A word vector representation method and system for annual report text based on semantic primitives

Abstract

The present invention discloses a word vector representation method and system for annual report text based on semantic primitives. The method comprises the following steps: obtaining accounting field dictionary text and financial report field stop word text, constructing a corpus based on the annual report text and performing word segmentation; performing text topic training, obtaining the vocabulary with the largest weight value in each topic classification as the subject word; performing word vector training, obtaining the vector representation of each vocabulary; calculating the semantic similarity between the subject words; selecting the subject word with the lowest semantic similarity as the semantic primitive; calculating the semantic similarity between each vocabulary and the semantic primitive as the weight in the corresponding semantic primitive dimension; performing normalization processing, dividing the similarity of each semantic primitive by the sum of the vocabulary similarities in all semantic primitives, and obtaining the feature vector representation of the vocabulary represented by each semantic primitive. The present invention uses a limited number of semantic primitives to represent word vectors, which can reduce the representation dimension of the word vector and improve the density of the vector space.

Description

Technical Field

[0001] The present invention relates to the technical field of annual report text data processing, and in particular to a semantic primitive-based word vector representation method and system for annual report text. Background Art

[0002] In view of the relatively free and rich content of annual report text information, it is urgent to introduce and apply computer text analysis programs to extract massive annual report text information and solve the problem of financial report terminology representation in the process of financial report text data processing;

[0003] Among existing solutions, there is a method that uses semantic primitives to represent terms. This method interprets financial reporting terms from the perspective of dictionary interpretability and provides a limited semantic primitive extraction method and results as a reference. However, this method is based on a dictionary and establishes a directed graph through the interpretation and interpretation relationship between words. It uses the uninterpretability of semantic primitives to extract a set of words that can interpret other words. New words that are not in the dictionary cannot be represented. It represents financial terms from the perspective of interpretation primitives. As a low-dimensional method for interpreting terms, it provides a semi-automatic semantic primitive extraction method. This method is limited by the choice of corpus. In addition, this technical solution does not provide a method for quantitatively representing words and does not further discuss how to quantitatively represent financial reporting terms using the extracted semantic primitives.

[0004] Some use the text vector space representation model to represent subject terms. This representation method uses the literature subject corresponding to the vocabulary as the dimension of the vector space and uses the improved term-lexical association weight matrix as the vector representation of the vocabulary in each dimension. However, the computational complexity of processing the relationship between vocabulary is too large, and vector data sparsity is inevitable.

[0005] Some approaches use a weighted fusion scheme that combines semantic vectors with word training representations to represent word vectors. This scheme incorporates semantic vectors as semantic features of terms based on the word vector representation trained on text. However, word vector representation is labor-intensive and lacks discussion of lexical expressions that have not been annotated with semantic vectors.

[0006] Some have proposed word vector representation methods for Tibetan. This method is a joint training method that adds the semantic primitives corresponding to words as context information to generate training materials during the training phase. Tibetan representations are then obtained through further word vector training. However, this method places high demands on the quality of the labeled data.

[0007] Some have proposed a method for representing word vectors by weighted averaging of sememe vectors. This method first trains the vector representation of sememes and then performs a weighted average on sememe-annotated words. However, this method only considers the semantic vectors of weighted sememes as word vectors. This is done on a set of sememe-annotated data, without discussing how to handle words that are not annotated with sememes. Furthermore, the vector representation of words is not interpretable.

[0008] Some have proposed a method of training semantic primitive vectors and then weighting them, using the contextual semantic similarity between words and the contribution of word structure as the weight of the primitive vector. However, this method is limited to the corpus. When a corpus from a non-dictionary contains words that cannot be matched with semantic primitives, it is impossible to represent the words with semantic primitives.

[0009] In summary, existing solutions have the defects of requiring high-quality labeled data, sparse word vectors, and uninterpretable vector representations of vocabulary. Summary of the Invention

[0010] In order to overcome the defects and shortcomings of the existing technology, the present invention provides a word vector representation method and system for annual report text based on semantic primitives. The present invention adopts semantic primitives as vector dimensions, takes the similarity with semantic primitives as the numerical value in the dimension, uses semantic primitives as explicit representations of the semantic meaning of words, and maps all words in the same set of semantic primitives to improve the interpretability of the word vector representation method; after obtaining the subject words and calculating the similarity between the subject words, based on the principle of semantic primitive independence, a limited number of words with the smallest semantic similarity are selected as semantic primitives. The use of a limited number of semantic primitives to represent word vectors can reduce the representation dimension of word vectors and improve the density of the vector space.

[0011] In order to achieve the above object, the present invention adopts the following technical solutions:

[0012] The present invention provides a method for representing annual report text word vectors based on semantic primitives, comprising the following steps:

[0013] Obtain accounting dictionary text and financial reporting stop word text, build a corpus based on the annual report text, and perform word segmentation on the corpus;

[0014] Perform text topic training based on the LDA training model and obtain the words with the largest weight in each topic classification as the topic words;

[0015] Perform word vector training based on the Word2Vec word vector model to obtain the vector representation of each word;

[0016] Calculate the semantic similarity between topic words based on the cosine similarity algorithm;

[0017] Select the topic word with the lowest semantic similarity as the semantic primitive;

[0018] Calculate the semantic similarity between each word and the semantic primitive as the weight in the corresponding semantic primitive dimension;

[0019] After calculating the similarity between a word and all semantic primitives, normalize it by dividing the similarity of each semantic primitive by the sum of the similarities of the word in all semantic primitives, so that the sum of the weights of the word on each semantic primitive is 1.

[0020] After calculating the weight of each word allocated to the semantic primitive, the feature vector representation of the word represented by each semantic primitive is obtained.

[0021] As an optimal technical solution, the full mode is used to segment the corpus.

[0022] As a preferred technical solution, text topic training is performed based on the LDA training model, specifically including:

[0023] Randomly select a topic from the topic distribution, randomly select a word based on the word distribution of the topic, calculate the correlation between the internal words of a topic word, calculate the text consistency value of different numbers of topics, select the value with the highest text consistency as the number of topics, take the average of the consistency indicators of all topics, label the text after training the text topics and print a set number of topics with higher weights under each topic.

[0024] As a preferred technical solution, the semantic similarity between subject words is calculated based on the cosine similarity algorithm, which is specifically expressed as:

[0025]

[0026] Among them, Them i ,Them j Indicates the similarity between two keywords, wv ik and wv jk Represents the weight of the topic word on the kth dimension after word vector training.

[0027] As a preferred technical solution, the similarity of each semantic primitive is divided by the sum of the similarities of the vocabulary in all semantic primitives, so that the sum of the weights of the vocabulary on each semantic primitive is 1. The final weight of the vocabulary on the semantic primitive is calculated as:

[0028]

[0029] Among them, P nk represents the final weight of word n ​​on semantic primitive k, X nk Represents the semantic similarity between word n ​​and semantic primitive k.

[0030] The present invention also provides a semantic primitive-based annual report text word vector representation system, comprising: a text acquisition module, a corpus construction module, a word segmentation module, a text topic training module, a word vector training module, a topic word semantic similarity calculation module, a semantic primitive construction module, a semantic primitive dimension weight calculation module, a normalization module, a weight distribution module, and a feature vector representation module;

[0031] The text acquisition module is used to acquire accounting field dictionary text and financial report field stop word text;

[0032] The corpus construction module is used to construct a corpus based on the annual report text;

[0033] The word segmentation module is used to segment the corpus;

[0034] The text topic training module is used to perform text topic training based on the LDA training model, and obtain the vocabulary with the largest weight value in each topic classification as the topic word;

[0035] The word vector training module is used to perform word vector training based on the Word2Vec word vector model to obtain a vector representation of each word;

[0036] The subject word semantic similarity calculation module is used to calculate the semantic similarity between subject words based on the cosine similarity algorithm;

[0037] The semantic primitive construction module is used to select the subject word with the lowest semantic similarity as the semantic primitive;

[0038] The semantic primitive dimension weight calculation module is used to calculate the semantic similarity between each word and the semantic primitive as the weight in the corresponding semantic primitive dimension;

[0039] The normalization module is used to perform normalization processing after calculating the similarity between a word and all semantic primitives;

[0040] The weight distribution module is used to divide the similarity of each semantic primitive by the sum of the similarities of the vocabulary in all semantic primitives, so that the sum of the weights of the vocabulary allocated to each semantic primitive is 1;

[0041] The feature vector representation module is used to construct a feature vector representation of a vocabulary, and after calculating the weight of each vocabulary allocated to the semantic primitive, obtains the feature vector representation of the vocabulary represented by each semantic primitive.

[0042] As an optimal technical solution, the full mode is used to segment the corpus.

[0043] As a preferred technical solution, the text topic training module is used to perform text topic training based on the LDA training model, specifically including:

[0044] Randomly select a topic from the topic distribution, randomly select a word based on the word distribution of the topic, calculate the correlation between the internal words of a topic word, calculate the text consistency value of different numbers of topics, select the value with the highest text consistency as the number of topics, take the average of the consistency indicators of all topics, label the text after training the text topics and print a set number of topics with higher weights under each topic.

[0045] As a preferred technical solution, the subject word semantic similarity calculation module is used to calculate the semantic similarity between subject words based on the cosine similarity algorithm, which is specifically expressed as:

[0046]

[0047] Among them, Them i ,Them j Indicates the similarity between two keywords, wv ik and wv jk Represents the weight of the topic word on the kth dimension after word vector training.

[0048] As a preferred technical solution, the weight allocation module is used to divide the similarity of each semantic primitive by the sum of the similarities of the vocabulary in all semantic primitives, so that the sum of the weights of the vocabulary on each semantic primitive is 1, and the final weight of the vocabulary on the semantic primitive is calculated as:

[0049]

[0050] Among them, P nk represents the final weight of word n ​​on semantic primitive k, X nk Represents the semantic similarity between word n ​​and semantic primitive k.

[0051] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0052] (1) Semantic primitive extraction methods in other fields are mostly manual extraction according to the extraction principles of semantic primitives, such as HowNet of CNKI. The results of such semantic primitive extraction do not have domain characteristics and are difficult to accurately represent the terms in the financial reporting field. The semi-automatic semantic primitive extraction method for the financial field is limited to the interpretation and interpretation relationship of the dictionary corpus. It is impossible to use semantic primitives to represent words that are not logged in the dictionary. The present invention realizes accurate word vector representation for the processing of unregistered words (words that have not been annotated with semantic primitives) by the semantic primitive method, thereby reducing the requirements for high-quality annotated data.

[0053] (2) In order to solve the problem of high dimensionality and sparse data in interpretable word vector representation methods, the present invention adopts semantic primitives as vector dimensions and uses the similarity with semantic primitives as the numerical value in the dimension, instead of the traditional method of using semantic primitive vector weighting. In order to represent the semantic features of words, the existing word co-occurrence method and traditional word vector training method usually require higher dimensionality (more than 300 dimensions) to represent words, and the non-word vector training method lacks connection between dimensions in order to clearly represent different word features, resulting in sparse data. In addition, traditional corpus training and existing semantic primitive representation methods (weighted and semantic primitive annotated corpus) cannot avoid the dimensional uninterpretability of word vectors. The present invention uses semantic primitives as an explicit representation of the semantic meaning of words, and maps all words into the same set of semantic primitives to improve the interpretability of word vector representation methods.

[0054] After obtaining the subject words and calculating the similarity between them, based on the principle of semantic primitive independence, a limited number of words with the smallest semantic similarity are selected as semantic primitives. Using a limited number of semantic primitives to represent word vectors can reduce the representation dimension of word vectors and improve the density of the vector space.

[0055] (3) The present invention adopts the LDA model to perform topic training on the text, and can extract subject words that can represent other words in the corpus as candidate semantic primitives, and utilize the independence characteristics of semantic primitives to extract subject words with low semantic similarity with other subject words as semantic primitives, thereby avoiding the problem of not being able to effectively represent unregistered words due to the inability to find the original meaning when extracting interpretation words from dictionary texts by manual or semi-automatic methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 Schematic diagram of the flow of the annual report text word vector representation method based on semantic primitives of the present invention;

[0057] Figure 2 Schematic diagram comparing the way the present invention and other methods use semantic primitives to represent vocabulary. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0059] Example 1

[0060] like Figure 1As shown, this embodiment provides a word vector representation method for annual report text based on semantic primitives, including semantic primitive extraction and word vector representation using semantic primitives. The annual report text is the text data disclosed by the current enterprise, and the word vector is the vector representation of each word after word segmentation processing of the text. Specifically, the method includes the following steps:

[0061] S1: Get text keywords;

[0062] S11: Word segmentation: Use Python version 3.7, load the jieba library, take out the vocabulary and stop words / dictionary including characters and cities from the accounting dictionary, obtain the TXT text of the accounting field dictionary and the TXT text of the stop words in the financial report field, build a corpus based on the annual report text, call the jieda.cut() method, and use the parameter cut_all=True to segment the corpus in full mode.

[0063] S12: Topic training: Load the mallet 2.0.8 package and use os.environ.update() to update environment variables and optimize the LDA training model. The LDA training model first defines the number of topics K in the text. Each topic contains a set of vocabulary distributions. Then, for each document, a topic is randomly selected from the topic distribution, and a word is randomly selected based on the word distribution of the topic. Finally, the entire process is repeated until a complete document is generated. The text is trained using the gensim.models.wrappers.LdaMallet() method, and the CoherenceModel() method calculates the text consistency value for different numbers of topics. Start, limit, and step are the initial number of topics, the maximum number of topics, and the step size parameters.

[0064] S13: Select the value with the highest text consistency as the number of topics.

[0065] In this embodiment, text consistency refers to text coherence, focusing on the degree of correlation between internal words in a topic. The greater the correlation, the better the consistency and the better the topic classification effect. Specifically, the method encapsulated in the mallet package is called;

[0066] The selection process is as follows: 1. Calculate the correlation between the internal words of a subject word; 2. Calculate the consistency index of all topics; 3. Take the average of the consistency index of all topics.

[0067] After training the text topics, label the text and print the top 10 topics with higher weights under each topic.

[0068] Among them, the top 10 topics with higher weights under each topic are printed, specifically:

[0069] (1) Loop through the dictionary enumerate() of LDA model training;

[0070] (2) Read the topics and the corresponding keywords stored in the LDA dictionary using the ldamodel.show_topic() method;

[0071] (3) storing the subject number and the subject words corresponding to the subject in a table-type data structure;

[0072] S2: Word vector training for annual report text

[0073] Select python 3.7, obtain the TXT text of the accounting field dictionary and the TXT text of the stop words in the financial report field, load the TXT file of the financial report note text corpus, cut the words and remove the stop words, load the Word2Vec word vector model, model = Word2Vec (size = XX, workers = XX, sg = 1) to generate the word vector training model, sg = 1 selects the skip word model.

[0074] Load the word vector training model:

[0075] model.train(word_clean, total_examples = model.corpus_count, epochs = model.iter) loads text, loads dictionaries, stops words, performs word segmentation on the text, trains the word vector model, and improves the accuracy of word vector training.

[0076] For word vector training, other solutions such as Glove and Bert can also be implemented;

[0077] S3: Calculate the cosine similarity between topic words;

[0078] Cosine similarity algorithm: Them i ,Them j Indicates the similarity between two keywords, wv ik and wv jk Represents the weight of the topic word on the kth dimension after word vector training.

[0079]

[0080] Call the word vector model:

[0081] new_model=gensim.models.Word2Vec.load('gensim_w2v_sg0_model')

[0082] Save the word vector training results in memory as a model file, and gensim.models.Word2Vec.load() represents the word vector training results;

[0083] Calculate the cosine similarity of the theme words print(new_model.similarity("Them i ","Them j ")), get the semantic similarity between the subject words;

[0084] S4: According to the independence principle of semantic primitives, semantic primitives are selected according to the rule of lowest semantic similarity with other subject words. ij Represents the semantic similarity between topic word i and topic word j.

[0085] As for the method of extracting the initial semantic primitives, a method such as TF-IDF can be used to extract the semantic primitives.

[0086] S5: Call the word vector model:

[0087] new_model=gensim.models.Word2Vec.load('gensim_w2v_sg0_model')

[0088] Save the results of word vector training as a model file, and use the load() method to load the model file that stores the word vector results.

[0089] S6: Calculate the semantic similarity between words and semantic primitives as the weight of the corresponding semantic primitive dimension: print(new_model.similarity("word i ","word j "));

[0090] After the word vector training is completed, this embodiment will initially obtain the vector representation of each word (including words of semantic primitives), and use the cosine similarity algorithm to perform vector operations to achieve similarity calculation.

[0091] S7: After calculating the similarity between a word and all semantic primitives, normalize the similarity between each primitive and the word by the sum of the similarities between the word and all semantic primitives, so that the sum of the weights of the word on each semantic primitive is 1. The final weight of the word on the semantic primitive is calculated as:

[0092]

[0093] Among them, P nk represents the final weight of word n ​​on semantic primitive k, Xnk Represents the semantic similarity between word n ​​and semantic primitive k;

[0094] like Figure 2 As shown, after calculating the weight of a word allocated to the semantic primitives, this embodiment can obtain a feature vector representation of the word represented by each semantic primitive. The annual report text word vector refers to the vector representation of each word in the annual report text.

[0095] Example 2

[0096] This embodiment provides a semantic primitive-based annual report text word vector representation system, comprising: a text acquisition module, a corpus construction module, a word segmentation module, a text topic training module, a word vector training module, a topic word semantic similarity calculation module, a semantic primitive construction module, a semantic primitive dimension weight calculation module, a normalization module, a weight assignment module, and a feature vector representation module;

[0097] In this embodiment, the text acquisition module is used to acquire dictionary text in the accounting field and stop word text in the financial report field;

[0098] In this embodiment, the corpus construction module is used to construct a corpus based on the annual report text;

[0099] In this embodiment, the word segmentation module is used to segment the corpus;

[0100] In this embodiment, the text topic training module is used to perform text topic training based on the LDA training model, and obtain the vocabulary with the largest weight value in each topic classification as the topic word;

[0101] In this embodiment, the word vector training module is used to perform word vector training based on the Word2Vec word vector model to obtain a vector representation of each word;

[0102] In this embodiment, the subject word semantic similarity calculation module is used to calculate the semantic similarity between subject words based on the cosine similarity algorithm;

[0103] In this embodiment, the semantic primitive construction module is used to select the subject word with the lowest semantic similarity as the semantic primitive;

[0104] In this embodiment, the semantic primitive dimension weight calculation module is used to calculate the semantic similarity between each word and the semantic primitive as the weight in the corresponding semantic primitive dimension;

[0105] In this embodiment, the normalization module is used to perform normalization processing after calculating the similarity between a word and all semantic primitives;

[0106] In this embodiment, the weight allocation module is used to divide the similarity of each semantic primitive by the sum of the similarities of the vocabulary in all semantic primitives, so that the sum of the weights of the vocabulary on each semantic primitive is 1;

[0107] In this embodiment, the feature vector representation module is used to construct the feature vector representation of the vocabulary. After calculating the weight of each vocabulary allocated to the semantic primitive, the feature vector representation of the vocabulary represented by each semantic primitive is obtained.

[0108] In this embodiment, the full mode is used to segment the corpus.

[0109] In this embodiment, the text topic training module is used to perform text topic training based on the LDA training model, specifically including:

[0110] Randomly select a topic from the topic distribution, randomly select a word based on the word distribution of the topic, calculate the correlation between the internal words of a topic word, calculate the text consistency value of different numbers of topics, select the value with the highest text consistency as the number of topics, take the average of the consistency indicators of all topics, label the text after training the text topics and print a set number of topics with higher weights under each topic.

[0111] In this embodiment, the subject word semantic similarity calculation module is used to calculate the semantic similarity between subject words based on the cosine similarity algorithm, which is specifically expressed as follows:

[0112]

[0113] Among them, Them i ,Them j Indicates the similarity between two keywords, wv ik and wv jk Represents the weight of the topic word on the kth dimension after word vector training.

[0114] In this embodiment, the weight allocation module is used to divide the similarity of each semantic primitive by the sum of the similarities of the vocabulary in all semantic primitives, so that the sum of the weights of the vocabulary on each semantic primitive is 1. The final weight of the vocabulary on the semantic primitive is calculated as:

[0115]

[0116] Among them, P nk represents the final weight of word n ​​on semantic primitive k, X nk Represents the semantic similarity between word n ​​and semantic primitive k.

[0117] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be considered as equivalent replacement methods and are included in the scope of protection of the present invention.

Claims

1. A word vector representation method for annual report text based on semantic primitives, characterized by: The steps include: Obtain accounting dictionary text and financial reporting stop word text, build a corpus based on the annual report text, and perform word segmentation on the corpus; Perform text topic training based on the LDA training model and obtain the words with the largest weight in each topic classification as the topic words; Perform word vector training based on the Word2Vec word vector model to obtain the vector representation of each word; Calculate the semantic similarity between topic words based on the cosine similarity algorithm; Select the topic word with the lowest semantic similarity as the semantic primitive; Calculate the semantic similarity between each word and the semantic primitive as the weight in the corresponding semantic primitive dimension; After calculating the similarity between a word and all semantic primitives, normalize it by dividing the similarity of each semantic primitive by the sum of the similarities of the word in all semantic primitives, so that the sum of the weights of the word on each semantic primitive is 1. After calculating the weight of each word allocated to the semantic primitive, the feature vector representation of the word represented by each semantic primitive is obtained.

2. The method for representing annual report text word vectors based on semantic primitives according to claim 1 is characterized in that: The full mode is used to segment the corpus.

3. The method for representing annual report text word vectors based on semantic primitives according to claim 1 is characterized in that: Text topic training is performed based on the LDA training model, specifically including: Randomly select a topic from the topic distribution, randomly select a word based on the word distribution of the topic, calculate the correlation between the internal words of a topic word, calculate the text consistency value of different numbers of topics, select the value with the highest text consistency as the number of topics, take the average of the consistency indicators of all topics, label the text after training the text topics and print a set number of topics with higher weights under each topic.

4. The method for representing annual report text word vectors based on semantic primitives according to claim 1, characterized in that: The semantic similarity between topic words is calculated based on the cosine similarity algorithm, which is specifically expressed as: Among them, Them i ,Them j Indicates two keywords, wv ik and wv jk Represents the weight of the topic word on the kth dimension after word vector training.

5. The method for representing annual report text word vectors based on semantic primitives according to claim 1, characterized in that: Divide the similarity of each semantic primitive by the sum of the similarities of the word in all semantic primitives, so that the sum of the weights of the word on each semantic primitive is 1. The final weight of the word on the semantic primitive is calculated as: Among them, P nk represents the final weight of word n ​​on semantic primitive k, X nk Represents the semantic similarity between word n ​​and semantic primitive k.

6. A semantic primitive-based annual report text word vector representation system, characterized by: include: Text acquisition module, corpus construction module, word segmentation module, text topic training module, word vector training module, topic word semantic similarity calculation module, semantic primitive construction module, semantic primitive dimension weight calculation module, normalization module, weight assignment module, feature vector representation module; The text acquisition module is used to acquire accounting field dictionary text and financial report field stop word text; The corpus construction module is used to construct a corpus based on the annual report text; The word segmentation module is used to segment the corpus; The text topic training module is used to perform text topic training based on the LDA training model, and obtain the vocabulary with the largest weight value in each topic classification as the topic word; The word vector training module is used to perform word vector training based on the Word2Vec word vector model to obtain a vector representation of each word; The subject word semantic similarity calculation module is used to calculate the semantic similarity between subject words based on the cosine similarity algorithm; The semantic primitive construction module is used to select the subject word with the lowest semantic similarity as the semantic primitive; The semantic primitive dimension weight calculation module is used to calculate the semantic similarity between each word and the semantic primitive as the weight in the corresponding semantic primitive dimension; The normalization module is used to perform normalization processing after calculating the similarity between a word and all semantic primitives; The weight distribution module is used to divide the similarity of each semantic primitive by the sum of the similarities of the vocabulary in all semantic primitives, so that the sum of the weights of the vocabulary allocated to each semantic primitive is 1; The feature vector representation module is used to construct a feature vector representation of a vocabulary, and after calculating the weight of each vocabulary allocated to the semantic primitive, obtains the feature vector representation of the vocabulary represented by each semantic primitive.

7. The annual report text word vector representation system based on semantic primitives according to claim 6 is characterized in that: The full mode is used to segment the corpus.

8. The annual report text word vector representation system based on semantic primitives according to claim 6 is characterized in that: The text topic training module is used to perform text topic training based on the LDA training model, specifically including: Randomly select a topic from the topic distribution, randomly select a word based on the word distribution of the topic, calculate the correlation between the internal words of a topic word, calculate the text consistency value of different numbers of topics, select the value with the highest text consistency as the number of topics, take the average of the consistency indicators of all topics, label the text after training the text topics and print a set number of topics with higher weights under each topic.

9. The annual report text word vector representation system based on semantic primitives according to claim 6 is characterized in that: The subject word semantic similarity calculation module is used to calculate the semantic similarity between subject words based on the cosine similarity algorithm, which is specifically expressed as: Among them, Them i ,Them j Indicates two keywords, wv ik and wv jk Represents the weight of the topic word on the kth dimension after word vector training.

10. The annual report text word vector representation system based on semantic primitives according to claim 6 is characterized in that: The weight distribution module is used to divide the similarity of each semantic primitive by the sum of the similarities of the vocabulary in all semantic primitives, so that the sum of the weights of the vocabulary on each semantic primitive is 1, and the final weight of the vocabulary on the semantic primitive is calculated as: Among them, P nk represents the final weight of word n ​​on semantic primitive k, X nk Represents the semantic similarity between word n ​​and semantic primitive k.

Images