Rule 2: For misaligned pruning, locate the original social comment text paragraph and determine whether to retain the path based on whether it conforms to the contextual semantics.

4. The data-driven automobile attribute viewpoint causal tracing method according to claim 1 is characterized in that: The causal relationship tracing of the attribute viewpoints corresponding to the attributes of the car divided in S4 includes: The Ochiia coefficient is used to screen out the strong connection paths between hierarchical viewpoints; If the Ochiia coefficient of attribute view a and attribute view b is greater than the preset threshold, it is determined that there is a strong connection path between attribute view a and attribute view b; Causal association strength measurement: In addition to direct connections, there may be multiple indirect connections between two user views. Assume that there are r strong connection paths between attribute view a and attribute view b, and each path has z i By T i A level sequence of viewpoints, represented as in Indicates the tth i point of view, o1=a, P(o j+1 |o j ) i is o on the i-th path j+1 With o j The conditional probability of obtaining the causal relationship strength T between attribute view a and attribute view b is:

5. The data-driven automobile attribute viewpoint causal tracing method according to any one of claims 1 to 4, characterized in that: Said S4 further comprises: The co-occurrence frequency of hierarchical opinion pairs with strong connection paths is calculated as the importance of the attribute opinion pairs. Combined with the correlation strength of the attribute opinion pairs, all attribute opinion pairs are divided into misleading opinion reason pairs, key opinion reason pairs, useless opinion reason pairs and minor opinion reason pairs.

6. A data-driven automobile attribute view causal tracing system, characterized by: The method for executing the automobile attribute viewpoint causal tracing method according to claim 1 comprises: An acquisition module, configured to acquire car attributes and corresponding attribute opinions based on the total text of social comments on the car to be analyzed; The recognition module is used to identify the relationship between car attributes, including: A judgment submodule, configured to obtain attribute pairs that may have potential hierarchical relationships based on the attributes of each vehicle; An extraction submodule, configured to extract hierarchical relationship features of various automobile attributes of the attribute pairs based on the total text of the social comment; A classification submodule, configured to use the hierarchical relationship features of the vehicle attributes as input to a pre-trained decision tree classification model to obtain the hierarchical relationship between the two attributes in each attribute pair; Building blocks for constructing a hierarchical network of car attributes, including: constructing a car attribute hierarchical network based on quadruple groups, wherein the quadruple groups are composed of attribute pairs, corresponding hierarchical relationships, and relationship weights; The tracing module is used to trace the causal relationship of attribute views, including: Based on the generated hierarchical network of automobile attributes, a hierarchical relationship between viewpoints is established to form a preliminary causal relationship tracing path of viewpoints, and the strong connection paths are screened and obtained; according to each strong connection path, the causal relationship strength between viewpoints is measured.

7. A storage medium, characterized in that: The computer program for data-driven causal tracing of automobile attribute viewpoints is stored therein, wherein the computer program enables a computer to execute the automobile attribute viewpoint causal tracing method according to any one of claims 1 to 5.

8. An electronic device, characterized in that: include: one or more processors; Memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including a method for executing the automobile attribute viewpoint causal tracing method as described in any one of claims 1 to 5.

Images