A domain patent quality grade prediction method fusing knowledge information

Abstract

This invention relates to research on patent quality level prediction in the field of natural language processing. Its main steps are as follows: 1. Identifying efficacy phrases using a multi-feature efficacy word extraction model; 2. Extracting topic words from the patent text based on the Albert-BiLSTM model; 3. Clustering the extracted efficacy phrases and topic words using the K-means algorithm to manually construct a technology efficacy matrix, obtaining the corresponding technology efficacy and technology scale; 4. Quantifying or combining the structured digital information contained in the patent individually, combining it with long text to obtain 132 evaluation indicators, and training a transfer learning model using US patent data, while simultaneously expanding the Chinese dataset using active learning techniques; 5. Combining the technology efficacy matrix with the 132 indicators for transfer training and updating parameters to obtain the final prediction model. This invention effectively improves the accuracy of patent quality assessment.

Description

Technical Field

[0001] This invention relates to research on patent quality level prediction in the field of natural language processing, and particularly to a method for extracting patent keywords and efficacy phrases and constructing a technical efficacy matrix. Background Technology

[0002] A patent includes a title, abstract, and specification. Early assessments of patent quality were primarily based on single-dimensional indicators such as technology, law, and economics. However, judging a patent's quality comprehensively based on a single dimension is insufficient. Therefore, experts and scholars have begun to combine these three dimensions to create more evaluation indicators. Currently, there is no clear definition of patent quality evaluation indicators both domestically and internationally. However, it is certain that multi-dimensional, multi-level, and multi-indicator assessments provide a more comprehensive and accurate analysis of the patent text. Consequently, research is no longer limited to three dimensions but is increasingly incorporating more dimensions, including strategic value, patent activity, demand levels, stock value, and conversion rates. As research deepens, more and more experts are moving beyond quantitative indicators and integrating them with the patent text content. Thus, the combination and selection of key indicators are becoming increasingly multi-dimensional, and quality assessment standards are diversifying. How to more accurately identify the indicators influencing quality assessment results has become a hot topic requiring further in-depth research.

[0003] In fact, quantifying the indicators and related data for patent quality rating is quite difficult and generally hard to obtain. Evaluation work largely relies on a combination of manual statistics and subjective predictions from experts in the relevant field. This inherent subjectivity and uncertainty in patent quality rating predictions mean that results may vary depending on the perspectives of different experts. With rapid technological innovation and development, it's easy to see that machine learning can be used to autonomously learn the features inherent in patent texts. Currently, methods integrating machine learning for patent quality assessment mainly include the analytic hierarchy process (AHP), fuzzy comprehensive evaluation, logistic regression, and decision tree methods. In recent years, techniques from neural networks and deep learning, such as transfer learning and active learning, have also been gradually applied to patent rating prediction tasks, achieving some improvement in prediction performance while still having significant room for further improvement. Summary of the Invention

[0004] To address the aforementioned technical problems, the purpose of this invention is to propose a method for predicting the quality level of domain patents by integrating knowledge information, based on existing dimensions and utilizing deep learning and transfer learning techniques.

[0005] The present invention provides a method for predicting the quality level of domain patents by integrating knowledge information, comprising the following steps:

[0006] 1. Obtain the data required for the experiment. This mainly involves screening sentences in the patent abstract that contain efficacy phrases. Through training, an extraction model that integrates multiple features (radicals, Wubi input method, word length, and part of speech) is obtained and used to identify efficacy words in the patent text.

[0007] 2. The keywords in the patent title and patent abstract are annotated using a self-built thesaurus, followed by multiple rounds of manual proofreading. An extraction model is then trained based on the Albert-BiLSTM model to extract technical topics from the patent text.

[0008] 3. The extracted keywords and efficacy phrases are clustered using the K-means algorithm, and after further manual review and supplementation, a technical efficacy matrix is ​​finally constructed based on the patent text. The size of the new energy patents in the matrix will be used for subsequent evaluation of the quality level of patents in this field.

[0009] 4. Using a relatively mature US patent quality assessment model, US patents were translated into Chinese for training. A small number of Chinese patent texts with quality labels were used to fine-tune the model. At the same time, long texts and numerical indicators were quantified and combined into 132 indicators to train a transfer learning model.

[0010] 5. Knowledge information (technical efficacy matrix) is used as a new dimension indicator and combined with 132 other indicators for patent quality rating. That is, all indicators are vectorized or normalized and then the vectors are concatenated. The prediction model is trained on the basis of a good transfer learning model and the patent quality level is predicted on the test set.

[0011] This invention discloses a method for predicting the quality level of domain patents by integrating knowledge information. In step 1, patent data from the Chinese patent database is first analyzed, and corresponding crawling rules are designed. Patent data required for the experiment is obtained by searching for "new energy vehicles." After removing stop words from the crawled content, the efficacy phrases contained in the text are labeled and categorized according to features, mainly including: radicals, Wubi input method, word length, and part of speech. The labeled patent text is preprocessed, and word vectors are trained using BERT to learn the semantic features of the patent text. The vectorized patent data is used to train an attention-based BiLSTM-CRF model. Finally, the feature model with the best efficacy phrase recognition performance is selected for the extraction task.

[0012] This invention discloses a method for predicting the quality level of domain patents by integrating knowledge information. In step 2, firstly, relevant patent texts in the field of new energy vehicles are obtained using web crawling technology, and patent titles and abstracts are selected for text preprocessing. Secondly, the processed data is passed through an Albert pre-trained model layer to perform word embedding training on the input text. The vectorized data is then fed into a BiLSTM layer for encoding to train a topic word extraction model.

[0013] This invention discloses a method for predicting the quality level of domain patents by integrating knowledge information. In step 3, the K-means clustering algorithm is used to classify words with high similarity into the same category. The k cluster core words are then used as the focus of research on efficacy phrases and topic words. Finally, 11 representative efficacy phrases are obtained. Through manual data annotation and statistics, a technology efficacy matrix is ​​constructed with the technology topic as the x-axis and the efficacy phrases as the y-axis. The data in the matrix is ​​used to represent the scale of patent texts with different efficacy results under the technology topic.

[0014] This invention discloses a method for predicting the quality level of domain patents by integrating knowledge information. In step 4, US patents are used as source data, and Chinese patents are used as target data. Layers 3-14, which exhibit the best performance, are selected for transfer learning. A small number of Chinese patent texts with quality labels are used to fine-tune the model. Then, long text information and numerical information are quantitatively combined and incorporated into the quality assessment indicators, which are divided into six dimensions and 132 indicators. Transfer learning utilizes BERT to concatenate the text vector with the indicator vector to obtain a feature vector containing patent text information, thus obtaining an automatic patent quality assessment transfer model.

[0015] This invention discloses a method for predicting the quality level of domain patents by integrating knowledge information. In step 5, the constructed technical efficacy matrix is ​​used as a new knowledge information dimension and concatenated with 132 other vectors to form a new feature vector. Since US patent quality assessment is relatively mature, the concatenated vector is used as input to a transfer-based quality assessment model. After passing through a fully connected neural network consisting of 512, 128, and 32 nodes and a SoftMax layer, the quality level prediction result of the patent text is obtained, effectively improving the accuracy of the result.

[0016] Compared with existing technologies, the beneficial effects of this invention are as follows: In previous studies, the biggest challenge in constructing the technology efficacy matrix was the poor quality of the resulting matrix, leading to errors in patent analysis. The root cause lies in the accuracy of judging and extracting the technical themes and efficacy. This invention utilizes a multi-feature recognition model and an extraction model based on Albert embedding BiLSTM to identify efficacy words and theme words in patent texts, ensuring the quality and accuracy of the technology efficacy matrix in experiments and laying a solid foundation for its use as a new evaluation indicator. Research shows that multi-dimensional and multi-level indicators can more comprehensively and accurately analyze patent texts when evaluating patent quality. This invention is the first to use the technology efficacy matrix as a knowledge information dimension for patent quality level prediction. A larger matrix number indicates more technological achievements in the field, representing, to some extent, the saturation of inventions and innovations in that field. Conversely, a smaller matrix number indicates fewer patent inventions in the field, suggesting potential technical difficulties requiring urgent innovation or that the invention lacks practical significance. Combining the innovativeness, technological scale, and future development reflected in the technology efficacy matrix with other indicators plays a positive role in patent quality level prediction. Attached Figure Description

[0017] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0018] Figure 1 This is a flowchart of a method for predicting the quality level of domain patents that integrates knowledge information, according to the present invention.

[0019] Figure 2 This is a diagram comparing the results of different models used for patent quality level prediction and prediction models that integrate knowledge information. Detailed Implementation

[0020] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the invention and to fully convey the scope of the invention to those skilled in the art.

[0021] Figure 1 The present invention provides a method and flowchart for predicting the quality level of domain patents by integrating knowledge information, comprising the following steps:

[0022] 1. Obtain the dataset required for the experiment, and label the data. The embedding layer consists of four features: radical, Wubi input method, word length, and part-of-speech. Use Word2vec to obtain data vectors for each of the four features, and concatenate the vectors of different features to obtain the vector input for model training. Represent a sentence of the input text as a sequence s = {s1, s2, s3, ..., s...}. n-1 ,s n}∈Vs, where V s Representing a character set, vectorized using Word2vec:

[0023]

[0024] The vectorization process using the BERT model is as follows:

[0025]

[0026]

[0027] The patent text obtains the radicals corresponding to each Chinese character through the newly compiled Xinhua Dictionary lexicon. If the character is not included in the lexicon, the interface is called to perform a Baidu search. The obtained radicals are then mapped to vectors using Word2vec, as shown below.

[0028]

[0029]

[0030] The patent text is processed using the official Wubi conversion table to obtain the Wubi code corresponding to each character. Then, the obtained Wubi code is processed using Word2vec to obtain its vectorized representation.

[0031]

[0032] This paper uses part-of-speech tagging to obtain the part-of-speech tag for each character in the patent text, and then uses Word2vec to obtain its vectorized representation:

[0033]

[0034] At the same time, each character is labeled with the number of characters contained in its corresponding functional word, that is, the word length. Then, the obtained numerical code is used to obtain its vectorized representation using Word2vec.

[0035]

[0036] By using BiLSTM as the encoding layer and CRF as the decoding layer, a multi-feature-integrated efficacy phrase recognition model is trained.

[0037] 2. First, web scraping technology is used to obtain relevant patent texts in the field of new energy vehicles, and patent titles and abstracts are selected for text preprocessing. Next, the processed data is passed through an Albert pre-trained model layer to train word embeddings on the input text. The vectorized data is then fed into a BiLSTM layer for encoding, and the semantic features and contextual information of the input samples are learned through training the BiLSTM model.

[0038] 3. The steps for completing K-means clustering are as follows: First, perform basic data processing to determine the number of clusters (k) based on the number of data categories. Currently, the "Elbow" algorithm and the Canopy algorithm are commonly used to determine the number of clusters. Then, locate k points as initial cluster centers. Points near the cluster centers are assigned to the same cluster as the center by calculating distances, completing the first initialization of clustering. Next, new cluster centers are determined by calculating the average coordinates of each associated point. Euclidean distance and cosine similarity are commonly used to calculate the distance from a point to its center. This process is iterated continuously, updating the cluster centers, until the center points no longer change. At this point, the optimal result is obtained, the iteration ends, and clustering is complete. Finally, a technical effectiveness matrix is ​​constructed using manually labeled data.

[0039]

[0040]

[0041] 4. By comparing and analyzing patent data from China and the United States, the similarities and differences between the two countries' patents were summarized. Quantitative indicators were aligned, and a multi-task learning network model trained on US patents with patent quality level labels was applied to Chinese patents. The transfer process mainly involved three parts: selection of transfer components, cross-language transfer, and data augmentation using active learning. Simultaneously, each data item underwent in-depth processing to obtain quantitative indicators related to patent quality. The experiment quantified these indicators across several key dimensions, including time, quantity, technology, legal, and inventor / agent dimensions, quantifying these dimensions individually or in combination.

[0042] 5. First, the constructed technical efficacy matrix is ​​used as a new knowledge information dimension and concatenated with other vectors to form a new feature vector. Since US patent quality assessment is relatively mature, the concatenated vector is used as input to a transfer-based quality assessment model. After passing through a fully connected neural network consisting of 512, 128, and 32 nodes and a SoftMax layer, the quality level prediction result of the patent text is obtained, effectively improving the accuracy of the result.

[0043] Example 1:

[0044] The experimental results in this embodiment are obtained by manually annotating a dataset based on the content of historical engineering consulting reports provided by an engineering consulting company and extracted title data, and then testing on this dataset. The experimental results demonstrate the technical effects of this invention as follows:

[0045] Figure 2 This diagram illustrates a comparison of results between different models used for patent quality level prediction and prediction models incorporating knowledge information. (The diagram shows the results of these models.)

[0046] To facilitate comparison of prediction results from different models and enhance readability, each model is defined as follows:

[0047] Model_1: A model that uses the Support Vector Machine (SVM) classifier in machine learning to predict the quality level of patents by combining text vectors and 132 quantitative indicators.

[0048] Model_2: A quality level prediction model trained using BiLSTM, which uses a concatenation of text vectors and vectors of 132 quantization metrics as input.

[0049] Model_3: A quality level prediction model trained using BiLSTM, which uses text vectors concatenated with vectors of 132 quantification indicators and knowledge mining dimensions (efficacy phrases, technical topic terms) proposed in the experiment as input.

[0050] Model_4: A quality level prediction model trained using BiLSTM, which uses text vectors concatenated with 132 quantification indicators and the vectors of the knowledge information dimension (technical efficacy matrix) proposed in the experiment as input.

[0051] Model_5: It uses a concatenation of text vectors and vectors of 132 quantification metrics as input, and utilizes a transfer learning model trained on US patent texts to Chinese patent texts.

[0052] Model_6: It uses a concatenation of text vectors with 132 quantitative indicators and knowledge mining dimensions (efficacy phrases, technical keywords) proposed in the experiment as input, and utilizes a transfer learning model trained on US patent texts to Chinese patent texts.

[0053] Model_7: It uses a concatenation of text vectors with 132 quantitative indicators and the vectors of the knowledge information dimension (technical efficacy matrix) proposed in the experiment as input, and utilizes a transfer learning model trained on US patent texts to Chinese patent texts.

[0054] The knowledge information dimension, or technical efficacy matrix, reflects the technological scale within the field in which the patent operates. It clearly indicates the technological focus and development trends of the relevant field. Through this matrix, one can further understand the patent application situation under a specific sub-category of a professional field, as well as the degree of patent innovation and saturation in that field. Figure 2 Experimental results show that using the generated matrix as a new evaluation metric, combined with other key indicators, in patent quality grading research has a positive effect on model training and effectively improves the accuracy of prediction tasks.

[0055] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting the quality grade of a field patent by fusing knowledge information, characterized by, Comprise the following steps: (I) obtain the data required for the experiment, screen the sentences involving efficacy phrases in the patent abstract, and obtain the extraction model by training the fusion of multiple features including radical, five-pin, word length and word features for the recognition of efficacy words in patent text; (II) For the subject words in the patent title and patent abstract, a round of annotation is carried out through the self-built word library, and multiple rounds of manual correction are carried out. The extraction model is obtained by training based on Albert-BiLSTM model for extraction of technical subjects in patent text; (III) The extracted subject words and efficacy phrases are clustered using K-means algorithm, and are manually reviewed and supplemented to construct a technology and efficacy matrix, wherein the size of the technology and efficacy matrix in the new energy patent field is used for subsequent patent quality level evaluation; (IV) Use the US patent quality evaluation model to translate US patents into Chinese for training, and fine-tune the model using a small number of Chinese patent texts with quality labels; At the same time, after quantifying and combining the long text and the digital indicators respectively, 132 indicators are divided, and a transfer learning model is trained; (V) Combine the technology and efficacy matrix as a new dimension indicator with the 132 indicators to evaluate the quality of the patent, including vectorizing or normalizing all indicators, then vector splicing, and training the quality level prediction model on the transfer learning model, and completing the prediction of the quality level of the patent on the test set.

2. The method of claim 1, wherein the method of predicting the quality grade of the field patent by fusing the knowledge information is characterized by: By analyzing the efficacy phrase features contained in the patent text, the characteristics of the efficacy phrase in the aspects of character structure, pronunciation, word color, word root, semantic relationship and word length are summarized, and are summarized as radical features, five-pin coding features, word features and word length features. These multiple features are fused to extract the efficacy phrase, laying a foundation for subsequent construction of technology and efficacy matrix.

3. The method of claim 2, wherein the method further comprises: determining the quality level of the patent based on the knowledge information. In view of the problem of large model parameter quantity and slow training speed of the subject word extraction model, the patent text is vectorized using Albert pre-training, which improves the accuracy of the model while improving the performance of the model by reducing the number of parameters, so that the model can be applied in the extraction task. Lay a foundation for the construction of technology and efficacy matrix in the future.

4. The method of claim 3, wherein the method further comprises: determining the quality level of the patent based on the knowledge information. The technology and efficacy matrix is constructed by taking the technical subject words as the horizontal coordinate and the efficacy phrases as the vertical coordinate. According to the data analysis of the technology cluster point and the blank point in the technology and efficacy matrix, the technology cluster point represents the hot field, and the technology blank point represents the future technology innovation direction. By introducing the technology and efficacy matrix as a knowledge information dimension into the patent quality evaluation index, the prediction effect of the patent quality level is improved.

5. The method of claim 4, wherein the method further comprises: determining the quality level of the patent based on the knowledge information. In the patent quality evaluation, the knowledge information dimension, i.e. the technology and efficacy matrix, is added to judge the quality level of the patent from the perspective of patent technology scale and functional utility, and based on the transfer learning model, a quality level prediction model is trained to finally improve the prediction effect.

Images