Automatic industry classification method and system

Abstract

An automatic industry classification method comprises: determining a scope of target patents, defining a target industry tree; generating marks on the target industry tree; performing a rough classification for the target patents by using the marks; performing a fine classification for the target patents according to a result of the rough classification. The automatic industry classification method and system provided by the present invention uses a transductive learning method, so that full mining of small annotation quantity information is realized. The automatic industry classification method and system uses information of IPC, so that information dimension is enriched, and calculation amount needed in the classification is reduced. The automatic industry classification method and system further uses the hierarchical vectors generated by the abstract, the claims and the description, so that the information of word order relation is reserved, and the patent text is deeply mined.

Description

CROSS REFERENCES TO THE RELATED APPLICATION

This application is a national phase of International Patent Application No. PCT/CN2020/073042 filed on Jan. 19, 2020, which claims priority based on Chinese patent application 201911358987.3 filed on Dec. 25, 2019. The disclosures of the aforementioned applications are hereby incorporated by reference in their entireties.

TECHNICAL FIELD

The present invention relates to the technical field of literature analysis, in particular to an automatic industry classification method and system.

BACKGROUND

A rapid development of science and technology brings about a proliferation of patent texts and continuous emergence of new industries. In order to analyze the technical development in an industry background, a patent needs to be labeled with an industry label. A manual labeling method is slow, expensive, but high in accuracy. Therefore, an automatic classification method with less labeling amount, higher computational efficiency and more fully mining labeling information is needed.

Existing methods either require a large amount of manual labeling, or do not need manual labeling at all and this causes a corresponding relationship with a target industry cannot be directly established. The existing methods generally use the patent text to carry out natural language processing, which is computationally expensive and omits important information of International Patent Classification (IPC) dimension. Natural language processing methods used in the existing methods generally mine information from an angle of a single word, which makes information of word order relation lost. The natural language processing methods used in the existing methods generally mine information in one or more of title, abstract, claims and description, but do not use hierarchical vectors generated by the abstract, the claims and the description, which makes deep information contained in the patent text omitted.

An invention patent with publication No. CN105808524A discloses a patent automatic classification method based on the abstract of patent documents, the method comprises dictionary construction, generation of category feature vectors at all levels of IPC, selection of patent text features, vectorization of patent text, construction of an SVM-based classification model and classification of a to-be-classified patent. The disadvantage of this method is that the natural language processing method used loses the information of word order relation, and does not use the hierarchical vectors generated by the abstract, the claims and the description, and the deep information contained in the patent text is omitted.

SUMMARY

In order to solve the above mentioned technical problems, an automatic industry classification method and system provided by the present invention uses a transductive learning method, so that full mining of small annotation quantity information is realized; uses information of IPC, so that information dimension is enriched, and calculation amount is reduced; uses the hierarchical vectors generated by the abstract, the claims and the description, so that the information of word order relation is reserved, and the patent text is deeply mined.

A first object of the present invention is to provide an automatic industry classification method. The method comprises determining a scope of target patents, and further comprises the following steps:

• • step 1: defining a target industry tree;
  • step 2: generating marks on the target industry tree;
  • step 3: performing a rough classification for the target patents by using the marks;
  • step 4: performing a fine classification for the target patents according to a result of the rough classification.

Preferably, the step 1 comprises that an industry tree I={i₁, . . . , i_j, . . . ., i_n} is defined as needed, wherein, i_j∈I and is a first level industry, j is a serial number of the first level industry, 1≤j≤n, n is the number of all leaf nodes of I.

In any of the above solutions, it is preferred that the step 1 further comprises setting i_jkl . . . ={i_{jkl . . . 1}, . . . , i_{jkl . . . t}} as any non-leaf node of I, degree of other nodes except the leaf nodes is greater than or equal to 2, wherein, k is a serial number of a second level industry, l is a serial number of a third level industry, and t is a serial number of a penultimate level industry.

In any of the above solutions, it is preferred that a step of determining a scope of target patents is to manually determine the scope of patents to be classified as needed.

In any of the above solutions, it is preferred that the step 2 comprises: according to resource constraints, determining the number p of patents which can be marked, p≥N, each leaf node of the industry tree should be marked with at least one patent belonging to the node, wherein, N is the number of the last level industry.

In any of the above solutions, it is preferred that the step 3 is determining nodes above the leaf node.

In any of the above solutions, it is preferred that the step 3 comprises the following sub-steps:

• • step 31: generating a node set V of a graph;
  • step 32: arranging the marks;
  • step 33: generating an edge set E of the graph;
  • step 34: generating an adjacency matrix;
  • step 35: performing node division.

In any of the above solutions, it is preferred that the step 31 comprises that IPC(s) of each target patent is defined as an IPC combination IPC_v={ipc₁, . . . , ipc_q} and all different IPC combinations of the target patents form the node set V.

In any of the above solutions, it is preferred that the step 32 comprises that the industry on the leaf node marked with patents is taken as a classification y_i∈ of the leaf node, the number of nodes which have been marked is set to be l, a sequence of the nodes is adjusted, and the marked nodes is adjusted to be the front, then 1≤i≤l.

In any of the above solutions, it is preferred that the step 32 further comprises verifying whether l<<the number of unmarked nodes u, and if not, adjusting the marked patent, otherwise V={IPC₁, . . . , IPC_l, IPC_l+1 . . . , IPC_l+u}.

In any of the above solutions, it is preferred that the edge set E is a matrix, and weight e_ij of edges between two vertices is the number of patents in a union IPC_i∪IPC_j of IPCs of the two vertices, wherein, e_ij is value in the matrix E.

In any of the above solutions, it is preferred that the step 34 comprises the following sub-steps:

• • step 341: generating a distance matrix S;
  • step 342: generating the adjacency matrix W by using the distance matrix S.

In any of the above solutions, it is preferred that a calculation formula of the distance matrix S is s_ij=∥e_i−e_j∥₂, wherein, e_i and e_j are respectively the i-th row and the j-th row of the edge set E.

In any of the above solutions, it is preferred that the step 35 comprises the following sub-steps:

• • step 351: generating a degree matrix D=diag(d₁, d₂, . . . , d_l+u), its diagonal element is d_i=Σ_j=1^l+u, wherein, u is the number of unmarked nodes, and W_ij is the adjacent matrix;
  • step 352: generating a marked matrix, a nonnegative (l+u)×|| marked matrix F=(F₁^T, F₂^T, . . . , F_l+u^T)^T, the element of the i-th row F_i=(F_i1, F_i2, . . . , ) is a marked vector of IPC_i in the node set, a classification rule is y_i=argmax F_ij, wherein, is a set of industries, T represents transpose of the matrix;
  • step 353: initializing the marked matrix F, for i=1, 2, . . . , m and j=1, 2, . . . ,

$❘𝒴❘,F\left(0\right)=Y_{ij}=\{\begin{array}{c}1,\mathrm{if}\left(1\le i\le l\right)\wedge \left(y_i=j\right)\\ 0,\mathrm{otherwise}\end{array};$ step 354: constructing a propagation matrix

$B=D^{-\frac{1}{2}}W{D}^{-\frac{1}{2}},$ wherein,

$D^{-\frac{1}{2}}=\mathrm{diag}\left(\frac{1}{\sqrt{d_1}},\frac{1}{\sqrt{d_2}},\dots ,\frac{1}{\sqrt{d_{l+u}}}\right),$ d represents diagonal elements of the degree matrix D;step 355: generating an iterative calculation formula F(t+1)=αBF(t)+(1−α)Y, wherein, α∈(0,1) is a parameter, F(t) is a result of the t-th iteration, and Y is an initial matrix;step 356: iterating the calculation formula to convergence to obtain a state

$F^{*}=\underset{t\to \infty }{\lim }F\left(t\right)=\left(1-\alpha \right){\left(M-\alpha B\right)}^{-1}Y$ under convergence, wherein, M is a unit matrix;step 357: performing a prediction of unmarked nodes y_i=argmaxF_ij*, wherein, l+1≤i≤l+u.

In any of the above solutions, it is preferred that the step 4 comprises the following sub-steps:

• • step 41: setting objects to be classified;
  • step 42: extracting text information of patents;
  • step 43: generating text sets to be trained;
  • step 44: performing text vectorization;
  • step 45: performing patent classification;
  • step 46: in any leaf node classed by the step 45, identifying the patent that does not belong to any industry of the leaf node on the tree.

In any of the above solutions, it is preferred that the step 41 comprises taking the patent nodes of each class divided in the step 3 as a group, that means patents corresponding to node marked as y_i∈ are a group, there are || groups.

In any one of the above solutions, it is preferred that the step 42 comprises extracting abstract, claims and description of each patent in each group, performing word segmentation of text information of patent by using an existing tool, and generating a text set G={g₁, . . . , g_n}, wherein g_i=(p_i1,p_i2,p_i3), p_i1, p_i2 and p_i3 are respectively word sequences obtained by word segmentation of the abstract, the claims, and the description of the i-th patent.

In any of the above solutions, it is preferred that the text sets to be trained comprise the text set G, a text set G₁={p₁₁, . . . , p_n1}, a text set G₂={p₁₂, . . . , p_n2} and a text set G₃={p₁₃, . . . , p_n3}, which are respectively composed of word segmentation results of the all-texts, the abstracts, the claims, and the descriptions of the patents in the group.

In any of the above solutions, it is preferred that the step 44 comprises the following sub-steps:

• • step 441: vectoring the text in the text sets to be trained;
  • step 442: generating a matrix of text.

In any of the above solutions, it is preferred that the step 441 comprises that: in each text set to be trained, an element P=(t₁, . . . , t_m) is a segmented word sequence with m elements, t_i∈P is determined by w words t_{i, context}={t_i−w, . . . , t_i−2, t_i−1, t_i+1, t_i+2, . . . , t_i+w} before and after it, and by maximizing

$\frac{1}{m}\sum _{i=w}^{m-w}\log p\left(t_i❘t_{i,\mathrm{context}},\mathrm{pid}\right)$ wherein, the pid is a paragraph number of t_i in p,

$p\left(t_i❘t_{i,\mathrm{context}},\mathrm{pid}\right)=\frac{e^{y_{t_i}}}{{\sum }_j{e}^{y_j}},$ γ_ti=b+UΦ(t_{i, context},pid), U and b are parameters of softmax, and a vector corresponding to P is obtained by training data to be trained using a stochastic gradient descent method.

In any one of the above solutions, it is preferred that the step 442 comprises that the vectorization results of G={g₁, . . . , g_n}, G₁={p₁₁, . . . , p_n1}, G₂={p₁₂, . . . , p_n2} and G₃={p₁₃, . . . , p_n3} are supposed to be respectively H₁={h₁₁, . . . , h_n1}, H₂={h₁₂, . . . , h_n2}, H₃={h₁₃, . . . , h_n3}, and H₄={h₁₄, . . . , h_n4}, then a generated set of matrix of text of target patents is H={h₁, . . . , h_n}, wherein h_i=(h_i1, h_i2, h_i3, h_i4).

In any of the above solutions, it is preferred that the step 45 comprises that: marked patents are set as S=∪_j=1^kS_j⊂H, wherein, S_j≠Ø is the marked patent of the j-th leaf node on the industry tree, j cluster centers of a k-means algorithm are initialized using the marked patents, and cluster membership of marked patents is not changed in an iterative updating process of clusters.

In any of the above solutions, it is preferred that the step 46 comprises the following sub-steps:

• • step 461: a k distance of a patent p is calculated, and a k-th distance of the patent p is set as k−distance(o), and in patents divided into a leaf node on the industry tree, there is a patent o, and the distance between the patent o and the patent p is d(p, o);
  • step 462: a k-th distance domain of the patent p is calculated: a patent set whose distance from the patent p is ≤k−distance(o) is called the k-th distance domain N_k(p) of the patent p;
  • step 463: a reachable distance reachdist(p, o)=max{k−distance(o), ∥p−o∥} of the patent p relative to the patent o is calculated;
  • step 464: a local reachable density

$lr{d}_k\left(p\right)=\frac{\left|N_k\left(p\right)\right|}{{\sum }_{o\in N_n\left(p\right)}{\mathrm{reach}}_d{\mathrm{ist}}_k\left(p,o\right)}$ is calculated;

• • step 465: a local outlier factor

$L0F_k\left(p\right)=\frac{{\sum }_{o\in N_n\left(p\right)}\frac{{\mathrm{lrd}}_k\left(o\right)}{{\mathrm{lrd}}_k\left(p\right)}}{❘N_k\left(p\right)❘}$ is calculated;

• • step 466: if LOF(p) is greater than a threshold, it is thought that p is an outlier, and does not belong to the leaf node.

In any of the above solutions, it is preferred that if the following two conditions are met, k−distance(o)=d(p, o):

• • {circle around (1)} in the leaf node, there are at least k patents q to make d(p, q)≤d(p, o);
  • {circle around (2)} in the leaf node, there are at most k−1 patents q to make d(p, q)<d(p, o).

A second object of the present invention is to provide an automatic industry classification system, the system comprises a confirmation module for determining a scope of target patents, and further comprises the following modules:

• • an industry tree generation module for defining a target industry tree;
  • a mark generation module for generating marks on the target industry tree;
  • a rough classification module for performing a rough classification for the target patents by using the marks;
  • a fine classification module for performing a fine classification for the target patents according to a result of the rough classification;
  • the system performs automatic industry classification by executing the method of the first object.

The present invention provides the automatic industry classification method and system which reduce annotation quantity, and improve calculation efficiency and the classification accuracy at the same time.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a flowchart of a preferred embodiment of an automatic industry classification method according to the present invention.

FIG. 1A is a flowchart of a rough classification method for target patents in the embodiment shown in FIG. 1 of industrial automatic classification method according to the present invention.

FIG. 1B is a flowchart of a generating method of an adjacency matrix in the embodiment shown in FIG. 1 of industrial automatic classification method according to the present invention.

FIG. 1C is a flowchart of a node division method in the embodiment shown in FIG. 1 of industrial automatic classification method according to the present invention.

FIG. 1D is a flowchart of a fine classification method for target patents in the embodiment shown in FIG. 1 of industrial automatic classification method according to the present invention.

FIG. 1E is a flowchart of a text vectorization method in the embodiment shown in FIG. 1 of industrial automatic classification method according to the present invention.

FIG. 1F is a flowchart of a patent classification method in the embodiment shown in FIG. 1 of industrial automatic classification method according to the present invention.

FIG. 2 is a block diagram of a preferred embodiment of an automatic industry classification system according to the present invention.

DETAILED DESCRIPTION OF THE EMBODIMENTS

The present invention is further described with reference to the drawings and specific embodiments.

Embodiment 1

As shown in FIG. 1 and FIG. 2, step 1000 is executed, and a target industry tree is defined by using an industry tree generation module 200, and a scope of patents to be classified is manually determined as needed.

Step 1100 is executed, and a scope of target patents is determined by using a confirmation module 210. An industry tree I={i₁, . . . , i_j, . . . , i_n} is defined as needed, wherein, i_j∈I and is a first level industry, j is a serial number of the first level industry, 1≤j≤n, n is the number of all leaf nodes of I. i_jkl . . . ={i_{jkl . . . 1}, . . . , i_{jkl . . . t}} is set as any non-leaf node of I, degree of other nodes except the leaf nodes is greater than or equal to 2, wherein, k is a serial number of a second level industry, l is a serial number of a third level industry, and t is a serial number of a penultimate level industry.

Step 1200 is executed, and a mark generation module 220 is used to generate marks on the target industry tree. The number p of patents which can be marked is determined according to resource constraints, p≥N, each leaf node of the industry tree should be marked with at least one patent belonging to the node, wherein, N is the number of the last level industry.

Step 1300 is executed, and a rough classification module 230 is used to perform a rough classification for the target patents by using the marks, and nodes above the leaf node are determined. As shown in FIG. 1A, step 1310 is executed, and a node set V of a graph is generated. IPC(s) of each target patent is defined as an IPC combination IPC_v={ipc₁, . . . , ipc_q} and all different IPC combinations of the target patents form the node set V.

Step 1320 is executed, and the marks are arranged. The industry on the leaf node marked with patents is taken as a classification γ_i∈ of the leaf node, the number of nodes which have been marked is set to be l, a sequence of the nodes is adjusted, and the marked nodes is adjusted to be the front, then 1≤i≤l. Whether l<<the number of unmarked nodes u is verified, and if not, adjusting the marked patent, otherwise V={IPC₁, . . . , IPC_l, IPC_l+1 . . . , IPC_l+u}.

Step 1330 is executed, and an edge set E of the graph is generated. The edge set E is a matrix, and weight e_ij of edges between two vertices is the number of patents in a union IPC_i∪IPC_j of IPCs of the two vertices, wherein, e_ij is value in the matrix E.

Step 1340 is executed, and an adjacency matrix is generated. As shown in FIG. 1B, step 1341 is executed, and a distance matrix S is generated. A calculation formula of the distance matrix S is s_ij=∥e_i−e_j∥₂, wherein, e_i and e_j are respectively the i-th row and the j-th row of the edge set E.

Step 1342 is executed, and the adjacency matrix W is generated by using the distance matrix S.

Step 1350 is executed, and node division is performed. As shown in FIG. 1C, step 1351 is executed, and a degree matrix D=diag(d₁, d₂, . . . , d_l+u) is generated, diagonal elements of the degree matrix is d_i=Σ_j=1^l+uW_ij, wherein, u is the number of unmarked nodes, and W_ij is the adjacent matrix.

Step 1352 is executed, and a marked matrix is generated, a nonnegative (l+u)×|| marked matrix F=(F₁^T, F₂^T, . . . , F_l+u^T)^T, the element of the i-th row F_i=(F_i1, F_i2, . . . , F_i||) is a marked vector of IPC_i in the node set, a classification rule is γ_i=argmaxF_ij, wherein, is a set of industries, T represents transpose of the matrix.

Step 1353 is executed, and the marked matrix F is initialized, for i=1, 2, . . . , m and j=1, 2, . . . , ||,

$F\left(0\right)=Y_{ij}=\{\begin{array}{c}1,\mathrm{if}\left(1\le i\le l\right)\wedge \left(y_i=j\right)\\ 0,\mathrm{otherwise}\end{array}.$

Step 1354 is executed, and a propagation matrix

$B=D^{-\frac{1}{2}}W{D}^{-\frac{1}{2}}$ is constructed, wherein,

$D^{-\frac{1}{2}}=\mathrm{diag}\left(\frac{1}{\sqrt{d_1}},\frac{1}{\sqrt{d_2}},\dots ,\frac{1}{\sqrt{d_{l+u}}}\right),$ d represents diagonal elements of the degree matrix D.

Step 1355 is executed, and an iterative calculation formula F(t+1)=αBF(t)+(1−α)Y is generated, wherein, α∈(0,1) is a parameter, F(t) is a result of the t-th iteration, and Y is an initial matrix.

Step 1356 is executed, and the calculation formula is iterated to convergence to obtain a state

$F^{*}=\underset{t\to \infty }{\lim }F\left(t\right)=\left(1-\alpha \right){\left(M-\alpha B\right)}^{-1}Y$ under convergence, wherein, M is a unit matrix.

Step 1357 is executed, and a prediction of unmarked nodes γ_i=argmaxF_ij* is performed, wherein, l+1≤i≤l+u.

Step 1400 is executed, and a fine classification module 240 is used to perform a fine classification for the target patents according to a result of the rough classification. As shown in FIG. 1D, step 1410 is executed, and objects to be classified is set. The patent nodes of each class divided in the step 1300 are taken as a group, that means patents corresponding to node marked as γ_i∈ are a group, there are || groups.

Step 1420 is executed, and text information of patents is extracted. Abstract, claims and description of each patent in each group are extracted, word segmentation of text information of patent is performed by using an existing tool, and a text set G={g₁, . . . , g_n} is generated, wherein g_i=(p_i1,p_i2,p_i3), p_i1, p_i2, and p_i3 are respectively word sequences obtained by word segmentation of the abstract, the claims, and the description of the i-th patent.

Step 1430 is executed, and text sets to be trained are generated. The text sets to be trained comprise the text set G, a text set G₁={p₁₁, . . . , p_n1}, a text set G₂={p₁₂, . . . , p_n2} and a text set G₃={p₁₃, . . . , p_n3}, which are respectively composed of word segmentation results of the all-texts, the abstracts, the claims, and the descriptions of the patents in the group.

Step 1440 is executed, and text vectorization is performed. As shown in FIG. 1E, step 1411 is executed, and the text in the text sets to be trained is vectored. In each text set to be trained, an element P=(t₁, . . . , t_m) is a segmented word sequence with m elements, t_i∈P is determined by w words t_{i, context}={t_i−w, . . . , t_i−2, t_i−1, t_i+1, t_i+2, . . . , t_i+w} before and after it, and by maximizing

$\frac{1}{m}\sum _{i=w}^{m-w}\log p\left(t_i❘t_{i,\mathrm{context}},\mathrm{pid}\right)$ wherein, the pid is a paragraph number of t_i in p,

$p\left(t_i❘t_{i,\mathrm{context}},\mathrm{pid}\right)=\frac{e^{y_{t_i}}}{{\sum }_j{e}^{y_j}},$ γ_ti=b+UΦ(t_{i, context},pid), U and b are parameters of softmax, and a vector corresponding to P is obtained by training data to be trained using a stochastic gradient descent method. Wherein, Φ is a mapping operation.

Step 1442 is executed, and a matrix of text is generated. The vectorization results of G={g₁, . . . , g_n}, G₁={p₁₁, . . . , p_n1}, G₂={p₁₂, . . . , p_n2} and G₃={p₁₃, . . . , p_n3} are supposed to be respectively H₁={h₁₁, . . . , h_n1}, H₂={h₁₂, . . . , h_n2}, H₃={h₁₃, . . . , h_n3}, and H₄={h₁₄, . . . , h_n4}, then a generated set of matrix of text of target patents is H={h₁, . . . , h_n}, wherein h_i=(h_i1, h_i2, h_i3, h_i4).

Step 1450 is executed, and patent classification is performed. Marked patents are set as S=∪_j=1^kS_j⊂H, wherein, S_j≠Ø is the marked patent of the j-th leaf node on the industry tree, j cluster centers of a k-means algorithm are initialized using the marked patents, and cluster membership of marked patents is not changed in an iterative updating process of clusters.

Step 1460 is executed, and in any leaf node classed by the step 45, the patent that does not belong to any industry of the leaf node on the tree is identified. As shown in FIG. 1F, step 1461 is executed, and a k distance of a patent p is calculated. A k-th distance of the patent p is set as k−distance(o), and in patents divided into a leaf node on the industry tree, there is a patent o, and the distance between the patent o and the patent p is d(p, o).

Step 1462 is executed, and a k-th distance domain of the patent p is calculated: a patent set whose distance from the patent p is ≤k−distance(o) is called the k-th distance domain N_k(p) of the patent p.

Step 1463 is executed, and a reachable distance reachdist(p, o)=max{k−distance(o), ∥p−∥} of the patent p relative to the patent o is calculated. If the following two conditions are met, k−distance(o)=d(p, o):

• • {circle around (1)} in the leaf node, there are at least k patents q to make d(p, q)≤d(p, o);
  • {circle around (2)} in the leaf node, there are at most k−1 patents q to make d(p, q)<d(p, o).

Step 1464 is executed, and a local reachable density

$lr{d}_k\left(p\right)=\frac{❘N_k\left(p\right)❘}{{\sum }_{o\in N_n\left(p\right)}{\mathrm{reach}}_d{\mathrm{ist}}_k\left(p,o\right)}$ is calculated.

Step 1465 is executed, and a local outlier factor

$\mathrm{L0}{F}_k\left(p\right)=\frac{{\sum }_{o\in N_n\left(p\right)}\frac{{\mathrm{lrd}}_k\left(o\right)}{{\mathrm{lrd}}_k\left(p\right)}}{❘N_k\left(p\right)❘}$ is calculated.

Step 1466 is executed, and if LOF(p) is greater than a threshold, it is thought that p is an outlier, and does not belong to the leaf node.

Embodiment 2

An automatic industry classification method comprises the following steps.

1. Defining a target industry tree. An industry tree I={i₁, . . . , i_j, . . . , i_n} is defined as needed, wherein, i_j∈I and is a first level industry, and i_j may be divided into second level industries, i_j={i_j1, . . . , i_jm}, and so on, any non-leaf node of I is i_jkl . . . ={i_{jkl . . . 1}, . . . , i_{jkl . . . t}}. According to a general practice of industry division, degree of other nodes except the leaf nodes is greater than or equal to 2. The number of leaf nodes under I is set as N.

2. Determining a scope of target patents. The scope of patents to be classified is manually determined as needed, such as applications in a certain country or applications in certain years.

3. Generating marks. The number p of patents which can be marked is determined according to resource constraints, p≥N, each leaf node of the industry tree should be marked with at least one patent belonging to the node.

4. Performing a rough classification for the target patents, that is determining nodes above the leaf node.

(1) Generating a node set V of a graph: IPC(s) of each target patent is defined as an IPC combination IPC_v={ipc₁, . . . , ipc_q}, and all different IPC combinations of the target patents form the node set V.

(2) Arranging marks: the industry on the leaf node marked with patents are taken as a classification γ_i∈ of the leaf node, the number of nodes which have been marked is set to be l, a sequence of the nodes is adjusted, and the marked nodes is adjusted to be the front, then 1≤i≤l, whether l<<the number of unmarked nodes u is verified, and if not, adjusting the marked patent until yes, otherwise V={IPC₁, . . . , IPC_l, . . . , IPC_l+1 . . . , IPC_l+u}.

(3) Generating an edge set E of the graph: the E may be expressed as a matrix, a union of IPCs of two vertices is IPC_i∪IPC_j, then weight of edges between the two vertices e_ij is equal to the number of patents with IPC in IPC_i∪IPC_j.

(4) Generating an adjacency matrix:

(4.1) a distance matrix S is generated using such as Euclidean distance, s_ij=∥e_i−e_j∥₂;

(4.2) the adjacency matrix W is generated by using the distance matrix S by using such as a full-connected method of Gaussian kernel function.

(5) Performing node division:

(5.1) a degree matrix D=diag(d₁, d₂, . . . , d_l+u) is generated, diagonal elements of the degree matrix is d_i=Σ_j=1^l+uW_ij;

(5.2) a marked matrix is generated, a nonnegative (l+u)×|| marked matrix F=(F₁^T, F₂^T, . . . , F_l+u^T)^T, the element of the i-th row F_i=(F_i1, F_i2, . . . , ) is a marked vector of IPC_i in the node set, a classification rule is γ_i=argmaxF_ij;

(5.3) the marked matrix F is initialized, for i=1, 2, . . . , m and j=1, 2, . . . , ||,

$F\left(0\right)=Y_{ij}=\{\begin{array}{c}1,\mathrm{if}\left(1\le i\le l\right)\wedge \left(y_i=j\right)\\ 0,otherwise\end{array};$ (5.4) a propagation matrix

$B=D^{-\frac{1}{2}}W{D}^{-\frac{1}{2}}$ is constructed, wherein,

$D^{-\frac{1}{2}}=\mathrm{diag}\left(\frac{1}{\sqrt{d_1}},\frac{1}{\sqrt{d_2}},...,\frac{1}{\sqrt{d_{l+u}}}\right);$ (5.5) an iterative calculation formula F(t+1)=αBF(t)+(1−α)Y is generated, wherein, α∈(0,1) is a parameter;(5.6) the calculation formula is iterated to convergence to obtain a state

$F^{*}=\underset{t\to \infty }{\lim }F\left(t\right)=\left(1-\alpha \right){\left(M-\alpha B\right)}^{-1}Y;$ (5.7) a classification prediction of unmarked nodes γ_i=argmaxF_ij* is performed, wherein, l+1≤i≤l+u.

5. Performing a fine classification for target patents, that is determining the leaf node.

(1) Setting objects to be classified: the patents corresponding to node of each class divided in the step 4 are taken as a group, that means patents corresponding to node marked as γ_i∈ are a group, there are || groups.

(2) Extracting text information of patents: abstract, claims and description (hereinafter referred to as “all-text”) of each patent in each group are extracted, word segmentation of text information of patent is performed by using an existing tool, and a text set G={g₁, . . . , g_n} is generated, wherein d_i=(p_i1,p_i2,p_i3), p_i1, p_i2, and p_i3 are respectively word sequences obtained by word segmentation of the abstract, the claims, and the description of the i-th patent.

(3) Generating 4 text sets to be trained: G, G₁={p₁₁, . . . , p_n1}, G₂={p₁₂, . . . , p_n2} and G₃={p₁₃, . . . , p_n3}, which are respectively composed of word segmentation results of the all-texts, the abstracts, the claims, and the descriptions of the patents in the group.

(4) Text vectorization is performed.

(4.1) the text in the four text sets to be trained is vectored. In each text set to be trained, an element P=(t₁, . . . , t_m) is a segmented word sequence with m elements, t_i∈P is determined by w words t_{i, context}={t_i−w, . . . , t_i−2, t_i−1, t_i+1, t_i+2, . . . , t_i+w} before and after it, and by maximizing

$\frac{1}{m}\sum _{i=w}^{m-w}\log p\left(t_i❘t_{i,\mathrm{context}},\mathrm{pid}\right)$ wherein, the pid is a paragraph number of t_i in p,

$p\left(t_i|t_{i,\mathrm{context}},\mathrm{pid}\right)=\frac{e^{y_{t_i}}}{{\sum }_j{e}^{y_j}},$ γ_ti=b+∪Φ(t_{i, context},pid), U and b are parameters of softmax, and a vector corresponding to P is obtained by training data to be trained using a stochastic gradient descent method.

(4.2) A matrix of text is generated. The vectorization results of G={g₁, . . . , g_n}, G₁={p₁₁, . . . , p_n1}, G₂={p₁₂, . . . , p_n2} and G₃={p₁₃, . . . , p_n3} are supposed to be respectively H₁={h₁₁, . . . , h_n1}, H₂={h₁₂, . . . , h_n2}, H₃={h₁₃, . . . , h_n3}, and H₄={h₁₄, . . . , h_n4}, then a generated set of matrix of text of target patents is H={h₁, . . . , h_n}, wherein h_i=h_i1, h_i2, h_i3, h_i4).

(5) Patent classification is performed. Marked patents are set as S=∪_j=1^kS_j⊂H, wherein, S_j≠Ø is the marked patent of the j-th leaf node on the industry tree, j cluster centers of a k-means algorithm are initialized using the marked patents, and cluster membership of marked patents is not changed in a iterative updating process of clusters.

(6) The patent that does not belong to any industry of the leaf node industry on the tree is identified, in any leaf node classed in (5):

• • (6.1) a k distance of a patent p:
  • for a positive integer k, the k-th distance of the patent p is set as k−distance(o), and in patents divided into a leaf node on the industry tree, there is a patent o, and the distance between the patent o and the patent p is d(p, o). If the following two conditions are met, k−distance(o)=d(p, o):
  • {circle around (1)} in the leaf node, there are at least k patents q to make d(p, q)≤d(p, o);
  • {circle around (2)} in the leaf node, there are at most k−1 patents q to make d(p, q)<d(p, o);
  • (6.2) a k-th distance domain of the patent p:
  • a patent set whose distance from the patent p is ≤k−distance(o) is called the k-th distance domain of the patent p, and is record as N_k(p);
  • (6.3) a reachable distance of the patent p relative to the patent o: reachdist(p, o)=max{k−distance(o), ∥p−o∥};
  • (6.4) a local reachable density:

$lr{d}_k\left(p\right)=\frac{\left|N_k\left(p\right)\right|}{{\sum }_{o\in N_n\left(p\right)}{\mathrm{reach}}_d{\mathrm{ist}}_k\left(p,o\right)};$

• • (6.5) a local outlier factor:

${\mathrm{LOF}}_k\left(p\right)=\frac{{\sum }_{o\in N_n\left(p\right)}\frac{{\mathrm{lrd}}_k\left(o\right)}{{\mathrm{lrd}}_k\left(p\right)}}{❘N_k\left(p\right)❘};$

• • (6.6) if LOF(p) is greater than a threshold, it is thought that p is an outlier, and does not belong to the leaf node.

In order to better understand the present invention, the detailed description is made above in conjunction with the specific embodiments of the present invention, but it is not a limitation of the present invention. Any simple modification to the above embodiments based on the technical essence of the present invention still belongs to the scope of the technical solution of the present invention. Each embodiment in this specification focuses on differences from other embodiments, and the same or similar parts between the various embodiments can be referred to each other. As for the system embodiment, since it basically corresponds to the method embodiment, the description is relatively simple, and the relevant part can refer to the part of the description of the method embodiment.

Claims

1. An automatic industry classification method, comprising determining a scope of target patents, wherein the automatic industry classification method further comprises following steps: step 1: defining a target industry tree;step 2: generating marks on the target industry tree;step 3: performing a rough classification for the target patents by using the marks; andstep 4: performing a fine classification for the target patents according to a result of the rough classification,wherein step 1 further comprises:defining an industry tree I={[i1, . . . , ij, . . . , in}, wherein ij ∈ I and is a first level industry, j is a serial number of the first level industry, 1≤j≤n, and n is a number of all leaf nodes of I; andsetting ijkl . . .={ijkl, . . . , ijkl . . . t} as any non-leaf node of I, wherein degree of other nodes except the leaf nodes is greater than or equal to 2, k is a serial number of a second level industry, l is a serial number of a third level industry, and t is a serial number of a penultimate level industry:wherein the determining of the scope of the target patents is to manually determine the scope of the target patents to be classified,

2. The automatic industry classification method according to claim 1, wherein step 2 further comprises: according to resource constraints, determining a number p of patents to be marked, wherein p≥N, each leaf node of the target industry tree is marked with at least one patent belonging to the node, and N is a number of a last level industry.

3. The automatic industry classification method according to claim 1, wherein step 31 further comprises: defining one or more International Patent Classification (IPC) of each target patent as an IPC combination wherein all different IPC combinations of the target patents form the node set V.

4. The automatic industry classification method according to claim 3, wherein step 31 further comprises: taking an industry on a leaf node marked with patents as a classification of the leaf node,setting a number of marked nodes to be l,adjusting a sequence of the leaf nodes, wherein the marked nodes is adjusted to be the front, then 1≤i≤l; andverifying whether l<< a number of unmarked nodes u, and if not, adjusting the leaf nodes marked with patents, otherwise V={IPC1, . . . , IPCl, IPCl+1 . . . , IPCl+u}.

5. The automatic industry classification method according to claim 3, wherein the edge set E is a matrix, and a weight eij of edges between two vertices is a number of patents in a union IPCi ∪ IPCj of IPCs of the two vertices, wherein, eij is value in the matrix E.

6. The automatic industry classification method according to claim 1, wherein step 34 further comprises following sub-steps: step 341: generating a distance matrix S, wherein a calculation formula of the distance matrix S is Sij=∥ei−ej∥2, wherein, e; and e; are respectively an i-th row and a j-th row of the edge set E;step 342: generating the adjacency matrix W by using the distance matrix S.

7. The automatic industry classification method according to claim 1, wherein step 4 further comprises following sub-steps: step 41: setting objects to be classified;step 42: extracting text information of patents;step 43: generating text sets to be trained;step 44: performing text vectorization;step 45: performing patent classification; andstep 46: in any leaf node classed by the step 45, identifying a patent, wherein the patent does not belong to any industry of the leaf nodes on the target industry tree.

8. The automatic industry classification method according to claim 7, wherein step 41 further comprises taking patent nodes of each class divided in step 3 as a group, wherein patents corresponding to a patent node marked as γi ∈ γ are in the group, and there are |γ| groups;wherein step 42 further comprisesextracting an abstract, claims and a description of each patent in each group,performing word segmentation of text information of patent by using an existing tool, andgenerating a text set G={g1, . . . , gn}, wherein gi=(pi1, Pi2, Pi3), pi1, Pi2, and pi3 are respectively word sequences obtained by word segmentation of an abstract, claims, and a description of an i-th patent; and wherein the text sets to be trained comprise the text set G, a text set G1={p11, . . . , Pni}, a text set G2={p12, . . . , Pn2} and a text set G3={p13, . . . , Pn3}, the text set G, the text set G1, the text set G2, the text set G3 are respectively composed of word segmentation results of all-texts, abstracts, claims, and descriptions of the patents in the group.

9. The automatic industry classification method according to claim 8, wherein step 44 further comprises following sub-steps: step 441: vectoring a text in the text sets to be trained, wherein in each text set to be trained, an element P=(t1, . . . , tm) is a segmented word sequence with m elements, ti ∈ P is determined by w words ti, context={ti-w . . . , t1-2, ti-1, ti+1, ti+2, . . . , ti+w} before and after ti, and by maximizing

10. The automatic industry classification method according to claim 9, wherein step 45 further comprises setting marked patents as S=∪j=1kSj⊂H, wherein, Sj≠Ø is the marked patent of a j-th leaf node on the industry tree, andinitializing j cluster centers of a k-means algorithm using the marked patents, wherein a cluster membership of the marked patents is not changed in an iterative updating process of clusters.

11. The automatic industry classification method according to claim 7, wherein step 46 further comprises following sub-steps: step 461: calculating a k distance of a patent p, and setting a k-th distance of the patent p-as k-distance(o), wherein in patents divided into a leaf node on the industry tree, there is a patent o, and the distance between the patent o and the patent p isd(p, o);step 462: calculating a k-th distance domain of the patent p, wherein a distance of a patent set from the patent p is ≤k-distance (o), and the patent set is called the k-th distance domain Nk(p) of the patent p;step 463: calculating a reachable distance reachdist (p, o)=max {k-distance(o), ∥p-o∥} of the patent p relative to the patent o;step 464: calculating a local reachable density