Multi-level value system construction method and system

Abstract

The invention relates to the technical field of natural language generation, and discloses a multi-level value system construction method and system, and the method comprises the steps: constructing a first layer of a value system based on a manually set value anchoring word; mining a value unit from a corpus of the real world: generating candidate phrases based on the part-of-speech tagging word sequence, training a phrase quality evaluator by adopting a remote supervision integrated training technology so as to screen the candidate phrases, and taking the screened phrases as the value unit; under the guidance of value anchor words, clustering the value units of similar semantics to form a third layer of the value system; and taking the generalized phrases summarized based on the clustering result as a second layer of the value system, so as to construct a multi-layer value system. The method provided by the invention is high in automation degree, multi-technology path fusion and real world corpus driving are emphasized, and the constructed value system is wide in coverage and rich in connotation.

Description

Technical Field

[0001] This invention relates to the field of natural language generation technology, specifically to a method and system for constructing a multi-level value system. Background Technology

[0002] A rich and profound value system can play a key role in many application scenarios such as value dissemination, education and training, and large-scale model alignment.

[0003] Early research on value system construction methods primarily relied on expert manual induction. While this method ensured the accuracy of values, it was inefficient and dependent on regular manual maintenance, making it difficult to respond in real-time to changing social trends. In recent years, some studies have focused on using topic models and word embedding techniques for value system construction. The Latent Dirichlet Allocation (LDA) model has been widely applied to document topic mining, but its generated parallel topic structures struggle to capture hierarchical semantic relationships (such as the hierarchy between abstract and concrete concepts). To overcome this limitation, Blei et al. proposed the Hierarchical Topic Model (hLDA), which constructs a tree-like topic structure through nesting, achieving hierarchical topic mining. However, hLDA still relies on pure data-driven approaches and lacks domain knowledge guidance, leading to semantic biases in knowledge systems constructed in specific domains such as values. Regarding lexical semantic representation, static word embedding models such as Word2Vec and GloVe learn lexical vectors through contextual co-occurrence relationships, but fail to adequately model word order and direction information. Subsequent research, such as variants of Skip-gram, attempts to introduce position vectors, but their characterization of complex semantic relationships remains insufficient. Furthermore, while current word embedding-based clustering methods (such as K-means) can generate category divisions, their intra-class semantic consistency and inter-class distinguishability are limited, making it difficult to support the construction of fine-grained knowledge systems. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method and system for constructing a multi-layered value system based on fine-grained value unit mining, emphasizing the integration of multiple technical paths and real-world corpus-driven approaches. Specifically, starting with a small number of manually set value anchor words, this invention designs a fine-grained value unit mining module and a value unit clustering learning module, used to construct the content of the third layer (connotation knowledge layer) and the second layer (domain subdivision layer) of the value system, respectively. In the fine-grained value unit mining module, this invention employs far-supervised ensemble training technology, introducing multiple weak supervision signals to construct various heteroproton models and fusing their prediction results to mitigate the interference of single-label noise, thereby obtaining a robust phrase quality evaluator that can mine fine-grained value units from massive amounts of relevant value corpus as candidate units for the third layer of the value system. In the value unit clustering learning module, this invention proposes an adaptive semantic clustering algorithm, using the semantic features of fine-grained value units for clustering learning, and summarizing and generalizing the clustering results through human-computer collaboration to form the second layer of the value system and its corresponding third layer content, thus constructing a rich and clearly defined multi-layered value system.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: Firstly, the present invention provides a method for constructing a multi-level value system, comprising: The first layer of the value system is constructed based on manually set value anchors; Mining value units from real-world corpora, wherein the mining of value units includes: generating candidate phrases based on part-of-speech tagging word sequences, and training a phrase quality evaluator using far-supervised ensemble training technology to filter from the candidate phrases, and using the filtered phrases as value units; Guided by value anchor words, value units with similar semantics are clustered to form the third layer of the value system: classification is performed based on the distance of the value unit representation to the cluster center, and then each centroid is calculated by the weighted sum of all value unit representations. The value units are then divided into different classes by clustering loss. The general phrases summarized based on the clustering results are used as the second layer of the value system, thus completing the construction of a multi-level value system.

[0006] In one embodiment, generating candidate phrases based on part-of-speech tagging word sequences specifically includes: The real-world corpus is processed into a sequence of words with part-of-speech tags. Phrase boundaries are delineated by calculating part-of-speech quality scores, and candidate phrases are generated.

[0007] In one embodiment, the process of processing a real-world corpus into a sequence of words with part-of-speech tags, dividing phrase boundaries by calculating part-of-speech quality scores, and generating candidate phrases specifically includes: Suppose the corpus is processed to a length of Part-of-speech tagging sequence ; Represents the i-th word and the part-of-speech tag corresponding to the i-th word The pair formed; The sequence of part-of-speech tags consisting of all part-of-speech tags , Any subsequence of tags consisting of consecutive part-of-speech tags , recorded as The part-of-speech quality score of the tag subsequence Defined as the corresponding word sequence It is the conditional probability of a complete semantic unit, in the following form: ; when When it exceeds the set threshold, The corresponding word sequence is used as a candidate phrase; For conditional probability, Used to mark the boundaries of word sequences.

[0008] In one embodiment, the part-of-speech quality score is calculated as follows: ; in, The part-of-speech tag representing the word at the beginning of the r-th phrase. The part-of-speech tag representing the word preceding the word at the beginning of the r-th phrase. express In the phrases of the corpus and Adjacent and in The previous probability, express In the phrases of the corpus and Adjacent and in The previous probability, These represent the part-of-speech tags corresponding to the (j-1)th word and the jth word, respectively.

[0009] In one embodiment, Defined as: ; ; Represents any pair of adjacent words in a corpus; function Used to generate part-of-speech tags for input words.

[0010] In one embodiment, the step of training a phrase quality evaluator using far-supervised ensemble training techniques to filter candidate phrases and using the filtered phrases as value units specifically includes: Keywords and phrases that appear more frequently than a set value in the public knowledge base are used as weak supervision signals to construct a positive phrase sample pool; the corpus is processed by the n-gram method to generate a phrase set, and phrases in the phrase set that do not appear in the public knowledge base are assigned to the negative phrase sample pool. A perturbation training set is constructed by randomly drawing phrase samples from positive and negative phrase sample pools in a stratified manner, and an unpruned decision tree is used as the initial heteroprotic model. Multiple independently trained heteroprotic models are then integrated to obtain the phrase quality evaluator. ; Input a candidate phrase into the phrase quality evaluator ,Will The proportion of heteroprotic models that predict the candidate phrase as a positive phrase sample out of all heteroprotic models is used as the quality score of the candidate phrase. : ; for The set of heteroprotic models that predict the same candidate phrase as a positive phrase sample. for The set of heteroprotic models that predict the same candidate phrase as a negative phrase sample; All satisfied Candidate phrases are used as value units, where, This is the preset phrase quality threshold.

[0011] In one embodiment, a phrase quality evaluator is provided. If a phrase quality estimator contains Y heteroproton models, then the prediction error is... Represented as: ; in, This represents the empirical error of the heteroproton model. Represents the number of combinations. As an intermediate variable; The optimal number of heteroproton models in the phrase quality evaluator is calculated using the prediction error.

[0012] In one embodiment, the clustering of value units with similar semantics to form the third layer of the value system involves: classifying value units based on their distance from the cluster center, calculating each centroid using a weighted sum of all value unit representations, and then dividing the value units into different classes using a clustering loss. Specifically, this includes: For a batch Each value unit Encoding as a value unit representation A set representing value units The set represented by the cluster center As input to the clustering module, the distance matrix between the value unit representation and the center representation is obtained. ; According to the distance matrix To classify the representation of value units: ; in, The representation of the k-th cluster center is initialized as a random vector; Represents the value unit representation The degree to which it does not belong to every category This refers to the argmin function. This represents the value unit representation. The index of the class to which it was assigned; Clustering loss for: ; For the first Representation of cluster centers; Represents Euclidean distance; These are the model parameters that the clustering module can learn; The weights of all value unit representations are obtained from the distance matrix, and the cluster center representations are updated based on the weighted sum of the value unit representations. ; It represents the number of domain sub-categories under each value anchor keyword; After the clustering module has been trained, it is calculated... This yields the clustering results for each value unit.

[0013] In one embodiment, the distance matrix between the value unit representation and the center representation The calculation method is as follows: ; ; elements in It is the first Each cluster center represents With the Each value unit represents The normalized distance between them.

[0014] In a second aspect, the present invention provides a computer system including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method of any embodiment of the first aspect.

[0015] Compared with the prior art, the beneficial technical effects of the present invention are: In the fine-grained value unit mining module, this invention employs far-supervised ensemble training technology. By introducing multiple weak supervision signals, it constructs various heterogeneous models and fuses their prediction results to mitigate the interference of single-label noise, thereby obtaining a robust phrase quality evaluator. In the value unit clustering learning module, this invention proposes an adaptive semantic clustering algorithm. It uses the semantic features of fine-grained value units for clustering learning, avoiding the dependence of traditional co-occurrence clustering methods on corpus quality. Furthermore, through human-computer collaboration, the clustering results are summarized and generalized to form the second layer of the value system and its corresponding third layer of fine-grained value unit content. Unlike previous value system construction techniques that heavily rely on human experts, the method proposed in this invention has a high degree of automation and emphasizes the integration of multiple technical paths and real-world corpus-driven approaches. The constructed value system is broad in coverage and rich in content, suitable for various application scenarios such as value dissemination, education and training, and large-scale model alignment. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method in an embodiment of the present invention; Figure 2 This is a schematic diagram of the overall architecture in an embodiment of the present invention. Detailed Implementation

[0017] A preferred embodiment of the present invention will now be described in detail with reference to the accompanying drawings.

[0018] like Figure 1 As shown, the present invention provides a method for constructing a multi-level value system, comprising the following steps: S1, the first layer of the value system built based on manually set value anchor words; S2, mining value units from a real-world corpus, wherein mining value units includes: generating candidate phrases based on part-of-speech tagging word sequences, and training a phrase quality evaluator using far-supervised ensemble training technology to filter from the candidate phrases, and using the filtered phrases as value units; S3, guided by value anchor words, clusters value units with similar semantics to form the third layer of the value system: classify according to the distance of the value unit representation to the cluster center, and then calculate each centroid by the weighted sum of all value unit representations, and then divide the value units into different classes by clustering loss; the general phrases summarized based on the clustering results are used as the second layer of the value system, thereby constructing a multi-level value system.

[0019] This invention uses highly condensed value anchor words set by humans as the first layer of the value system, then mines the corresponding value connotations from real-world corpora as the value units of the third layer of the value system, and clusters these value connotations to learn more condensed general phrases as the second layer of the value system, thereby constructing a multi-layered value system with rich connotations and distinct levels.

[0020] The overall architecture of this invention is as follows Figure 2 As shown, it mainly includes a fine-grained value unit mining module and a value unit clustering learning module. The fine-grained value unit mining module uses far-supervised ensemble training technology to efficiently and robustly extract high-quality phrases from the text as value units. The value unit clustering learning module, guided by value anchor words, further clusters semantically related value units as the third layer of the value system, and summarizes more concise general phrases for each category in the third layer as the second layer of the value system.

[0021] 1. Fine-grained value unit mining module.

[0022] The goal of this module is to automatically extract fine-grained value units from massive real-world corpora. Fine-grained value units are defined as phrases hidden within the corpus that may embody value connotations, forming an important component of the third layer of the value system. To address this task, this module first generates candidate phrases based on part-of-speech tagging. Then, it trains a phrase quality evaluator using far-supervised ensemble learning, a technique that effectively reduces the training cost and error of the phrase quality evaluator. Finally, the trained phrase quality evaluator efficiently and robustly selects high-quality phrases from the candidate phrases as fine-grained value units. Specifically, the fine-grained value unit mining module mainly includes the following: (1) Candidate phrase generation based on part-of-speech tagging: Part-of-speech tagging provides shallow, language-specific knowledge, which helps improve the accuracy of candidate phrase generation, especially at the boundaries of syntactic components in the language. Part-of-speech-based candidate phrase segmentation can result in more complete value units.

[0023] Suppose the corpus is processed to a length of Part-of-speech tagging sequence ,in Refers to a word and its part-of-speech tags The pairs are formed. Part-of-speech-guided phrase segmentation is achieved through phrase boundary index sequences. Divide this sequence into Each segment satisfies . No. Each segment refers to .

[0024] Given the total length is The part-of-speech tag sequence of the corpus is , containing tag subsequences (For clarity, it is written as) The part-of-speech quality score of this tagged subsequence is defined as the conditional probability that its corresponding word sequence is a complete semantic unit. Formally: .

[0025] Part-of-speech quality score The aim is to reward phrases that correctly identify the part-of-speech sequence. (Selected) The specific form is: ; in, It is a part-of-speech tag Phrases in the corpus with part-of-speech tags The probability of being adjacent and preceding. In this formula, the first term represents the probability of being at index 1. and The probability of phrase boundaries between words, and the subsequent product represents... The probability that all part-of-speech tags are in the same phrase. This part-of-speech quality score can naturally offset the bias towards longer segments because for all... Regardless of how the corpus is segmented, and There happens to be one that is always multiplied.

[0026] Mathematically speaking, Defined as: ; Because it depends on how the document is segmented into phrases, It is uniformly initialized and will be learned during phrase segmentation.

[0027] (2) Phrase quality evaluator based on far-supervised ensemble training: In constructing the training set, this invention employs a far-supervised training method to build a large amount of training data at extremely low cost. Public knowledge bases (such as Wikipedia and policy and regulatory knowledge bases) are widely considered sources of high-quality information due to the authority and accuracy of their content. Therefore, this invention considers frequently occurring keywords and phrases in the knowledge base as high-quality phrases and uses them as weak supervision signals to construct a positive phrase sample pool. For the negative phrase sample pool, this invention generates a phrase set from the corpus using the n-gram method, where phrases not appearing in the knowledge base are included in the negative phrase sample pool. While this method avoids manual annotation, it introduces noise issues, meaning that some high-quality phrases may be incorrectly included in the negative phrase sample pool because they were not included in the knowledge base.

[0028] To reduce the impact of noise, this invention employs an ensemble learning method to train the phrase quality estimator. First, the method randomly extracts phrase samples stratified from both positive and negative phrase sample pools to construct a perturbed training set, and uses an unpruned decision tree as the initial heterogeneous model to ensure a complete fit to the phrase samples. Then, the results of multiple independently trained heterogeneous models are ensembled to obtain a robust phrase quality estimator. This invention uses the proportion of positive predictions across all heteroproton models as the quality score for candidate phrases. : ; If phrase quality evaluator If a model (decision tree) contains Y heteroatoms, then its prediction error is... It can be represented as: ; in, This represents the empirical error of the heteroproton model. This represents the number of combinations. Using the above formula, the present invention can calculate the optimal number of heteroproton models to construct.

[0029] Using the method described above, this module first uses the part-of-speech quality score of the tag subsequence. A large number of candidate phrases with strong integrity were extracted from the corpus, and then a robust phrase quality evaluator was used. High-quality phrases are selected and used as fine-grained value units in the third layer of the value system.

[0030] 2. Value Unit Clustering Learning Module.

[0031] The goal of this module is to cluster similar value units in the semantic space and summarize them into more concise general phrases as the second layer of the value system. This invention uses a two-layer fully connected neural network as the clustering module and designs a clustering algorithm to cluster value units with similar semantics. The clustering algorithm classifies value units based on the distance from their representations to the cluster centers (centroids). Then, it calculates each centroid by weighting the sum of all value unit representations. Finally, a clustering loss is designed to optimize the clustering network to divide value units into different classes.

[0032] In detail, for a batch Each value unit is analyzed using a pre-trained language model (such as the BERT model). Encoding yields a d-dimensional value unit representation. The clustering module represents the set of value units. The set represented by the cluster center As input, here It is the number of domain sub-segments under each value anchor keyword. It is a randomly initialized vector. Then, this invention can obtain the distance matrix between the value unit representation and the cluster center representation. : ; ; here It is the first Each cluster center represents With the Each value unit represents The normalized distance between them represents the value unit representation. Not belonging to the first The degree of class; These are learnable clustering module parameters. It is Euclidean distance.

[0033] Next, the present invention can be based on the distance matrix To classify the representation of value units: .

[0034] in Represents the value unit representation The degree to which it does not belong to every category The function represents the value unit based on the minimum distance. It is then assigned to the appropriate class. Therefore, This indicates that the value unit represents the class index to which it is divided.

[0035] To train clustering modules to obtain the optimal clustering paradigm, this invention proposes a clustering loss to minimize the distance between value unit representations and their corresponding cluster centers. The goal of the clustering loss is to find an optimal set of clustering module parameters. This allows for the calculation of optimal cluster centers, such that after a value unit representation passes through this clustering module, its distance to its corresponding cluster center is smaller, while its distance to other cluster centers is larger. Clustering loss. As shown below: .

[0036] Subsequently, this invention obtains the weights of all value unit representations based on the distance matrix, and updates the cluster center representations based on the weighted sum of the value unit representations. The update formula is as follows: .

[0037] After the clustering module is fully trained, it is calculated This yields the clustering results for each value unit, which constitutes the third layer of the value system. Furthermore, this invention utilizes an advanced large language model to summarize the general phrases for each category in the third layer, serving as the content for the second layer, the domain classification layer, thereby constructing a rich and multi-layered value system.

[0038] This invention can be applied to large language model training platforms, automatically constructing a deep value system that aligns with contemporary social values, laying a solid foundation for value alignment within large language models. At the implementation level, it can be embedded as software into intelligent systems, efficiently integrating into various application scenarios and comprehensively enhancing the value orientation and decision-making quality of intelligent systems.

[0039] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0040] It should be understood that although the steps in the flowcharts of the accompanying drawings are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the flowcharts of the accompanying drawings may include multiple steps or stages, which are not necessarily completed at the same time, but may be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but may be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0041] In one embodiment, the present invention also provides a computer system, which may be a server. The computer system includes a processor, memory, and a network interface connected via a system bus. The processor of the computer system provides computing and control capabilities. The memory of the computer system includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the computer system stores data used in the methods described above. The network interface of the computer system is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements the methods described above.

[0042] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0043] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.

[0044] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for constructing a multi-level value system, characterized in that, include: The first layer of the value system is constructed based on manually set value anchors; Mining value units from real-world corpora, wherein the mining of value units includes: generating candidate phrases based on part-of-speech tagging word sequences, and training a phrase quality evaluator using far-supervised ensemble training technology to filter from the candidate phrases, and using the filtered phrases as value units; Guided by value anchor words, value units with similar semantics are clustered to form the third layer of the value system: classification is performed based on the distance of the value unit representation to the cluster center, and then each centroid is calculated by the weighted sum of all value unit representations. The value units are then divided into different classes by clustering loss. The general phrases summarized based on the clustering results are used as the second layer of the value system, thus completing the construction of a multi-level value system.

2. The method for constructing a multi-level value system according to claim 1, characterized in that, The generation of candidate phrases based on part-of-speech tagging word sequences specifically includes: The real-world corpus is processed into a sequence of words with part-of-speech tags. Phrase boundaries are delineated by calculating part-of-speech quality scores, and candidate phrases are generated.

3. The method for constructing a multi-level value system according to claim 2, characterized in that, The process of processing a real-world corpus into a part-of-speech tagging sequence, dividing phrase boundaries by calculating part-of-speech quality scores, and generating candidate phrases specifically includes: Suppose the corpus is processed to a length of Part-of-speech tagging sequence ; Represents the i-th word and the part-of-speech tag corresponding to the i-th word The pair formed; The sequence of part-of-speech tags consisting of all part-of-speech tags , Any subsequence of tags consisting of consecutive part-of-speech tags , recorded as The part-of-speech quality score of the tag subsequence Defined as the corresponding word sequence It is the conditional probability of a complete semantic unit, in the following form: ； when When it exceeds the set threshold, The corresponding word sequence is used as a candidate phrase; For conditional probability, Used to mark the boundaries of word sequences.

4. The method for constructing a multi-level value system according to claim 3, characterized in that, The specific calculation method for the part-of-speech quality score is as follows: ； in, The part-of-speech tag representing the word at the beginning of the r-th phrase. The part-of-speech tag representing the word preceding the word at the beginning of the r-th phrase. express In the phrases of the corpus and Adjacent and in The previous probability, express In the phrases of the corpus and Adjacent and in The previous probability, These represent the part-of-speech tags corresponding to the (j-1)th word and the jth word, respectively.

5. The method for constructing a multi-level value system according to claim 4, characterized in that, Defined as: ； ； Represents any pair of adjacent words in a corpus; function Used to generate part-of-speech tags for input words.

6. The method for constructing a multi-level value system according to claim 1, characterized in that, The phrase quality evaluator is trained using remotely supervised ensemble training technology to filter candidate phrases, and the selected phrases are used as value units. Specifically, this includes: Keywords and phrases that appear more frequently than a set value in the public knowledge base are used as weak supervision signals to construct a positive phrase sample pool; the corpus is processed by the n-gram method to generate a phrase set, and phrases in the phrase set that do not appear in the public knowledge base are assigned to the negative phrase sample pool. A perturbation training set is constructed by randomly drawing phrase samples from positive and negative phrase sample pools in a stratified manner, and an unpruned decision tree is used as the initial heteroprotic model. Multiple independently trained heteroprotic models are then integrated to obtain the phrase quality evaluator. ; Input a candidate phrase into the phrase quality evaluator ,Will The proportion of heteroprotic models that predict the candidate phrase as a positive phrase sample out of all heteroprotic models is used as the quality score of the candidate phrase. : ； for The set of heteroprotic models that predict the same candidate phrase as a positive phrase sample. for The set of heteroprotic models that predict the same candidate phrase as a negative phrase sample; All satisfied Candidate phrases are used as value units, where, This is the preset phrase quality threshold.

7. The method for constructing a multi-level value system according to claim 6, characterized in that, Set up a phrase quality evaluator If a phrase quality estimator contains Y heteroproton models, then the prediction error is... Represented as: ； in, This represents the empirical error of the heteroproton model. Represents the number of combinations. As an intermediate variable; The optimal number of heteroproton models in the phrase quality evaluator is calculated using the prediction error.

8. The method for constructing a multi-level value system according to claim 1, characterized in that, The clustering of value units with similar semantics forms the third layer of the value system: Classification is performed based on the distance from the value unit representation to the cluster center. Then, each centroid is calculated using the weighted sum of all value unit representations, and finally, the value units are divided into different classes using clustering loss. Specifically, this includes: For a batch Each value unit Encoding as a value unit representation A set representing value units The set represented by the cluster center As input to the clustering module, the distance matrix between the value unit representation and the center representation is obtained. ; According to the distance matrix To classify the representation of value units: ； in, The representation of the k-th cluster center is initialized as a random vector; Represents the value unit representation The degree to which it does not belong to every category This refers to the argmin function. This represents the value unit representation. The index of the class to which it was assigned; Clustering loss for: ; For the first Representation of cluster centers; Represents Euclidean distance; These are the model parameters that the clustering module can learn; The weights of all value unit representations are obtained from the distance matrix, and the cluster center representations are updated based on the weighted sum of the value unit representations. ； It represents the number of domain sub-categories under each value anchor keyword; After the clustering module has been trained, it is calculated... This yields the clustering results for each value unit.

9. The method for constructing a multi-level value system according to claim 8, characterized in that, The distance matrix between the value unit representation and the center representation The calculation method is as follows: ； ； elements in It is the first Each cluster center represents With the Each value unit represents The normalized distance between them.

10. A computer system comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 9.

Images