A semantic alignment method for XBRL financial information elements based on domain ontology

Abstract

The present invention discloses a semantic alignment method for XBRL financial information elements based on domain ontology. The method comprises the following steps: extracting candidate domain concepts and constructing a domain dictionary; obtaining financial report annotation text as a corpus; performing word segmentation, part-of-speech tagging, and named entity recognition on the corpus to obtain all part-of-speech tagging results and constructing a stop word vocabulary for the financial report domain; expanding the corpus based on an LDA model and semantically expanding the financial report annotation text based on the expanded corpus; converting the vocabulary in the new corpus into vocabulary feature vectors; extracting hierarchical relationships and non-classification relationships; establishing an alignment method and performing semantic alignment of financial information elements. The present invention uses domain ontology to formally represent financial report terms and calculates the semantic similarity between tags in a tag link library, thereby achieving semantic alignment between financial information elements and maintaining consistency between XBRL reports.

Description

Technical Field

[0001] The present invention relates to the technical field of financial data processing, and in particular to a domain ontology-based XBRL financial information element semantic alignment method. Background Art

[0002] The financial information elements in the XBRL taxonomy technical system represent the financial reporting terms used in the disclosure of corporate financial information. In the tag link library of the XBRL taxonomy, tags are the display names of financial information elements. XBRL is the basis for the structuring and semanticization of financial data. However, XBRL itself has some limitations. Inconsistencies between XBRL reports disclosed by companies are a common problem. To solve these problems, more flexible financial data structuring and semanticization methods are needed, and artificial intelligence technology must be relied upon to automatically extract, cross-validate, and search for information.

[0003] Among existing solutions, one approach uses the OWL ontology language to transform XBRL taxonomies and performs semantic mapping between two ontologies to achieve semantic alignment of inconsistent information elements. This approach fully utilizes the technical information of the taxonomies, but the weak semantics of the taxonomies themselves cannot be avoided. Furthermore, the process model of converting taxonomies into ontologies and establishing mappings between two ontologies cannot reduce the complexity of data processing.

[0004] One approach uses an X-IM framework, using an investor ontology as a shared ontology. The investor ontology encapsulates a set of equivalent investor terms, maps tag elements in the XBRL tag link library to a class in the index ontology, and aligns equivalent terms with the investor ontology, thereby addressing the semantic heterogeneity of XBRL in enterprise applications. This approach targets automated reporting for specific XBRL application requirements and does not address methods for constructing ontologies in the financial reporting domain. Furthermore, the information element alignment method of vector computing makes it difficult to avoid the lack of semantic description information for terms.

[0005] There is a method that uses ant colony algorithm instead of domain ontology to align the financial information elements between two instance documents based on bipartite graph modeling of the information of each XBRL link library. The disadvantage is that it can only process two instance documents at a time, and the efficiency of data processing needs to be solved.

[0006] There is an ontology-enhanced XBRL model that builds a domain-shared ontology based on the XBRL classification standard. The path distance between concepts in the upper-level ontology is used as the calculation result to measure the similarity between concepts, thereby achieving semantic alignment of terms. However, the details of the construction of the financial reporting domain ontology are not mentioned, and only considering the path distance between concepts cannot accurately characterize the semantic information of terms.

[0007] Some methods use the link library relationship between classification standards to convert the relationship between concepts into is-a relationship, and use the M3 hierarchical similarity algorithm to align financial report information elements. However, this method has limitations in mapping different relationships between concepts into is-a relationships, and the problem of poor mapping accuracy is difficult to avoid.

[0008] Some papers use the OFXD decision framework as an ontology component to classify XBRL elements and achieve mapping between information elements, but they do not provide a method on how to achieve semantic mapping between concepts based on the ontology;

[0009] Some approaches convert XBRL taxonomies into ontology models and align financial information elements in a rule-based manner. This approach relies on mapping templates. As the number of taxonomies increases, the cost of creating data conversion templates for mapping between taxonomies continues to increase.

[0010] In summary, the existing solutions have the following shortcomings:

[0011] 1. Ontologies for the financial reporting field usually describe a framework, but do not actually discuss the ontology construction methods for the financial reporting field;

[0012] 2. Most solutions map classification standards to ontologies, which makes it difficult to avoid the problem of weak semantics of the classification standards themselves. At the same time, the cost of creating data conversion templates based on rules to map classification standards to each other continues to increase.

[0013] 3. It is difficult to achieve mapping between multiple documents by using heuristic algorithms and classification standard component information, and data processing efficiency needs to be improved;

[0014] 4. The use of vector calculation and path distance to align information elements lacks the use of other semantic information, and the accuracy of data processing needs to be improved;

[0015] Therefore, there is an urgent need for a technical solution that can provide specific financial reporting domain ontology construction to solve the inconsistency problem of different XBRL files caused by the weak semantics of XBRL itself, thereby reducing data processing costs and improving data processing efficiency and accuracy. Summary of the Invention

[0016] In order to overcome the defects and shortcomings of the existing technology, the present invention provides a semantic alignment method for XBRL financial information elements based on domain ontology. The present invention adopts domain ontology to formally represent financial reporting terms, and uses a fuzzy semantic similarity calculation method to calculate the similarity between tags in a tag link library, thereby achieving semantic alignment between financial information elements and maintaining consistency between XBRL reports.

[0017] In order to achieve the above object, the present invention adopts the following technical solutions:

[0018] The present invention provides a method for semantic alignment of XBRL financial information elements based on domain ontology, comprising the following steps:

[0019] Extract the words with explanations in the accounting dictionary as candidate domain concepts and store them in the domain dictionary with one concept word per line;

[0020] Obtain the text of financial report notes as the corpus to be processed;

[0021] Perform word segmentation, part-of-speech tagging, and named entity recognition on the corpus to obtain all part-of-speech tagging results. Use preset screening rules to obtain multiple stop words, and store the multiple stop words as a stop word list for the financial reporting field.

[0022] Expand the corpus based on the LDA model, expand the semantics of the financial report notes text based on the expanded corpus, remove stop words from the text based on the stop word list, and use the expanded financial report notes text as a new corpus;

[0023] Convert the vocabulary in the new corpus into vocabulary feature vectors;

[0024] The final number of clusters is determined based on the change in the minimum value of inter-class differences. The maximum number of concepts in a class is used as the criterion for stopping iteration. When the number of concepts within a class is less than the criterion for stopping iteration, the iteration is stopped. The hierarchical classification of the domain ontology is obtained through iterative clustering.

[0025] The similarity between concepts within each class is calculated based on the cosine similarity algorithm, and the concept with the largest average similarity between the concept within the class and other concepts is used as the class label;

[0026] Perform word segmentation and part-of-speech tagging on the domain dictionary and the stop word list in the financial reporting field, and perform dependency syntax analysis to extract the current concept word and another concept word that has a corresponding relationship with the current concept word. Extract the verbs between the concept pairs as object attribute relationships;

[0027] Extract verbs that have syntactic dependencies with the vocabulary in the financial report notes text;

[0028] Establishing alignment between the financial information elements of the XBRL taxonomy and the financial reporting concept ontology;

[0029] The similarity between the concept words corresponding to the financial information elements is calculated, and the semantic similarity of the information elements is calculated and aligned based on the fuzzy semantic similarity.

[0030] As a preferred technical solution, multiple stop words are obtained through preset screening rules, and the preset screening rules are expressed as:

[0031] The words with part-of-speech tags of morphemes, quantifiers, adverbs, and prepositions are deleted, and the entity recognition results are: organization names, personal names, and place names.

[0032] As a preferred technical solution, the semantics of the financial report notes text is expanded based on the extended corpus, specifically including:

[0033] Calculate the semantic consistency of texts with different numbers of topics, select the value with the largest semantic consistency result as the number of topics, and build a topic dictionary;

[0034] Taking a sentence of the financial report notes text as a unit, we search for words in a unit. When a word appears in the subject dictionary, we expand the entire subject vocabulary to the financial report notes text. When multiple words appear in the same subject dictionary, we expand them only once and use the expanded financial report notes text as the new corpus.

[0035] As a preferred technical solution, the verbs between concept pairs are extracted as object attribute relationships, specifically including:

[0036] When the dependency syntactic relationship is a subject-predicate relationship, a verb-object relationship, an indirect object relationship, a prepositioned object, or a conjunctive modifier, and the parent node and child node vocabulary of the dependency syntax are both extracted concepts, the predicate verb between the concept pairs is extracted as the object attribute relationship.

[0037] As a preferred technical solution, verbs that have syntactic dependencies with the vocabulary in the notes to the financial report are extracted, specifically including:

[0038] Search the vocabulary in the concept dictionary to determine whether the concept dependency syntactic relationship is a pre-set syntactic dependency relationship. If one of the relationships exists, extract the attribute words corresponding to the relationship.

[0039] As a preferred technical solution, the alignment of financial information elements of the XBRL taxonomy with the financial reporting concept ontology is established, specifically including:

[0040] The financial information element is used as the tag name in the XBRL link library, the financial reporting domain ontology is stored in OWL language, and the financial information element exists as an instance of the financial reporting domain ontology. The relationship between the financial information element and each concept in the domain ontology through word segmentation is an instance relationship.

[0041] As a preferred technical solution, the similarity between the concept words corresponding to the financial information elements is calculated, specifically including:

[0042] Concept similarity is expressed as a triple C = {R, Dep, Den}, where R includes the path distance similarity, relationship similarity, and attribute similarity between concepts, Dep represents the concept distance similarity, and Den represents the concept density similarity. The specific calculation formula is expressed as follows:

[0043]

[0044]

[0045]

[0046]

[0047]

[0048] α+β+μ+v+η=1

[0049] Among them, RP sim Represents the path distance similarity between concepts, RR sim Represents the similarity between concepts, RA sim Indicates the attribute similarity between concepts, Dep sim Indicates the similarity between concepts, Den sim represents the density similarity between concepts, Indicates the similarity between concepts. Represents the concepts s and t corresponding to the two terms i and j, θ is an adjustable parameter greater than 0, and α, β, μ, v, and η represent the weight values ​​corresponding to each similarity.

[0050] As a preferred technical solution, the semantic similarity of information elements is calculated and aligned based on fuzzy semantic similarity, specifically including:

[0051] Constructing financial information elements and Represents the concept words corresponding to the last position of two financial information elements, FRT i and FRT j represents two financial information elements, and n represents the number of concepts of financial information elements;

[0052] The semantic alignment formula of financial information elements is specifically expressed as:

[0053]

[0054]

[0055] Among them, RP represents the path distance between two concepts in the domain ontology. The superscript t indicates other concepts in the financial information element except the last position concept. The superscript 0 refers to the concept at the end position, n ≥ 1 means that concepts also exist at other positions except the end position, and λ and ∈ are adjustable parameters greater than 0.

[0056] The present invention also provides a domain ontology-based XBRL financial information element semantic alignment system, comprising: a candidate domain concept extraction module, a domain dictionary construction module, a corpus construction module, a stop word vocabulary construction module, a corpus expansion module, a vocabulary feature vector construction module, a hierarchical relationship extraction module, a non-classification relationship extraction module, an alignment mode construction module, and a semantic alignment module;

[0057] The candidate domain concept extraction module is used to extract words with explanations in the accounting dictionary as candidate domain concepts;

[0058] The domain dictionary building module is used to store one concept word per line to build a domain dictionary;

[0059] The corpus building module is used to obtain the financial report notes text as the corpus to be processed;

[0060] The stop word vocabulary building module is used to perform word segmentation, part-of-speech tagging and named entity recognition on the corpus, obtain all part-of-speech tagging results, obtain multiple stop words through preset screening rules, and store the multiple stop words as a stop word vocabulary in the financial reporting field;

[0061] The corpus expansion module is used to expand the corpus based on the LDA model, expand the semantics of the financial report notes text based on the expanded corpus, remove stop words from the text based on a stop word list, and use the expanded financial report notes text as a new corpus;

[0062] The vocabulary feature vector construction module is used to convert the vocabulary in the new corpus into vocabulary feature vectors;

[0063] The hierarchical relationship extraction module is used to extract hierarchical relationships, specifically including:

[0064] The final number of clusters is determined based on the change in the minimum value of inter-class differences. The maximum number of concepts in a class is used as the criterion for stopping iteration. When the number of concepts within a class is less than the criterion for stopping iteration, the iteration is stopped. The hierarchical classification of the domain ontology is obtained through iterative clustering.

[0065] The similarity between concepts within each class is calculated based on the cosine similarity algorithm, and the concept with the largest average similarity between the concept within the class and other concepts is used as the class label;

[0066] The non-classification relationship extraction module is used to extract non-classification relationships, specifically including:

[0067] Perform word segmentation and part-of-speech tagging on the domain dictionary and the stop word list in the financial reporting field, and perform dependency syntax analysis to extract the current concept word and another concept word that has a corresponding relationship with the current concept word. Extract the verbs between the concept pairs as object attribute relationships;

[0068] Extract verbs that have syntactic dependencies with the vocabulary in the financial report notes text;

[0069] The alignment construction module is used to establish an alignment between the financial information elements of the XBRL taxonomy and the financial reporting concept ontology;

[0070] The semantic alignment module is used to perform semantic alignment of financial information elements, calculate the similarity between concept words corresponding to financial information elements, and calculate and align the semantic similarity of information elements based on fuzzy semantic similarity.

[0071] The present invention also provides a computer device comprising a processor and a memory for storing a program executable by the processor. When the processor executes the program stored in the memory, the method for semantic alignment of XBRL financial information elements based on domain ontology is implemented.

[0072] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0073] (1) The present invention realizes semantic alignment of financial information elements by constructing and utilizing a financial reporting domain ontology. First, in the XBRL field, discussions on manually constructing a financial reporting domain ontology and framework have been conducted, but there has been no discussion on how to automatically or semi-automatically construct a financial reporting domain ontology from a corpus. In addition, domain ontology construction methods in other general fields lack the characteristics of the financial reporting domain. In response to the problem of short financial report texts, the present invention utilizes an LDA topic model to expand the corpus, semi-automatically enhance the semantic relationship between similar words, and improve the feature representation effect of word vectors.

[0074] Secondly, in the field of XBRL semantic alignment, in response to the current problems of high XBRL semantic alignment data processing costs and the inability to process multiple files simultaneously, the financial reporting domain ontology of the present invention acts as an intermediate medium to enable multiple instance documents of multiple companies to be processed together. This process can be carried out by mapping financial information elements with different company representations to instances in the ontology, and further, on the basis of fuzzy semantic similarity calculation methods, the semantic information of financial information elements is integrated to enhance the semantic description ability of financial information elements, thereby improving the accuracy of semantic alignment of financial information elements.

[0075] (2) The present invention adopts a set of domain-shared concepts to describe the tags in the XBRL link library. Its financial reporting domain ontology avoids the problem that the method of mapping classification standards to ontology needs to re-match data templates due to the increase in financial information elements. At the same time, it avoids the problem that the method based on heuristic algorithm and ontology mapping cannot perform semantic alignment of more than two XBRL files at the same time, thereby reducing data processing costs and improving data processing efficiency.

[0076] (3) The present invention integrates the semantic information of financial report information elements from the perspectives of path, relationship, attribute, depth and density between the concepts corresponding to the terms, thereby avoiding the problem of insufficient semantic alignment accuracy caused by insufficient semantic information. Furthermore, based on the semantic characteristics of financial information elements, a fuzzy semantic similarity algorithm is designed to improve the efficiency of data processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 Schematic diagram of the process of the domain ontology-based semantic alignment method of XBRL financial information elements of the present invention;

[0078] Figure 2 A schematic diagram of the contents of the extended dictionary of the present invention;

[0079] Figure 3 Schematic diagram of calculation results of similarity between concepts within each class of the present invention;

[0080] Figure 4 A schematic diagram of the results of extracting verbs that have syntactic dependencies with terms in the present invention;

[0081] Figure 5 A schematic diagram of the system for describing financial reporting terms in the financial reporting domain ontology of the present invention;

[0082] Figure 6 This is a schematic diagram of the hierarchical relationships and non-classified relationships contained in the financial information elements of the financial reporting domain ontology of the present invention. DETAILED DESCRIPTION

[0083] 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.

[0084] Example 1

[0085] like Figure 1 As shown, this embodiment provides a method for semantic alignment of XBRL financial information elements based on domain ontology, including constructing a financial report domain ontology and semantic alignment of information elements, specifically including the following steps:

[0086] S1: Concept extraction and representation, from the accounting dictionary, the words with explanations are stored as domain concepts in a txt file with one word per line;

[0087] S11: Extract words with detailed explanations from the accounting dictionary as candidate concepts and store them in a TXT document with one concept word per line;

[0088] S12: Processing of the financial report notes text corpus, specifically including:

[0089] S121: Obtaining stop words in the financial reporting field. Using Python version 3.6.4, load the pyltp library, call the TXT document storing concepts as the domain dictionary, and load the TXT text storing the financial report notes as the corpus to be processed. Call the segmantor() method in the pyltp library to perform word segmentation on the financial report notes text, call the postagger() and recongnizer() methods in the pyltp library to perform part-of-speech tagging and named entity recognition on the corpus respectively, and obtain the results of all part-of-speech tagging after word segmentation, part-of-speech tagging, and named entity recognition. Obtain stop words through screening rules. The specific rules are: delete words with part-of-speech tags as morphemes, quantifiers, adverbs, and prepositions. The entity recognition results are: organization names, personal names, and place names. Add words of the type [S-Ns|S-Nh|S-Ni] in the vocabulary to the general stop word TXT file as the stop word vocabulary in the financial reporting field, where Ns, Nh, and Ni represent place names, personal names, and organization names, respectively.

[0090] S122: Expand the corpus based on the LDA model and establish semantic associations between words. This embodiment uses the expanded corpus to expand the text semantics and uses the stop word list to remove stop words in the text, thereby improving the effect of word vector training and making the word vector training results more accurate, that is, the extracted word vectors are more accurate.

[0091] Based on the LDA model, a certain number of topics and the first 10 words that make up a topic are obtained as a semantic expansion dictionary. The expanded dictionary is then used to expand the corpus. The expansion rules are as follows: text segmentation, stop word removal, and word search in the text. If a word is in the expanded dictionary, the entire dictionary entry consisting of 10 words is expanded after the corresponding word. If the same dictionary word appears multiple times in a text, it is expanded only once, thereby establishing semantic relationships between the words.

[0092] Use Python 3.7, load the jieba library, load the accounting field dictionary TXT and the stop words in the financial reporting field. Call the jieda.cut() method to use the full mode to segment the corpus. Load mallet2.0.8 to optimize the LDA training model, starting with 100 topics, with a step size of 10 and a maximum domain of 1500. Calculate the semantic consistency of texts with different numbers of topics, and select the value with the largest semantic consistency result as the number of topics. After obtaining the number of topics, recalculate the LDA model with the optimal number. The topic content is the vocabulary that makes up each topic. Take the 10 words with the largest vocabulary weights in a topic as the topic content, that is, a topic includes 10 words, which is equivalent to an entry in an extended dictionary.

[0093] The number of topics affects the number of entries in the semantic extension dictionary. This embodiment selects topic consistency as an indicator and ultimately determines the number of topics to be 1440. Figure 2 As shown, the text is trained through the LDA model, and the result of the topic training is a series of topics. The topics are composed of words. The results of the topic training are saved in the first Excel document, that is, the extended dictionary, which is used for the subsequent semantic expansion of the corpus and the expansion of the financial report notes text. Expansion rules: take a sentence of the financial report notes text as a unit, query the words in a unit, if the word appears in the subject dictionary, then expand the entire subject vocabulary to the financial report notes text, when multiple words appear in the same subject dictionary, only expand once, and store the expanded financial report notes text in TXT as a new corpus. This embodiment expands the financial report notes text in order to establish semantic relationships between words and improve the word vector training effect, because the financial report notes text is relatively short after word segmentation, which affects the word vector training model that establishes connections based on the word context relationship.

[0094] S123: word vector training and extraction of concept word vectors;

[0095] Load the TXT corpus after topic expansion. Load the Word2Vec() method to obtain feature vectors to represent vocabulary. The word vector dimension is set to 250 dimensions, the window is set to 5, and the word vector training is started using the skip-gram model [size=250|workers=5|sg=1]. sg=1 indicates that the skip-gram model in Word2Vec is selected during the word vector training phase. Word2Vec() is a tool that converts words into vector form. Through this conversion, the processing of text content can be simplified to vector operations in the vector space, and the similarity in the vector space can be calculated to represent the semantic similarity of the text. After word segmentation and stop word removal, the word vector training model is loaded, and the feature vector of each word is output according to the word vector training model.

[0096] The training of word vectors can be done by increasing the size of the corpus, using deep learning methods such as tensorflow, BERT, or various weighted sum methods.

[0097] Use Python to read the feature vector of each word and save it in model form. By looping through the concept words in the concept dictionary TXT file, obtain the feature vector of each concept word in each row of the concept dictionary TXT file from the memory and save it to the second Excel;

[0098] S2: Hierarchical relationship extraction;

[0099] S21: Load word vectors in Excel using Clementine 12.0. Select the two-step clustering algorithm. Based on the analysis of the conceptual structure of the financial report, set the maximum and minimum values ​​of the first clustering to 20 and 13, respectively. The two-step clustering algorithm determines the final number of clusters based on the change in the minimum value of the inter-class difference.

[0100] S22: In the following iterative clustering process, the maximum number of concepts in the class, MaxFR, is used as the criterion for stopping the iteration. When the number of concepts in the class is less than MaxFR, the iteration is stopped. In this process, different MaxFR values ​​are set to select appropriate results, and the hierarchical classification of the domain ontology is obtained through iterative clustering.

[0101] S23: Selection of class labels, such as Figure 3 As shown in the figure, the cosine similarity algorithm is used to calculate the similarity between concepts within each class, and the concept with the largest average similarity between the concept within the class and other concepts is used as the class label;

[0102] S3: Non-classification relationship extraction. Non-classification relationships include object attribute relationships and numerical attribute relationships, which are described by predicate verbs. Specifically, they include:

[0103] S31: The predicate verb is usually the verb in the text. In order to improve data processing efficiency, this embodiment uses Python to count the word frequencies of all verbs identified in the part-of-speech tagging step in the corpus before extracting the predicate verb, and deletes verbs with a word frequency lower than 100.

[0104] S32: Object attribute extraction: Object attributes are the relationship between two concepts. Specific implementation methods:

[0105] Python version 3.6.4 loaded the pyltp library. A corpus of TXT text from the financial report notes of 18 companies across 18 industries was randomly selected. A dictionary and stop word list were loaded, and word segmentation and part-of-speech tagging were performed. Dependency parsing was also performed. The parse() method in the pyltp library was used to obtain the vocabulary's relation. If a vocabulary had a relation such as SBV or VOB, the concept and its related concepts were extracted.

[0106] When the dependency syntax relationship is subject-verb relationship SBV, verb-object relationship VOB, indirect object relationship IOB, prepositioned object FOB, and conjunctive word DBL, and the parent node and child node vocabulary of the dependency syntax are both extracted concepts, the verb between the concept pair is extracted as the object attribute relationship;

[0107] The specific implementation method is the parse() method in the pyltp library. i and ii represent the position information of the word in the text. Define the words, relation and heads lists, and obtain the concept of the dependency analysis result, the dependency relationship corresponding to the concept, and the verb corresponding to the relationship from the parse method respectively;

[0108] The logical process of implementation is:

[0109] (1) First, traverse the vocabulary in a text;

[0110] (2) Check whether the vocabulary is in the concept dictionary;

[0111] (3) When the word is in the concept dictionary, determine whether the concept dependency syntactic relation is the defined relation (SBV, VOB, IOB, FOB, DBL);

[0112] (4) When one of the relationships exists, skip the current word position in the current word loop and repeat steps (2) and (3). At this time, the current position word and the predicate verb leads corresponding to the relationship are obtained. The result extraction formula is as follows:

[0113] FRO_.append("Term: "+words[i]+""+"Verb: "+heads[i]+"Term: "+words[ii])

[0114] After loop traversal and relationship judgment, a possible non-classification relationship is obtained: word[i] and word[ii] are a concept pair, and heads[i] is a non-classification relationship between concept pairs. The final result needs to be manually judged later.

[0115] S33: extracting numerical attribute relationships;

[0116] Numerical attributes describe the attribute relationship of a single term. This process is carried out by extracting verbs that have syntactic dependencies with the term. The parse() method is also used. The extraction process is as follows: (1) First, traverse the vocabulary in a text. (2) Find whether the vocabulary is in the concept dictionary. (3) If the vocabulary is in the concept dictionary, determine whether the concept dependency syntactic relationship relation is a defined relationship (SBV, VOB, IOB, FOB, DBL, ATT, ADV, CMP, POB, COO, LAD, RAD). If one of the relationships exists, extract the attribute word head corresponding to the relationship.

[0117] like Figure 4 As shown, the extraction syntax is:

[0118] Verbs of [SBV|VOB|IOB|FOB|DBL|ATT|ADV|CMP|POB|COO|LAD|RAD] and save in excel.

[0119] S4: Establish the alignment of the financial information elements of the XBRL taxonomy with the financial reporting concept ontology;

[0120] like Figure 5 As shown, the financial reporting domain ontology actually includes a system that can describe terms (in this embodiment, the terms refer to financial information elements, that is, the tag names in the XBRL link library), which can be used to represent financial reporting terms;

[0121] like Figure 6 As shown, the above steps obtain concepts, concept hierarchical relationships, concept non-classification relationships and concept numerical attributes as a domain ontology as a whole, that is, the financial reporting domain ontology describes the hierarchical relationships and non-classification relationships contained in financial information elements;

[0122] In the protégé software, the financial reporting domain ontology is stored in the OWL language, and each company's financial information element exists as an instance of the financial reporting domain ontology. The relationship between the financial information element and each concept in the domain ontology through word segmentation is an instance_of relationship.

[0123] The correspondence between financial information elements and concepts is achieved by segmenting the financial information elements, such as "financial assets" corresponds to the concept "assets".

[0124] S5: Use the concept ontology (i.e., the financial reporting domain ontology) to achieve semantic alignment of financial information elements. This process is divided into two steps: first, calculate the similarity between the concepts corresponding to the information elements, and finally calculate and align the semantic similarity of the information elements based on fuzzy semantic similarity;

[0125] S51: Concept similarity calculation: Concept similarity can be expressed as a triple C = {R, Dep, Den}, where R includes the path distance RP similarity between concepts and the relationship similarity RR (when the relationship is an equivalence relationship, RR sim =1, parent-child relationship RR sim =0.75, whole-part relationship RR sim =0.25) and attribute similarity RA, Dep represents concept distance similarity, and Den represents concept density similarity. Other formulas are as follows, where Represents the concepts s and t corresponding to two terms i and j. θ is an adjustable parameter greater than 0.

[0126]

[0127]

[0128]

[0129]

[0130]

[0131] α+β+μ+v+η=1

[0132] Among them, RP sim Represents the path distance similarity between concepts, RR sim Represents the similarity between concepts, RA sim Indicates the attribute similarity between concepts, Dep sim Indicates the similarity between concepts, Den sim represents the density similarity between concepts, Represents the similarity between concepts, and is the result of weighted summation of similarities such as path distance and density;

[0133] S52: Information element semantic alignment method based on fuzzy semantics, specifically including:

[0134] Financial Information Elements and Represents the concept words corresponding to the last position of two financial information elements, FRT i and FRT j It represents two financial information elements, and the superscript n represents the n concepts included in one financial information element.

[0135] λ and ∈ are adjustable parameters greater than 0, FR(FRT i ,FRT j ) is the final semantic alignment formula for financial information elements.

[0136]

[0137]

[0138] Among them, RP represents the path distance between two concepts in the domain ontology (concept ontology), which can be regarded as a graph, counting the length connecting the two concepts. The superscript t indicates other concepts in the financial information element except the last position concept. The superscript 0 refers to the concept at the last position. The n in the formula is the number of concepts in the financial information element, except for the concept at the last position. n ≥ 1 means that there are concepts in other positions except the last position. Concept s and concept t in financial information element i and financial information j need to first perform the first similarity calculation to obtain the correspondence between the concept pairs, and then accumulate the similarity results between the concept pairs and take the average value 1 / n.

[0139] Example 2

[0140] This embodiment provides a domain ontology-based XBRL financial information element semantic alignment system, comprising: a candidate domain concept extraction module, a domain dictionary construction module, a corpus construction module, a stop word vocabulary construction module, a corpus expansion module, a vocabulary feature vector construction module, a hierarchical relationship extraction module, a non-classification relationship extraction module, an alignment mode construction module, and a semantic alignment module;

[0141] In this embodiment, the candidate domain concept extraction module is used to extract words with explanations in the accounting dictionary as candidate domain concepts;

[0142] In this embodiment, the domain dictionary construction module is used to store one concept word per line to construct a domain dictionary;

[0143] In this embodiment, the corpus construction module is used to obtain the financial report note text as the corpus to be processed;

[0144] In this embodiment, the stop word vocabulary building module is used to perform word segmentation, part-of-speech tagging, and named entity recognition on the corpus, obtain all part-of-speech tagging results, obtain multiple stop words through preset screening rules, and store the multiple stop words as a stop word vocabulary in the financial reporting field;

[0145] In this embodiment, the corpus expansion module is used to expand the corpus based on the LDA model, expand the semantics of the financial report notes text based on the expanded corpus, remove stop words from the text based on a stop word list, and use the expanded financial report notes text as a new corpus;

[0146] In this embodiment, the vocabulary feature vector construction module is used to convert vocabulary in the new corpus into vocabulary feature vectors;

[0147] In this embodiment, the hierarchical relationship extraction module is used to extract hierarchical relationships, specifically including:

[0148] The final number of clusters is determined based on the change in the minimum value of inter-class differences. The maximum number of concepts in a class is used as the criterion for stopping iteration. When the number of concepts within a class is less than the criterion for stopping iteration, the iteration is stopped. The hierarchical classification of the domain ontology is obtained through iterative clustering.

[0149] The similarity between concepts within each class is calculated based on the cosine similarity algorithm, and the concept with the largest average similarity between the concept within the class and other concepts is used as the class label;

[0150] In this embodiment, the non-classification relationship extraction module is used to extract non-classification relationships, specifically including:

[0151] Perform word segmentation and part-of-speech tagging on the domain dictionary and the stop word list in the financial reporting field, and perform dependency syntax analysis to extract the current concept word and another concept word that has a corresponding relationship with the current concept word. Extract the verbs between the concept pairs as object attribute relationships;

[0152] Extract verbs that have syntactic dependencies with the vocabulary in the financial report notes text;

[0153] In this embodiment, the alignment construction module is used to establish the alignment between the financial information elements of the XBRL taxonomy and the financial reporting concept ontology;

[0154] In this embodiment, the semantic alignment module is used to perform semantic alignment of financial information elements, calculate the similarity between concept words corresponding to the financial information elements, and calculate and align the semantic similarity of information elements based on fuzzy semantic similarity.

[0155] Example 3

[0156] This embodiment provides a computing device, which may be a desktop computer, a laptop computer, a smart phone, a PDA handheld terminal, a tablet computer, or other terminal device with a display function. The computing device includes a processor and a memory, wherein the memory stores one or more programs. When the processor executes the programs stored in the memory, the domain ontology-based XBRL financial information element semantic alignment method of embodiment 1 is implemented.

[0157] 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 domain ontology-based semantic alignment method for XBRL financial information elements, characterized by: The steps include: Extract the words with explanations in the accounting dictionary as candidate domain concepts and store them in the domain dictionary with one concept word per line; Obtain the text of financial report notes as the corpus to be processed; Perform word segmentation, part-of-speech tagging, and named entity recognition on the corpus to obtain all part-of-speech tagging results. Use preset screening rules to obtain multiple stop words, and store the multiple stop words as a stop word list for the financial reporting field. Expand the corpus based on the LDA model, expand the semantics of the financial report notes text based on the expanded corpus, remove stop words from the text based on the stop word list, and use the expanded financial report notes text as a new corpus; Convert the vocabulary in the new corpus into vocabulary feature vectors; The final number of clusters is determined based on the change in the minimum value of inter-class differences. The maximum number of concepts in a class is used as the criterion for stopping iteration. When the number of concepts within a class is less than the criterion for stopping iteration, the iteration is stopped. The hierarchical classification of the domain ontology is obtained through iterative clustering. The similarity between concepts within each class is calculated based on the cosine similarity algorithm, and the concept with the largest average similarity between the concept within the class and other concepts is used as the class label; Perform word segmentation and part-of-speech tagging on the domain dictionary and the stop word list in the financial reporting field, and perform dependency syntax analysis to extract the current concept word and another concept word that has a corresponding relationship with the current concept word. Extract the verbs between the concept pairs as object attribute relationships; Extract verbs that have syntactic dependencies with the vocabulary in the financial report notes text; Establishing alignment between the financial information elements of the XBRL taxonomy and the financial reporting concept ontology; The similarity between the concept words corresponding to the financial information elements is calculated, and the semantic similarity of the information elements is calculated and aligned based on the fuzzy semantic similarity.

2. The domain ontology-based XBRL financial information element semantic alignment method according to claim 1, characterized in that: Multiple stop words are obtained through preset filtering rules. The preset filtering rules are expressed as follows: The words with part-of-speech tags of morphemes, quantifiers, adverbs, and prepositions are deleted, and the entity recognition results are: organization names, personal names, and place names.

3. The domain ontology-based XBRL financial information element semantic alignment method according to claim 1, characterized in that: The semantics of financial report notes are expanded based on the extended corpus, specifically including: Calculate the semantic consistency of texts with different numbers of topics, select the value with the largest semantic consistency result as the number of topics, and build a topic dictionary; Taking a sentence of the financial report notes text as a unit, we search for words in a unit. When a word appears in the subject dictionary, we expand the entire subject vocabulary to the financial report notes text. When multiple words appear in the same subject dictionary, we expand them only once and use the expanded financial report notes text as the new corpus.

4. The domain ontology-based XBRL financial information element semantic alignment method according to claim 1, characterized in that: Extract verbs between concept pairs as object attribute relationships, including: When the dependency syntactic relationship is a subject-predicate relationship, a verb-object relationship, an indirect object relationship, a prepositioned object, or a conjunctive modifier, and the parent node and child node vocabulary of the dependency syntax are both extracted concepts, the predicate verb between the concept pairs is extracted as the object attribute relationship.

5. The method for semantic alignment of XBRL financial information elements based on domain ontology according to claim 1, characterized in that: Extract verbs that have syntactic dependencies with the vocabulary in the financial report notes, including: Search the vocabulary in the concept dictionary to determine whether the concept dependency syntactic relationship is a pre-set syntactic dependency relationship. If one of the relationships exists, extract the attribute words corresponding to the relationship.

6. The domain ontology-based XBRL financial information element semantic alignment method according to claim 1, characterized in that: Establish alignment between the financial information elements of the XBRL taxonomy and the financial reporting concept ontology, specifically including: The financial information element is used as the tag name in the XBRL link library, the financial reporting domain ontology is stored in OWL language, and the financial information element exists as an instance of the financial reporting domain ontology. The relationship between the financial information element and each concept in the domain ontology through word segmentation is an instance relationship.

7. The method for semantic alignment of XBRL financial information elements based on domain ontology according to claim 1, characterized in that: Calculate the similarity between the concept words corresponding to the financial information elements, including: Concept similarity is expressed as a triple C = {R, Dep, Den}, where R includes the path distance similarity, relationship similarity, and attribute similarity between concepts, Dep represents the concept distance similarity, and Den represents the concept density similarity. The specific calculation formula is expressed as follows: α+β+μ+ν+η=1 Among them, RP sim Represents the path distance similarity between concepts, RP represents the path distance between two concepts in the domain ontology, RR sim Represents the similarity between concepts, RA sim Indicates the attribute similarity between concepts, Dep sim Indicates the similarity between concepts, Den sim represents the density similarity between concepts, Indicates the similarity between concepts. Represents the concepts s and t corresponding to the two terms i and j, θ is an adjustable parameter greater than 0, and α, β, μ, v, and η represent the weight values ​​corresponding to each similarity.

8. The domain ontology-based XBRL financial information element semantic alignment method according to claim 7, characterized in that: Calculate and align the semantic similarity of information elements based on fuzzy semantic similarity, specifically including: Constructing financial information elements and Represents the concept words corresponding to the last position of two financial information elements, FRT i and FRT j represents two financial information elements, and n represents the number of concepts of financial information elements; The semantic alignment formula of financial information elements is specifically expressed as: Among them, RP represents the path distance between two concepts in the domain ontology. The superscript t indicates other concepts in the financial information element except the last position concept. The superscript 0 refers to the concept at the end position, n ≥ 1 means that concepts also exist at other positions except the end position, and λ and ∈ are adjustable parameters greater than 0.

9. A domain ontology-based XBRL financial information element semantic alignment system, characterized by: include: Candidate domain concept extraction module, domain dictionary construction module, corpus construction module, stop word vocabulary construction module, corpus expansion module, vocabulary feature vector construction module, hierarchical relationship extraction module, non-classification relationship extraction module, alignment construction module, semantic alignment module; The candidate domain concept extraction module is used to extract words with explanations in the accounting dictionary as candidate domain concepts; The domain dictionary building module is used to store one concept word per line to build a domain dictionary; The corpus building module is used to obtain the financial report notes text as the corpus to be processed; The stop word vocabulary building module is used to perform word segmentation, part-of-speech tagging and named entity recognition on the corpus, obtain all part-of-speech tagging results, obtain multiple stop words through preset screening rules, and store the multiple stop words as a stop word vocabulary in the financial reporting field; The corpus expansion module is used to expand the corpus based on the LDA model, expand the semantics of the financial report notes text based on the expanded corpus, remove stop words from the text based on a stop word list, and use the expanded financial report notes text as a new corpus; The vocabulary feature vector construction module is used to convert the vocabulary in the new corpus into vocabulary feature vectors; The hierarchical relationship extraction module is used to extract hierarchical relationships, specifically including: The final number of clusters is determined based on the change in the minimum value of inter-class differences. The maximum number of concepts in a class is used as the criterion for stopping iteration. When the number of concepts within a class is less than the criterion for stopping iteration, the iteration is stopped. The hierarchical classification of the domain ontology is obtained through iterative clustering. The similarity between concepts within each class is calculated based on the cosine similarity algorithm, and the concept with the largest average similarity between the concept within the class and other concepts is used as the class label; The non-classification relationship extraction module is used to extract non-classification relationships, specifically including: Perform word segmentation and part-of-speech tagging on the domain dictionary and the stop word list in the financial reporting field, and perform dependency syntax analysis to extract the current concept word and another concept word that has a corresponding relationship with the current concept word. Extract the verbs between the concept pairs as object attribute relationships; Extract verbs that have syntactic dependencies with the vocabulary in the financial report notes text; The alignment construction module is used to establish an alignment between the financial information elements of the XBRL taxonomy and the financial reporting concept ontology; The semantic alignment module is used to perform semantic alignment of financial information elements, calculate the similarity between concept words corresponding to financial information elements, and calculate and align the semantic similarity of information elements based on fuzzy semantic similarity.

10. A computer device comprising a processor and a memory for storing a program executable by the processor, characterized in that: When the processor executes the program stored in the memory, the domain ontology-based XBRL financial information element semantic alignment method according to any one of claims 1 to 8 is implemented.

Images