CN107122382B - A description-based patent classification method - Google Patents

Abstract

The invention discloses a patent classification method based on a description, which belongs to the field of text processing and data mining. Firstly, the text preprocessing is performed on the patent specification; then an inverted index file is constructed, and the feature selection method combining information gain and word frequency is used to select the feature words; the improved TF-IDF formula is further used to calculate the feature word weight, and construct Patent feature vector; then build a set of training patent fields; finally use the optimized KNN classifier to classify patents. This research provides a new idea for the classification of patent documents, and also lays a foundation for further research on intelligent retrieval of patent documents.

Description

A description-based patent classification method

technical field

The invention belongs to the application of computer analysis technology in patent documents, and specifically relates to a patent classification method using patent specifications.

Background technique

Patents are the concrete manifestation of technological innovation and enterprise value, and one of the important carriers, achievements and sources of knowledge development and innovation. Many inventions and creations only appear in patent documents. According to the statistics of the World Intellectual Property Organization (WIPO), 70% to 90% of the inventions in the world first appear in patent documents, rather than in the documents of other carriers such as journals and papers. In addition, in order to protect their own interests, enterprises will apply for patents as early as possible, and patents often focus on the most active and advanced technologies, including 90% to 95% of the technical information in the world. At the same time, for the convenience of examination, patent documents are often written in more detail. Compared with other types of documents, patent documents can provide more information and are the most common technological innovation achievements, recording the complete process of patent activities. It not only reflects the status quo of technical activities in various technical fields, but also reflects the development history of technical activities in a specific technical field. Patent documents contain specific technical solutions for each patented invention and creation, which plays a very important role in enterprise innovation. It not only enables enterprises to understand the latest scientific research trends, avoids repeated research, saves research time and research funds, but also enables enterprises. Enlighten the innovative ideas of enterprise researchers, improve the starting point of innovation, learn from past inventions, and greatly shorten the progress of scientific research.

With the continuous emergence of new research results and inventions in my country, the number of patents has shown rapid growth. As of October 5, 2016, the number of published invention patents in my country has exceeded 5.98 million, of which the total number of authorized invention patents is 2.2385 million. If the average size of each patent is 2M, the capacity of patent data is as high as several hundred terabytes. In order to manage these patent document data scientifically and to search related patent documents quickly and conveniently, the classification of patent documents is particularly important. At present, most countries in the world use the International Patent Classification (IPC) to classify patent documents. IPC is classified according to five levels, namely Section, Class, Subclass, Main group (Main Grop), group (Grop), the part of which is the highest classification layer in the classification table, according to different fields, is divided into eight major parts, marked with one-digit English letters, respectively A-H, each part Subordinates have multiple categories, which are composed of two-digit numbers, and there are different numbers of categories under each subordinate. For example: G06F21/00 stands for Physical-Electrical Digital Data Processing-Security Devices for Protecting Computers, Its Components, Programs or Data from Unauthorized Actions.

It can be seen that for the invention patent that has been or will be published, one or more corresponding classification numbers must be assigned. For example, the classification number of the invention patent "a method for protecting privacy data in association rule mining" is G06F21/ 00. For the patent application to be submitted, its classification number is unknown and needs to be determined. In this regard, the common practice is to determine it according to the field or patent content of the patent description object. It is necessary to rely on relevant experts to manually read the application and follow the instructions. With the rapid increase in the number of patent applications (the number of patent applications per year is close to 1 million), this method requires a lot of manpower and material resources, and the limitations of experts' own knowledge make it difficult to ensure the consistency and accuracy of the classification results. To this end, the present invention proposes a patent classification method based on the specification of patent documents. The method utilizes the information in the published invention patent specification to construct a classifier or a classification function, and uses this to determine the category of the patent application, thereby realizing Automatic classification of patents.

SUMMARY OF THE INVENTION

The purpose of the present invention is to propose a patent classification method based on the description of the patent document, which can make full use of the description contained in the published invention patent, in view of the fact that the existing patent classification method cannot fully and effectively utilize the description information in the published invention patent. Information and corresponding categories are used to construct classifiers or classification functions, so as to determine the category of the submitted patent application, and propose corresponding optimization solutions for the feature extraction and selection of the specification and the determination of the classifier during the construction process.

The technical solution adopted in the present invention is: the patent classification method based on the patent document description mainly includes the following steps:

(1) Patent data preprocessing

Collection of patent sample data, sample IPC number, extraction of description, Chinese word segmentation, part-of-speech tagging. Remove the symbols and numbers in the specification (there are a lot of paragraph marks in the specification). Use regular matching to filter out stop words, function words, conjunctions and other words that are not useful for patent classification, and only retain keywords such as nouns, adjectives, and verbs.

(2) Build an inverted index file

Count the word frequency, position information, part-of-speech weight and distribution information between classes of each word, and use these statistical values and patent text information to construct an inverted index file.

(3) Feature selection of patent text

The feature selection method combining information gain and word frequency is used to calculate the feature values of the words in step (2), sort the feature values, and select a certain number of feature words to represent the patent text.

Let A _ij be the number of documents that contain the feature word t _i and belong to c _j , B _ij be the number of documents that contain the feature word t _i and the category does not belong to c _j , and C _ij is the number of documents that do not contain the feature word t _i and the category belongs to c _j The number of documents, D _ij is the number of documents that do not contain the feature word t _i and the category belongs to not c _j , then the calculation of the feature value is shown in formula (1).

Among them, TF represents the degree of influence of the word frequency in the patent on the selection of patent features. Let m be the total number of categories in the training patents, N _j the total number of patents in the c _j category, TF _jk the word frequency of the feature word t _i in the c _j category of the patent P _k , then the calculation of TF is shown in formula (2) .

表示特征词t_i在所有类中出现的频数平均值，则计算如式(3)所示。IC in formula (1) represents the degree of dispersion of feature words among categories. The more dispersed, the less representative the word is, and the smaller the value is. Let TF _j (t _i ) represent the frequency of the feature word t _i in the class c _j , TF(t _i ) represent the total frequency of the feature word t _i ,

Represents the average frequency of the feature word t _i appearing in all classes, then the calculation is shown in formula (3).

(4) Vectorization of patent text

This step will include:

①Weight calculation, the calculation is shown in formula (4).

表示特征词t在文本

中出现的频率，N表示全部专利样本集中所有专利的个数，n表示全部专利样本集中出现特征词t的专利个数，C_t表示特征词词性所对应的词性权重系数，P_t表示特征词的位置权重系数。in,

Represents the feature word t in the text

The frequency of occurrence in all patent sample sets, N represents the number of all patents in all patent sample sets, n represents the number of patents in which feature word t appears in all patent sample sets, C _t represents the part-of-speech weight coefficient corresponding to the feature word part-of-speech, and P _t represents the feature word The position weight coefficient of .

②Sort, sort according to the weight in descending order, construct the space model vector V _i (w _i1 ,w _i2 ,...,w _in ) of the patent text, so as to represent the content of each patent text.

(5) Generate the feature vectors of each level of the IPC

This step includes:

① Combine the category description of each subgroup into the category description of the main group to which it belongs, and perform word segmentation and stop word removal processing.

② Combine the descriptions of each main group and perform feature selection to construct the category feature vector of the IPC subclass level, which is represented as {V _A01B1/00 ,V _A01B3/00 ,...,V _H99Z99/00 }. Among them, A01B1/00 is the first main group in IPC, and H99Z99/00 is the last main group in IPC.

③ Combine all the basic descriptions under the same sub-category and perform feature selection to construct the category feature vector of the IPC category level, which is represented as {V _A01B ,V _A01C ,...,V _H99Z }. Among them, A01B is the first subclass in IPC, and H99Z is the last subclass in IPC.

④ Combine all the basic descriptions under the same category and perform feature selection to construct a category feature vector at the IPC level, which is represented as {V _A01 ,V _A21 ,...,V _H99 }. Among them, A01 is the first category in IPC, and H99Z is the last category in IPC.

(6) Construct patent sample neighborhood

This step includes:

① Calculate the similarity between the patents in the patent training set. Similarity can be obtained by calculating the cosine of the angle between the vectors. Suppose sim(d _i , d _j ) represents the similarity between the patent texts d _i and d _j , then the calculation formula is shown in formula (5).

Among them, _Wik and _Wjk represent the weight of the corresponding special testimony in the patent vector, and n represents the dimension of the vector.

② Sort the similarity between d _i and all other patent samples d _j in descending order, select the first K patent samples to form a set D _i , and D _i is called the neighborhood of patent d _i , and the value of K depends on the specific situation.

(7) Calculation of similarity of patents to be classified

This step includes:

①Extracting the description, Chinese word segmentation, part-of-speech tagging, and removing stop words for the patent to be classified.

②Patent feature selection and vectorization.

③ Calculate the cosine similarity S _ai of the feature vector of the patent B _j to be classified and the feature vector of each IPC category.

④ Calculate the cosine similarity S _bj between the patent B _j to be classified and each patent in the patent training set.

⑤ Sort the above training patents in descending order of similarity value S _bj , and select the top K patents as its neighborhood set.

(8) Classification decision

This step includes:

① Calculate the shared field size L(B _j ,d _i ) between the patent B _j to be classified and the sample patent d _i , that is, the number of identical patents in the two field sets.

② Calculate the final weighted similarity between the patent to be classified and each IPC category, and the calculation formula is shown in formula (6).

Among them, I represents the category, and p, k, α, and β are adjustable parameters. By default, p is 0.8, k is 0.95, α is 0.6, and β is 5.

③ Classify the patents to be classified into the class with the largest similarity S(i).

The main beneficial effects of the present invention are shown in:

(1) Feature selection of patent text

Compared with the title and abstract of the patent, the patent specification is richer in content and contains a larger amount of information. Because of this, the patent specification contains a lot of noise data, especially in the classification below the IPC sub-category, the similar information contained between different patents is more, which is not conducive to the classification. Therefore, the present invention improves the method of feature extraction and feature vector quantization in the patent specification, reduces noise interference, and improves the classification accuracy of patents.

(2) Design of patent classification method

Due to the huge amount of patent data and the large number of patent categories, the training speed of the classification model is too slow, which is obviously not suitable for patent classification. Therefore, the present invention proposes a new nearest neighbor classification algorithm, and adds IPC description information in the classification process, which further improves the accuracy of patent classification on the premise of ensuring the classification speed.

Description of drawings

Fig. 1 is a structural block diagram in an embodiment of the present invention

Fig. 2 is the construction flow of the patent vector space in the embodiment of the present invention

Fig. 3 is the classification flow chart based on improved KNN in the embodiment of the present invention

Detailed ways

The patent classification method of the present invention is described in detail below with the patent documentation as an embodiment, and the specific execution process is as follows:

Step 1: Obtain the data of the patent text, and perform text preprocessing on the patent specification, mainly including word segmentation and removal of stop words.

① Obtain the description of the IPC category, perform word segmentation and part-of-speech tagging on the description, remove stop words, and manually correct the word segmentation results to construct a user dictionary.

(2) Perform format conversion and description extraction on the patent samples extracted above, add the user dictionary constructed in (1) to the word segmentation program, and then perform Chinese word segmentation and part-of-speech tagging on the description.

③Using regular expressions, remove stop words, function words, conjunctions and other words that are not useful for patent classification in the patent specification, and only retain nouns, adjectives, and verbs.

Step 2: Count the word frequency, position information, part-of-speech weight and inter-class distribution information of each word, and use these statistical values and patent text information to construct an inverted index file.

其中n表示词汇在说明书中出现的总次数，l_i表示词汇第i次出现所处位置的权重，实例中设技术领域权重1，背景技术0.8，其他位置0.5。词性权重设定为名词2.5，动词和形容词均为1，具体结果如表1所示。An inverted index file is constructed according to the words filtered out in step 1. The index file structure includes a vocabulary table and an event table. Each vocabulary corresponds to an event table. The event table stores the word frequency and position of the patent number in the patent document where the vocabulary appears. weights and part-of-speech weights. The formula for calculating the position weight here is:

Among them, n represents the total number of times the word appears in the specification, and li represents the weight of the position where the word appears for the _i -th time. The part-of-speech weight is set to 2.5 for nouns, and 1 for both verbs and adjectives. The specific results are shown in Table 1.

Table 1 User dictionary and inverted index merge

Step 3: Use the feature selection method combining information gain and word frequency to calculate the feature values of words, sort the feature values, and select a certain number of feature words to represent the patent text.

Due to the defect of low-frequency words in information gain, the applicant often repeats some special words in order to emphasize an innovative point, and these high-frequency words are beneficial for classification. Therefore, the present invention adopts a feature selection method combining information gain and word frequency. , first calculate the eigenvalues of the words in each patent according to formula (1), then sort these words in descending order according to the eigenvalues, and select the first 20 words as the characteristic words of the patent.

Let A _ij be the number of documents that contain the feature word t _i and belong to c _j , B _ij be the number of documents that contain the feature word t _i and the category does not belong to c _j , and C _ij is the number of documents that do not contain the feature word t _i and the category belongs to c _j The number of documents, D _ij is the number of documents that do not contain the feature word t _i and the category belongs to not c _j , then the calculation of the feature value is shown in formula (1).

Among them, TF represents the degree of influence of the word frequency in the patent on the selection of patent features. Let m be the total number of categories in the training patents, N _j the total number of patents in the c _j category, TF _jk the word frequency of the feature word t _i in the c _j category of the patent P _k , then the calculation of TF is shown in formula (2) .

表示特征词t_i在所有类中出现的频数平均值，则计算如式(3)所示。IC in formula (1) represents the degree of dispersion of feature words among categories. The more dispersed, the less representative the word is, and the smaller the value is. Let TF _j (t _i ) represent the frequency of the feature word t _i in the class c _j , TF(t _i ) represent the total frequency of the feature word t _i ,

Represents the average frequency of the feature word t _i appearing in all classes, then the calculation is shown in formula (3).

Step 4: Use the inverted index file to calculate the weight of each patent feature word, then use the improved TF-IDF formula to calculate the feature word weight, and finally construct the patent feature vector.

This step specifically includes:

①Weight calculation, the calculation is shown in formula (4).

表示特征词t在文本

中出现的频率，N表示全部专利样本集中所有专利的个数，n表示全部专利样本集中出现特征词t的专利个数，C_t表示特征词词性所对应的词性权重系数，P_t表示特征词的位置权重系数。in,

Represents the feature word t in the text

The frequency of occurrence in all patent sample sets, N represents the number of all patents in all patent sample sets, n represents the number of patents in which feature word t appears in all patent sample sets, C _t represents the part-of-speech weight coefficient corresponding to the feature word part-of-speech, and P _t represents the feature word The position weight coefficient of .

②Sort, sort according to the weight in descending order, construct the space model vector V _i (w _i1 ,w _i2 ,...,w _in ) of the patent text, so as to represent the content of each patent text.

The word frequency, position weight, and part-of-speech weight of the feature word have been recorded in the inverted index file, so it is only necessary to count the number of texts in which the feature word also appears. The total number of texts is also known. The specific results are shown in Table 2.

Table 2 Patent feature vector

Step 5: Generate the category feature vectors of each level of IPC. On the basis of step 1, start from the small category and go up layer by layer, calculate the category weight of each vocabulary at the corresponding level, use TF-IDF to calculate the weight, and regard a category description as a text, and then construct the category feature vector at each level.

This step specifically includes:

① Combine the category description of each subgroup into the category description of the main group to which it belongs, and perform word segmentation and stop word removal processing.

② Combine the descriptions of each main group and perform feature selection to construct the category feature vector of the IPC subclass level, which is represented as {V _A01B1/00 ,V _A01B3/00 ,...,V _H99Z99/00 }. Among them, A01B1/00 is the first main group in IPC, and H99Z99/00 is the last main group in IPC.

③ Combine all the basic descriptions under the same sub-category and perform feature selection to construct the category feature vector of the IPC category level, which is represented as {V _A01B ,V _A01C ,...,V _H99Z }. Among them, A01B is the first subclass in IPC, and H99Z is the last subclass in IPC.

④ Combine all the basic descriptions under the same category and perform feature selection to construct a category feature vector at the IPC level, which is represented as {V _A01 ,V _A21 ,...,V _H99 }. Among them, A01 is the first category in IPC, and H99Z is the last category in IPC.

For example, combine the vocabulary of all groups under the A01B sub-category into an A01B vocabulary set, as well as the sub-categories under the other A01 categories, then calculate the weight of each word in the A01B vocabulary set, and finally construct the feature vector of the A01B sub-category .

Step 6: Construct the patent sample neighborhood, use the patent feature vector in step 4 to calculate the similarity between each patent and other patents, sort the similarity of these patents, and select the 100 patents with the highest similarity to form the Neighborhood set of patents.

This step specifically includes:

① Calculate the similarity between the patents in the patent training set. Similarity can be obtained by calculating the cosine of the angle between the vectors. Suppose sim(d _i , d _j ) represents the similarity between the patent texts d _i and d _j , then the calculation formula is shown in formula (5).

Among them, _Wik and _Wjk represent the weight of the corresponding special testimony in the patent vector, and n represents the dimension of the vector.

② Sort the similarity between d _i and all other patent samples d _j in descending order, select the first K patent samples to form a set D _i , and D _i is called the neighborhood of patent d _i , and the value of K depends on the specific situation.

The specific results are shown in Table 3.

Table 3 Collection of patent fields

Step 7: Calculate the cosine similarity value between the patent vector to be classified, the IPC category feature vector, and the training set patent, and also calculate the neighborhood set of the patent to be classified.

This step includes:

①Preprocessing, feature selection, vectorization and data format conversion of patents to be classified.

②Patent feature selection and vectorization.

③ Calculate the cosine similarity S _ai of the feature vector of the patent B _j to be classified and the feature vector of each IPC category.

④ Calculate the cosine similarity S _bj between the patent B _j to be classified and each patent in the patent training set.

⑤ Sort the above training patents in descending order of similarity value S _bj , and select the top K patents as its neighborhood set.

Step 8: Classification decision, first calculate the size of the shared field between the patents to be classified and the patents in the training set, that is, calculate the number of identical patents in the neighborhood set. Then, the weighted sum of the similarity between the patent to be divided and the patent category is calculated, and after the weighted sum is sorted, the patent to be divided is divided into the class with the largest value.

This step specifically includes:

① Calculate the shared field size L(B _j ,d _i ) between the patent B _j to be classified and the sample patent d _i , that is, the number of identical patents in the two field sets.

② Calculate the final weighted similarity between the patent to be classified and each IPC category, and the calculation formula is shown in formula (6).

Among them, I represents the category, and p, k, α, and β are adjustable parameters. By default, p is 0.8, k is 0.95, α is 0.6, and β is 5.

③ Classify the patents to be classified into the class with the largest similarity S(i).

In the description of this specification, reference to the terms "one embodiment," "some embodiments," "exemplary embodiment," "example," "specific example," or "some examples", etc., is meant to incorporate the embodiments A particular feature, structure, material, or characteristic described by an example or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the particular features, structures, materials or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

Although embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and alterations can be made in these embodiments without departing from the principles and spirit of the invention, The scope of the invention is defined by the claims and their equivalents.

Claims (3)

1. A patent classification method based on an instruction book is characterized by comprising the following steps:

step 1, acquiring data of a patent text, and performing text preprocessing on a patent specification;

step 2, counting word frequency, position information, part-of-speech weight and inter-class distribution information of each word, and constructing an inverted index file by using the statistical values and the text information of the patent specification;

step 3, calculating the characteristic values of the words by using a characteristic selection method combining information gain and word frequency, sequencing the characteristic values, and selecting a certain number of characteristic words to represent the text of the patent specification;

the calculation process of the characteristic value in the step 3 is as follows:

let A_ijFor containing the feature word t_iAnd belong to c_jNumber of documents, B_ijFor containing the feature word t_iAnd the category does not belong to c_jNumber of documents of C_ijTo not contain the feature word t_iAnd the category belongs to c_jNumber of documents, D_ijTo not contain the feature word t_iAnd the category belongs to_jThe feature value is calculated as shown in equation (1):

wherein, TF represents the influence degree of the word frequency in the patent on the patent feature selection; let m be the total number of classes in the training patent, N_jDenotes c_jTotal number of patents in class, TF_jkRepresentation feature word t_iAt c_jMiddle class patent P_kThe term frequency in (1), then the calculation of TF is as shown in equation (2):

IC in the formula (1) represents the dispersion degree of the characteristic words among the categories, and the dispersion indicates that the words are less representative and the value is smaller; let TF_j(t_i) Representation feature word t_iIn class c_jFrequency of middle, TF (t)_i) Representation feature word t_iThe total frequency of the frequency of (c) is,

representation feature word t_iThe average of the frequencies occurring in all classes is calculated as shown in equation (3):

step 4, calculating the weight of each patent feature word by using the inverted index file, then calculating the weight of the feature word by using an improved TF-IDF formula, and finally constructing a patent feature vector;

the specific process of the step 4 is as follows:

step 4.1, weight calculation, wherein the calculation is shown as the formula (4):

wherein,

representing the characteristic word t in the text

The frequency of occurrence in (1), N represents the number of all patents in all patent sample sets, N represents the number of patents in all patent sample sets with characteristic words t, C_tRepresenting part-of-speech weight coefficients, P, corresponding to the part-of-speech of the feature word_tRepresenting characteristic wordsA position weight coefficient;

step 4.2, sorting according to weight descending order, and constructing a space model vector V of the text of the patent specification_i(w_i1,w_i2,...,w_in) The content of the text of each patent specification is expressed as such;

step 5, generating IPC class characteristic vectors of each level, calculating the class weight of each vocabulary in the corresponding level layer by layer from subclasses upwards on the basis of the step 1, calculating the weight by using TF-IDF, regarding a class description as a text, and then constructing the class characteristic vectors of each level;

the specific process of the step 5 is as follows:

step 5.1, merging the category description of each subgroup into the category description of the main group to which the subgroup belongs, and performing word segmentation and word stop processing;

and 5.2, combining the descriptions of each main group, then selecting features, constructing a class feature vector of an IPC subclass layer, and expressing the vector as { V }_A01B1/00,V_A01B3/00,...,V_H99Z99/00}; wherein, A01B1/00 is the first main group in IPC, H99Z99/00 is the last main group in IPC;

step 5.3, merging all basic descriptions under the same subclass, then performing feature selection, constructing a class feature vector of an IPC (IPC) major class hierarchy, and expressing the vector as { V }_A01B,V_A01C,...,V_H99Z}; wherein, A01B is the first subclass in IPC, H99Z is the last subclass in IPC;

step 5.4, merging all basic descriptions under the same large class, then performing feature selection, constructing a class feature vector of an IPC part level, and expressing the vector as { V }_A01,V_A21,...,V_H99}; wherein, A01 is the first major class in IPC, H99Z is the last major class in IPC;

step 6, constructing a neighborhood of a patent sample, calculating the similarity between each patent and other patents by using the patent feature vectors in the step 4, sequencing the patent similarities, and selecting K patents with the maximum similarity to form a neighborhood set of the patent;

the specific process of the step 6 is as follows:

step 6.1, calculating the similarity among the patents in the patent training set; the similarity can be obtained by calculating the cosine of an included angle between vectors; let sim (d)_i,d_j) Text d representing patent specification_iAnd d_jThe calculation formula is shown in formula (5):

wherein, W_ikAnd W_jkRepresenting the weight of a corresponding special token in the patent vector, wherein n represents the dimension of the vector;

step 6.2, mixing d_iSample d of all other patents_jThe similarity is sorted in descending order, and the first K patent samples are selected to form a set D_i，D_iReferred to as patent d_iThe value of K is case specific;

step 7, calculating cosine similarity values between the patent vectors to be classified, the IPC class characteristic vectors and the patents in the training set, and calculating neighborhood sets of the patents to be classified;

step 8, firstly, calculating the size of a shared field between the patents to be classified and the patents in the training set, namely calculating the number of the same patents in the neighborhood set; then calculating the similarity weighted sum between the patents to be classified and the patent categories, and after the weighted sum is sorted, dividing the patents to be classified into the category with the largest value;

the specific process of the step 8 is as follows:

step 8.1, calculate the patent B to be classified_jAnd sample patent d_iSize L (B) of shared field between_j,d_i) I.e. the number of identical patents in the two domain sets;

and 8.2, calculating the final weighted similarity between the patents to be classified and each IPC class, wherein the calculation formula is shown as the formula (6):

wherein, I represents the category, p, k, alpha and beta are adjustable parameters, and under the default condition of the system, p is 0.8, k is 0.95, alpha is 0.6 and beta is 5;

and 8.3, classifying the patents to be classified into the class with the maximum similarity S (i).

2. A specification-based patent classification method according to claim 1, characterized in that: the step 1 specifically comprises:

collecting patent sample data, sampling IPC (International patent Classification) numbers, extracting specifications, dividing Chinese words, labeling parts of speech, and removing symbols and numbers in the specifications; regular matching is used for filtering out words with little use for patent classification, such as stop words, dummy words and connecting words, and only nouns, adjectives and verb keywords are reserved.

3. A specification-based patent classification method according to claim 1, characterized in that: the specific process of the step 7 is as follows:

7.1, extracting a specification, dividing Chinese words, labeling parts of speech and removing stop words of the patent to be classified;

step 7.2, selecting and vectorizing patent features;

step 7.3, calculating the patent B to be classified_jCosine similarity S of eigenvector and each IPC class eigenvector_ai；

Step 7.4, calculate the patent B to be classified_jThe cosine similarity S with each patent in the patent training set_bj；

Step 7.5, the above training patents are processed according to the similarity value S_bjSorting in descending order, and selecting the top K patents as the neighborhood set.

Images