Automatic recognition method of multilingual accounting terms based on neural network

Abstract

The present invention discloses a multilingual accounting term automatic recognition method based on a neural network, which belongs to the field of semantic recognition technology. The present invention ensures a comprehensive understanding of each word in a specific context by constructing a phrase set of a left neighborhood, a right neighborhood and a full neighborhood, shortens the scope of the considered text, and realizes three combinations of the word at the same time, thereby improving the accuracy of semantic recognition. The allocation of weight coefficients enables the model to recognize the influence of different parts of speech, thereby more accurately capturing the characteristics of professional terms. The semantic features are extracted from the phrase vector by a semantic extraction unit, and by applying weights to the features, the accuracy of semantic recognition is improved, and the ability to judge accounting terms is enhanced. Then, the left neighborhood feature vector, the right neighborhood feature vector and the full neighborhood feature vector are matched in pairs to improve the accuracy of accounting term score prediction, thereby improving the accuracy of accounting term recognition.

Description

Technical Field

[0001] The present invention relates to the technical field of semantic recognition, and in particular to a multilingual accounting term automatic recognition method based on neural network. Background Art

[0002] In a globalized economy, cross-language communication and understanding of accounting information are becoming increasingly important. As a core element of financial reporting and description of economic activities, accurate recognition of accounting terms is crucial for international business, cross-border investment, and financial analysis. However, accounting terms are often highly professional and precise in meaning, and may have subtle differences in different languages, which poses a huge challenge to automatic recognition. Relying solely on the characteristics of the word itself is often not enough to accurately identify accounting terms, especially when dealing with polysemous or context-sensitive terms. In the accounting context, whether a word constitutes a professional term often depends on the words around it. Although existing machine learning methods take into account the overall semantics of the text, their scope of consideration is often too lengthy, resulting in low semantic recognition accuracy. Summary of the invention

[0003] In view of the above-mentioned deficiencies in the prior art, the present invention provides a neural network-based multilingual accounting terminology automatic recognition method that solves the problem of low semantic recognition accuracy in the prior art.

[0004] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is: a method for automatic recognition of multilingual accounting terms based on neural network, comprising the following steps:

[0005] S1. read the multilingual accounting text to be analyzed, and construct three phrase sets with each word as the center, wherein the three phrase sets specifically include: a phrase set of the left neighborhood, a phrase set of the right neighborhood, and a phrase set of the entire neighborhood;

[0006] S2, assigning weight coefficients to various parts of speech;

[0007] S3, encode each phrase set to obtain a phrase vector, and construct a word importance vector based on the weight coefficient of the part of speech;

[0008] S4, using a semantic extraction unit to extract semantic features from the phrase vector, and using the corresponding word importance vector to apply a weight coefficient to the semantic feature, to obtain a left neighborhood feature vector, a right neighborhood feature vector and a full neighborhood feature vector respectively;

[0009] S5, matching the left neighborhood feature vector, the right neighborhood feature vector and the whole neighborhood feature vector in pairs, and using a multi-path scoring unit to predict the matched vectors to obtain the accounting term score of the word when the word is taken as the center;

[0010] S6. Output the accounting terms according to the accounting term score of each word.

[0011] Furthermore, S1 is specifically:

[0012] S11, read the multilingual accounting text to be analyzed, focusing on each word;

[0013] S12, the center word and the left neighbor words form a left neighbor word set;

[0014] S13, the center word and the right neighbor words form a right neighbor word set;

[0015] S14. The center word and the left and right neighbor words form a phrase set of the entire neighborhood, wherein the number of words included in the phrase set of the left neighborhood, the phrase set of the right neighborhood, and the phrase set of the entire neighborhood are the same.

[0016] Furthermore, S3 includes the following sub-steps:

[0017] S31, encode each phrase set to obtain a phrase vector, wherein the phrase vector is composed of word encoding sub-vectors of multiple words;

[0018] S32, constructing a part-of-speech sub-vector with the same length as the word encoding sub-vector for each word encoding sub-vector, wherein each element in the part-of-speech sub-vector is filled with the weight coefficient of the part-of-speech of the word;

[0019] S33. Concatenate the part-of-speech sub-vectors corresponding to a phrase vector into a word importance vector.

[0020] Furthermore, the semantic extraction units in S4 include: a left-neighborhood phrase semantic extraction unit, a right-neighborhood phrase semantic extraction unit and a full-neighborhood phrase semantic extraction unit; the left-neighborhood phrase semantic extraction unit is used to extract semantic features from the phrase vector of the left neighborhood, and the word importance vector of the left neighborhood is used to apply a weight coefficient to the semantic features to obtain a left-neighborhood feature vector; the right-neighborhood phrase semantic extraction unit is used to extract semantic features from the phrase vector of the right neighborhood, and the word importance vector of the right neighborhood is used to apply a weight coefficient to the semantic features to obtain a right-neighborhood feature vector; the full-neighborhood phrase semantic extraction unit is used to extract semantic features from the phrase vector of the full neighborhood, and the word importance vector of the full neighborhood is used to apply a weight coefficient to the semantic features to obtain a full-neighborhood feature vector.

[0021] Further, the left neighborhood phrase semantic extraction unit, the right neighborhood phrase semantic extraction unit and the full neighborhood phrase semantic extraction unit all include: a plurality of LSTM layers, a Concat layer, a weight application layer and a BiLSTM layer;

[0022] Each Cell unit in each LSTM layer processes a coding value in a word coding sub-vector, and the input end of each LSTM layer is used to input a word coding sub-vector, and the word coding sub-vectors input by multiple LSTM layers constitute a phrase vector; the input end of the Concat layer is respectively connected to the output ends of multiple LSTM layers, and its output end is connected to the input end of the weight application layer; the input end of the BiLSTM layer is connected to the output end of the weight application layer, and its output end serves as the output end of the left neighborhood phrase semantic extraction unit, the right neighborhood phrase semantic extraction unit or the full neighborhood phrase semantic extraction unit.

[0023] Furthermore, the expression of the weight application layer is: , where R w is the output sequence of the weight application layer, R is the output sequence of the Concat layer, W is the word importance vector, is element-wise multiplication.

[0024] Further, the left neighborhood feature vector, the right neighborhood feature vector and the whole neighborhood feature vector are matched in pairs in S5 as follows: the left neighborhood feature vector is matched with the whole neighborhood feature vector to obtain a first matching vector; the left neighborhood feature vector is matched with the right neighborhood feature vector to obtain a second matching vector; the whole neighborhood feature vector is matched with the right neighborhood feature vector to obtain a third matching vector;

[0025] The multi-path scoring unit in S5 includes: a first Sigmoid layer, a second Sigmoid layer, a third Sigmoid layer and an output layer;

[0026] The input end of the first Sigmoid layer is used to input and process the first matching vector; the input end of the second Sigmoid layer is used to input and process the second matching vector; the input end of the third Sigmoid layer is used to input and process the third matching vector; the input end of the output layer is respectively connected to the output end of the first Sigmoid layer, the output end of the second Sigmoid layer and the output end of the third Sigmoid layer, and its output end serves as the output end of the multi-path scoring unit.

[0027] Furthermore, the expression of the first Sigmoid layer is: , the expression of the second Sigmoid layer is: , the expression of the third Sigmoid layer is: , where h1 is the output of the first Sigmoid layer, x 1,n is the nth element in the first matching vector, ω 1,n For x 1,n The weight of b 1,n For x 1,n The bias of h2 is the output of the second Sigmoid layer, x2,n is the nth element in the second matching vector, ω 2,n For x 2,n The weight of b 2,n For x 2,n The bias of h3 is the output of the third Sigmoid layer, x 3,n is the nth element in the third matching vector, ω 3,n For x 3,n The weight of b 3,n For x 3,n The bias of , e is a natural constant, N is the length of the matching vector, and n is the number of the element;

[0028] The expression of the output layer is: , where y is the accounting term score of the word at the center of the output layer.

[0029] Further, S6 includes the following sub-steps:

[0030] S61, traversing each word in the multilingual accounting text to be analyzed, marking the words with accounting terminology scores higher than a threshold, and obtaining candidate words;

[0031] S62, searching the left neighborhood of the current candidate word to obtain a left potential word set;

[0032] S63, searching the right neighborhood of the current candidate word to obtain a right potential word set;

[0033] S64, concatenating the left potential word set, the current candidate word, and the right potential word set to obtain a concatenated set. When the number of words in the concatenated set is greater than or equal to 2, the concatenated set is a valid set;

[0034] S65. Extract a continuous word sequence with the same part of speech from the valid set to obtain an accounting term.

[0035] Further, S62 includes the following sub-steps:

[0036] S621, set the iteration parameter m, where m is a positive integer and the initial value of m is 1;

[0037] S622, determine whether there is a candidate word at the mth position of the left neighborhood of the current candidate word, if so, proceed to step S623, if not, the left neighborhood traversal ends and proceeds to step S624;

[0038] S623, mark the candidate word at the mth position as a left potential word, increase the value of m by 1, and jump to step S622;

[0039] S624, all left potential words form a left potential word set, wherein when there is no left potential word, the left potential word set is an empty set;

[0040] S63 includes the following sub-steps:

[0041] S631, set iteration parameter k, k is a positive integer, and the initial value of k is 1;

[0042] S632, determine whether there is a candidate word at the k-th position of the right neighborhood of the current candidate word, if so, proceed to step S633, if not, the right neighborhood traversal ends and proceeds to step S634;

[0043] S633, mark the candidate word at the kth position as a right potential word, increase the value of k by 1, and jump to step S632;

[0044] S634: All right-side potential words are combined to form a right-side potential word set, wherein when there is no right-side potential word, the right-side potential word set is an empty set.

[0045] In summary, the beneficial effects of the present invention are as follows: by constructing a phrase set of the left neighborhood, the right neighborhood, and the entire neighborhood, the present invention ensures a comprehensive understanding of each word in a specific context, shortens the scope of the considered text, and simultaneously realizes three combinations of the word, thereby improving the accuracy of semantic recognition. The allocation of weight coefficients enables the model to recognize the influence of different parts of speech, thereby more accurately capturing the characteristics of professional terms. The semantic features are extracted from the phrase vector by the semantic extraction unit, and by applying weights to the features, the accuracy of semantic recognition is improved, and the ability to judge accounting terms is enhanced. The left neighborhood feature vector, the right neighborhood feature vector, and the entire neighborhood feature vector are then matched in pairs to improve the accuracy of accounting term score prediction, thereby improving the accuracy of accounting term recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 A flowchart of a method for automatic recognition of multilingual accounting terms based on neural networks;

[0047] Figure 2 It is a schematic diagram of the structure of a left neighborhood phrase semantic extraction unit, a right neighborhood phrase semantic extraction unit and a full neighborhood phrase semantic extraction unit;

[0048] Figure 3 It is a connection relationship diagram of the left neighborhood phrase semantic extraction unit, the right neighborhood phrase semantic extraction unit, the full neighborhood phrase semantic extraction unit and the multi-path scoring unit. DETAILED DESCRIPTION

[0049] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0050] like Figure 1 As shown, the neural network-based multilingual accounting terminology automatic recognition method includes the following steps:

[0051] S1. read the multilingual accounting text to be analyzed, and construct three phrase sets with each word as the center, wherein the three phrase sets specifically include: a phrase set of the left neighborhood, a phrase set of the right neighborhood, and a phrase set of the entire neighborhood;

[0052] S2, assigning weight coefficients to various parts of speech;

[0053] S3, encode each phrase set to obtain a phrase vector, and construct a word importance vector based on the weight coefficient of the part of speech;

[0054] S4, using a semantic extraction unit to extract semantic features from the phrase vector, and using the corresponding word importance vector to apply a weight coefficient to the semantic feature, to obtain a left neighborhood feature vector, a right neighborhood feature vector and a full neighborhood feature vector respectively;

[0055] S5, matching the left neighborhood feature vector, the right neighborhood feature vector and the whole neighborhood feature vector in pairs, and using a multi-path scoring unit to predict the matched vectors to obtain the accounting term score of the word when the word is taken as the center;

[0056] S6. Output the accounting terms according to the accounting term score of each word.

[0057] In accounting terms, words of different parts of speech contribute differently to the formation and meaning of the terms. Generally, nouns and adjectives play a more important role in accounting terms, while function words such as prepositions and conjunctions are relatively less important. By assigning weight coefficients, this difference can be more accurately reflected, thereby improving the accuracy of recognition. At the same time, by introducing part-of-speech weights, the structure and composition of the language can be better understood, making the recognition process more in line with linguistic principles and improving the linguistic sensitivity of the model.

[0058] In this embodiment, the part of speech of each word in the multilingual accounting text can be annotated using NLP tools.

[0059] In this embodiment, the weight coefficients in step S2 are manually set. For example, the weight coefficient of nouns is 1, the weight coefficient of adjectives is 0.8, the weight coefficient of verbs is 0.7, the weight coefficient of adverbs is 0.5, the weight coefficient of prepositions is 0.3, the weight coefficient of conjunctions is 0.2, and the weight coefficient of articles is 0.1.

[0060] In this embodiment, the multilingual accounting text is an accounting text containing two or more languages.

[0061] In this embodiment, S1 is specifically:

[0062] S11, read the multilingual accounting text to be analyzed, focusing on each word;

[0063] S12, the center word and the left neighbor words form a left neighbor word set;

[0064] S13, the center word and the right neighbor words form a right neighbor word set;

[0065] S14. The center word and the left and right neighbor words form a phrase set of the entire neighborhood, wherein the number of words included in the phrase set of the left neighborhood, the phrase set of the right neighborhood, and the phrase set of the entire neighborhood are the same.

[0066] In this embodiment, the length of the phrase set of the left neighborhood, the phrase set of the right neighborhood, and the phrase set of the entire neighborhood is 5. In the phrase set of the left neighborhood, 4 words are selected from the left range and the word at the center to form the phrase set of the left neighborhood; in the phrase set of the right neighborhood, 4 words are selected from the right range and the word at the center to form the phrase set of the right neighborhood; in the phrase set of the entire neighborhood, 2 words are selected from the left range, 2 words are selected from the right range, and the word at the center to form the phrase set of the entire neighborhood. The length of the phrase set is adjustable, and each word in the phrase set is continuous in the text, that is, the relative position of the word is not changed.

[0067] When there is no neighbor word on the left or right side, the phrase vector is filled with 0.

[0068] The present invention combines the word at the center with the left and right sides to obtain three phrase sets, which can reflect the different semantics of the word at the center.

[0069] In this embodiment, S3 includes the following sub-steps:

[0070] S31, encode each phrase set to obtain a phrase vector, wherein the phrase vector is composed of word encoding sub-vectors of multiple words;

[0071] In computers, a word consists of multiple encoding values, so a word has a word encoding subvector;

[0072] S32, constructing a part-of-speech sub-vector with the same length as the word encoding sub-vector for each word encoding sub-vector, wherein each element in the part-of-speech sub-vector is filled with the weight coefficient of the part-of-speech of the word;

[0073] S33. Concatenate the part-of-speech sub-vectors corresponding to a phrase vector into a word importance vector.

[0074] Through the processing of step S32, the constructed word importance vector is made to have the same length as the phrase vector, so that weight coefficients are applied to the elements corresponding to the feature vector output by the Concat layer, so that the feature values ​​belonging to the same word have the same weight coefficient, and the feature values ​​of words with different parts of speech have different weight coefficients.

[0075] In this embodiment, the semantic extraction unit in S4 includes: a left-neighborhood phrase semantic extraction unit, a right-neighborhood phrase semantic extraction unit and a full-neighborhood phrase semantic extraction unit; the left-neighborhood phrase semantic extraction unit is used to extract semantic features from the phrase vector of the left neighborhood, and the word importance vector of the left neighborhood is used to apply a weight coefficient to the semantic features to obtain a left-neighborhood feature vector; the right-neighborhood phrase semantic extraction unit is used to extract semantic features from the phrase vector of the right neighborhood, and the word importance vector of the right neighborhood is used to apply a weight coefficient to the semantic features to obtain a right-neighborhood feature vector; the full-neighborhood phrase semantic extraction unit is used to extract semantic features from the phrase vector of the full neighborhood, and the word importance vector of the full neighborhood is used to apply a weight coefficient to the semantic features to obtain a full-neighborhood feature vector.

[0076] In the present invention, there are three phrase sets, and therefore, there are three corresponding word importance vectors.

[0077] The present invention captures the contextual information of words in different directions through the left-neighborhood phrase semantic extraction unit, the right-neighborhood phrase semantic extraction unit and the full-neighborhood phrase semantic extraction unit, and can also perform differentiated feature extraction according to the importance of the words, so as to more accurately identify and understand complex accounting terms.

[0078] like Figure 2 As shown, the left neighborhood phrase semantic extraction unit, the right neighborhood phrase semantic extraction unit and the full neighborhood phrase semantic extraction unit all include: multiple LSTM layers, a Concat layer, a weight application layer and a BiLSTM layer;

[0079] Each Cell unit in each LSTM layer processes a coding value in a word coding sub-vector, and the input end of each LSTM layer is used to input a word coding sub-vector, and the word coding sub-vectors input by multiple LSTM layers constitute a phrase vector; the input end of the Concat layer is respectively connected to the output ends of multiple LSTM layers, and its output end is connected to the input end of the weight application layer; the input end of the BiLSTM layer is connected to the output end of the weight application layer, and its output end serves as the output end of the left neighborhood phrase semantic extraction unit, the right neighborhood phrase semantic extraction unit or the full neighborhood phrase semantic extraction unit.

[0080] The present invention uses an LSTM layer to process a word encoding sub-vector, thereby considering the semantic features of a single word, and then splices the outputs of multiple LSTM layers to obtain the semantic features of a phrase vector. A weight application layer is used to apply a weight coefficient to the semantic features of a phrase vector, increase the importance of parts of speech of different words, adjust the attention of words of different parts of speech, and then a BiLSTM layer is used to consider the overall semantic features after applying the weight coefficient.

[0081] In this embodiment, the expression of the weight application layer is: , where R w is the output sequence of the weight application layer, R is the output sequence of the Concat layer, W is the word importance vector, is element-wise multiplication.

[0082] In S32, each element in the part-of-speech sub-vector is filled with the weight coefficient of the part-of-speech of the word, that is, the elements in a part-of-speech sub-vector are the same. Therefore, in the weight application layer, the weight coefficients applied to the feature values ​​of words with different parts of speech are different, and the weight coefficients applied to the feature values ​​of words with the same part of speech are the same.

[0083] like Figure 3 As shown, in S5, the left neighborhood feature vector, the right neighborhood feature vector and the full neighborhood feature vector are matched in pairs as follows: the left neighborhood feature vector is matched with the full neighborhood feature vector to obtain a first matching vector; the left neighborhood feature vector is matched with the right neighborhood feature vector to obtain a second matching vector, and the full neighborhood feature vector is matched with the right neighborhood feature vector to obtain a third matching vector.

[0084] like Figure 3 As shown, the multi-path scoring unit in S5 includes: a first Sigmoid layer, a second Sigmoid layer, a third Sigmoid layer and an output layer;

[0085] The input end of the first Sigmoid layer is used to input and process the first matching vector; the input end of the second Sigmoid layer is used to input and process the second matching vector; the input end of the third Sigmoid layer is used to input and process the third matching vector; the input end of the output layer is respectively connected to the output end of the first Sigmoid layer, the output end of the second Sigmoid layer and the output end of the third Sigmoid layer, and its output end serves as the output end of the multi-path scoring unit.

[0086] In this embodiment, the expression of the first Sigmoid layer is: , the expression of the second Sigmoid layer is: , the expression of the third Sigmoid layer is: , where h1 is the output of the first Sigmoid layer, x 1,n is the nth element in the first matching vector, ω 1,n For x 1,n The weight of b 1,n For x 1,n The bias of h2 is the output of the second Sigmoid layer, x 2,n is the nth element in the second matching vector, ω 2,n For x 2,n The weight of b 2,n For x 2,n The bias of h3 is the output of the third Sigmoid layer, x 3,n is the nth element in the third matching vector, ω 3,n For x 3,n The weight of b 3,n For x 3,n , e is a natural constant, N is the length of the matching vector, and n is the number of the element.

[0087] In this embodiment, the expression of the output layer is: , where y is the accounting term score of the word at the center of the output layer.

[0088] The present invention matches the left neighborhood feature vector, the right neighborhood feature vector and the full neighborhood feature vector in pairs to achieve matching of different semantic features, and then uses a Sigmoid layer to calculate each matching situation. The outputs of the three Sigmoid layers are combined to improve the calculation accuracy of accounting terminology scores. At the same time, the use of three Sigmoid layers can achieve a more complex mapping.

[0089] In this embodiment, the model includes: a left-neighborhood phrase semantic extraction unit, a right-neighborhood phrase semantic extraction unit, a full-neighborhood phrase semantic extraction unit and a multi-path scoring unit. The training samples are: three phrase sets centered on each word in a multilingual accounting text and a word importance vector. The label is the accounting term score of the word. Multiple training samples are input into the model, and the gradient descent method is used to train the various weights and biases in the model until the output accounting term score is the same or similar to the label.

[0090] In this embodiment, S6 includes the following sub-steps:

[0091] S61, traversing each word in the multilingual accounting text to be analyzed, marking the words with accounting terminology scores higher than a threshold, and obtaining candidate words;

[0092] S62, searching the left neighborhood of the current candidate word to obtain a left potential word set;

[0093] S63, searching the right neighborhood of the current candidate word to obtain a right potential word set;

[0094] S64, concatenating the left potential word set, the current candidate word, and the right potential word set to obtain a concatenated set. When the number of words in the concatenated set is greater than or equal to 2, the concatenated set is a valid set;

[0095] S65. Extract a continuous word sequence with the same part of speech from the valid set to obtain an accounting term, wherein the number of words contained in the accounting term is greater than or equal to 2.

[0096] When extracting accounting terms, the present invention not only considers the accounting term score of a single word, but also splices the left and right neighboring words, which can effectively identify complex accounting terms composed of multiple words. It requires that the number of words in the spliced ​​set is greater than or equal to 2, which helps to reduce the situation where a single common word is mistakenly identified as a term. By extracting a continuous sequence of words with the same part of speech, the grammatical consistency and integrity of the identified terms are ensured, avoiding classifying phrases with different parts of speech into one accounting term.

[0097] In this embodiment, S62 includes the following sub-steps:

[0098] S621, set the iteration parameter m, where m is a positive integer and the initial value of m is 1;

[0099] S622, determine whether there is a candidate word at the mth position of the left neighborhood of the current candidate word, if so, proceed to step S623, if not, the left neighborhood traversal ends and proceeds to step S624;

[0100] S623, mark the candidate word at the mth position as a left potential word, increase the value of m by 1, and jump to step S622;

[0101] S624: All left potential words form a left potential word set, wherein when there is no left potential word, the left potential word set is an empty set.

[0102] The mth position in the left neighborhood, that is, the current candidate word is the 0th position, and counts to the left in sequence.

[0103] In this embodiment, S63 includes the following sub-steps:

[0104] S631, set iteration parameter k, k is a positive integer, and the initial value of k is 1;

[0105] S632, determine whether there is a candidate word at the k-th position of the right neighborhood of the current candidate word, if so, proceed to step S633, if not, the right neighborhood traversal ends and proceeds to step S634;

[0106] S633, mark the candidate word at the kth position as a right potential word, increase the value of k by 1, and jump to step S632;

[0107] S634: All right-side potential words are combined to form a right-side potential word set, wherein when there is no right-side potential word, the right-side potential word set is an empty set.

[0108] In the present invention, a search is performed on the left and right sides of the current candidate word to find out candidate words that exist continuously within the left and right ranges, thereby further improving the recognition accuracy of accounting terms.

[0109] The present invention ensures a comprehensive understanding of each word in a specific context by constructing a phrase set of the left neighborhood, the right neighborhood and the entire neighborhood, shortens the scope of the considered text, and realizes three combinations of the word at the same time, improves the accuracy of semantic recognition, and the allocation of weight coefficients enables the model to recognize the influence of different parts of speech, thereby more accurately capturing the characteristics of professional terms. The semantic features are extracted from the phrase vector by the semantic extraction unit, and by applying weights to the features, the accuracy of semantic recognition is improved, and the ability to judge accounting terms is enhanced. Then, the left neighborhood feature vector, the right neighborhood feature vector and the entire neighborhood feature vector are matched in pairs to improve the accuracy of accounting term score prediction, thereby improving the accuracy of accounting term recognition.

[0110] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for automatic recognition of multilingual accounting terms based on neural networks, characterized in that: The following steps are involved: S1. read the multilingual accounting text to be analyzed, and construct three phrase sets with each word as the center, wherein the three phrase sets specifically include: a phrase set of the left neighborhood, a phrase set of the right neighborhood, and a phrase set of the entire neighborhood; S2, assigning weight coefficients to various parts of speech; S3, encode each phrase set to obtain a phrase vector, and construct a word importance vector based on the weight coefficient of the part of speech; S4, using a semantic extraction unit to extract semantic features from the phrase vector, and using the corresponding word importance vector to apply a weight coefficient to the semantic feature, to obtain a left neighborhood feature vector, a right neighborhood feature vector and a full neighborhood feature vector respectively; S5, matching the left neighborhood feature vector, the right neighborhood feature vector and the whole neighborhood feature vector in pairs, and using a multi-path scoring unit to predict the matched vectors to obtain the accounting term score of the word when the word is taken as the center; S6. Output accounting terms according to the accounting term score of each word; The S3 comprises the following sub-steps: S31, encode each phrase set to obtain a phrase vector, wherein the phrase vector is composed of word encoding sub-vectors of multiple words; S32, constructing a part-of-speech sub-vector with the same length as the word encoding sub-vector for each word encoding sub-vector, wherein each element in the part-of-speech sub-vector is filled with the weight coefficient of the part-of-speech of the word; S33, concatenating the part-of-speech sub-vectors corresponding to a phrase vector into a word importance vector; The semantic extraction unit in S4 includes: a left-neighborhood phrase semantic extraction unit, a right-neighborhood phrase semantic extraction unit and a full-neighborhood phrase semantic extraction unit; the left-neighborhood phrase semantic extraction unit is used to extract semantic features from the phrase vector of the left neighborhood, and the word importance vector of the left neighborhood is used to apply a weight coefficient to the semantic features to obtain a left-neighborhood feature vector; the right-neighborhood phrase semantic extraction unit is used to extract semantic features from the phrase vector of the right neighborhood, and the word importance vector of the right neighborhood is used to apply a weight coefficient to the semantic features to obtain a right-neighborhood feature vector; the full-neighborhood phrase semantic extraction unit is used to extract semantic features from the phrase vector of the full neighborhood, and the word importance vector of the full neighborhood is used to apply a weight coefficient to the semantic features to obtain a full-neighborhood feature vector; The left neighborhood phrase semantic extraction unit, the right neighborhood phrase semantic extraction unit and the full neighborhood phrase semantic extraction unit all include: multiple LSTM layers, a Concat layer, a weight application layer and a BiLSTM layer; Each Cell unit in each of the LSTM layers processes a code value in a word coding sub-vector, and the input end of each of the LSTM layers is used to input a word coding sub-vector, and the word coding sub-vectors input by multiple LSTM layers constitute a phrase vector; the input end of the Concat layer is respectively connected to the output ends of multiple LSTM layers, and its output end is connected to the input end of the weight application layer; the input end of the BiLSTM layer is connected to the output end of the weight application layer, and its output end serves as the output end of the left neighborhood phrase semantic extraction unit, the right neighborhood phrase semantic extraction unit or the full neighborhood phrase semantic extraction unit; The expression of the weight application layer is: , where R w is the output sequence of the weight application layer, R is the output sequence of the Concat layer, W is the word importance vector, is element-wise multiplication; The step of performing pairwise matching of the left neighborhood feature vector, the right neighborhood feature vector and the whole neighborhood feature vector in S5 is as follows: matching the left neighborhood feature vector with the whole neighborhood feature vector to obtain a first matching vector; matching the left neighborhood feature vector with the right neighborhood feature vector to obtain a second matching vector; matching the whole neighborhood feature vector with the right neighborhood feature vector to obtain a third matching vector; The multi-path scoring unit in S5 includes: a first Sigmoid layer, a second Sigmoid layer, a third Sigmoid layer and an output layer; The input end of the first Sigmoid layer is used to input and process the first matching vector; the input end of the second Sigmoid layer is used to input and process the second matching vector; the input end of the third Sigmoid layer is used to input and process the third matching vector; the input end of the output layer is respectively connected to the output end of the first Sigmoid layer, the output end of the second Sigmoid layer and the output end of the third Sigmoid layer, and its output end serves as the output end of the multi-path scoring unit.

2. The method for automatic recognition of multilingual accounting terms based on neural network according to claim 1 is characterized in that: The S1 is specifically: S11, read the multilingual accounting text to be analyzed, focusing on each word; S12, the center word and the left neighbor words form a left neighbor word set; S13, the center word and the right neighbor words form a right neighbor word set; S14. The center word and the left and right neighbor words form a phrase set of the entire neighborhood, wherein the number of words included in the phrase set of the left neighborhood, the phrase set of the right neighborhood, and the phrase set of the entire neighborhood are the same.

3. The method for automatic recognition of multilingual accounting terms based on neural network according to claim 1 is characterized in that: The expression of the first Sigmoid layer is: , the expression of the second Sigmoid layer is: , the expression of the third Sigmoid layer is: , where h1 is the output of the first Sigmoid layer, x 1,n is the nth element in the first matching vector, ω 1,n For x 1,n The weight of b 1,n For x 1,n The bias of h2 is the output of the second Sigmoid layer, x 2,n is the nth element in the second matching vector, ω 2,n For x 2,n The weight of b 2,n For x 2,n The bias of h3 is the output of the third Sigmoid layer, x 3,n is the nth element in the third matching vector, ω 3,n For x 3,n The weight of b 3,n For x 3,n The bias of , e is a natural constant, N is the length of the matching vector, and n is the number of the element; The expression of the output layer is: , where y is the accounting term score of the word at the center of the output layer.

4. The method for automatic recognition of multilingual accounting terms based on neural network according to claim 1 is characterized in that: The S6 comprises the following sub-steps: S61, traversing each word in the multilingual accounting text to be analyzed, marking the words with accounting terminology scores higher than a threshold, and obtaining candidate words; S62, searching the left neighborhood of the current candidate word to obtain a left potential word set; S63, searching the right neighborhood of the current candidate word to obtain a right potential word set; S64, concatenating the left potential word set, the current candidate word, and the right potential word set to obtain a concatenated set. When the number of words in the concatenated set is greater than or equal to 2, the concatenated set is a valid set; S65. Extract a continuous word sequence with the same part of speech from the valid set to obtain an accounting term.

5. The method for automatic recognition of multilingual accounting terms based on neural network according to claim 4 is characterized in that: The S62 The following steps are included: S621, set the iteration parameter m, where m is a positive integer and the initial value of m is 1; S622, determine whether there is a candidate word at the mth position of the left neighborhood of the current candidate word, if so, proceed to step S623, if not, the left neighborhood traversal ends and proceeds to step S624; S623, mark the candidate word at the mth position as a left potential word, increase the value of m by 1, and jump to step S622; S624, all left potential words form a left potential word set, wherein when there is no left potential word, the left potential word set is an empty set; The S63 includes the following sub-steps: S631, set iteration parameter k, k is a positive integer, and the initial value of k is 1; S632, determine whether there is a candidate word at the k-th position of the right neighborhood of the current candidate word, if so, proceed to step S633, if not, the right neighborhood traversal ends and proceeds to step S634; S633, mark the candidate word at the kth position as a right potential word, increase the value of k by 1, and jump to step S632; S634: All right-side potential words are combined to form a right-side potential word set, wherein when there is no right-side potential word, the right-side potential word set is an empty set.

Images