Power text information extraction method and device, computer equipment, computer readable storage medium and computer program product

Abstract

The invention relates to an information extraction method and device of a power text, computer equipment, a computer readable storage medium and a computer program product, and relates to the technical field of natural language processing. The method comprises the steps of obtaining font features and character pronunciation features of statements of a power text and vocabulary features in the power field, and fusing the font features and the character pronunciation features to obtain fused features; determining each target span in the statements according to the fusion features; performing feature extraction on each target span to obtain span features of each target span; inputting the span feature corresponding to each target span into a pre-trained joint recognition model to obtain the category of each target span and the association relationship between the target spans; the categories comprise entities, numerical values and measurement units; and obtaining measurable quantity information according to the numerical comparison word in the statement, each target span, the category of each target span and the association relationship between the target spans. By adopting the method, the reliability of the extracted information can be improved.

Description

Technical Field

[0001] The present application relates to the technical field of natural language processing, and particularly to a method, device, computer device, computer-readable storage medium, and computer program product for extracting information from power texts. Background Art

[0002] Measurable Quantitative Information (MQI) describes the connections between entities, quantities (numerical values and measurement units), such as dependency relationships or comparison relationships. Usually, measurable quantitative information is extracted from text through a pre-trained model based on deep learning (such as Bidirectional Encoder Representations from Transformers, BERT).

[0003] However, on the one hand, the professional terms in the power field are complex; on the other hand, the glyph structures of characters (such as Chinese characters) are complex, and there are sometimes polyphonic characters. Therefore, if a pre-trained model based on deep learning is used to extract measurable quantitative information from power texts, it is likely to result in poor reliability of the extracted information. Summary of the Invention

[0004] Based on this, in view of the technical problem of the poor reliability of the above-extracted information, it is necessary to provide a method, device, computer device, computer-readable storage medium, and computer program product for extracting information from power texts that can improve the reliability of the extracted information.

[0005] In a first aspect, the present application provides a method for extracting information from power texts, including:

[0006] Performing feature extraction processing on the sentences of the power text to obtain the glyph features, pronunciation features, and lexical features in the power field of the sentence, and performing fusion processing on the glyph features, pronunciation features, and lexical features of the sentence to obtain the fusion features of the sentence;

[0007] Determining, according to the fusion features, each target span corresponding to the power field in the sentence; the target span is used to represent the part of the sentence that belongs to an entity, a numerical value, or a measurement unit;

[0008] Performing feature extraction processing on each target span to obtain the span features of each target span; the span features are at least used to represent the position, context semantics, and syntactic structure of the target span in the sentence;

[0009] Input the span features corresponding to the respective target spans into a pre-trained joint recognition model to obtain the categories of the respective target spans and the association relationships between the respective target spans; the categories include the entity, the numerical value, and the measurement unit;

[0010] According to the numerical comparison words in the statement, the respective target spans, the categories of the respective target spans, and the association relationships between the respective target spans, obtain the measurable quantity information extracted from the statement.

[0011] In one embodiment, the numerical comparison words in the statement are obtained by the following method:

[0012] Obtain an extended numerical comparison dictionary; the extended numerical comparison dictionary includes a plurality of target preset numerical comparison words;

[0013] Identify, from the statement, a word that matches any one of the target preset numerical comparison words as the numerical comparison word.

[0014] In one embodiment, the extended numerical comparison dictionary is obtained by the following method:

[0015] Obtain an initial numerical comparison dictionary; the initial numerical comparison dictionary includes a plurality of initial preset numerical comparison words;

[0016] Perform an expansion process on the initial numerical comparison dictionary through a masking operation on the plurality of initial preset numerical comparison words to obtain the extended numerical comparison dictionary.

[0017] In one embodiment, the power text is a Chinese power text;

[0018] The feature extraction process for the statement of the power text to obtain the glyph feature, the pronunciation feature, and the vocabulary feature in the power field of the statement includes:

[0019] Perform a feature extraction process on the glyph structure of each character in the statement to obtain the glyph feature of the statement;

[0020] Perform a feature extraction process on the pinyin of each character to obtain the pronunciation feature of the statement; wherein, in the case where the character is a polyphonic character, perform a feature extraction process on the pinyin of the character in combination with the context of the character;

[0021] Identify, from the statement, a word that matches any one of the pre-determined power field vocabulary, and perform feature extraction on the identified word to obtain the vocabulary feature of the statement.

[0022] In one embodiment, the joint recognition model includes a category classifier and a relationship classifier;

[0023] Inputting the span features corresponding to the respective target spans into a pre-trained joint recognition model to obtain the categories of the respective target spans and the association relationships between the respective target spans includes:

[0024] Inputting the span features corresponding to the respective target spans into the category classifier;

[0025] Based on the category classifier and the span features corresponding to the respective target spans, determining the category to which each target span belongs among the entity, the numerical value, and the measurement unit, to obtain the category of each target span;

[0026] Inputting the span features corresponding to the respective target spans and the categories of the respective target spans into the relationship classifier;

[0027] Based on the relationship classifier, the span features corresponding to the respective target spans, and the categories of the respective target spans, determining the association relationship between any entity in the respective target spans and any numerical value in the respective target spans, and determining the association relationship between any numerical value in the respective target spans and any measurement unit in the respective target spans.

[0028] In one embodiment, before performing feature extraction processing on the statements of the power text, it further includes:

[0029] Obtaining an initial power text to be information-extracted;

[0030] Performing preprocessing on the initial power text to obtain the power text; the preprocessing includes at least removing redundant characters processing and format normalization processing;

[0031] Performing text cutting processing on the power text to obtain at least one statement of the power text.

[0032] In a second aspect, the present application further provides an information extraction of a power text, including:

[0033] A fusion feature extraction module, configured to perform feature extraction processing on the statements of the power text to obtain the glyph feature, the pronunciation feature, and the lexical feature in the power field of the statement, and perform fusion processing on the glyph feature, the pronunciation feature, and the lexical feature of the statement to obtain the fusion feature of the statement;

[0034] A target span recognition module, configured to determine, according to the fusion feature, respective target spans corresponding to the power field in the statement; the target span is used to represent the part of the statement that belongs to an entity, a numerical value, or a measurement unit;

[0035] A span feature extraction module, configured to perform feature extraction processing on each of the target spans to obtain span features of each of the target spans; the span features are at least used to characterize the position, context semantics, and syntactic structure of the target span in the statement.

[0036] A category relationship recognition module, configured to input the span features corresponding to each of the target spans into a pre-trained joint recognition model to obtain the categories of each of the target spans and the association relationships between each of the target spans; the categories include the entity, the numerical value, and the measurement unit.

[0037] A target information determination module, configured to obtain measurable quantity information extracted from the statement according to the numerical comparison words in the statement, each of the target spans, the categories of each of the target spans, and the association relationships between each of the target spans.

[0038] In a third aspect, the present application further provides a computer device, including a memory and a processor, where the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0039] Perform feature extraction processing on the statement of the power text to obtain the glyph feature, the pronunciation feature, and the lexical feature in the power field of the statement, and perform fusion processing on the glyph feature, the pronunciation feature, and the lexical feature of the statement to obtain the fusion feature of the statement.

[0040] According to the fusion feature, determine each target span corresponding to the power field in the statement; the target span is used to characterize the part of the statement that belongs to an entity, a numerical value, or a measurement unit.

[0041] Perform feature extraction processing on each of the target spans to obtain span features of each of the target spans; the span features are at least used to characterize the position, context semantics, and syntactic structure of the target span in the statement.

[0042] Input the span features corresponding to each of the target spans into a pre-trained joint recognition model to obtain the categories of each of the target spans and the association relationships between each of the target spans; the categories include the entity, the numerical value, and the measurement unit.

[0043] According to the numerical comparison words in the statement, each of the target spans, the categories of each of the target spans, and the association relationships between each of the target spans, obtain measurable quantity information extracted from the statement.

[0044] Fourthly, the present application further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented:

[0045] Perform feature extraction processing on the statements of the power text to obtain the glyph features, pronunciation features, and vocabulary features of the statements in the power field, and perform fusion processing on the glyph features, pronunciation features, and vocabulary features of the statements to obtain the fusion features of the statements;

[0046] According to the fusion features, determine each target span corresponding to the power field in the statement; the target span is used to represent the part of the statement that belongs to an entity, a numerical value, or a measurement unit;

[0047] Perform feature extraction processing on each target span to obtain the span features of each target span; the span features are at least used to represent the position, context semantics, and syntactic structure of the target span in the statement;

[0048] Input the span features corresponding to each target span into a pre-trained joint recognition model to obtain the categories of each target span and the association relationship between each target span; the categories include the entity, the numerical value, and the measurement unit;

[0049] According to the numerical comparison words in the statement, each target span, the categories of each target span, and the association relationship between each target span, obtain the measurable quantity information extracted from the statement.

[0050] Fifthly, the present application further provides a computer program product, including a computer program. When the computer program is executed by a processor, the following steps are implemented:

[0051] Perform feature extraction processing on the statements of the power text to obtain the glyph features, pronunciation features, and vocabulary features of the statements in the power field, and perform fusion processing on the glyph features, pronunciation features, and vocabulary features of the statements to obtain the fusion features of the statements;

[0052] According to the fusion features, determine each target span corresponding to the power field in the statement; the target span is used to represent the part of the statement that belongs to an entity, a numerical value, or a measurement unit;

[0053] Perform feature extraction processing on each target span to obtain the span features of each target span; the span features are at least used to represent the position, context semantics, and syntactic structure of the target span in the statement;

[0054] Input the span features corresponding to the respective target spans into a jointly trained recognition model obtained in advance to obtain the categories of the respective target spans and the association relationships between the respective target spans; the categories include the entity, the numerical value, and the measurement unit;

[0055] Based on the numerical comparison words in the statement, the respective target spans, the categories of the respective target spans, and the association relationships between the respective target spans, obtain the measurable quantity information extracted from the statement.

[0056] The above-mentioned information extraction method, device, computer equipment, computer-readable storage medium and computer program product for power texts. First, perform feature extraction processing on the sentences of the power texts to obtain the glyph features, pronunciation features and vocabulary features corresponding to the power field of the sentences, and perform fusion processing on the glyph features, pronunciation features and vocabulary features of the sentences to obtain the fusion features of the sentences; then, according to the fusion features, determine each target span corresponding to the power field in the sentences; the target span is used to represent the part of the sentence that belongs to an entity, a numerical value or a measurement unit; then, perform feature extraction processing on each target span to obtain the span features of each target span; the span features are at least used to represent the position, context semantics and syntactic structure of the target span in the sentence; then, input the span features corresponding to each target span into a pre-trained joint recognition model to obtain the category of each target span and the association relationship between each target span; the category includes entities, numerical values and measurement units; finally, according to the numerical comparison words, each target span, the category of each target span and the association relationship between each target span in the sentence, obtain the measurable quantity information extracted from the sentence. In this way, through the fusion of the glyph features, pronunciation features and vocabulary features of the sentence, the feature expression of the sentence can be enriched from multiple aspects such as glyphs, pronunciation and the professionalism of the power field; based on the span, the positions of the entities, numerical values and measurement units to be extracted can be located in the sentence; based on the span features, the feature expression of the target span can be enriched from multiple aspects such as position, context semantics and syntactic structure; through the joint classification model, the category of each target span and the association relationship between each target span can be determined, and then the entities, numerical values and measurement units in the sentence, as well as the association relationship between the entities, numerical values and measurement units can be identified; based on the numerical comparison words and the association relationship between the entities, numerical values, measurement units and entities, numerical values, measurement units identified from the sentence, the measurable quantity information used to describe the entity, quantity (numerical value and measurement unit) and their previous association relationship can be obtained. The information extraction method for power texts based on the above process fully considers the influence of glyphs, pronunciation and power field-specific terms on information extraction through the fusion of glyph features, pronunciation features and vocabulary features, and fully considers the influence of position, context semantics and syntactic structure on information extraction through the application of spans in information extraction. Therefore, the reliability of the extracted information is improved. Brief Description of the Drawings

[0057] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following will briefly introduce the drawings required to be used in the description of the embodiments of the present application or related technologies. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, other related drawings can be obtained based on these drawings without creative efforts.

[0058] Figure 1 It is a schematic flowchart of a method for extracting information from power texts in an embodiment;

[0059] Figure 2 It is a schematic flowchart of the steps for identifying numerical comparison words in a sentence in an embodiment;

[0060] Figure 3 It is a schematic flowchart of the steps for obtaining an extended numerical comparison dictionary in an embodiment;

[0061] Figure 4 It is a schematic flowchart of the steps for performing feature extraction processing on a sentence to obtain the glyph features, pronunciation features, and vocabulary features of the sentence in the power field in an embodiment;

[0062] Figure 5 It is a schematic flowchart of the steps for inputting the span features corresponding to each target span into a pre-trained joint recognition model to obtain the category of each target span and the association relationship between each target span in an embodiment;

[0063] Figure 6 It is a schematic flowchart of a method for extracting information from power texts in another embodiment;

[0064] Figure 7 It is a schematic flowchart of a method for extracting measurable quantity information from power field texts of spans in an embodiment;

[0065] Figure 8 It is a schematic flowchart of a method for extracting measurable quantity information from power field texts of spans in another embodiment;

[0066] Figure 9 It is a structural block diagram of an information extraction device for power texts in an embodiment; [[ID=3)]

[0067] Figure 10 It is an internal structure diagram of a computer device in an embodiment. Detailed implementation manners

[0068] In order to make the objectives, technical solutions, and advantages of the present application clearer, the present application 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 application and are not used to limit the present application.

[0069] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or fully authorized by all parties, and the collection, use, and processing of relevant data need to comply with relevant regulations.

[0070] It should also be noted that the information extraction method, system, device, computer device, storage medium, and computer program product for power texts provided in this application can be applied to the field of power technology to extract measurable quantity information from texts in the power field; it can also be used in other related fields to extract measurable quantity information from texts in other fields.

[0071] In one embodiment, as Figure 1 shown, an information extraction method for power texts is provided. In this embodiment, an example is given where this method is applied to a server. It can be understood that this method can also be applied to a terminal, and can also be applied to a system including a server and a terminal, and is implemented through the interaction between the server and the terminal; among them, the server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services; the terminal can be, but is not limited to, various personal computers, laptop computers, smart phones, tablet computers, Internet of Things devices, and portable wearable devices, etc. In this embodiment, the method includes the following steps S102 to step S110:

[0072] Step S102, perform feature extraction processing on the sentences of the power text to obtain the glyph features, phonetic features, and lexical features in the power field of the sentences, and perform fusion processing on the glyph features, phonetic features, and lexical features of the sentences to obtain the fusion features of the sentences.

[0073] Among them, the power text is a text in the power field, including but not limited to technical standards, operation logs and device manuals of power equipment, industry reports, academic papers, news reports, patent documents, etc.

[0074] In specific applications, the server performs text cutting processing on the power text to obtain a sentence sequence of the power text; the sentence sequence includes multiple sentences; for each sentence in the sentence sequence, the server performs information extraction on the sentence and extracts the corresponding measurable quantity information from the sentence. Further, the server can perform information extraction on each sentence in sequence, or can perform information extraction on multiple sentences simultaneously.

[0075] Among them, the measurable quantity information MQI describes the entities, quantities (numerical values and measurement units), and the connection relationships between them, such as dependency relationships or comparison relationships.

[0076] Among them, the glyph feature is used to represent the glyph structure of each character in the sentence.

[0077] Among them, the phonetic feature is used to represent the pronunciation of each character in the sentence.

[0078] Among them, lexical features are used to represent the words in the sentence that are highly relevant to the power field.

[0079] Specifically, for each sentence, the server performs glyph feature extraction processing, phonetic feature extraction processing, and lexical feature extraction processing on the sentence respectively, obtains the glyph features, phonetic features, and lexical features of the sentence, and fuses the above-mentioned glyph features, phonetic features, and lexical features to obtain the fused features of the sentence.

[0080] In one embodiment, the server can fuse the glyph features, phonetic features, and lexical features based on weighted fusion.

[0081] In another embodiment, the glyph features, phonetic features, and lexical features are all vector features; the server splices the glyph features, phonetic features, and lexical features to obtain the fused features.

[0082] Step S104: According to the fused features, determine each target span corresponding to the power field in the sentence.

[0083] Among them, in the field of information extraction technology, a span refers to a continuous segment of characters or words in the text.

[0084] Among them, the target span is used to represent the part of the sentence that belongs to an entity, a numerical value, or a measurement unit. For example, in the sentence "The device current shall not exceed 30A", the target span includes the entity "device current", the numerical value "30", and the measurement unit "A".

[0085] Specifically, for each sentence, the server determines, according to the fused features of the sentence, the part of the sentence that may belong to the entity, numerical value, or measurement unit corresponding to the power field, that is, the part that may be to be extracted, so as to obtain each target span in the sentence.

[0086] It should be noted that in step S104, for the target span, the server only knows that it may be the part to be extracted, but does not know whether it is an entity, a numerical value, or a measurement unit specifically.

[0087] Step S106: Perform feature extraction processing on each target span to obtain the span features of each target span.

[0088] Among them, the span features are at least used to represent the position, context semantics, and syntactic structure of the target span in the sentence. Further, the span features at least include the position feature, context semantic feature, and syntactic structure feature of the target span; the position feature is used to represent the position of the target span in the sentence, the context semantic feature is used to represent the context semantics of the target span in the sentence, and the syntactic structure is used to represent the syntactic structure of the target span in the sentence.

[0089] Specifically, the server extracts the span features of each target span through max pooling operation.

[0090] It is easy to understand that the span features can help the server understand which continuous word sequences in the sentence constitute a complete object. Based on the span features, the server can process nested, overlapping, and long-length information more flexibly, significantly improving the information extraction performance in complex scenarios.

[0091] Step S108: Input the span features corresponding to each target span into the pre-trained joint recognition model to obtain the category of each target span and the association relationship between each target span.

[0092] Among them, the categories include entities, numerical values, and measurement units.

[0093] Specifically, for each sentence, the server inputs the span features of each target span of the sentence into the pre-trained joint recognition model, and through the processing of the span features by the joint recognition model, it identifies the category of each target span, that is, determines which category among entity, numerical value, and measurement unit each target span belongs to, and also identifies the association relationship between each target span, that is, determines whether there is an association relationship between two target spans.

[0094] Among them, there is an association relationship between two target spans, which means that the first target span among the two target spans is used to describe the second target span among the two target spans; the first target span is any one of the two target spans, and the second target span is the target span other than the first target span among the two target spans. It is easy to understand that in practical applications, there may be multiple entities, multiple numerical values, and multiple measurement units in a sentence. Therefore, the server needs to determine which numerical values are used to describe which entities, and which measurement units are used to describe which numerical values.

[0095] Furthermore, after the server identifies the association relationship between each target span, it can also record the entities and numerical values with an association relationship among each target span as the first information pair (entity - numerical value), and record the numerical values and measurement units with an association relationship among each target span as the second information pair (numerical value - measurement unit).

[0096] Step S110: Obtain the measurable quantity information extracted from the sentence according to the numerical comparison words in the sentence, each target span, the category of each target span, and the association relationship between each target span.

[0097] Among them, numerical comparison words are words used to represent numerical comparison relationships. For example, greater than, less than, equal to, greater than or equal to, less than or equal to, not exceeding, etc. For example, in the statement "The device current shall not exceed 30 A", the numerical comparison word is "shall not exceed". In specific applications, there is at least one numerical comparison word.

[0098] Specifically, for each statement, the server identifies, from the statement, the words that hit the extended numerical comparison dictionary as numerical comparison words. Since numerical comparison words are usually next to numerical values, for each numerical comparison word, the server can determine which numerical value has an associated relationship with the numerical comparison word based on the distance between the numerical comparison word and each numerical value. Further, the server records the numerical comparison word and the numerical value with an associated relationship as a third information pair (numerical value - numerical comparison word). Then, based on the associated relationships between the target spans, and the associated relationships between each numerical comparison word and each numerical value, the server records the entities, numerical values, measurement units, and numerical comparison relationships with associated relationships in the form of a triple (entity - quantity - numerical comparison relationship) or a quadruple (entity - numerical value - measurement unit - numerical comparison relationship) to obtain the measurable quantity information extracted from the statement; for example, the server integrates each first information pair, second information pair, and third information pair based on the associated relationships, and integrates the mutually associated first information pair, second information pair, and third information into the above triple and quadruple forms.

[0099] Among them, the extended numerical comparison dictionary includes multiple target preset numerical comparison words. In specific applications, hitting the extended numerical comparison dictionary means matching any one of the target preset numerical comparison words in the extended numerical comparison dictionary, for example, the similarity with any one of the target numerical comparison words is greater than or equal to the first preset similarity threshold.

[0100] For example, for the statement "The device current shall not exceed 30 A", the server outputs the following measurable quantity information: Entity: device current; Numerical value: 30; Measurement unit: A; Numerical comparison relationship: shall not exceed.

[0101] In the above information extraction method for power texts, first, the server performs feature extraction processing on the sentences of the power text to obtain the glyph features, pronunciation features, and vocabulary features corresponding to the power field of the sentences, and performs fusion processing on the glyph features, pronunciation features, and vocabulary features of the sentences to obtain the fusion features of the sentences; then, the server determines each target span corresponding to the power field in the sentence according to the fusion features; the target span is used to represent the part of the sentence that belongs to an entity, a numerical value, or a measurement unit; then, the server performs feature extraction processing on each target span to obtain the span features of each target span; the span features are at least used to represent the position, context semantics, and syntactic structure of the target span in the sentence; then, the server inputs the span features corresponding to each target span into a pre-trained joint recognition model to obtain the category of each target span and the association relationship between each target span; the category includes entities, numerical values, and measurement units; finally, the server obtains the measurable quantity information extracted from the sentence according to the numerical comparison words, each target span, the category of each target span, and the association relationship between each target span in the sentence. In this way, through the fusion of the glyph features, pronunciation features, and vocabulary features of the sentence, the feature expression of the sentence can be enriched from multiple aspects such as glyph, pronunciation, and the professionalism of the power field; based on the span, the positions of the entities, numerical values, and measurement units to be extracted can be located in the sentence; based on the span features, the feature expression of the target span can be enriched from multiple aspects such as position, context semantics, and syntactic structure; through the joint classification model, the category of each target span and the association relationship between each target span can be determined, and then the entities, numerical values, and measurement units in the sentence, as well as the association relationship between the entities, numerical values, and measurement units, can be recognized; based on the numerical comparison words and the association relationship between the entities, numerical values, measurement units, and entities, numerical values, measurement units recognized from the sentence, the measurable quantity information used to describe the entities, quantities (numerical values and measurement units), and their previous association relationship can be obtained. The information extraction method for power texts based on the above process fully considers the influence of glyphs, pronunciation, and power-related professional terms on information extraction through the fusion of glyph features, pronunciation features, and vocabulary features, and fully considers the influence of position, context semantics, and syntactic structure on information extraction through the application of spans in information extraction. Therefore, the reliability of the extracted information is improved.

[0102] In an exemplary embodiment, as Figure 2 shown, the numerical comparison words in the sentence are obtained by the following method:

[0103] Step S202, obtain an extended numerical comparison dictionary.

[0104] Step S204, identify the words that match any one of the target preset numerical comparison words from the sentence as the numerical comparison words.

[0105] Among them, the expanded numerical comparison dictionary includes multiple target preset numerical comparison words.

[0106] Specifically, the server obtains the expanded numerical comparison words and determines the similarity between each word in the sentence and each target preset numerical comparison word; for each word, if the similarity between the word and any one of the target preset numerical comparison words is greater than the first preset similarity, the server can consider that the word matches the any one of the target preset numerical comparison words.

[0107] In this embodiment, based on the expanded numerical comparison dictionary, the server can identify numerical comparison words from the sentence.

[0108] In an exemplary embodiment, as Figure 3 shown, the expanded numerical comparison dictionary is obtained through the following method:

[0109] Step S302, obtain the initial numerical comparison dictionary.

[0110] Step S304, through the masking operation on multiple initial preset numerical comparison words, perform an expansion process on the initial numerical comparison dictionary to obtain the expanded numerical comparison dictionary.

[0111] Among them, the initial numerical comparison dictionary includes multiple initial preset numerical comparison words; in specific applications, the initial preset numerical comparison words are common initial preset numerical comparison words. For example, power personnel can obtain multiple power texts, obtain common numerical comparison words from the multiple power texts, and obtain multiple initial preset numerical comparison words.

[0112] Specifically, the server obtains the initial numerical comparison dictionary; the server combines the masking strategy of BERT to perform a masking operation on multiple initial preset numerical comparison words, and predicts possible numerical comparison words through BERT, thereby expanding the initial numerical comparison dictionary to obtain the expanded numerical comparison dictionary.

[0113] For example, the server performs a masking operation on multiple initial preset numerical comparison words to obtain one or more expanded numerical comparison words corresponding to the multiple initial preset numerical comparison words; then, the server combines each initial preset numerical comparison word and each expanded numerical comparison word to obtain the expanded numerical comparison dictionary.

[0114] In this embodiment, based on the masking strategy of BERT, the server can expand the initial numerical comparison dictionary and obtain a richer expanded numerical comparison dictionary; based on the richer expanded numerical comparison dictionary, the server can more accurately identify numerical comparison words from power texts, avoiding misidentification or omission of numerical comparison words.

[0115] In an exemplary embodiment, the power text is a Chinese power text. In a specific application, the Chinese power text refers to a power text mainly in Chinese.

[0116] As Figure 4 shown, in the above step S102, feature extraction processing is performed on the sentences of the power text to obtain the glyph features, phonetic features, and lexical features in the power field of the sentences, which specifically include the following steps:

[0117] Step S402, perform feature extraction processing on the glyph structure of each character in the sentence to obtain the glyph features of the sentence.

[0118] Step S404, perform feature extraction processing on the pinyin of each character to obtain the phonetic features of the sentence.

[0119] Step S406, identify the words in the sentence that match any one of the pre-determined power field words, and perform feature extraction on the identified words to obtain the lexical features of the sentence.

[0120] Among them, in the case where the character is a polyphonic character, feature extraction processing is performed on the pinyin of the character in combination with the context of the character.

[0121] Specifically, when performing glyph feature extraction, for each character in the sentence, the server encodes the image of the character through a CNN (Convolutional Neural Networks) to capture the semantics of the character structure and obtain the glyph features of the character; then, the server splices the glyph features of each character in the sentence to obtain the glyph features of the sentence.

[0122] When performing phonetic feature extraction, for each character in the sentence, the server encodes the pinyin of the character. For example, the information such as the initials, finals, and tones of the pinyin is converted into a numerical form, and then mapped to a corresponding vector through an embedding layer to obtain the phonetic features of the character; then, the server splices the phonetic features of each character in the sentence to obtain the phonetic features of the sentence. Among them, for polyphonic characters, the server performs feature extraction on the pinyin of the character in combination with the context semantics or phonetic features of the context.

[0123] When performing lexical feature extraction, the server first obtains a dictionary of the power field. There are multiple power field words commonly used in the power field in this dictionary of the power field. Then, the server identifies the words that hit the dictionary of the power field from this sentence, and encodes the words in the power field to obtain the lexical features of the words. Finally, the server concatenates the lexical features of each word that hits the dictionary of the power field in the sentence to obtain the lexical features of the sentence. For example, for each word in the sentence, if the similarity between the word and any power field word is greater than the second preset similarity, the server can consider that the word matches the any power field word, and the word hits the dictionary of the power field.

[0124] In this embodiment, through glyph feature extraction, the server can capture the structural features of Chinese characters; through phonetic feature extraction, the server can capture the pronunciation features of Chinese characters, and for polyphonic characters, an idea of combining context for phonetic feature extraction is proposed to perform more accurate feature extraction for the phenomenon of polyphonic characters; through lexical feature extraction, the server can optimize the feature expression of words in the professional field in combination with the power professional dictionary; based on the above process of feature extraction, it is beneficial to enrich the feature expression of the sentence, thereby improving the reliability of the extracted information.

[0125] In an exemplary embodiment, the joint recognition model includes a category classifier and a relationship classifier.

[0126] In specific applications, the category classifier and the relationship classifier in the joint recognition model are optimized through a joint loss function of binary cross-entropy.

[0127] In specific applications, the joint recognition model introduces an attention mechanism to enhance the model's ability to capture complex associated information.

[0128] As Figure 5 shown, in step S108 above, inputting the span features corresponding to each target span into the pre-trained joint recognition model to obtain the categories of each target span and the association relationships between each target span specifically includes the following steps:

[0129] Step S502, inputting the span features corresponding to each target span into the category classifier.

[0130] Step S504, based on the category classifier and the span features corresponding to each target span, determining the category to which each target span belongs among entities, numerical values, and measurement units to obtain the category of each target span.

[0131] Step S506, inputting the span features corresponding to each target span and the categories of each target span into the relationship classifier.

[0132] Step S508: Based on the relation classifier, the span features corresponding to each target span, and the category of each target span, determine the association relationship between any entity in each target span and any value in each target span, and determine the association relationship between any value in each target span and any measurement unit in each target span.

[0133] Specifically, the server inputs the span features corresponding to each target span into the category classifier. The category classifier determines whether each target span belongs to an entity, a value, or a measurement unit by learning the span features, thereby obtaining the category of each target span.

[0134] Then, the server inputs the span features corresponding to each span and the category of each target span into the relation classifier. The relation classifier determines the association relationship between any entity in each target span and any value in each target span, and determines the association relationship between any value in each target span and any measurement unit in each target span by learning the span features and categories of each target span. For example, determine whether there is an association relationship between any entity in each target span and any value in each target span, and determine whether there is an association relationship between any value in each target span and any measurement unit in each target span.

[0135] In this embodiment, the server can perform category recognition and relation recognition synchronously through the joint recognition model, avoiding error propagation in the traditional pipeline method, thereby improving the reliability of the extracted information.

[0136] In an exemplary embodiment, in the above step S102, before performing feature extraction processing on the statements of the power text, the following specific content is further included: obtaining the initial power text to be information-extracted; preprocessing the initial power text to obtain the power text; performing text cutting processing on the power text to obtain at least one statement of the power text.

[0137] Among them, the preprocessing includes at least redundant character removal processing and format normalization processing; in specific applications, redundant character removal refers to clearing garbled characters, redundant spaces, and line breaks, and format normalization processing refers to normalizing numerical values and their measurement units. For example, "20A" is unified as "20 amperes". In specific applications, garbled characters include special symbols and encoding errors.

[0138] Specifically, the server reads the initial power text to be information-extracted; then, the server first clears the garbled characters, redundant spaces, and line breaks in the power text, and normalizes the numerical values and their measurement units in the initial power text to obtain the power text; next, the server performs text cutting on the power text to obtain at least one sentence of the power text, and the sentences of the power text form the sentence sequence of the power text.

[0139] In this embodiment, through the preprocessing of the initial power text, the server can clear the garbled characters, redundant spaces, and line breaks in the power text, and normalize the numerical values and their measurement units in the initial power text, which can reduce the workload of subsequent data processing.

[0140] In an exemplary embodiment, as Figure 6 shown, another method for information extraction of power text is provided. Taking this method applied to the server as an example for illustration, it includes the following steps:

[0141] Step S602, obtain the initial power text to be information-extracted, and perform preprocessing on the initial power text to obtain the power text.

[0142] Step S604, perform text cutting on the power text to obtain at least one sentence of the power text.

[0143] Step S606, obtain the expanded numerical comparison dictionary, and identify the words that match any one of the target preset numerical comparison words in the expanded numerical comparison dictionary from the sentences as the numerical comparison words.

[0144] Step S608, perform feature extraction on the glyph structure of each character in the sentence to obtain the glyph features of the sentence.

[0145] Step S610, perform feature extraction on the pinyin of each character to obtain the phonetic features of the sentence. In the case where the character is a polyphonic character, perform feature extraction on the pinyin of the character in combination with the context of the character.

[0146] Step S612, identify the words that match any one of the pre-determined power domain words from the sentence, and perform feature extraction on the identified words to obtain the lexical features of the sentence.

[0147] Step S614, perform fusion processing on the glyph features, phonetic features, and lexical features of the sentence to obtain the fusion features of the sentence.

[0148] Step S616, determine each target span corresponding to the power domain in the sentence according to the fusion features.

[0149] Step S618, perform feature extraction on each target span to obtain the span features of each target span.

[0150] Step S620: Input the span features corresponding to each target span into the category classifier in the joint classification model. Based on the category classifier and the span features corresponding to each target span, determine the category to which each target span belongs among entities, numerical values, and measurement units, and obtain the category of each target span.

[0151] Step S622: Input the span features corresponding to each target span and the category of each target span into the relationship classifier in the joint classification model. Based on the relationship classifier, the span features corresponding to each target span, and the category of each target span, determine the association relationship between any entity in each target span and any numerical value in each target span, and determine the association relationship between any numerical value in each target span and any measurement unit in each target span.

[0152] Step S624: Obtain the measurable quantity information extracted from the sentence according to the numerical comparison words, each target span, the category of each target span, and the association relationship between each target span.

[0153] In this embodiment, through the fusion of the glyph features, pronunciation features, and lexical features of the sentence, the feature expression of the sentence can be enriched from multiple aspects such as glyphs, pronunciation, and the professionalism of the power field; based on the span, the positions of the entities, numerical values, and measurement units to be extracted can be located in the sentence; based on the span features, the feature expression of the target span can be enriched from multiple aspects such as position, context semantics, and syntactic structure; through the joint classification model, the category of each target span and the association relationship between each target span can be determined, thereby identifying the entities, numerical values, and measurement units in the sentence, as well as the association relationship between the entities, numerical values, and measurement units; based on the numerical comparison words and the association relationship between the entities, numerical values, measurement units, and between the entities, numerical values, measurement units identified from the sentence, the measurable quantity information used to describe the entities, quantities (numerical values and measurement units), and their previous association relationship can be obtained. The information extraction method for power texts based on the above process fully considers the influence of glyphs, pronunciation, and professional terms in the power field on information extraction through the fusion of glyph features, pronunciation features, and lexical features, and fully considers the influence of position, context semantics, and syntactic structure on information extraction through the application of spans in information extraction. Therefore, the reliability of the extracted information is improved.

[0154] To more clearly illustrate the information extraction method for power texts provided in the embodiments of the present application, the following uses a specific embodiment to specifically describe the information extraction method for power texts. However, it should be understood that the embodiments of the present application are not limited thereto. As Figure 7As shown, in one exemplary embodiment, the present application also provides a method for extracting measurable quantity information from power domain texts based on spans, specifically including the following steps:

[0155] 1. Patent text preprocessing: Clean redundant characters, standardize the format, and generate sentence sequences for power texts.

[0156] 1) Cleaning redundant characters: Remove garbled characters, meaningless spaces, and format noises; further, garbled characters include special symbols and encoding errors.

[0157] 2) Normalizing numerical units: Unify full-width and half-width characters, and standardize the expression of numerical values and units.

[0158] 3) Generating sentence sequences: Split the preprocessed text into independent sentence sequences.

[0159] 2. Identifying comparison relationships: Based on a predefined comparison relationship dictionary and combined with the masking strategy of BERT to expand the dictionary, identify numerical comparison relationships in sentences.

[0160] 3. Vector feature fusion: Extract the glyph features, pinyin features, and domain vocabulary features of sentences, generate multi-dimensional feature vector representations, and perform vector fusion.

[0161] 1) Glyph feature encoding: Encode the Chinese character structure features through a CNN model;

[0162] 2) Pinyin feature encoding: Analyze the semantic information of polyphonic characters and homophonic characters.

[0163] 3) Domain vocabulary embedding: Optimize the vocabulary representation in combination with a professional dictionary in the power domain.

[0164] 4) Vector fusion representation: Perform vector concatenation fusion on the feature vectors to form a new vector.

[0165] 4. Span representation modeling: Based on span representation modeling and max pooling operation, construct a joint representation of nested entities and association relationships.

[0166] 1) Generate span combinations of potential entities, numerical values, and units based on the character sequence.

[0167] 2) Span embedding representation: Extract the semantic information of span boundaries through max pooling operation.

[0168] 5. Joint task optimization: Synchronously optimize the entity-numerical value-unit recognition and its association relationship classification tasks through a joint classifier, and adopt a joint loss function to improve the overall performance of the joint classifier.

[0169] 1) Entity-numerical value-unit recognition classifier: Perform multi-class classification on candidate spans (classified into entities, numerical values, or units);

[0170] 2) Association relationship classifier: Determine the "entity - value" and "value - unit" associations for candidate span pairs;

[0171] 3) Joint loss function: Synchronously optimize the binary cross - entropy loss for the recognition and association tasks.

[0172] 4) The joint classifier introduces an attention mechanism to enhance the model's ability to capture complex association relationships.

[0173] Such as Figure 8 shown is another schematic diagram of this embodiment.

[0174] In this embodiment, through the evaluation of experimental metrics such as recognition accuracy, recognition recall rate, association accuracy, association recall rate, and comprehensive F1 - score, the method provided in this embodiment can achieve efficient and accurate extraction of MQI in power texts by fusing feature vectors and optimizing joint tasks, and has the following advantages: Efficiency: The joint model reduces the computational overhead of task separation; Robustness: The feature vectors enhance the model's adaptability to complex semantics; Scalability: It can be extended to scenarios such as standardized documents and technical specifications.

[0175] It should be understood that although the steps in the flowcharts involved in the above - mentioned embodiments are sequentially shown according to the arrows, these steps do not necessarily need to be executed in the order indicated by the arrows. Unless there is a clear description in this article, the execution of these steps has no strict order limit, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above - mentioned embodiments may include multiple steps or multiple stages. These steps or stages do not necessarily need to be executed at the same moment, but can be executed at different moments. The execution order of these steps or stages does not necessarily need to be sequential, but can be executed alternately or in turn with at least a part of other steps or steps or stages in other steps.

[0176] Based on the same inventive concept, the embodiments of the present application also provide an information extraction device for power texts for implementing the information extraction method for power texts involved above. The implementation solutions for solving problems provided by this device are similar to the implementation solutions recorded in the above - mentioned method. Therefore, the specific limitations in one or more embodiments of the information extraction device for power texts provided below can refer to the limitations for the information extraction method for power texts in the above text, and will not be elaborated here.

[0177] In an exemplary embodiment, such as Figure 9As shown in the figure, an information extraction device for power texts is provided, including: a fusion feature extraction module 902, a target span recognition module 904, a span feature extraction module 906, a category relationship recognition module 908, and a target information determination module 910, where:

[0178] The fusion feature extraction module 902 is configured to perform feature extraction processing on the sentences of the power text to obtain the glyph features, phonetic features, and lexical features in the power field of the sentences, and perform fusion processing on the glyph features, phonetic features, and lexical features of the sentences to obtain the fusion features of the sentences.

[0179] The target span recognition module 904 is configured to determine, according to the fusion features, each target span corresponding to the power field in the sentence; the target span is used to represent the part of the sentence that belongs to an entity, a numerical value, or a measurement unit.

[0180] The span feature extraction module 906 is configured to perform feature extraction processing on each target span to obtain the span features of each target span; the span features are at least used to represent the position, context semantics, and syntactic structure of the target span in the sentence.

[0181] The category relationship recognition module 908 is configured to input the span features corresponding to each target span into a jointly trained recognition model to obtain the categories of each target span and the association relationships between each target span; the categories include entities, numerical values, and measurement units.

[0182] The target information determination module 910 is configured to obtain the measurable quantity information extracted from the sentence according to the numerical comparison words, each target span, the categories of each target span, and the association relationships between each target span in the sentence.

[0183] In an exemplary embodiment, the device further includes a numerical comparison recognition module, configured to obtain an extended numerical comparison dictionary; the extended numerical comparison dictionary includes a plurality of target preset numerical comparison words; and identify, from the sentence, the words that match any one of the target preset numerical comparison words as numerical comparison words.

[0184] In an exemplary embodiment, the device further includes a comparison dictionary extension module, configured to obtain an initial numerical comparison dictionary; the initial numerical comparison dictionary includes a plurality of initial preset numerical comparison words; and perform extension processing on the initial numerical comparison dictionary through a masking operation on the plurality of initial preset numerical comparison words to obtain the extended numerical comparison dictionary.

[0185] In an exemplary embodiment, the power text is a Chinese power text.

[0186] The fusion feature extraction module 902 is further configured to perform feature extraction processing on the glyph structure of each character in the statement to obtain the glyph features of the statement; perform feature extraction processing on the pinyin of each character to obtain the phonetic features of the statement; wherein, in the case where the character is a polyphonic character, the pinyin of the character is subjected to feature extraction processing in combination with the context of the character; identify a vocabulary in the statement that matches any one of the pre-determined power domain vocabularies, and perform feature extraction on the identified vocabulary to obtain the vocabulary features of the statement.

[0187] In one exemplary embodiment, the joint recognition model includes a category classifier and a relationship classifier.

[0188] The category relationship recognition module 908 is further configured to input the span features corresponding to each target span into the category classifier; determine the category to which each target span belongs in terms of entity, numerical value, and measurement unit based on the category classifier and the span features corresponding to each target span, to obtain the category of each target span; input the span features corresponding to each target span and the category of each target span into the relationship classifier; determine the association relationship between any one entity in each target span and any one numerical value in each target span, and determine the association relationship between any one numerical value in each target span and any one measurement unit in each target span based on the relationship classifier, the span features corresponding to each target span, and the category of each target span.

[0189] In one exemplary embodiment, the apparatus further includes a power text processing module, configured to obtain an initial power text to be information-extracted; preprocess the initial power text to obtain a power text; the preprocessing at least includes redundant character removal processing and format normalization processing; perform text cutting processing on the power text to obtain at least one statement of the power text.

[0190] Each module in the above information extraction apparatus for power text can be implemented in whole or in part by software, hardware, and their combination. The above modules can be embedded in the processor in the computer device in the form of hardware or be independent of the processor, or can be stored in the memory in the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above respective modules.

[0191] In one exemplary embodiment, a computer device is provided. The computer device can be a server, and its internal structure diagram can be as Figure 10As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O), and a communication interface. Among them, the processor, the memory, and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store power text data and measurable quantity information extracted from the power text. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, it implements a method for extracting information from power text.

[0192] Those skilled in the art can understand that Figure 10 the structure shown in is only a block diagram of some structures related to the solution of this application, and does not constitute a limitation on the computer device to which the solution of this application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements.

[0193] In an exemplary embodiment, a computer device is further provided, including a memory and a processor. A computer program is stored in the memory, and when the processor executes the computer program, it implements the steps in the above method embodiments.

[0194] In an exemplary embodiment, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by the processor, it implements the steps in the above method embodiments.

[0195] In an exemplary embodiment, a computer program product is provided, including a computer program, and when the computer program is executed by the processor, it implements the steps in the above method embodiments.

[0196] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, database, or other medium used in the embodiments provided in the present application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The databases involved in the embodiments provided in the present application can include at least one of relational databases and non-relational databases. Non-relational databases can include distributed databases based on blockchain, etc., without limitation. The processors involved in the embodiments provided in the present application can be general-purpose processors, central processors, graphics processors, digital signal processors, programmable logic devices, data processing logics based on quantum computing, artificial intelligence (AI) processors, etc., without limitation.

[0197] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered to be within the scope recorded in the present application.

[0198] The above-described embodiments merely represent several implementation manners of the present application. The description thereof is relatively specific and detailed, but it should not be construed as a limitation to the patent scope of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all fall within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the appended claims.

Claims

1. An information extraction method for power texts, characterized in that, The method includes: Performing feature extraction processing on the statements of the power text to obtain the glyph features, pronunciation features, and vocabulary features in the power field of the statements, and performing fusion processing on the glyph features, pronunciation features, and vocabulary features of the statements to obtain the fusion features of the statements; According to the fusion features, determining each target span corresponding to the power field in the statement; the target span is used to represent the part of the statement that belongs to an entity, a numerical value, or a measurement unit; Performing feature extraction processing on each of the target spans to obtain the span features of each of the target spans; the span features are at least used to represent the position, context semantics, and syntactic structure of the target span in the statement; Inputting the span features corresponding to each of the target spans into a pre-trained joint recognition model to obtain the categories of each of the target spans and the association relationships between each of the target spans; the categories include the entity, the numerical value, and the measurement unit; According to the numerical comparison words in the statement, each of the target spans, the categories of each of the target spans, and the association relationships between each of the target spans, obtaining the measurable quantity information extracted from the statement.

2. The method according to claim 1, wherein The numerical comparison words in the statement are obtained by the following method: Obtaining an extended numerical comparison dictionary; the extended numerical comparison dictionary includes a plurality of target preset numerical comparison words; Identifying, from the statement, a word that matches any one of the target preset numerical comparison words as the numerical comparison word.

3. The method according to claim 2, wherein The extended numerical comparison dictionary is obtained by the following method: Obtaining an initial numerical comparison dictionary; the initial numerical comparison dictionary includes a plurality of initial preset numerical comparison words; Performing an expansion process on the initial numerical comparison dictionary through a masking operation on the plurality of initial preset numerical comparison words to obtain the extended numerical comparison dictionary.

4. The method according to claim 1, wherein The power text is a Chinese power text; The performing feature extraction processing on the statements of the power text to obtain the glyph features, pronunciation features, and vocabulary features in the power field of the statements includes: Performing feature extraction processing on the glyph structure of each character in the statement to obtain the glyph features of the statement; Performing feature extraction processing on the pinyin of each character to obtain the pronunciation features of the statement; wherein, when the character is a polyphonic character, performing feature extraction processing on the pinyin of the character in combination with the context of the character; Identifying, from the statement, a word that matches any one of the pre-determined power field vocabulary, and performing feature extraction on the identified word to obtain the vocabulary features of the statement.

5. The method according to claim 1, wherein The joint recognition model includes a category classifier and a relationship classifier; The inputting the span features corresponding to each of the target spans into a pre-trained joint recognition model to obtain the categories of each of the target spans and the association relationships between each of the target spans includes: Inputting the span features corresponding to each of the target spans into the category classifier; Based on the category classifier and the span features corresponding to each target span, determine the category to which each target span belongs among the entity, the numerical value, and the measurement unit, and obtain the category of each target span; Input the span features corresponding to the respective target spans and the categories of the respective target spans into the relationship classifier; Based on the relationship classifier, the span features corresponding to the respective target spans, and the categories of the respective target spans, determine the association relationship between any entity in the respective target spans and any numerical value in the respective target spans, and determine the association relationship between any numerical value in the respective target spans and any measurement unit in the respective target spans.

6. The method according to any one of claims 1 to 5, characterized in that Before performing feature extraction processing on the statements of the power text, it further includes: Obtain the initial power text to be information-extracted; Perform preprocessing on the initial power text to obtain the power text; the preprocessing includes at least redundant character removal processing and format normalization processing; Perform text cutting processing on the power text to obtain at least one statement of the power text.

7. An information extraction device for power texts, characterized in that The device includes: A fusion feature extraction module, configured to perform feature extraction processing on the statements of the power text to obtain the glyph feature, the phonetic feature, and the lexical feature in the power field of the statement, and perform fusion processing on the glyph feature, the phonetic feature, and the lexical feature of the statement to obtain the fusion feature of the statement; A target span recognition module, configured to determine, according to the fusion feature, each target span corresponding to the power field in the statement; the target span is used to represent the part of the statement that belongs to an entity, a numerical value, or a measurement unit; A span feature extraction module, configured to perform feature extraction processing on the respective target spans to obtain the span features of the respective target spans; the span feature is at least used to represent the position, the context semantics, and the syntactic structure of the target span in the statement; A category relationship recognition module, configured to input the span features corresponding to the respective target spans into a pre-trained joint recognition model to obtain the categories of the respective target spans and the association relationships between the respective target spans; the categories include the entity, the numerical value, and the measurement unit; A target information determination module, configured to obtain the measurable quantity information extracted from the statement according to the numerical comparison words in the statement, the respective target spans, the categories of the respective target spans, and the association relationships between the respective target spans.

8. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the method according to any one of claims 1 to 6.

Images