A Patent Retrieval System Based on Neural Network

Abstract

The present application discloses a patent search system based on a neural network, which relates to the field of machine recognition technology, and includes: a data acquisition module, which is used to acquire product images and product titles; a first search module, which is used to match each word in the product title based on the acquired product title, and extract a first keyword related to the product; a second search module, which is used to extract a second keyword from the acquired first keyword, wherein the second keyword is a professional synonym in the product title identified by reordering the first keyword; a vocabulary search result module, which is used to obtain a first search result of the first keyword and a second search result of the second keyword in sequence according to the first keyword and the second keyword, and obtain a first matching result about the product title according to the obtained first search result and the second search result; and the accuracy of patent search is improved.

Description

Technical Field

[0001] The present invention relates to the field of machine recognition technology, and in particular to a patent retrieval system based on neural network. Background Art

[0002] Since online shopping has the limitation that buyers cannot touch the physical goods, buyers can often only select goods based on the product parameter information or product pictures provided by the merchant. Therefore, in order to attract consumers, merchants often put a large number of promotional terms or promotional pictures on the corresponding shopping pages to attract consumers to buy. Since consumers cannot try out the goods, it is difficult to distinguish the quality of the goods, and it is often difficult to make a decision when comparing and weighing multiple similar goods. At this time, consumers are often prone to choose lower-priced goods due to factors such as the price of the goods, which can easily lead to consumers purchasing some poor quality or counterfeit goods, affecting the consumer's shopping experience.

[0003] For example, Chinese patent publication number CN108984577A discloses a method and system for obtaining online product patents based on machine recognition, including: using a web crawler to automatically crawl product titles and pictures in shopping web pages, and performing word segmentation on the product titles, identifying the brand name and product name of the product as search factors, and retrieving the patent document corresponding to the product; then selecting the corresponding image library according to the product name, and using a convolutional neural network based on color features to match the template image with the highest similarity to the product from the image library, and then extracting the contour image of the product based on the contour features of the template image, and matching the contour image of the product with the drawings in the retrieved patent document to find the corresponding patent search results.

[0004] In the prior art, a specific product is identified by recognizing the overall edge contour of the image, and the corresponding color and shape of the product are also identified to determine whether the product has a patent. However, when searching, if the patent of the product cannot be determined only by the shape of the product, or if the patent is searched only by keywords related to the product, since the keywords set for the product tend to be commercial, their interpretations and related descriptions may be related to the original name of the product, but not completely consistent. In this case, it is difficult to accurately identify the product through similar images found by keywords. Summary of the invention

[0005] This application solves the technical problems of low retrieval efficiency, low accuracy and inability to fully utilize product image information for retrieval in the prior art by providing a patent retrieval system based on a neural network, thereby improving the accuracy and efficiency of patent retrieval.

[0006] This application provides a patent search system based on neural network, including:

[0007] Data acquisition module, used to obtain product images and product titles;

[0008] A first search module is used to match each word in the product title based on the obtained product title, and extract a first keyword related to the product;

[0009] A second search module is used to extract a second keyword from the acquired first keyword, where the second keyword is a professional synonym in the product title identified by reordering the first keyword;

[0010] A vocabulary search result module, used to obtain a first search result of the first keyword and a second search result of the second keyword in sequence according to the first keyword and the second keyword, and obtain a first matching result about the product title according to the obtained first search result and the second search result;

[0011] An image retrieval module is used to identify the six-sided product image corresponding to the product image based on the acquired product image, and obtain a second matching result about the product image according to the six-sided product image of the product image;

[0012] The search and evaluation module is used to determine the patent list to be output to the user based on the first matching result and the second matching result obtained.

[0013] One or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages:

[0014] Through the collaborative work of the data acquisition module, the first search module, the second search module, the vocabulary search result module, the image search module and the search evaluation module, a comprehensive search of product titles and images is achieved, making the search process more detailed and comprehensive; the first search module and the second search module extract and reorder the keywords in the product titles to identify more accurate professional synonyms, thereby improving the accuracy and professionalism of the search; the image retrieval module uses neural network technology to perform six-sided image recognition on product images, and searches according to the importance and overlapping scores of each view, effectively utilizing image information and improving the accuracy and efficiency of the search; the search evaluation module compares and comprehensively evaluates the title-based search results and the image-based search results, further improving the accuracy and reliability of the search results; in the comparison process, by comparing the loss value of the fused feature vector, the output patent list is dynamically adjusted, so that the final output result is more in line with the actual needs of users. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a system schematic diagram of a neural network-based patent retrieval system of the present invention. DETAILED DESCRIPTION

[0016] To facilitate the understanding of the present invention, the present application will be described more comprehensively below with reference to the relevant drawings; the drawings show preferred embodiments of the present invention, but the present invention can be implemented in many different forms and is not limited to the embodiments described herein; on the contrary, the purpose of providing these embodiments is to enable a more thorough and comprehensive understanding of the disclosed content of the present invention.

[0017] It should be noted that the terms “vertical”, “horizontal”, “up”, “down”, “left”, “right” and similar expressions used in this document are only for illustrative purposes and do not represent the only implementation method.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by technicians in the technical field to which the present invention belongs; the terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention; the term "and / or" used herein includes any and all combinations of one or more related listed items.

[0019] like Figure 1 As shown, the present application is a patent search system based on a neural network, comprising:

[0020] The data acquisition module is used to acquire product images and product titles. The acquired product images include but are not limited to all the product-related drawings acquired for sale and all the names marking the product.

[0021] The first search module is used to match each word in the product title based on the obtained product title, and extract the first keyword related to the product. The first keyword is obtained by matching the words in the product title with the product title to determine the word in the product title that is most relevant to the current title.

[0022] The second search module is used to extract the second keyword from the acquired first keyword, where the second keyword is a professional synonym in the product title identified by reordering the first keyword.

[0023] The vocabulary search result module is used to obtain the first search result of the first keyword and the second search result of the second keyword in turn according to the first keyword and the second keyword, and obtain the first matching result about the product title according to the first and second search results; the first matching result is used to determine the most relevant search results about professional vocabulary and product titles. The matching results obtained at this time are to limit the search to the corresponding labels and categories of the current product, so that when searching in combination with product pictures, the most relevant patents can be found, thereby obtaining more accurate recognition results.

[0024] The image retrieval module is used to identify the six-sided product image corresponding to the product image based on the acquired product image, and obtain a second matching result about the product image according to the six-sided product image of the product image.

[0025] In the acquisition process of the present invention, the first keyword and the second keyword are obtained by pre-marking the product title to obtain the corresponding first keyword and the second keyword, so as to obtain the corresponding first matching result according to the obtained first keyword and the second keyword.

[0026] The search and evaluation module is used to determine the patent list to be output to the user based on the first matching result and the second matching result obtained.

[0027] The first keyword is obtained by processing the original product title. After the original product title is segmented and stop words are removed, the recognized vocabulary is extracted to obtain the first keyword. The second keyword is obtained by comparing the first keyword with professional vocabulary. At the same time, the probability of the first keyword and the second keyword appearing together is determined based on the obtained vocabulary to select the patent document most similar to the current search.

[0028] The second keyword is an extension of the first keyword to find relevant professional synonyms. At the same time, according to the number of professional synonyms that can be identified for each word in the product title, the professional synonyms are sorted according to the number of times they are queried and the order set by the first keyword, so as to select the second keyword.

[0029] When obtaining the first matching result, a first search result is made based on the keywords in the product title, and a second search result is made based on the professional terms most similar to the product title. The first search result and the second search result are compared to determine the first matching result about the title.

[0030] The first matching result is obtained by comparing the text obtained in the first search and the second search with the patent document to determine the number of times each word appears in the patent document and the importance of the word in the corresponding search.

[0031] The calculation method for the first search result includes:

[0032] The occurrence probability of the first keyword in the preset document and the inverse document probability of the first keyword are obtained, and the first search result is obtained according to the occurrence probability of the first keyword in the preset document and the inverse document probability of the first keyword.

[0033] The probability of the first keyword appearing in the preset document is expressed as:

[0034] ;

[0035] Among them, is the occurrence probability of the first keyword in the preset document, is the first keyword in the preset document the number of occurrences; is the preset document the total number of all words in.

[0036] The inverse document probability of the first keyword is expressed as obtaining the total number of documents in the corresponding set of the preset document and the number of documents containing the first keyword, and obtaining the inverse document probability of the first keyword:

[0037] ;

[0038] Among them, is the inverse document probability of the first keyword, is the total number of documents in the corresponding set of the preset document, is the number of documents containing the first keyword.

[0039] Therefore, the representation of the first search result is:

[0040] ;

[0041] Among them, is represented as the first search result, is the occurrence probability of the first keyword in the preset document, is the inverse document probability of the first keyword.

[0042] The finally obtained first search result will reflect the importance of the first keyword used in the preset document, thereby being able to reflect the importance score of each word. According to the score obtained at this time, it is possible to know which vocabulary extracted from the product title is more important and which vocabulary should be selected for patent retrieval.

[0043] Assume the product title is: "New smart watch, waterproof and dustproof, suitable for sports and fitness". Tokenize the title into: "new", "smart", "watch", "waterproof", "dustproof", "suitable", "for", "sports", "fitness".

[0044] First, preprocess the recognized vocabulary to remove stop words: Remove "for" (assuming "for" is a common stop word).

[0045] Select the first keyword: Assume "smart", "watch", "waterproof", "dustproof", "sports", "fitness" are selected.

[0046] The calculation of the first keyword is: the assumed preset document set contains 100 documents, 20 of which contain the word "smart", 30 contain the word "watch", and so on.

[0047] Calculate TF (Term Frequency):

[0048] TF(t,d): Calculates the frequency of each keyword in the product title.

[0049] "Intelligence" appears once, and the total number of words in the title is 6, so TF (Intelligence, )=1 / 6.

[0050] "Watch" appears once, so TF(watch, )=1 / 6.

[0051] "Waterproof" appears once, so TF (waterproof, )=1 / 6.

[0052] "Dustproof" appears once, so TF (dustproof, )=1 / 6.

[0053] "Sports" appears once, so TF(Sports, )=1 / 6.

[0054] "Fitness" appears once, so TF(fitness, )=1 / 6.

[0055] Calculate IDF (Inverse Document Frequency)

[0056] IDF(t): Calculates the inverse document frequency of each keyword.

[0057] The document frequency of "smart" is 20, so IDF(smart) = .

[0058] The document frequency of "watch" is 30, so IDF(watch) = .

[0059] The document frequency of "waterproof" is 5, so IDF(waterproof) = .

[0060] The document frequency of "dustproof" is 3, so IDF(dustproof) = .

[0061] The document frequency of "sports" is 15, so IDF(sports) = .

[0062] The document frequency of "fitness" is 10, so IDF(fitness) = .

[0063] Calculate the first search result :

[0064] Assumptions , first calculate the TF-IDF value of each keyword, and then add ϵ .

[0065] TF1(intelligence)=0.087+0.001=0.088.

[0066] TF1(watch)=0.075+0.001=0.076.

[0067] TF1(waterproof)=0.115+0.001=0.116.

[0068] TF1(dustproof)=0.125+0.001=0.126.

[0069] TF1(motion)=0.094+0.001=0.095.

[0070] TF1(Fitness)=0.103+0.001=0.104.

[0071] From the calculation results, we can see that "dustproof" has the highest value, so "dustproof" may be one of the most important keywords when searching for patents. Based on these calculation results, we can know which words are more important and conduct patent searches accordingly.

[0072] In the present invention, the calculation method for the second search result is:

[0073] The first keyword and the second keyword are converted into word vectors, and a second search result is determined according to the word vectors of the first keyword and the second keyword.

[0074] The second search result is represented as:

[0075] ;

[0076] in, is the second search result, is the number of times the first keyword i and the second keyword j appear together, is the vector representation of the first keyword, is the vector representation of the second keyword, is the bias term for the first keyword i, is the bias term for the second keyword j, It is a weighting function used to assign different weights to word pairs containing the first keyword and the second keyword. Usually, low-frequency words are given higher weights to balance the influence of high-frequency words. The value range of i and j is from 1 to the number of all word pairs containing words i and j. It represents the difference between the word vector of the first keyword and the second keyword and the statistical information of the first keyword and the second keyword appearing together. T represents the transposition operation, that is, the word vector The transpose of a vector refers to the process of changing a row vector into a column vector, or changing a column vector into a row vector.

[0077] The weighting function for the setting will be adjusted according to the set hyperparameters.

[0078] Therefore It is expressed as:

[0079] ;

[0080] in, and is a hyperparameter, usually Set to 100, Set to 0.75.

[0081] When searching for product titles, a degenerate method is used to compare the words that appear in sequence, reducing the probability of using the word query each time. At the same time, the first search and the second search are merged based on the query degeneration method to obtain the most appropriate first matching result.

[0082] According to the obtained first search result and the second search result, the first matching result is determined as follows:

[0083] Get the keyword set corresponding to the maximum value of the first search result and the second search result.

[0084] According to the keyword set corresponding to the maximum value of the first search result and the second search result, the keyword sets of the first search result and the second search result are combined, the keyword sets are weighted and then merged, and the merged keyword set is output as the first matching result.

[0085] At this time, the corresponding contents of the first search result and the second search result are merged in a weighted manner to obtain a combination of the most appropriate keywords currently used, thereby reflecting the accuracy of patent retrieval from product titles.

[0086] For example, when the maximum value of the first search result is obtained, the keyword set is {A, B, C, D, E}, and the keyword set of the maximum value of the second search result is {C, D, E, F, G}. The keywords that appear are selected and recorded according to the number of occurrences of the keyword set. At this time, {C} or {D} or {E} appears more times, which means that it is more relevant to the currently queried patent. There is also a relationship between the other queried patent descriptions. At this time, this part is to output the keywords that take the maximum value by calculating the first search result and the second search result in sequence, so as to reflect the currently queried patent and improve the accuracy of the search.

[0087] When obtaining the second matching result, the corresponding processing means include:

[0088] The meaning of the six-sided view of the product is the images corresponding to the six sides of the currently acquired product, namely the front view, rear view, left view, right view, top view and bottom view; the acquired product image is divided into the form of the six-sided view of the product. If there is only one side of the product in the identified product image, the identified image will be searched for six sides, and the view of the corresponding position of the current product during the search will be determined. The two images will be compared to determine whether the currently identified patent image is the image of the current product.

[0089] If the current product image contains multiple faces, the acquired products will be sorted according to the acquired faces, and the search results corresponding to each face will be obtained. The search results of each face will be compared to determine whether the search is for the same patent. If there are multiple identified patents during the comparison, the weights will be set according to the features identified in each face, and each face will be searched in an interlaced manner, and the search results will be output as the second matching results.

[0090] When obtaining the second matching result, the product images are searched through staggered scheduling. For each product image obtained, after the product classification and type are determined according to the part of the product image and identification, cross-search is performed in turn according to the weight of the image in the picture. Finally, it is verified whether the retrieved patent is the same, and a detailed comparison is performed.

[0091] At the same time, in order to verify whether the retrieved image is the real corresponding image, it is also necessary to calculate the vectors of different dimensions in the image, perform contamination analysis on all image retrieval calculations, determine the differences corresponding to each feature retrieved, and identify the reasons that cause the image retrieval to differ from the current product. Use 8-gram overlap to score the examples and determine the impact on retrieval accuracy when retrieval recognition differences occur, thereby determining the features that can affect retrieval during retrieval.

[0092] The implementation method for staggered scheduling retrieval is:

[0093] According to the number of faces in the six-sided product image, the six-sided product image is divided into multiple views according to each face, and each view corresponds to a face in the six-sided product image. The initial features of each view in the six-sided product image in the product image are determined, and the initial features of each view are matched to obtain the relative importance of the initial features between each view. The initial features of each view are cross-compared according to the relative importance to obtain the relative importance of each view. The priority of each view is calculated based on the relative importance of each view, so as to determine which image is used as the basis when performing a search, and which image can be used to accurately identify the information of the patent that needs to be queried currently.

[0094] According to the priority of each view, the search is carried out in sequence to determine the image set after the search. The image set after the search is represented by the patent information set retrieved for each surface view after the current product image is divided into six surface views. At this time, the image set needs to judge the features between each image to determine the accuracy of the image when searching.

[0095] The relative importance of the initial features of each view is obtained by calculating the cosine similarity of the initial features in the acquired product images in advance, thereby determining the relative importance of the current view when performing retrieval.

[0096] The method of setting weights for the features recognized in each face is to obtain the initial number of features recognized in each face, and based on the relative importance of the initial features, obtain the relative importance of each view, that is, to obtain the weight of each face.

[0097] ;

[0098] in, is the relative importance of each initial feature, is the relative importance of each view, is the number of initial features, The value range is 1 to .

[0099] The priority of each view is expressed as

[0100] ;

[0101] in, Represents the priority of each view, For the The relative importance of each view, It is The weight coefficient of each view, is the number of views, The value range is 1 to .

[0102] Based on the retrieved image set, the overlap score between the image set and the product image is determined. The overlap score is used to express the similarity of the feature vectors between the product image and the retrieved patent image. The image set data with an overlap score greater than a preset score value is extracted as the second matching result output.

[0103] The overlap score is calculated by generating a feature series from the product image and the retrieved image set to determine the calculated overlap score.

[0104] ;

[0105] in, is the overlapping score, is the feature sequence of the product image, is the feature sequence of the retrieved image set.

[0106] When the image set data with an overlap score greater than a preset score value is extracted as the second matching result outputted, the processing method for the feature sequence in the image set further includes:

[0107] The feature sequence in the image set is taken as the first feature sequence. For each first feature sequence, the distance metric between the first feature sequences is determined. The set distance metric is determined using Manhattan distance. Manhattan distance is obtained by taking the difference of all features in the first feature sequence and summing up the differences of all features to obtain the distance metric between the first feature sequences.

[0108] The first feature sequence contains the feature sequence and , whose characteristics are and , the eigenvalues ​​are and ,feature and The Manhattan distance can be expressed as , at this time, the distance metric of the corresponding feature in the first feature sequence is the sum of the Manhattan distances of all features.

[0109] The average value and standard deviation of the distance measurement between the first feature sequences are calculated, and the correlation coefficient between the first feature sequences is calculated according to the average value and standard deviation of the distance measurement between the first feature sequences.

[0110] The correlation coefficient is expressed as the ratio of the covariance and standard deviation of the distance metric of the features in the first feature series; the covariance of the distance metric can be calculated by the mean and standard deviation.

[0111] The output second matching results are sorted according to the magnitude of the correlation coefficients, and the image set data having a correlation coefficient greater than a preset correlation coefficient is output.

[0112] At this time, the correlation coefficient is used to further limit the output image set data, so that the image set data finally outputted can reflect the existence of high correlation, thereby improving the accuracy of image retrieval.

[0113] The processing of the search evaluation module also includes comparing the first matching result and the second matching result obtained, and outputting the comparison result as a patent list to the user.

[0114] The first matching result and the second matching result are compared as follows:

[0115] Determine whether the retrieved image and title correspond to the same feature sequence. If the feature sequences after final comparison are not the same, adjust the comparison content of the first matching result and the second matching result.

[0116] The method for adjusting the comparison content of the first matching result and the second matching result is to obtain the text feature vector corresponding to the first matching result and the image feature vector corresponding to the second matching result, fuse the text feature vector corresponding to the first matching result and the image feature vector corresponding to the second matching result, obtain the fused feature vector, input the fused feature vector into the neural network, and determine the loss value of the fused feature vector.

[0117] The loss after alignment is expressed as;

[0118] ;

[0119] in, is the loss value of the fused feature vector, is the number of fused feature vectors, is the predicted value of the fused feature vector, is the value of the fused feature vector.

[0120] At the same time, when the text feature vector corresponding to the first matching result and the image feature vector corresponding to the second matching result are fused, based on the keyword set existing in the first matching result, the image feature vector identified in the second matching result is classified according to the obtained keyword set, and the text feature vector corresponding to the first matching result and the image feature vector corresponding to the second matching result are fused in sequence according to the sorting order of the keyword set.

[0121] After obtaining the loss value of the fused feature vector, compare the loss probability of the corresponding feature in the first matching result and the loss probability of the corresponding feature in the second matching result when the loss occurs. If the loss probability of the corresponding feature in the first matching result is greater than the loss probability of the corresponding feature in the second matching result, the second matching result is used as the basis, and the first matching result combination with the highest similarity to the second matching result is selected as the output patent list; if the loss probability of the corresponding feature in the first matching result is less than the loss probability of the corresponding feature in the second matching result, the first matching result is used as the basis, and the second matching result combination with the highest similarity to the first matching result is selected as the output patent list.

[0122] The calculation method of the loss probability of the corresponding feature in the first matching result and the loss probability of the corresponding feature in the second matching result is to calculate the ratio of the number of patents in the currently queried patent list that are completely inconsistent with the current search classification and content to the total number of patents.

[0123] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A patent search system based on neural network, characterized in that: include: Data acquisition module, used to obtain product images and product titles; A first search module is used to match each word in the product title based on the obtained product title, and extract a first keyword related to the product; A second search module is used to extract a second keyword from the acquired first keyword, where the second keyword is a professional synonym in the product title identified by reordering the first keyword; A vocabulary search result module, used to obtain a first search result of the first keyword and a second search result of the second keyword in sequence according to the first keyword and the second keyword, and obtain a first matching result about the product title according to the obtained first search result and the second search result; An image retrieval module is used to identify the six-sided product image corresponding to the product image based on the acquired product image, and obtain a second matching result about the product image according to the six-sided product image of the product image; A search evaluation module, used to determine a patent list output to a user based on the obtained first matching result and the second matching result; When obtaining the second matching result, the corresponding processing means include: According to the number of faces in the six-sided image of the product, the six-sided image of the product is divided into multiple views according to each face, each view corresponds to a face in the six-sided image of the product, the initial features of each view in the six-sided image of the product in the product image are determined, the initial features of each view are matched, the relative importance of the initial features of each view is obtained, the initial features of each view are cross-compared according to the relative importance, the relative importance of each view is obtained, and the priority of each view is calculated based on the relative importance of each view; According to the priority of each view, the retrieval is performed in sequence to determine the retrieved image set; Based on the retrieved image set, determine the overlap score between the image set and the product image, and extract the image set data whose overlap score is greater than a preset score value as the output second matching result; Get the number of initial features identified in each face, and based on the relative importance of the initial features, get the relative importance of each view: ； in, is the relative importance of each initial feature, is the relative importance of each view, is the number of initial features, The value range is 1 to ; The overlap score is calculated by generating a feature series from the product image and the retrieved image set to determine the calculated overlap score; ； in, is the overlapping score, is the feature sequence of the product image, is the feature sequence of the retrieved image set.

2. A patent search system based on a neural network as claimed in claim 1, characterized in that: The calculation method for the first search result includes: The occurrence probability of the first keyword in the preset document and the inverse document probability of the first keyword are obtained, and the first search result is obtained according to the occurrence probability of the first keyword in the preset document and the inverse document probability of the first keyword.

3. A patent search system based on neural network as claimed in claim 2, characterized in that: The probability of the first keyword appearing in the preset document is expressed as: ； in, is the probability of the first keyword appearing in the preset document, Is the first keyword In the default document The number of times it appears in It is a preset document The number of all words in ; The inverse document probability of the first keyword is expressed as: the total number of documents in the preset document corresponding set and the number of documents containing the first keyword are obtained to obtain the inverse document probability of the first keyword: ； in, is the inverse document probability of the first keyword, is the total number of documents in the collection corresponding to the preset document, is the number of documents containing the first keyword; The first search result is represented as: ； in, Represents the first search result, is the probability of the first keyword appearing in the preset document, is the inverse document probability of the first keyword, .

4. A patent search system based on neural network as claimed in claim 1, characterized in that: The calculation method for the second search result is: Convert the first keyword and the second keyword into word vectors, and determine a second search result according to the word vectors of the first keyword and the second keyword; The second search result is represented as: ； in, is the second search result, is the number of times the first keyword i and the second keyword j appear together, is the vector representation of the first keyword, is the vector representation of the second keyword, is the bias term for the first keyword i, is the bias term for the second keyword j, is a weighting function, and T represents a transpose operation.

5. A patent search system based on neural network as claimed in claim 1, characterized in that: According to the obtained first search result and the second search result, the first matching result is determined as follows: Obtain the keyword set corresponding to the maximum value of the first search result and the second search result; According to the keyword set corresponding to the maximum value of the first search result and the second search result, the keyword sets of the first search result and the second search result are combined, the keyword sets are weighted and then merged, and the merged keyword set is output as the first matching result.

6. A patent search system based on neural network as claimed in claim 1, characterized in that: The processing of the search evaluation module also includes comparing the first matching result and the second matching result, and outputting the comparison result as a patent list to the user; The first matching result and the second matching result are compared as follows: Determine whether the retrieved image and title correspond to the same feature sequence. If the feature sequences after final comparison are not the same, adjust the comparison content of the first matching result and the second matching result. The method for adjusting the comparison content of the first matching result and the second matching result is to obtain the text feature vector corresponding to the first matching result and the image feature vector corresponding to the second matching result, fuse the text feature vector corresponding to the first matching result and the image feature vector corresponding to the second matching result, obtain the fused feature vector, input the fused feature vector into the neural network, and determine the loss value of the fused feature vector.

7. A patent search system based on neural network as claimed in claim 6, characterized in that: After obtaining the loss value of the fused feature vector, compare the loss probability of the corresponding feature in the first matching result and the loss probability of the corresponding feature in the second matching result when the loss occurs. If the loss probability of the corresponding feature in the first matching result is greater than the loss probability of the corresponding feature in the second matching result, the second matching result is used as a basis to select the first matching result combination with the highest similarity to the second matching result as the output patent list; if the loss probability of the corresponding feature in the first matching result is less than the loss probability of the corresponding feature in the second matching result, the first matching result is used as a basis to select the second matching result combination with the highest similarity to the first matching result as the output patent list; The calculation method of the loss probability of the corresponding feature in the first matching result and the loss probability of the corresponding feature in the second matching result is to calculate the ratio of the number of patents in the currently queried patent list that are completely inconsistent with the current search classification and content to the total number of patents.

Images