A patent automatic classification method, device, terminal and storage medium

Abstract

The application provides a patent automatic classification method and device, a terminal and a storage medium, and comprises the following steps: performing dependency syntax analysis on a patent document to be classified to obtain pretreatment information; performing word embedding on the pretreatment information to obtain document representation; and obtaining a flow category prediction result, a function category prediction result and a meta-function category prediction result based on the document representation and trained flow classifiers, function classifiers and meta-function classifiers. Through dependency syntax analysis and word embedding on the patent document to be classified, the flow category prediction result, the function category prediction result and the meta-function category prediction result are obtained by using machine learning technology, the comprehensiveness of the classification of the patent document to be classified is realized, and the problem that the patent automatic classification method in the prior art is very limited and cannot comprehensively classify patents is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of document processing, and particularly relates to a patent automatic classification method and device, a terminal and a storage medium. BACKGROUND

[0002] As the product of innovative design, the technical information contained in patent documents can expand the knowledge space of designers, and then inspire them to create new products or technical systems. Automatic classification of patents can effectively mine useful information in patents, thereby promoting the reuse of patent knowledge by designers.

[0003] The existing patent automatic classification methods can be divided into three categories: patent classification methods based on engineering fields, patent classification methods based on the theory of invention problem solving, and patent classification methods based on functions. However, the patent classification based on engineering fields can only support the acquisition of patents in a specific engineering field, the patent classification based on the theory of invention problem solving can only acquire cross-domain patents related to invention principles, and the patent classification based on functions can only acquire cross-domain patents related to functions. Therefore, the existing patent automatic classification methods are very limited and cannot comprehensively classify patents.

[0004] Therefore, the prior art has defects and needs to be improved and developed. SUMMARY

[0005] The technical problem to be solved by the present application is that, in view of the above defects of the prior art, a patent automatic classification method, device, terminal and storage medium are provided, aiming at solving the problem that the existing patent automatic classification method is very limited and cannot comprehensively classify patents.

[0006] The technical solution adopted by the present application to solve the technical problem is as follows:

[0007] A patent automatic classification method, wherein the method comprises:

[0008] Performing dependency syntax analysis on the patent document to be classified to obtain pre-processing information;

[0009] Performing word embedding on the pre-processing information to obtain document representation;

[0010] Obtaining flow class prediction results, function class prediction results and meta-function class prediction results based on the document representation and the trained flow classifier, function classifier and meta-function classifier.

[0011] In an implementation manner, the dependency syntax analysis on the patent document to be classified to obtain pre-processing information comprises:

[0012] Analyze the part of speech of each word in each sentence of the patent document to be classified, and analyze the grammatical relationship between different words to obtain an analysis result corresponding to each sentence;

[0013] According to the analysis result, the noun object, the verb predicate and the verb pair maintaining the dependency relationship of each sentence are extracted to obtain preprocessing information.

[0014] In an implementation manner, the preprocessing information is subjected to word embedding to obtain a document representation, including:

[0015] Each of the noun objects is subjected to semantic vector representation to obtain a noun vector, and an average value of each noun vector is calculated to obtain a noun vector representation of the patent document to be classified;

[0016] Each of the verb predicates is subjected to semantic vector representation to obtain a verb vector, and an average value of each verb vector is calculated to obtain a verb vector representation of the patent document to be classified;

[0017] The two verbs in the verb pair maintaining the dependency relationship are subjected to semantic vector representation to obtain a first verb vector and a second verb vector, the difference between the first verb vector and the second verb vector is taken as a verb relationship vector, and an average value of each of the verb relationship vectors is calculated to obtain a verb relationship vector representation of the patent document to be classified;

[0018] The noun vector representation, the verb vector representation and the verb relationship vector representation are saved as the document representation.

[0019] In an implementation manner, the document representation and the trained flow classifier, function classifier and meta-function classifier are used to obtain flow category prediction results, function category prediction results and meta-function category prediction results, including:

[0020] The noun vector representation in the document representation is input into the trained flow classifier to obtain the flow category prediction results;

[0021] The verb vector representation in the document representation is input into the trained function classifier to obtain the function category prediction results;

[0022] The verb relationship vector representation in the document representation is input into the trained meta-function classifier to obtain the meta-function category prediction results.

[0023] In an implementation manner, before the dependency syntax analysis of the patent document to be classified is performed to obtain the preprocessing information, the method further includes:

[0024] Extracting the title and the abstract of the patent to be classified;

[0025] A period is added at the end of the title, and the title after the period and the abstract form a patent document to be classified.

[0026] In an implementation, the training steps of the flow classifier, the function classifier and the meta-function classifier include:

[0027] A set of training patent documents is obtained, including a plurality of training patent documents and corresponding flow term labels, function term labels and meta-function term labels for each training patent document;

[0028] Each of the training patent documents is subjected to dependency syntax analysis and word embedding processing to obtain training document representations, including noun vector training representations, verb vector training representations and verb relationship vector training representations;

[0029] Based on the noun vector training representations and the flow term labels, an initially constructed initial flow classifier is trained to obtain a trained flow classifier;

[0030] Based on the verb vector training representations and the function term labels, an initially constructed initial function classifier is trained to obtain a trained function classifier;

[0031] Based on the verb relationship vector training representations and the meta-function term labels, an initially constructed initial meta-function classifier is trained to obtain a trained meta-function classifier.

[0032] In an implementation, the dependency syntax analysis and word embedding processing of each of the training patent documents to obtain training document representations include:

[0033] The part of speech of each word in each sentence of each of the training patent documents is analyzed, and the grammatical relationship between different words is analyzed;

[0034] According to the part of speech of each word and the grammatical relationship between different words, the noun object training data, the verb predicate training data and the verb pair training data that maintain the dependency relationship of each sentence are extracted;

[0035] Each of the noun object training data is subjected to semantic vector representation to obtain each noun vector training data, and the average value of each noun vector training data is calculated to obtain the noun vector training representation of the patent document to be classified;

[0036] Each of the verb predicate training data is subjected to semantic vector representation to obtain each verb vector training data, and the average value of each verb vector training data is calculated to obtain the verb vector training representation of the patent document to be classified;

[0037] The two verbs in the training data are semantically represented by a verb pair that maintains a dependency relationship to obtain a third verb vector and a fourth verb vector, and the difference between the third verb vector and the fourth verb vector is taken as verb relationship vector training data, and the average value of each verb relationship vector training data is calculated to obtain the verb relationship vector training representation of the patent document to be classified.

[0038] The application further provides a patent automatic classification device, wherein the device comprises:

[0039] An analysis module is configured to perform dependency syntax analysis on the patent document to be classified to obtain pre-processing information.

[0040] A word embedding module is configured to perform word embedding on the pre-processing information to obtain document representation.

[0041] A classification module is configured to obtain flow category prediction results, function category prediction results and meta-function category prediction results based on the document representation and the trained flow classifier, function classifier and meta-function classifier.

[0042] The application further provides a terminal, which comprises a memory, a processor and a patent automatic classification program stored in the memory and executable on the processor, and the patent automatic classification program is used to implement the steps of the patent automatic classification method.

[0043] The application further provides a computer readable storage medium, which stores a computer program capable of being executed to implement the steps of the patent automatic classification method.

[0044] The patent automatic classification method provided by the application comprises the following steps: performing dependency syntax analysis on a patent document to be classified to obtain pre-processing information; performing word embedding on the pre-processing information to obtain document representation; and obtaining flow category prediction results, function category prediction results and meta-function category prediction results based on the document representation and the trained flow classifier, function classifier and meta-function classifier. The application realizes the comprehensiveness of the classification of the patent document to be classified by performing dependency syntax analysis and word embedding on the patent document to be classified and then obtaining the flow category prediction results, function category prediction results and meta-function category prediction results by using machine learning technology, and solves the problem that the patent automatic classification method in the prior art is very limited and cannot comprehensively classify patents. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 is a flow chart of a preferred embodiment of the patent automatic classification method in the application;

[0046] Figure 2The classification system and design target definition of the automatic patent classification method of the present application;

[0047] Figure 3 The patent automatic classification flow chart based on function and meta-function of the automatic patent classification method of the present application;

[0048] Figure 4 The function principle block diagram of the preferred embodiment of the patent automatic classification device in the present application;

[0049] Figure 5 The function principle block diagram of the preferred embodiment of the terminal in the present application. DETAILED DESCRIPTION

[0050] In order to make the object, technical scheme and advantages of the present application clearer and more explicit, the present application is further described in detail below with reference to the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application.

[0051] The existing patent automatic classification method can be divided into three categories: patent classification method based on engineering field, patent classification method based on invention principle in the theory of invention problem solving (TRIZ) and patent classification method based on function.

[0052] For the patent classification method based on engineering field, the patents are classified according to the engineering / technology field to which the patent description object belongs. The current world-wide classification system is the International Patent Classification (IPC). About 95% of existing patents are classified according to the IPC. The IPC is a hierarchical classification system that covers 8 sections, 128 classes, 648 subclasses, about 7200 groups and about 72000 subgroups, describing the field of the patent. For example, a three-stage classification method combining vector space model and traditional machine learning automatically classifies patent documents into the "subgroup" of IPC. In view of the shortcomings of traditional machine learning methods, some methods integrate word embedding, convolutional neural network, attention mechanism and other deep learning technologies, and consider text order, context and other features, to automatically classify Chinese patents into the "section" of IPC. Some technologies design a patent classification method based on word embedding and convolutional neural network, to automatically classify English patent documents into the "subclass" of IPC. Some automatic classification methods use latent Dirichlet allocation to represent the theme of patent text, as the input of support vector machine, to classify patent documents into the "section", "class", "subclass", "group" and "subgroup" of IPC. Some technologies propose a classification method based on deep learning converter and graph embedding, to realize automatic classification of patents according to IPC code. Some technologies build an integrated machine learning classification framework, to automatically classify patents into the "subclass" of IPC. Some technologies classify patents based on IPC, combining information gain and word frequency features and K nearest neighbor classifier. Some technologies extract and train a keyword dictionary library from the IPC information of patents, and then automatically classify patents according to the keyword group. Some technologies propose a deep learning classification method based on dual-channel feature fusion, to automatically classify patents according to the IPC system. Some technologies construct a technology classification system similar to IPC, and apply a neural network classifier to classify patents. Some technologies fuse phrase and context features, and apply deep learning technologies such as convolutional neural network, long short-term memory network and attention mechanism, to automatically classify patent documents according to IPC. Some technologies combine patent feature maps and neural network models to classify patent documents into the IPC hierarchical system. Some technologies extract financial field classification labels from the IPC information of patents, and then automatically classify patent text through a bidirectional encoder representation from transformers model. Some technologies propose a patent classification method based on network representation learning and hierarchical label embedding, to automatically classify patents according to the CPC (Cooperative Patent Classification) classification system similar to IPC. From the perspective of product innovation, patent classification based on engineering / technology field can only support the acquisition of patents in a specific engineering field, which will limit the divergent thinking of designers.To obtain innovative inspiration, designers need to draw analogies from cross-domain patents, not only from patents in the same field.

[0053] For the patent classification method based on the invention principles in the theory of inventive problem solving (TRIZ), patents are classified according to the invention principles in the theory of inventive problem solving (TRIZ). TRIZ theory has summarized and refined 40 invention principles that can be used to solve technical contradictions from a large number of patents. According to the TRIZ invention principle, the classification of patents can provide example references for abstract invention principles, and can make designers or TRIZ users more efficiently convert invention principles into special solutions. Some technologies redivide the 40 invention principles into 22 groups as classification categories, and automatically classify 674 patents. Some technologies select 10 invention principles from TRIZ as classification categories, and automatically classify 293 patents. Some technologies propose a collaborative filtering method to automatically classify 2868 patents based on TRIZ invention principles. Some technologies propose a multi-feature fusion Chinese patent classification method based on TRIZ invention principles, use ALBERT model, BiATextCNN model, BiGRU model and SelfAttention mechanism to obtain word vectors containing local key character information features and context deep semantic information feature sentence vectors, and fuse to obtain multi-feature representation of text vector representation. Finally, use the Softmax classifier to classify patents according to TRIZ invention principles. From the perspective of supporting product innovation, the patent classification method based on TRIZ invention principles helps designers obtain cross-domain patents to inspire inspiration; however, from the patents classified by this method, only invention principle related cross-domain patents can be obtained, which can inspire designers to solve technical contradictions in products or systems, but it is difficult to effectively inspire designers to creatively realize target functions and target meta-functions in product systems.

[0054] For the function-based patent classification method, the patent is classified according to the function realized by the patent description object. Function is the basic characteristic of product or technology, and the value of the patent is embodied in its originality of realizing a certain function. By classifying the patent by function, it is helpful for designers to obtain various forms of realizing the target function from the patent, thereby generating design ideas. Some technologies propose a method based on a function co-reference model, a vector space model and machine learning for automatic function classification of patents. Some technologies combine function bases, term frequency-inverse document frequency and naive Bayes to automatically classify Chinese patents by function. Some technologies propose a new patent feature weight calculation method for function-based patent classification, which improves the function classification effect of the patent. From the perspective of supporting product innovation, the function-based patent classification method helps designers to obtain cross-domain patent inventions, thereby stimulating design inspiration. However, from the patents classified by this method, only function-related cross-domain patents can be obtained, which can inspire designers to creatively realize the target function in the product system, but it is difficult to effectively inspire designers to creatively realize the target meta-function in the product system.

[0055] The present application provides a kind of patent automatic classification method based on function and meta function, the classification system of the method provided combines meta function, function and flow terminology. Intelligent technology such as dependency syntax analysis, information extraction, word embedding, machine learning is used to realize patent representation and classification. Through the method provided, the function and meta function attributes of patent document can be automatically identified, thereby solving the problem of automatic classification of patents based on function and meta function, and then enabling designers to effectively obtain function and meta function associated cross-domain patent inventions.

[0056] Please see Figure 1 , Figure 1 It is the flow chart of the patent automatic classification method in the present application. As Figure 1 Indicated, the patent automatic classification method described in the embodiment of the present application comprises:

[0057] Step S100, dependency syntax analysis is carried out on the patent document to be classified, and preprocessing information is obtained.

[0058] In the embodiment of the present application, the step S100 specifically comprises:

[0059] Step S110, the part of speech of each word in each sentence of the patent document to be classified is analyzed, and the grammatical relationship between different words is analyzed, and the analysis result corresponding to each sentence is obtained;

[0060] Step S120, according to the analysis result, the noun object, the verb predicate and the verb pair maintaining the dependency relationship of each sentence are extracted, and the preprocessing information is obtained.

[0061] Specifically, in the Python environment, a dependency syntax analysis tool Standford CoreNLP is called to perform dependency syntax analysis on the patent document to be classified. For each sentence in the patent document to be classified, the part-of-speech (such as noun, verb, adjective, adverb, etc.) of each word and the grammatical relationship (such as subject-predicate relationship, verb-object relationship, subject-predicate-object relationship, appositive relationship, etc.) between different words are analyzed. Noun objects and verb predicates are extracted from the analyzed sentences by object and verb extraction rules. For each verb, the directly connected word in the dependency relationship, i.e., the word directly dependent on the verb, is found, and it is determined whether the dependency relationship is also a verb. If so, the pair of verbs is retained.

[0062] The embodiment of the present application provides a basis for realizing the automatic function and meta-function classification of patent text through dependency syntax analysis and information extraction.

[0063] As shown in Figure 1 The patent automatic classification method provided by the embodiment further includes:

[0064] In step S200, the preprocessed information is subjected to word embedding to obtain document representation.

[0065] In the embodiment of the present application, the step S200 specifically includes:

[0066] In step S210, each noun object is subjected to semantic vector representation to obtain a noun vector, and the average value of each noun vector is calculated to obtain the noun vector representation of the patent document to be classified.

[0067] In step S220, each verb predicate is subjected to semantic vector representation to obtain a verb vector, and the average value of each verb vector is calculated to obtain the verb vector representation of the patent document to be classified.

[0068] In step S230, the two verbs in the verb pair maintaining the dependency relationship are subjected to semantic vector representation to obtain a first verb vector and a second verb vector, the difference between the first verb vector and the second verb vector is taken as a verb relationship vector, and the average value of each verb relationship vector is calculated to obtain the verb relationship vector representation of the patent document to be classified.

[0069] In step S240, the noun vector representation, the verb vector representation, and the verb relationship vector representation are saved as document representation.

[0070] Specifically, 300-dimensional semantic vector representations are performed on the extracted nominal objects, verbal predicates, and verb pairs that maintain dependency relationships by a word embedding tool GloVe. The nominal vectors are averaged to obtain a nominal vector representing the document. The verbal vectors are averaged to obtain a verbal vector representing the document. The verb relationship vectors representing the verb pairs that maintain dependency relationships are obtained by differencing the two verbal vectors representing the verb pairs that maintain dependency relationships, and then the verb relationship vectors are averaged to obtain a verb relationship vector representing the document.

[0071] The embodiment of the present application provides a basis for implementing automatic function and meta-function classification of patent texts by performing word embedding on preprocessed information.

[0072] As shown in Figure 1 The patent automatic classification method provided by the embodiment further includes:

[0073] In step S300, a flow category prediction result, a function category prediction result, and a meta-function category prediction result are obtained based on the document representation and the trained flow classifier, function classifier, and meta-function classifier.

[0074] Specifically, in the classifier application, the patent document to be classified is preprocessed and vectorized to represent the document, which is input into the trained flow, function, and meta-function classifiers, so as to implement flow, function, and meta-function term category prediction of the unannotated patent document. The function and flow term prediction results constitute the function classification result of the patent document. The meta-function, function, and flow term prediction results constitute the meta-function classification result of the patent document.

[0075] The embodiment of the present application provides a basis for implementing automatic function and meta-function classification of patent texts by utilizing machine learning technology.

[0076] In the embodiment of the present application, the step S300 specifically includes:

[0077] In step S310, a nominal vector representation in the document representation is input into the trained flow classifier to obtain a flow category prediction result.

[0078] In step S320, a verbal vector representation in the document representation is input into the trained function classifier to obtain a function category prediction result.

[0079] In step S330, a verb relationship vector representation in the document representation is input into the trained meta-function classifier to obtain a meta-function category prediction result.

[0080] The embodiment of the present application improves the comprehensiveness of patent classification by simultaneously predicting the flow category prediction result, the function category prediction result and the meta-function category prediction result of the patent document to be classified, and solves the problem that the existing patent automatic classification method is very limited and cannot comprehensively classify patents.

[0081] In an embodiment of the present application, the step S100 further includes: extracting the title and the abstract of the patent to be classified; adding a period at the end of the title, and the title after adding the period and the abstract form the patent document to be classified.

[0082] Specifically, the embodiment of the present application adds a period at the end of the title and connects the abstract to form a complete text, and the text is the patent document. In this way, the embodiment of the present application can greatly reduce the workload and improve the classification efficiency by extracting the title and the abstract of the patent to be classified.

[0083] In the embodiment of the present application, the training steps of the flow classifier, the function classifier and the meta-function classifier include:

[0084] Step A10, obtaining a training patent document set, the training patent document set including a plurality of training patent documents and corresponding flow term labels, function term labels and meta-function term labels of each training patent document;

[0085] Step A20, performing dependency syntax analysis and word embedding processing on each training patent document to obtain a training document representation, the training document representation including a noun vector training representation, a verb vector training representation and a verb relationship vector training representation;

[0086] Step A30, training the pre-constructed initial flow classifier based on the noun vector training representation and the flow term label to obtain a trained flow classifier;

[0087] Step A40, training the pre-constructed initial function classifier based on the verb vector training representation and the function term label to obtain a trained function classifier;

[0088] Step A50, training the pre-constructed initial meta-function classifier based on the verb relationship vector training representation and the meta-function term label to obtain a trained meta-function classifier.

[0089] Specifically, the classification system in the embodiments of the present application combines general meta-function terms, function terms and flow terms. The function terms and flow terms constitute function concepts. The meta-function terms, function terms and flow terms constitute meta-function concepts. The above three types of terms are applied by design experts to label patent documents collected from patent websites to obtain a labeled patent document set. The patent document set labeled by each type of term is divided into a training patent document set and a test patent document set. The training documents and test documents are preprocessed based on dependency syntax analysis and information extraction. The extracted information is word embedding to convert the training documents and test documents into vector form. Based on the training document vectors and machine learning classifiers, meta-function classifiers, function classifiers and flow classifiers are trained and tested on test documents to build applicable classifiers. Based on the built classifiers, the meta-function term, function term and flow term labels of non-labeled patent documents collected from patent websites are predicted to obtain function and meta-function classified patents.

[0090] The classification system of the patent classification method provided by the present application is shown in Figure 2 The function concept is described by the combination of function terms and flow terms, i.e. “<function term>+<flow term>”, such as “<transmission>+<solid>”, which means “changing the spatial position of solid”. The meta-function term describes the relationship between different functions, such as “prevention” means “the implementation of one function avoids the failure of another function during operation”.

[0091] Specifically, the meta-function term describes the interdependence between different functions, including eight types: providing, driving, enabling, improving, enhancing, allowing, preventing and controlling. “Providing” refers to the matter or energy generated or transmitted by one function, which is the object specially processed or operated by another function; “driving” refers to the energy generated or transmitted by one function supporting the action process of another function whose processing or operating object is matter; “enabling” refers to the occurrence of a function whose processing or operating object is a matter attribute directly leading to the occurrence of another function whose processing or operating object is matter or energy; “improving” refers to the implementation of one function improving the action effect of another function without increasing additional energy input; “enhancing” refers to the implementation of one function improving the action effect of another function with additional energy input; “allowing” refers to the increase of the processing or operating object related attribute of one function leading to the decrease of the side effects of another function on the third function or the entire system; “preventing” refers to the failure of another function if one function is not implemented; “controlling” refers to maintaining the value of the action object related attribute of another function by changing the value of the action object related attribute of one function.

[0092] The meta-function concept is described by the combination of the meta-function term, the function term and the flow term, i.e., "<meta-function term>+<function term>+<flow term>", such as "<prevention>+<transmission>+<solid>", which means "avoiding failure in the process of transmitting solid".

[0093] As shown in Figure 3 When constructing the patent classifier, first, the patent documents are obtained from the patent data platform in batches. For example, the product or technology system patents from various fields are obtained from the website of the patent and trademark office in batches at random, and the title and abstract of the patent are extracted as the patent document. Among them, the title is added with a period at the end and connected with the abstract to form a complete text, which is the patent document.

[0094] The collected patent documents are labeled, for example, read and understood by a design expert, and the main functions and meta-functions realized by the patent description object are extracted, and then the meta-function term, the function term and the flow term are used to label the patent document to obtain a labeled patent document set. The labeled patent document set is used for training and testing, for example, the patent document set labeled with each type of term is divided into training and testing patent document sets according to a ratio of 4:1.

[0095] In the Python environment, the noun vector representing the training document and the flow term label are taken as the input of the machine learning classifier to train the flow classifier. The verb vector representing the training document and the function term label are taken as the input of the machine learning classifier to train the function classifier. The verb relationship vector representing the training document and the meta-function term label are taken as the input of the machine learning classifier to train the meta-function classifier. The flow classifier, the function classifier and the meta-function classifier of the embodiment of the application can all be support vector machines.

[0096] In the Python environment, the pre-processing and vectorization representation of the test document set are performed, and are taken as the classification input to test the performance of the trained flow, function and meta-function classifiers. The performance evaluation index adopts the classification accuracy (Accuracy). If the test performance is not high, the overall data size or the data size under a specific category of the training document set is adjusted to cope with it, aiming to build flow, function and meta-function classifiers with performance reaching an applicable degree.

[0097] The embodiment of the application covers the patent classification system of meta-function, function and flow terms, which is helpful for cross-field patent association and organization in the aspects of function and meta-function.

[0098] In an embodiment of the application, the step A20 specifically comprises:

[0099] The step A21 comprises analyzing the part of speech of each word in each sentence of each training patent document, and analyzing the grammatical relationship between different words.

[0100] Step A22, according to the part of speech of each word and the grammatical relationship between different words, extract the nominal object training data, the verbal predicate training data and the verb pair training data of keeping the dependency relationship of each sentence;

[0101] Step A23, the semantic vector representation is carried out on each of the nominal object training data, and each nominal vector training data is obtained, the average of each nominal vector training data is calculated, and the nominal vector training representation of the to-be-classified patent document is obtained.

[0102] Step A24, the semantic vector representation is carried out on each of the verbal predicate training data, and each verbal vector training data is obtained, the average of each verbal vector training data is calculated, and the verbal vector training representation of the to-be-classified patent document is obtained.

[0103] Step A25, the semantic vector representation is carried out on the two verbs in the verb pair training data of keeping the dependency relationship, and the third verb vector and the fourth verb vector are obtained, the difference between the third verb vector and the fourth verb vector is taken as the verb relationship vector training data, the average of each of the verb relationship vector training data is calculated, and the verb relationship vector training representation of the to-be-classified patent document is obtained.

[0104] Specifically, in the Python environment, the dependency syntax analysis tool Standford CoreNLP is called, and the dependency syntax analysis is carried out on the training document set. For each sentence in each document, the part of speech of each word (such as noun, verb, adjective, adverb, etc.) and the grammatical relationship between different words (such as subject-predicate relationship, verb-object relationship, subject-predicate-object relationship, appositive relationship, etc.) are analyzed. Through the object and verb extraction rule, the nominal object and the verbal predicate are extracted from the analyzed sentence. For each verb, find its directly connected word in the dependency relationship, that is, its directly dependent word), and judge whether this dependency relationship is also a verb. If so, keep the pair of verbs. For the screened verb pair, extract them as the two verbs of keeping the dependency relationship.

[0105] The nominal object, the verbal predicate and the verb pair of keeping the dependency relationship extracted from the training document are subjected to 300-dimensional semantic vector representation by the word embedding tool GloVe. The average of each nominal vector is taken to obtain the nominal vector representing the document. The average of each verb vector is taken to obtain the verb vector representing the document. The difference between the two verb vectors representing the verb pair of keeping the dependency relationship is taken to obtain the verb relationship vector representing the verb pair of keeping the dependency relationship, and then the average of each verb relationship vector is taken to obtain the verb relationship vector representing the document.

[0106] In one embodiment, asFigure 4 As shown, based on the above patent automatic classification method, the application also correspondingly provides a patent automatic classification device, comprising:

[0107] The analysis module 100 is configured to perform dependency syntax analysis on the patent document to be classified to obtain pre-processing information.

[0108] The word embedding module 200 is configured to perform word embedding on the pre-processing information to obtain document representation.

[0109] The classification module 300 is configured to obtain the flow category prediction result, the function category prediction result and the meta-function category prediction result based on the document representation, the trained flow classifier, the function classifier and the meta-function classifier.

[0110] Figure 5 The terminal provided by the embodiment of the application is shown in a structural schematic diagram. The terminal can include:

[0111] The memory 501, the processor 502 and the computer program stored in the memory 501 and executable on the processor 502.

[0112] The processor 502 implements the patent automatic classification method provided in the above embodiment when executing the program.

[0113] Further, the terminal also includes:

[0114] The communication interface 503 is configured to communicate in the memory 501 and the processor 502.

[0115] The memory 501 is configured to store the computer program executable on the processor 502.

[0116] The memory 501 can include a high-speed RAM memory, and can also include a non-volatile memory, for example, at least one disk memory.

[0117] If the memory 501, the processor 502 and the communication interface 503 are implemented independently, the communication interface 503, the memory 501 and the processor 502 can be connected with each other through a bus and complete communication between each other. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For the convenience of representation, only one line is used to represent the bus in the figure, but it does not mean that there is only one bus or only one type of bus.

[0118] Optionally, in a specific implementation, if the memory 501, the processor 502 and the communication interface 503 are integrated on a chip, the memory 501, the processor 502 and the communication interface 503 can complete communication between each other through an internal interface.

[0119] The processor 502 can be a central processing unit (CPU), or an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present application.

[0120] The embodiments also provide a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the above patent automatic classification method.

[0121] In the description of the specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example" or "some examples" means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or N embodiments or examples in a suitable manner. In addition, the person skilled in the art can combine and combine the different embodiments or examples described in the specification and the features of the different embodiments or examples without contradiction.

[0122] Moreover, the terms "first", "second", "third", etc. are used herein only to describe different instances, and do not imply or suggest relative importance or a number of instances indicated. Thus, features defined with "first", "second" can explicitly or implicitly include at least one of the features. In the description of the present application, the meaning of "N" is at least two, for example, two, three, etc., unless otherwise explicitly and specifically limited.

[0123] Any process or method descriptions or blocks in flow charts herein, and elsewhere, can be understood as representing modules, segments, or portions of code which include one or more executable instructions for implementing specific logical functions or steps in the process. Alternate implementations are included within the scope of the preferred embodiments of the present application in which additional functionality can be added or some functionality can be removed, by, for example, adding one or more steps performing a similar or different function, rearranging the order of the steps, etc.

[0124] Logic and / or steps represented in flow charts herein, and elsewhere, can be embodied in computer-readable media, for use by or in connection with an instruction execution system, apparatus, or device, such as a computer-based system, processor- containing system, or other system that can fetch the instructions from the instruction execution system, apparatus, or device and execute the instructions. In the context of this specification, a "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device. The computer-readable medium can be a machine-readable storage device (e.g., magnetic, optical, or other storage device) a machine-readable storage substrate, a machine-readable signal, or any combination thereof. Other examples of a computer- readable medium include, but are not limited to: an electronic connection (e.g., a conventional computer-readable medium storing a program) having one or more wires, portable computer diskette, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber, and a portable compact disc read-only memory (CDROM). Note that the computer-readable medium can even be paper or another suitable medium upon which the program is printed, as the program can be electronically captured, via, for instance, optical scanning of the paper or other medium, then compiled, interpreted, or otherwise processed in a suitable manner, if necessary, and then stored in a computer memory.

[0125] It should be understood that portions of the present application can be realized with hardware, software, firmware or a combination thereof. In the above embodiments, the N steps or methods can be realized with software or firmware stored in a memory and executed by a suitable instruction execution system. If realized with hardware and as in another embodiment, any one or a combination of the following technologies known in the art can be used: discrete logic circuit with logic gate circuit for implementing logic functions on data signals, application specific integrated circuit with suitable combination logic gate circuit, programmable gate array (PGA), field programmable gate array (FPGA) and the like.

[0126] Those skilled in the art of the present technology can understand that all or part of the steps carried out by the above-mentioned embodiment methods can be completed by a program instructing the relevant hardware, which can be stored in a computer readable storage medium. When the program is executed, it includes one of the steps of the method embodiment or a combination thereof.

[0127] In addition, the functional units in each embodiment of the present application can be integrated into one processing module, or each unit can be physically present separately, or two or more units can be integrated into one module. The above integrated module can be realized in the form of hardware or in the form of a software function module. The integrated module, if realized in the form of a software function module and sold or used as an independent product, can also be stored in a computer readable storage medium.

[0128] The storage medium mentioned above can be a read-only memory, a magnetic disk or an optical disk, etc. Although the embodiments of the present application have been shown and described above, it should be understood that the above embodiments are exemplary and cannot be understood as limiting the present application. Those skilled in the art can make changes, modifications, replacements and variations to the above embodiments within the scope of the present application.

[0129] In summary, the patent automatic classification method, device, terminal and storage medium disclosed by the present application comprise: performing dependency syntax analysis on the patent document to be classified to obtain pre-processing information; performing word embedding on the pre-processing information to obtain document representation; and obtaining flow category prediction results, function category prediction results and meta-function category prediction results based on the document representation and trained flow classifier, function classifier and meta-function classifier. The present application realizes the comprehensiveness of the classification of the patent document to be classified by performing dependency syntax analysis and word embedding on the patent document to be classified, and then using machine learning technology to obtain flow category prediction results, function category prediction results and meta-function category prediction results, solves the problem that the existing patent automatic classification method is very limited and cannot comprehensively classify patents.

[0130] It is to be understood that the application is not limited to the examples described above, which can be modified or adapted in several ways by those skilled in the art without departing from the scope of the present application, as defined by the appended claims.

Claims

1. A method for automatic classification of patents, characterized in that, The method comprises: performing dependency syntax analysis on the patent document to be classified to obtain preprocessing information; performing word embedding on the preprocessing information to obtain document representation; obtaining flow category prediction results, function category prediction results and meta-function category prediction results based on the document representation and trained flow classifier, function classifier and meta-function classifier; The method comprises: performing dependency syntax analysis on the patent document to be classified to obtain preprocessing information; analyzing the part of speech of each word in each sentence of the patent document to be classified, and analyzing the grammatical relationship between different words to obtain the analysis result corresponding to each sentence; extracting the noun object, verb predicate and verb pair maintaining the dependency relationship of each sentence according to the analysis result to obtain the preprocessing information; performing word embedding on the preprocessing information to obtain document representation, comprising: performing semantic vector representation on each noun object to obtain each noun vector, calculating the average value of each noun vector to obtain the noun vector representation of the patent document to be classified; performing semantic vector representation on each verb predicate to obtain each verb vector, calculating the average value of each verb vector to obtain the verb vector representation of the patent document to be classified; performing semantic vector representation on the two verbs in the verb pair maintaining the dependency relationship to obtain a first verb vector and a second verb vector, taking the difference between the first verb vector and the second verb vector as a verb relationship vector, calculating the average value of each verb relationship vector to obtain the verb relationship vector representation of the patent document to be classified; saving the noun vector representation, verb vector representation and verb relationship vector representation as document representation; The method comprises: inputting the noun vector representation in the document representation into the trained flow classifier to obtain the flow category prediction results; inputting the verb vector representation in the document representation into the trained function classifier to obtain the function category prediction results; 2. The method of claim 1, wherein, inputting the verb relationship vector representation in the document representation into the trained meta-function classifier to obtain the meta-function category prediction results. Before the dependency syntax analysis on the patent document to be classified to obtain the preprocessing information, the method further comprises: extracting the title and abstract of the patent to be classified; 3. The method of claim 1, wherein, adding a period at the end of the title, and the title after adding the period and the abstract form the patent document to be classified. The training steps of the flow classifier, function classifier and meta-function classifier comprise: obtaining a training patent document set, the training patent document set comprising a plurality of training patent documents and flow terminology labels, function terminology labels and meta-function terminology labels corresponding to each training patent document; performing dependency syntax analysis and word embedding processing on each training patent document to obtain training document representation, the training document representation comprising noun vector training representation, verb vector training representation and verb relationship vector training representation; training an initial flow classifier pre-constructed based on the noun vector training representation and the flow term label to obtain a trained flow classifier; training an initial function classifier pre-constructed based on the verb vector training representation and the function term label to obtain a trained function classifier; training an initial meta-function classifier pre-constructed based on the verb relation vector training representation and the meta-function term label to obtain a trained meta-function classifier.

4. The method of claim 3, wherein, The dependency syntax analysis and word embedding processing of each of the training patent documents to obtain training document representations include: analyzing the part-of-speech of each word in each sentence of each of the training patent documents, and analyzing the grammatical relations between different words; extracting the noun object training data, the verb predicate training data, and the verb pair training data that maintains the dependency relation of each sentence according to the part-of-speech of each word and the grammatical relations between different words; performing semantic vector representation on each of the noun object training data to obtain each noun vector training data, and calculating the average of each noun vector training data to obtain the noun vector training representation of the patent document to be classified; performing semantic vector representation on each of the verb predicate training data to obtain each verb vector training data, and calculating the average of each verb vector training data to obtain the verb vector training representation of the patent document to be classified; performing semantic vector representation on the two verbs in the verb pair training data that maintains the dependency relation to obtain a third verb vector and a fourth verb vector, taking the difference between the third verb vector and the fourth verb vector as the verb relation vector training data, and calculating the average of each of the verb relation vector training data to obtain the verb relation vector training representation of the patent document to be classified.

5. An apparatus for automatically classifying patents, characterized by The apparatus includes: an analysis module configured to perform dependency syntax analysis on the patent document to be classified to obtain pre-processing information; a word embedding module configured to perform word embedding on the pre-processing information to obtain a document representation; a classification module configured to obtain a flow category prediction result, a function category prediction result, and a meta-function category prediction result based on the document representation and the trained flow classifier, the trained function classifier, and the trained meta-function classifier. The dependency syntax analysis on the patent document to be classified to obtain pre-processing information includes: analyzing the part-of-speech of each word in each sentence of the patent document to be classified, and analyzing the grammatical relations between different words to obtain an analysis result corresponding to each sentence; extracting the noun object, the verb predicate, and the verb pair that maintains the dependency relation of each sentence according to the analysis result to obtain the pre-processing information; performing word embedding on the pre-processing information to obtain a document representation includes: performing semantic vector representation on each of the noun objects to obtain each noun vector, and calculating the average of each noun vector to obtain the noun vector representation of the patent document to be classified; performing semantic vector representation on each of the verb predicates to obtain each verb vector, and calculating the average of each verb vector to obtain the verb vector representation of the patent document to be classified; The two verbs in a verb pair maintaining a dependency relationship are represented by semantic vectors to obtain a first verb vector and a second verb vector, the difference between the first verb vector and the second verb vector is taken as a verb relationship vector, the average of each verb relationship vector is calculated to obtain a verb relationship vector representation of the patent document to be classified; The noun vector representation, the verb vector representation and the verb relationship vector representation are saved as a document representation; The flow class prediction result, the function class prediction result and the meta-function class prediction result are obtained based on the document representation and the trained flow classifier, function classifier and meta-function classifier, and the method comprises the following steps: The noun vector representation in the document representation is input into the trained flow classifier to obtain the flow class prediction result; The verb vector representation in the document representation is input into the trained function classifier to obtain the function class prediction result; The verb relationship vector representation in the document representation is input into the trained meta-function classifier to obtain the meta-function class prediction result.

6. A terminal, characterized by comprising: The patent automatic classification method comprises the following steps: A memory, a processor and a patent automatic classification program stored in the memory and executable on the processor, wherein the patent automatic classification program is executed by the processor to implement the steps of the patent automatic classification method according to any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program can be executed to implement the steps of the patent automatic classification method according to any one of claims 1 to 4.

Images