Information processing device, information processing method and computer program

Abstract

To provide a similar document retrieval system capable of enhancing retrieval accuracy of a retrieval document and documents to be retrieved in retrieving documents to be retrieved having high similarity to a retrieval sentence.SOLUTION: According to the present disclosure, a similar document retrieval device acquires a retrieval sentence and documents to be retrieved including a sentence to be retrieved having a paragraph number and a paragraph unit, calculates sentence vectors of the retrieval sentence and the sentence to be retrieved, calculates similarity between the sentence vector of the retrieval sentence and the sentence vector of the sentence to be retrieved, and acquires a document having high similarity including a sentence to be retrieved having high similarity on the basis of the calculated similarity.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present disclosure relates to an information processing device, an information processing method, and a computer program for searching documents. [Background technology]

[0002] It has been common practice to search for documents containing target information among documents stored in a database. For example, to find out whether a certain technology is already known, a database containing accumulated documents is searched using a search formula or the like, and documents describing the target technology or similar technologies are extracted. It is also known to use machine learning models to assist in the search of documents stored in a database.

[0003] Patent Document 1 discloses a system that automatically generates a summary of a document to be summarized (e.g., a publicly known document) that reflects the reference information when searching for publicly known documents using application documents (e.g., a specification) that are the subject of patent examination as reference information. In the learning process of a neural network to which the reference document and the document to be summarized are input, the system divides each of the reference document and the document to be summarized into blocks of paragraph units, calculates a value indicating the importance of the user's research perspective for each block of the reference document based on the search query, and calculates a score indicating the relevance of the user's research perspective for each block of the document to be summarized based on the value.

[0004] Patent Document 2 discloses a technology relating to a machine learning-based system, such as a neural network-based system, for searching, comparing, or analyzing documents containing natural language, in which pairs of claim blocks and description blocks are used as training data sets to form a neural network model trained for the purpose of patent search, and the vector angle between the claim graph of a patent application and the description graph of a patent document used as a novelty barrier thereto is minimized or the similarity score is maximized.

[0005] Patent Document 3 discloses a technology relating to concept base generation or concept search for searching for search target documents that conceptually match a search query entered by a user, in which a search target document X is updated and the set of words in the updated search target document X is made to include the set of words in the corresponding search query p, so that when a new search query g that is conceptually close to the corresponding search query p is entered in a search means, the conceptual vector of the search target document X is made closer to the conceptual vector of the new search query g than before the update. [Prior art documents] [Patent documents]

[0006] [Patent Document 1] Patent Publication No. 2021-174251 [Patent Document 2] Japanese Patent Publication No. 2022-513353 [Patent Document 3] Japanese Patent Application Publication No. 2017-072884 Summary of the Invention [Problem to be solved by the invention]

[0007] In the prior art, to generate a summary of a document to be summarized, a neural network is input with a reference document and the document to be summarized, each divided into blocks of paragraphs, and a learning process is performed. However, no consideration has been given to using this neural network to search documents divided into blocks of paragraphs. Furthermore, while machine learning is performed on a neural network base using pairs of claim blocks and description blocks, training using pairs other than claims and descriptions is unknown. Furthermore, concept base generation or concept search is performed by updating the data of the document to be searched to include each of the search queries, but the application of machine learning to concept base generation or concept search has not been disclosed.

[0008] When searching for target documents using a search phrase, there is a need to accurately search for similar documents by comparing the search phrase with the details described in the target document. Furthermore, when searching for similar documents using a machine learning learning model, there is a need to improve the estimation accuracy of the learning model.

[0009] The present disclosure has been made in consideration of the above circumstances, and aims to provide an information processing device, an information processing method, and a computer program that, when searching for documents that are highly similar to a search text, improves the accuracy of searching for similar documents by taking into account the sentence structure of the document to be searched, and enables effective use of a learning model. [Means for solving the problem]

[0010] An information processing device of a first aspect of the present disclosure is an information processing device for searching for documents that are highly similar to a search text, and is characterized by comprising: a search text acquisition unit that acquires the search text; a searched text acquisition unit that acquires searched documents including paragraph numbers and searched text in paragraph units; a text vector calculation unit that calculates text vectors of the search text and the searched text; a similarity calculation unit that calculates the similarity between the text vector of the search text calculated by the text vector calculation unit and the text vector of the searched text calculated by the text vector calculation unit; and a high-similarity document acquisition unit that acquires high-similarity documents that include highly similar searched text based on the similarity.

[0011] The information processing device of the second aspect of the present disclosure further includes a searchable sentence creation unit that concatenates searchable sentences of consecutive paragraphs so that the number of tokens is within a predetermined number, and creates searchable sentences of normalized paragraphs, and the searchable sentence acquisition unit acquires the searchable sentences of the normalized paragraphs.

[0012] An information processing device of a third aspect of the present disclosure is characterized in that the searchable sentence creation unit concatenates the searchable sentences of consecutive paragraphs so that the boundary of the normalized paragraph is approximately at the center and so that the number of tokens is within the specified number, thereby creating a searchable sentence of a normalized auxiliary paragraph, and the searchable sentence acquisition unit acquires the searchable sentence of the normalized auxiliary paragraph.

[0013] In the information processing device of a fourth aspect of the present disclosure, the sentence vector calculation unit calculates sentence vectors of the search sentence and the to-be-searched sentence in advance.

[0014] The information processing device of the fifth aspect of the present disclosure is characterized in that the search sentence is one or more sentences among sentences described in the claims, the problem that the invention aims to solve, the means for solving the problem, the gist of the form for implementing the invention, the effects of the invention, or the summary of the invention.

[0015] The information processing device of the sixth aspect of the present disclosure is characterized in that the gist of the form for implementing the invention is to find multiple paragraphs of text described in the form for implementing the invention that are highly similar to the text of the claims by concatenating them so that they are within a predetermined number of tokens.

[0016] The information processing device of the seventh aspect of the present disclosure is characterized by further comprising a multiple search unit that performs searches using multiple search sentences, and a composite processing unit that acquires highly similar documents based on the results of the search by the multiple search unit.

[0017] An information processing device of an eighth aspect of the present disclosure is characterized in that the composite processing unit searches using the multiple search unit using the multiple search sentences, obtains document numbers of a predetermined number of searched documents in order of highest similarity for each of the multiple search sentences, assigns evaluation points to the document numbers according to their rankings of similarity, adds up the evaluation points for each document number, and obtains searched documents with high added evaluation points as highly similar documents.

[0018] In the information processing device of a ninth aspect of the present disclosure, the composite processing unit assigns the evaluation points based on different weights for each of the plurality of search sentences.

[0019] An information processing device according to a tenth aspect of the present disclosure is characterized in that the text vector calculation unit uses a text vector calculation model that is fine-tuned using patent document data. [Effects of the Invention]

[0020] According to the first aspect of the present disclosure, by focusing on paragraph units having coherent content in a sentence structure and calculating the similarity between a search sentence and a sentence in paragraph units described in a searched document, the accuracy of similarity determination can be improved and the accuracy of searching for similar documents can be increased. Furthermore, when sentences are structured in paragraph units, such as in patent documents, the unique structure of the document can be utilized in searching for similar documents.

[0021] According to the second and tenth aspects of the present disclosure, even when there is a limit to the number of words (tokens) that can be input into a learning model that outputs sentence vectors of a search sentence and a searched sentence, by setting a normalized paragraph, it is possible to increase the amount of information in the searched sentence that can be input within the limit, thereby improving the estimation accuracy and utilization efficiency of the learning model.

[0022] According to the third aspect of the present disclosure, even if a coherent description is broken up by generating a normalized paragraph, the broken up description can be complemented by setting a normalized auxiliary paragraph, thereby further improving the estimation accuracy of the learning model.

[0023] According to the fourth aspect of the present disclosure, by performing the time-consuming calculation of sentence vectors in advance, search results can be quickly presented when a user searches for similar documents.

[0024] According to the fifth aspect of the present disclosure, by selecting a portion of the search document that includes features of the invention as the search sentence, it is possible to efficiently search for searched documents that are similar to the invention.

[0025] According to the sixth aspect of the present disclosure, since the matters described in the form for implementing the invention among the search texts are expected to be long, by extracting the parts that are highly similar to the invention and making a similarity judgment, it is possible to efficiently search for searched documents that are similar to the invention.

[0026] According to the seventh aspect of the present disclosure, by setting multiple descriptions related to the features of the invention as multiple search sentences, it is possible to evaluate the similarity of the searched documents from different perspectives, thereby enabling accurate searches for similar documents.

[0027] According to the eighth and ninth aspects of the present disclosure, by assigning evaluation points to the retrieved similar documents in descending order of similarity, it is possible to easily determine the ranking of highly similar documents for each search sentence. Furthermore, by weighting the evaluation points for each of multiple search sentences, it is possible to obtain a judgment result that reflects the importance of the search sentence. [Brief explanation of the drawings]

[0028] [Figure 1] FIG. 1 is a block diagram showing an example of the overall configuration of a similar document search system 1 to which information processing according to the first embodiment is applied. [Figure 2] FIG. 2 is a block diagram showing an example of the functional configuration of the similar document search device 10 of the first embodiment. [Figure 3] FIG. 3 is a diagram showing an example of the function of the text vector calculation unit 203 in the first embodiment. [Figure 4] FIG. 4 is a block diagram showing an overview of the similarity calculation unit 204 and the high-similarity document acquisition unit 205 of the first embodiment. [Figure 5] FIG. 5 is a flowchart illustrating the operation of the similar document search device 10 of the first embodiment. [Figure 6] FIG. 6 is a diagram for explaining the normalization process of the second embodiment. [Figure 7] FIG. 7 is a flowchart illustrating the operation of the to-be-searched sentence creation unit 206 of the second embodiment. [Figure 8] FIG. 8 is a block diagram showing an example of the functional configuration of the text vector calculation unit 203 according to the third embodiment. [Figure 9] FIG. 9 is a diagram for explaining the operation of the search sentence acquisition unit 201 of the fourth embodiment. [Figure 10] FIG. 10 is a block diagram showing an overview of a multiple search unit 1010 and a combined processing unit 1020 according to the fifth embodiment. [Figure 11] FIG. 11 is a block diagram showing an outline of the composite processing unit 1020 of the sixth embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0029] Hereinafter, several embodiments of the present disclosure will be described in detail with reference to the drawings. In the following description, configurations and elements that are the same as or similar to configurations that have already been described will be assigned the same reference numerals and description thereof will be omitted. In the following embodiments, a case where a patent document is searched will be described as a non-limiting example, but the present disclosure may also be applied to a case where a document other than a patent document is searched.

[0030] First Embodiment (1) Overall structure of the similar document search system 1 FIG. 1 is a bird's-eye view showing an example of the overall configuration of a similar document search system 1 to which an information processing device, an information processing method, and a computer program according to a first embodiment of the present disclosure are applied.

[0031] The similar document search system 1 shown in FIG. 1 includes a similar document search device 10 and one or more search terminals 11, 12, etc. (since search terminals 11, 12, etc. are terminals with the same functions, only search terminal 11 will be described below), and similar document search device 10 is communicably connected to search terminal 11. Although FIG. 1 illustrates similar document search device 10 and search terminal 11 as separate devices, these devices may be configured as a combined device, or each device may be further configured as multiple devices. Furthermore, these devices may be communicably connected via a network.

[0032] The similar document search device 10 is an information processing device for searching for searched documents that contain searched sentences that are highly similar to a search sentence. Here, the search sentence is a sentence that is contained in the searched document and contains the target information, i.e., the information that is to be searched. The searched document is a document that is the object of search, and is usually stored in a database and can be searched by a query, etc. The searched document is a document that includes paragraph numbers and sentences in paragraph units, and one or more sentences in paragraph units with paragraph numbers corresponds to one searched sentence.

[0033] If the searched document is a patent document, the specification includes paragraph numbers and paragraph-based sentences, with the paragraph numbers consisting of black brackets and numbers. The claims include claim numbers and paragraph-based sentences, with the claim numbers consisting of black brackets, letters, and numbers. The abstract includes paragraph-based sentences separated by [Problem] and [Means for solving the problem] (hereinafter, black brackets will be written as [ ]). In this disclosure, paragraph numbers include those expressed in black brackets and numbers or letters, such as [paragraph number], [claim number], [Problem], and [Means for solving the problem]. Paragraph numbers may also be composed of numbers or letters that identify sentences separated into arbitrary blocks.

[0034] The similar document search device 10 acquires a search sentence from a search document, acquires one or more search sentences from searched documents to be searched, and calculates a sentence vector for each of the search sentence and one or more search sentences. A sentence vector is a numerical representation of the characteristics of a sentence in natural language.

[0035] The similar document search device 10 calculates the similarity between the sentence vector of the search sentence and the sentence vector of the sentence to be searched. The similarity represents the degree of similarity between the items described in the search sentence and the items described in the sentence to be searched. The similar document search device 10 calculates the similarity between the sentence vector of the search sentence and the sentence vector of the sentence to be searched for all combinations of the sentence vector of the search sentence and the sentence vector of each sentence to be searched. Methods such as Euclidean distance and cosine similarity can be used to calculate the similarity.

[0036] The similar document search device 10 acquires searched documents that have a high similarity to the searched text as highly similar documents based on the similarity between the text vectors. The similar document search device 10 sorts the calculated similarities for all combinations of the text vectors of the searched text and the text vectors of the searched text in descending order, determines the document numbers to which the searched text belongs, eliminates duplicates, and extracts highly similar documents. Alternatively, all the calculated similarities may be aggregated for each searched document to extract the similarity of the searched text contained in each searched document, and the similarity for each searched document may be aggregated to obtain the similarity for each searched document (hereinafter referred to as "total similarity"). The similar document search device 10 may extract highly similar documents from the searched documents in descending order of total similarity.

[0037] The search terminal 11 is an information processing device operated by a searcher who searches for documents or a person involved in the document search (hereinafter also referred to as a "user"). When the search terminal 11 is started up by a user's operation, it is communicatively connected to the similar document search device 10 and displays a user interface for searching transmitted from the similar document search device 10 on a display device. The user issues a search instruction to the similar document search device 10 via the user interface displayed on the display device 115, and the search results from the similar document search device 10 are displayed.

[0038] (2) Hardware configuration The similar document search device 10 includes a memory 102 that stores a computer program, a processor 101 that executes the computer program stored in the memory, a storage 103, a communication interface 104, and an input / output interface 105. The processor 101 is typically a central processing unit (CPU) and / or a graphics processing unit (GPU), but may also be a microcomputer, a field programmable gate array (FPGA), a digital signal processor (DSP), or the like. The memory 102 temporarily stores programs executed by the processor 101 and data used by the processor. The storage 103 is a storage device also called an external storage device, and may include, for example, a non-volatile storage medium such as a hard disk drive (HDD) or a solid state drive (SSD).

[0039] The search terminal 11 includes a processor 111, a memory 112, a storage 113, an input device 114, a display device 115, and a communication interface 116. The input device 114 is a device operated by, for example, a user to input various operational commands to the similar document search device 10, and may include, for example, a mouse, a keyboard, a touch panel, a microphone, etc. The display device 115 includes a flat display device such as a liquid crystal display device or an organic EL display device to display the user interface output from the similar document search device 10.

[0040] (3) Functional configuration Fig. 2 is a block diagram showing an example of the functional configuration of the similar document search device 10 according to the first embodiment. The similar document search device 10 executes a computer program stored in a memory to function as a search management unit 20, a storage unit 21, and a communication unit 22. The communication unit 22 controls communication between the similar document search device 10 and the search terminal 11. The to-be-searched sentence creation unit 206 in Fig. 2 will be described in detail in the second embodiment.

[0041] The memory unit 21 includes a document accumulation unit 211 for storing data processed by the search management unit 20. The document accumulation unit 211 accumulates search documents that include search sentences and searched documents that include searched sentences. The search documents may be specified by the user each time a search is performed, but if multiple searches are planned, the search documents and searched documents to be used for the multiple searches may be stored in the memory unit 21 in advance.

[0042] Patent documents include published publications such as unpublished patent gazettes, published patent gazettes, and patent gazettes, as well as unpublished patent applications. Patent documents are divided into sections by inked brackets and document titles, such as [Claims], [Specification], and [Abstract]. Claims are divided into inked brackets and claim numbers, such as [Claim 1], and specifications are composed of sentences divided into inked brackets and numbers, such as

[0001] .

[0043] The search management unit 20 functions as a search sentence acquisition unit 201, a searched sentence acquisition unit 202, a sentence vector calculation unit 203, a similarity calculation unit 204, and a high similarity document acquisition unit 205 by the processor 101 executing a computer program stored in the memory 102.

[0044] (3-1) Search sentence acquisition unit 201 The search text acquisition unit 201 acquires from the user the document number of the search document containing the target information, and acquires the information of the search document corresponding to the document number from the document accumulation unit 211. The search text acquisition unit 201 acquires text data of the search text from the acquired information of the search document. When the search document is a patent document, the search text corresponds to the matters described in the entire [claims], [specification], and [abstract], for example. Note that the search text acquisition unit 201 may acquire text data of the search text.

[0045] (3-2) Searched text acquisition unit 202 The searched text acquisition unit 202 extracts searched documents from the document accumulation unit 211. If the searched documents are classified and a search range for the searched documents is specified based on the classification, the searched documents that satisfy the specified search range are extracted. Alternatively, the searched documents can be searched using a search query and the retrieved documents can be acquired. The searched text acquisition unit 202 acquires paragraph numbers and paragraph-by-paragraph text from the extracted searched documents. In the case of patent documents, like searched documents, they include specifications, claims, and abstracts, and each document consists of paragraph numbers and paragraph-by-paragraph text. A searched text is one or more paragraphs with paragraph numbers. Searched texts may be grouped by category or theme, assigned unique searched text numbers, and a list may be created containing source data such as the corresponding searched document numbers and paragraph numbers.

[0046] (3-3) Text Vector Calculation Unit 203 The sentence vector calculation unit 203 calculates sentence vectors of the search sentence and the sentence to be searched. Known techniques can be used to calculate sentence vectors, but for example, each sentence can be broken down into morphemes to obtain tokens (abbreviated words) (Tokenizer), these can be embedded in a multidimensional vector space to obtain token vectors (Embedding), and further the vectors of each token can be aggregated using an average value or the like to obtain sentence vectors (Pooling).

[0047] To obtain sentence vectors, a sentence vector calculation model fine-tuned with patent document data can be used. For example, Sentence BERT (Bidirectional Encoder Representations from Transformers) fine-tuned with patent document data can be used as a sentence vector calculation model fine-tuned with patent document data. That is, by inputting two sentences to Sentence BERT, it is possible to have Sentence BERT output the respective sentence vectors.

[0048] 3 is a diagram showing an example of the function of the sentence vector calculation unit 203, illustrating the calculation of sentence vectors between one search sentence and n searched sentences 1 to n. The sentence vector calculation unit 203 is configured with a machine-learned learning model consisting of, for example, an input layer, an intermediate layer (hidden layer), and an output layer. The learning model is a trained learning model that has been trained so that when the search sentence and the searched sentences 1 to n of the search literature text are input, the learning model outputs the sentence vector u of the search sentence and the sentence vectors v1 to vn of the searched sentences.

[0049] 3 is the search document text described in the search document, and in the case of a patent document, it is composed of the text described in the description, claims, and abstract. On the other hand, the searched sentences are the texts described in the searched documents 1 to 4, and are composed of the number of searched sentences (n) that are the number of paragraphs contained in the searched documents 1 to 4. If the searched documents 1 to 4 are patent documents, each searched document is composed of a description, claims, and abstract, just like the searched document.

[0050] (3-4) Similarity calculation unit 204 4 is a block diagram showing an overview of the similarity calculation unit 204 and the high-similarity document acquisition unit 205. The similarity calculation unit 204 receives the sentence vector u of the search sentence output by the sentence vector calculation unit 203 and the sentence vectors v1 to vn of the to-be-searched sentences 1 to n. The similarity calculation unit 204 calculates the similarity between the sentence vector u of the search sentence output by the sentence vector calculation unit 203 and the sentence vectors v1 to vn of the to-be-searched sentences. For example, cosine similarity can be used as the similarity. The similarity calculation unit 204 outputs similarities 1 to n for each to-be-searched sentence as the calculation results of the similarity between the sentence vector u of the search sentence and each of the sentence vectors v1 to vn of the to-be-searched sentences.

[0051] The sentence vector calculation unit 203 may calculate sentence vectors for a huge number of sentences to be searched. For example, in the case of patent documents, there are 100,000 documents to be searched, and the total number of paragraphs is 8 million, so it is expected that it will take several hours to calculate sentence vectors of 768 dimensions. Therefore, it is preferable to calculate sentence vectors of sentences to be searched in advance.

[0052] (3-5) High similarity document acquisition unit 205 The high-similarity document acquisition unit 205 acquires similarities 1 to n for each searched sentence from the similarity calculation unit 204 and acquires searched documents with high similarity as high-similarity documents. The high-similarity document acquisition unit 205 sorts the similarities 1 to n between the searched sentence and the searched sentences 1 to n in descending order of similarity, obtains the document number to which each searched sentence belongs from the source data, and extracts highly similar documents by eliminating duplicates. Alternatively, the high-similarity document acquisition unit 205 may aggregate the similarities 1 to n between the searched sentence and the searched sentences 1 to n for each of the searched sentences 1 to 4 and calculate a total similarity for each of the searched sentences 1 to 4. For example, since searched document 1 is composed of searched sentence 1 and searched sentence 2, the similarity 1 value between the searched sentence and searched sentence 1 and the similarity 2 value between the searched sentence and searched sentence 2 are added together to calculate the total similarity 1 value for searched document 1. Similarly, the total similarities 2 to 4 are calculated for searched documents 2 to 4. If the calculated total similarity values ​​for searched documents 1 to 4 are 2, 105, 17, and -10, respectively, the high-similarity document acquisition unit 205 outputs searched documents 1 to 4 with positive total similarities in order of total similarity. In Fig. 4, searched document 2 (total similarity 105), searched document 3 (total similarity 17), and searched document 2 (total similarity 2) are output as high-similarity documents. The high-similarity document acquisition unit 205 displays the total similarity for each searched document on the display device 115 of the search terminal 11 as the similar document search results.

[0053] (4) Operation 5 is a flowchart illustrating an example of the operation of the similar document search device 10 in the first embodiment. A computer program stored in the memory 102 of the similar document search device 10 causes the processor to execute each step illustrated in FIG.

[0054] The similar document search device 10 is started by receiving a similar document search request from a user from the search terminal 11. The similar document search device 10 receives information for identifying a search sentence, such as the patent application number of a patent application or the patent document number, from the search terminal 11. The similar document search device 10 acquires text data of the corresponding search sentence from the document accumulation unit 211 using the information for identifying the search sentence (step S51). The similar document search device 10 may acquire the text data of the search sentence, or it may acquire the document number of a search document that includes the search sentence.

[0055] The similar document search device 10 then acquires text data of the sentences to be searched (step S52). The similar document search device 10 extracts the descriptions in the documents to be searched by paragraph, and generates sentences to be searched by paragraph by assigning unique sentence numbers to the extracted paragraph numbers and sentences in paragraph units. A sentence to be searched may be generated by concatenating multiple paragraphs, and a unique sentence number to be searched may be assigned. When search conditions are set by a search query or a search range, the sentences to be searched of the documents to be searched that satisfy the search conditions are acquired.

[0056] The similar document search device 10 obtains the sentence vectors of the acquired search sentence and searched sentence. The sentence vector calculation process involves, for example, a cleaning process to remove data unnecessary for sentence vector calculation from the target sentence, a morphological analysis to divide the sentence into words, and normalization to correct spelling variations, followed by a sentence vector calculation process to convert words, which are character strings, into sentence vectors. The sentence vector calculation process calculates the sentence vector u of the search sentence and the sentence vectors v1 to vn of the searched sentence. Alternatively, the search sentence and the searched sentence may be input to a learning model, which outputs the sentence vector u of the search sentence and the sentence vectors v1 to vn of the searched sentence.

[0057] The similar document search device 10 calculates the similarity between the sentence vector u of the search sentence and the sentence vectors v1 to vn of the sentences to be searched. The similarity can be found from the cosine similarity between the sentence vector u of the search sentence and the sentence vector vj (j = 1 to n) of the sentences to be searched. The cosine similarity is an index that measures the similarity based on the direction of the sentence vectors. The cosine similarity is calculated using the following formula. Cosine similarity j = cos(text vector u, text vector vj) (j = 1 to n) When using cosine similarity, the similarity can be calculated from the angle between two text vectors. Cosine similarity takes values ​​between -1 and 1, with the closer to 1 the greater the similarity between the two text vectors, and the closer to -1 the smaller the similarity.

[0058] The similar document search device 10 sorts the similarities of the searched sentences in descending order, finds the document numbers to which the searched sentences belong, eliminates duplicates, and extracts highly similar documents. For example, if searched sentences 1 and 2 are included in searched document 1, searched sentences 3 and 4 are included in searched document 2, searched sentences 5 and 6 are included in searched document 3, and searched sentences 7 and 8 are included in searched document 4, their respective sentence vectors are v1 to v8, and when their similarities with the sentence vector u of the searched sentence are calculated as s1 to s8, respectively, if these similarities are sorted in descending order as s5, s6, s1, s3, s2, s4, s7, and s8, the order of highly similar documents output is searched document 3, searched document 1, searched document 2, and searched document 4.

[0059] Alternatively, the similar document search device 10 may tally the similarity for each searched sentence for each searched document and calculate a total similarity for each searched document. For example, when the similarity between the sentence vector u of the searched sentence and the sentence vectors v1 and v2 of searched sentences 1 and 2 included in searched document 1 is calculated as s1 and s2, the total similarity for searched document 1 is calculated as s1+s2. Similarly, the total similarity for each searched document is calculated for searched documents 2 to 4.

[0060] The similar document search device 10 outputs, as high-similarity documents, searched documents with high total similarities from among the searched documents 1 to 4. For example, searched documents with positive total similarities may be output in order of total similarity, or only documents with high total similarities may be output.

[0061] According to the first embodiment, the similarity between the items described in paragraph units of the searched document and the searched sentence can also be determined, making it possible to search the descriptions in the searched document without any omissions, and enabling highly accurate similarity searches.

[0062] Second Embodiment In the sentence vector calculation unit 203 of the first embodiment, when a sentence vector is calculated using a learning model, the upper limit of the number of words that can be input to the learning model in the searched sentence may be limited to a predetermined number of tokens (hereinafter, a word in a sentence is referred to as a "token," and the number of words is referred to as the "token count"). For example, when Sentence BERT is used as the learning model, the number of tokens that can be input is limited to 512. In such cases, the second embodiment performs a normalization process on the searched sentence to improve the accuracy of the sentence vector calculation by the learning model and efficiently use the learning model. Note that Sentence BERT is preferably fine-tuned using patent document data. For example, a dataset can be created using the sentence of claim 1 in a past office action as an anchor and the sentence citing the reason for office action as a positive, and fine-tuned using known techniques. Furthermore, multiple Sentence BERT models can be created by grouping sections or themes with similar patent content into one dataset and fine-tuning using that dataset. This enables sentence vectors to be generated using models corresponding to groups with similar patent content, which can lead to higher calculation accuracy.

[0063] The sentence vector calculation unit 203 of the second embodiment further includes a searched sentence creation unit 206 that concatenates consecutive paragraphs of searched sentences so that the number of tokens falls within a predetermined number, as a preprocessing step for inputting the searched sentences into the learning model, and creates a normalized paragraph of searched sentences, and the searched sentence acquisition unit 202 is configured to acquire the normalized paragraph of searched sentences from the searched sentence creation unit 206.

[0064] Figure 6 is a diagram for explaining the normalization process of a sentence to be searched. The upper part of Figure 6 illustrates the paragraph numbers of the original paragraphs of the sentence to be searched and the number of tokens contained in each original paragraph. If the upper limit of the number of tokens that can be input to the learning model is limited to 512, the original paragraphs included in the sentence to be searched are aggregated so that the cumulative number of tokens is within 512, and a 512-token normalized paragraph with a normalized paragraph number assigned is generated.

[0065] In the example of FIG. 6, the cumulative number of tokens for original paragraphs 0001 to 0003 is 495. Adding the 321 tokens for paragraph 0004 makes the total number exceed 512, so the group of original paragraphs 0001 to 0003 is set as a 512-token normalized paragraph and assigned normalized paragraph number 1. Similarly, the group of original paragraphs 0004 and 0005 is assigned normalized paragraph number 2, and the group of original paragraphs 0006 to 0008 is assigned normalized paragraph number 3. This process allows the searched sentences to be grouped by normalized paragraph so that they fit within the upper limit of the number of tokens in the learning model. Each normalized paragraph is input to the learning model as a searched sentence via the searched sentence acquisition unit 202.

[0066] Incidentally, when original paragraphs of searched documents are grouped so that they fit within a predetermined number of tokens, the boundary of the 512-token normalized paragraph may be set in the middle of a coherent statement written in multiple original paragraphs. In such a case, the 512-token normalized paragraph can only include a portion of the coherent statement, so the content of the original paragraph may not be reflected in the calculation of the sentence vector or similarity.

[0067] Therefore, in order to reflect the contents of the original paragraphs existing before and after the boundary of the normalized paragraph in the output of the learning model, a normalized auxiliary paragraph is generated by the normalization auxiliary process. The searched sentence creation unit 206 creates a searched sentence of the normalized auxiliary paragraph by concatenating the searched sentences of consecutive paragraphs so that the boundary of the normalized paragraph is approximately at the center and the number of tokens is within a predetermined number. For example, a normalized auxiliary paragraph is generated by concatenating two original paragraphs existing before and after the boundary of two 512-token normalized paragraphs so that the number of tokens of one normalized auxiliary paragraph is within a predetermined number, and then repeating the operation of concatenating other original paragraphs adjacent to the two original paragraphs. The searched sentence acquisition unit 202 acquires the searched sentence of the normalized auxiliary paragraph from the searched sentence creation unit 206.

[0068] In the example of Figure 6, first, original paragraph 0003 (161 tokens) and original paragraph 0004 (321 tokens), which are located at the boundary between normalized paragraphs 1 and 2, are concatenated to generate a provisional supplementary paragraph with a total of 482 tokens. Next, when the number of tokens in paragraph 0002 (212 tokens) or paragraph 0005 (102 tokens), which are adjacent to original paragraphs 0003 and 0004, are added to the total of 482 tokens, the total becomes 694 and 584, respectively, which exceed the specified number of tokens of 512. Therefore, the provisional supplementary paragraphs of paragraphs 0003 and 0004 are set as normalized supplementary paragraph 1-1.

[0069] Next, similar processing is performed at the boundary between normalized paragraphs 2 and 3, between normalized paragraphs 3 and 4, and between normalized paragraphs 4 and 5. For example, at the boundary between normalized paragraphs 3 and 4, original paragraph 0008 (102 tokens) and original paragraph 0009 (182 tokens) are first set as provisional supplementary paragraphs (total token count: 284). A check is then performed to determine whether adjacent original paragraphs 0007 (263 tokens) and 0010 (26 tokens) can be added before and after the provisional supplementary paragraphs. In this case, concatenating original paragraph 0007 would exceed the specified number of 512, but original paragraph 0010 can be concatenated. Therefore, original paragraph 0010 is concatenated to the provisional supplementary paragraph, creating a new provisional supplementary paragraph (total token count: 310). Similarly, original paragraph 0011 (145 tokens) is concatenated, ultimately generating normalized supplementary paragraph 3-1 (total token count: 455). Note that, as in the case of normalized auxiliary paragraphs 3-1 and 4-1, the original paragraph 0011 may be included in two normalized auxiliary paragraphs 3-1 and 4-1.

[0070] FIG. 7 is a diagram for explaining the operation of the to-be-searched sentence creation unit 206. The to-be-searched sentence creation unit 206 acquires the to-be-searched sentence before it is input to the to-be-searched sentence acquisition unit 202 (step S70). The to-be-searched sentence creation unit 206 calculates the number of tokens for each original paragraph of the to-be-searched sentence, generates a table that associates the original paragraph number with the number of tokens, and concatenates multiple paragraphs in order from the first original paragraph of the to-be-searched sentence so that the number of tokens falls within a predetermined number, thereby setting a 512-token normalized paragraph (step S71). The to-be-searched sentence creation unit 206 repeats the process of setting normalized paragraphs for all original paragraphs, and generates a to-be-searched sentence of 512-token normalized paragraphs (step S72). Here, each of the 512-token normalized paragraphs is composed of a normalized paragraph number and a sentence in normalized paragraph units, and corresponds to one to-be-searched sentence.

[0071] Next, the to-be-searched sentence creation unit 206 concatenates the two original paragraphs existing at the boundary between the two 512-token normalized paragraphs to form a provisional auxiliary paragraph collapse, adds original paragraphs before and after the provisional auxiliary paragraphs until the upper limit of the number of tokens is reached to set normalized auxiliary paragraphs, and generates a to-be-searched sentence for each normalized auxiliary paragraph (step S73). The to-be-searched sentence creation unit 206 stores the to-be-searched sentence of the generated normalized paragraph and the to-be-searched sentence of the normalized auxiliary paragraph in the document storage unit 211 (step S74). Note that in step S74, the generated to-be-searched sentence may be stored and also output to the to-be-searched sentence acquisition unit 202.

[0072] In step S73, adjacent original paragraphs before and after the provisional auxiliary paragraph are linked in order until the cumulative number of tokens reaches the upper limit, but for example, when selecting one of two adjacent original paragraphs, an original paragraph that has continuity in the contents of the searched sentence may be linked preferentially. For example, if one of the two adjacent original paragraphs is a paragraph relating to a different embodiment, the other paragraph relating to the same embodiment may be selected and linked.

[0073] As described above, by providing the searched sentence creation unit 206 that generates normalized paragraphs and normalized auxiliary paragraphs, it is possible to prepare a searched sentence that maintains the contents of the original paragraph of the searched sentence even when the number of tokens that can be input to the learning model is limited. Furthermore, since a searched sentence is constructed with a number of tokens that is roughly close to the number input to the learning model, the learning model can be used efficiently.

[0074] <Third embodiment> In the third embodiment, sentences that characterize the present invention are preferentially extracted from the searched documents as search sentences. The search sentences in the third embodiment are any one or more sentences from the claims, the problem to be solved by the invention, the means for solving the problem, the gist of the mode for carrying out the invention, the effects of the invention, or the abstract of the invention. Note that, when extracting search sentences, any delimiters that make up the sentences in the searched documents, such as document names or item names, may be used.

[0075] FIG. 8 is a diagram showing an example of the sentence vector calculation unit 203 of the third embodiment. In the third embodiment, sentences that characterize the invention among the specifications, claims, and abstracts contained in the searched documents are set as search sentences. For example, the matters described in the [Claims] section of the searched documents, the matters described in the [Means for Solving the Problem] section of the specification, and some or all of the matters described in the "Abstract" section are extracted as sentences that characterize the present invention, and are used as a single search sentence. Note that the matters described that characterize the invention are not limited to the above-mentioned matters, and may also include, for example, matters described in the [Problem to be Solved by the Invention] and the [Effects of the Invention].

[0076] The sentence vector calculation unit 203 calculates sentence vectors between the search sentence containing features of the invention and the searched sentences, and outputs sentence vector u of the search sentence and sentence vectors v1 to vn of the searched sentences 1 to n. In FIG. 8, the search sentences are set as a group consisting of [Claims], [Means for Solving the Problem (Invention...Problem)], and [Abstract (Summary of the Invention)], and the sentence vector calculation unit 203 acquires this group as a single search sentence. Then, as in the first embodiment, the sentence vector calculation unit 203 calculates sentence vectors between the search sentence and the searched sentences 1 to n to obtain sentence vector u and sentence vectors v1 to vn. Similarity calculation and acquisition of highly similar documents are also performed in the same way as in the first embodiment.

[0077] In this way, by selecting a sentence containing the characteristics of the invention as the search sentence, the searched documents can be narrowed down to items similar to the characteristics of the invention, thereby improving the accuracy of searching for similar documents.

[0078] <Fourth embodiment> In the fourth embodiment, the "Summary of the form for implementing the invention" in the third embodiment is obtained by concatenating multiple paragraphs of text described in the form for implementing the invention that are highly similar to the text of the claims so that the number of tokens is within a predetermined range.

[0079] In patent documents, the "Mode for Carrying Out the Invention" in the specification describes specific embodiments of the invention and is the section with the most paragraphs in the entire specification. Therefore, rather than including the entire "Mode for Carrying Out the Invention" as the "Summary of the Mode for Carrying Out the Invention" in the search query, it is preferable to select and include in the search query descriptions the descriptions that characterize the present invention. Furthermore, if an upper limit is set on the number of tokens that can be input to the sentence vector calculation unit 203, the "Mode for Carrying Out the Invention" must be divided into paragraphs so that the number of tokens is roughly the upper limit, and then input into the learning model.

[0080] 9 is a diagram for explaining the operation of the search sentence acquisition unit 201 of the fourth embodiment, and shows an example in which the "Summary of the Mode for Carrying Out the Invention" is extracted from the section "Mode for Carrying Out the Invention" within the range of the upper limit number of tokens and used as the search sentence. In FIG. 9, the sentences described in the "Mode for Carrying Out the Invention" of the specification are exemplified by paragraph numbers 0001 to 0015 of the embodiment, the number of tokens described in the paragraphs corresponding to the paragraph numbers, and the similarity between the description of the paragraphs corresponding to the paragraph numbers and the description of claim 1. The similarity may be obtained by calculating a sentence vector between the description of the paragraph and the description of claim 1, or may be calculated using a learning model that outputs the similarity between a claim and an embodiment.

[0081] In FIG. 9, in generating the "Summary of the Embodiments," paragraphs of the embodiments are concatenated in descending order of similarity with claim 1 so that the cumulative total of the tokens is approximately the upper limit. In the example of FIG. 9, paragraphs 0005, 0008, 0011, and 0009 of the embodiments are concatenated in this order. The number of tokens contained in these paragraphs is 102, 102, 145, and 182, respectively. As a result, the sentence concatenating paragraphs 0005, 0008, 0011, and 0009 of the embodiments has a cumulative token count of 531, which is approximately the upper limit (512). Therefore, these paragraphs are extracted as search sentences corresponding to the "Summary of the Embodiments." Note that the "Summary of the Embodiments" may be extracted by repeating the generation of the "Summary of the Embodiments" a predetermined number of times.

[0082] In this way, even if a maximum number of tokens that can be input to the sentence vector calculation unit 203 is set, it is possible to form a sentence equivalent to the "Summary of a form for implementing the invention" while satisfying the maximum number of tokens, and therefore it is possible to prepare a search sentence suitable for input to the learning model.

[0083] Fifth Embodiment In the first to fourth embodiments, a single search sentence is used to search for documents to be searched, but in the fifth embodiment, multiple search sentences are used to search for documents to be searched. The fifth embodiment further includes a multiple search unit that performs searches using multiple search sentences, and a combined processing unit that acquires highly similar documents based on the results of the search by the multiple search unit. The multiple search unit has functions equivalent to the sentence vector calculation unit 203, similarity calculation unit 204, and highly similar document acquisition unit 205 in the first embodiment.

[0084] 10 is a diagram showing an overview of the multiple search unit 1010 and the combined processing unit 1020 of the fifth embodiment. The multiple search unit 1010 inputs a plurality of search sentences and a plurality of searched sentences, and outputs a similarity for each searched document for each of the plurality of search sentences. The combined processing unit 1020 determines and outputs searched documents that are highly similar to the searched document based on the similarity for each search sentence output by the multiple search unit 1010.

[0085] In FIG. 10, three search sentences, namely, [Claims], [Means for Solving the Problems of the Invention (The Invention is...Problem)], and [Abstract (Abstract of the Invention)], are input to the multiple search unit 1010 as search sentences. On the other hand, search sentences 1 to n contained in search documents 1 to m are input as searched sentences. The multiple search unit 1010 performs sentence vector calculation and similarity calculation, and calculates the similarity (hereinafter referred to as "subtotal similarity") for each search sentence, namely, [Claims], [Means for Solving the Problems of the Invention (The Invention is...Problem)], and [Abstract (Abstract of the Invention)], for each searched document. The multiple search unit 1010 outputs the document number and subtotal similarity for each searched document as the multiple search result for each search sentence.

[0086] Here, the sentence vector calculation and similarity calculation in the multiple search unit 1010 are performed in the same manner as in the first embodiment. That is, for each of the three search sentences, the multiple search unit 1010 calculates the sentence vector of the search sentence and the sentence vector of the searched sentences 1 to m, calculates the similarity between the sentence vector of the search sentence and the sentence vector of the searched sentences 1 to m, totals the calculated similarity for each of the searched documents 1 to m, and calculates a subtotal similarity for each of the searched documents.

[0087] The composite processing unit 1020 acquires the document numbers and subtotal similarities of searched documents 1 to m output for each search sentence. The composite processing unit 1020 aggregates the subtotal similarities for each searched document for each search sentence across the three search sentences for each searched document, and calculates the similarity for each searched document (hereinafter referred to as "total similarity"). The composite processing unit 1020 determines the total similarity value for each searched document 1 to m, and outputs the searched document with the highest total similarity value (for example, searched documents 105, 2, and 17).

[0088] Here, the multiple search unit 1010 outputs the document number and subtotal similarity for each searched document as the multiple search results for each search sentence, and determines and outputs searched documents that are highly similar to the searched document by calculating the total similarity for each searched document.However, it is also possible to determine searched documents that are highly similar without calculating the subtotal similarity or total similarity, as follows.

[0089] For example, the composite processing unit 1020 acquires the similarities of searched documents 1 to m output for each search sentence. Then, for the first search sentence, the composite processing unit 1020 sorts the acquired similarities in descending order, finds the document numbers to which each searched sentence belongs from the source data, extracts highly similar documents, and eliminates duplicates to determine the highly similar documents. The composite processing unit 1020 then assigns a discovery order to the documents with the highest similarity. The same process is performed for the second and third search sentences. Furthermore, to aggregate the three search results, evaluation points may be assigned according to their respective discovery orders, and the documents with the highest total evaluation points may be ranked in descending order of similarity. For example, the first place searched document may be assigned 10 points, the second place searched document may be assigned 9 points, and the third place searched document may be assigned 7 points, and the totals may be aggregated for each searched document. The composite processing unit 1020 determines the total evaluation points for each searched document 1 to m and outputs the top-ranked searched documents (e.g., searched documents 105, 102, 103, 105-107).

[0090] In the fifth embodiment, the searched documents are searched using multiple search sentences, so that the searched documents can be searched from different perspectives using multiple search sentences related to the features of the invention, thereby improving search accuracy.

[0091] Sixth Embodiment The composite processing unit of the sixth embodiment searches the document numbers and subtotal similarities of searched documents 1 to m for each search sentence output from the multiple search unit 1010 in the fifth embodiment using multiple search sentences, obtains the document numbers of a predetermined number of searched documents in order of highest similarity (subtotal similarity) for each of the multiple search sentences, assigns evaluation points to the document numbers according to the ranking of similarity (subtotal similarity), adds up the evaluation points for each document number, and obtains the searched documents with the highest added evaluation points as high-similarity documents.

[0092] Fig. 11 is a diagram showing an overview of the composite processing unit 1020 of the sixth embodiment. As in the fifth embodiment, the composite processing unit 1020 of Fig. 11 obtains the document numbers and subtotal similarities of each of the searched documents 1 to m for each of the searched sentences from the multiple search unit 1010, and arranges the document numbers in descending order of subtotal similarity. Next, the composite processing unit 1020 assigns evaluation points according to the ranking of the subtotal similarity. For example, evaluation points are assigned such that 0.6 is assigned to the document number with the highest subtotal similarity, 0.5 is assigned to the document number with the second highest subtotal similarity, and so on.

[0093] Here, the composite processing unit 1020 obtains the document number and subtotal similarity of each searched document 1 to m, arranges the document numbers in descending order of subtotal similarity, and then assigns evaluation points according to the ranking of subtotal similarity. However, it is also possible to output highly similar documents without calculating the subtotal similarity as follows.

[0094] For example, as in the fifth embodiment, the composite processing unit 1020 obtains the similarity between each search sentence and the searched sentence from the multiple search unit 1010, sorts the obtained similarities in descending order, finds the document number to which each searched sentence belongs from the source data, extracts highly similar documents, and eliminates duplicates to determine the highly similar documents. The composite processing unit 1020 then assigns a discovery order to the documents in descending order of similarity, and assigns evaluation points to the searched documents according to their discovery order. Similar processing can be performed for similarities with other search sentences. Evaluation points are assigned, for example, to the document number with the highest discovery order, 0.6, to the document number with the second highest discovery order, and so on.

[0095] In the example of Figure 11, for [Claims], evaluation points of 0.6 and 0.5 are assigned to document numbers 2 and 4, for [Means for Solving the Problems of the Invention], evaluation points of 0.3 and 0.2 are assigned to document numbers 4 and 1, and for [Abstract], evaluation points of 0.4 and 0.3 are assigned to document numbers 4 and 2.

[0096] When assigning evaluation points to document numbers, the composite processing unit 1020 may set different weights for each of multiple search sentences and assign weighted evaluation points to the document numbers by multiplying the evaluation points by the weights for each search sentence. For example, if the weight coefficients for the search sentences of the claims, the means for solving the problem of the invention, and the abstract are set to 3, 2, and 1, respectively, the document that obtains the highest similarity with respect to the claims will be calculated as a weighted evaluation point = 0.6 * 3 = 1.8. On the other hand, the document that obtains the highest similarity with respect to the abstract will be calculated as a weighted evaluation point = 0.6 * 1 = 0.6.

[0097] The composite processing unit 1020 totals the evaluation points for each searched document for each search sentence across the three search sentences for each searched document, and calculates a total evaluation point for each searched document, which is the sum of the evaluation points for each searched document. In Figure 11, the total evaluation point for searched document 1 is 32, the total evaluation point for searched document 2 is 89, and so on, and these are totaled and listed. The composite processing unit 1020 ranks the searched documents based on their evaluation point totals, and outputs the searched documents with the highest evaluation point totals as highly similar documents. In Figure 11, the highly similar documents output in this order are searched document 17, searched document 2, and searched document 105.

[0098] In the sixth embodiment, by assigning evaluation points to each search sentence according to the similarity ranking of the searched documents, it is possible to easily determine the ranking of highly similar documents for each search sentence. In addition, by making it possible to set weights according to the search sentence, it is possible to obtain search results that reflect the viewpoints that the user values.

[0099] The present disclosure is not limited to the above-described embodiments, and in the implementation stage, the components can be modified and embodied without departing from the spirit of the present disclosure. Furthermore, various inventions can be formed by appropriately combining multiple components disclosed in the above-described embodiments. For example, some components may be omitted from all the components shown in the embodiments. Furthermore, components from different embodiments may be appropriately combined. [Explanation of symbols]

[0100] 1...similar document search system, 10...similar document search device, 11...search terminal, 12...search terminal, 20...search management unit, 21...storage unit, 22...communication unit, 101...processor, 102...memory, 103...storage, 104...communication I / F, 105...input / output I / F, 111...processor, 112...memory, 113...storage, 114...input device, 115...display device, 116...communication I / F, 201...search text acquisition unit, 202...searched text acquisition unit, 203...text vector calculation unit, 204...similarity calculation unit, 205...high similarity document acquisition unit, 206...searched text creation unit, 211...document storage unit, 1010...multiple search unit, 1020...complex processing unit.

Claims

1. In an information processing device for searching for documents having high similarity to a search sentence, a search sentence acquisition unit that acquires a search sentence; a searched sentence acquisition unit for acquiring searched documents including paragraph numbers and searched sentences in paragraph units; a sentence vector calculation unit that calculates sentence vectors of the search sentence and the to-be-searched sentence; a similarity calculation unit that calculates a similarity between the sentence vector of the search sentence calculated by the sentence vector calculation unit and the sentence vector of the to-be-searched sentence calculated by the sentence vector calculation unit; a high similarity document acquisition unit that acquires high similarity documents including highly similar searched sentences based on the similarity; Information processing device.

2. a searched sentence creation unit that connects consecutive searched sentences of paragraphs so that the number of tokens falls within a predetermined number, and creates a searched sentence of a normalized paragraph; 2. The information processing apparatus according to claim 1, wherein the searched sentence acquisition unit acquires the searched sentence of the normalized paragraph.

3. the searched sentence creation unit connects the searched sentences of consecutive paragraphs so that the boundary of the normalized paragraph is approximately at the center and the number of tokens is within the predetermined number, thereby creating a searched sentence of a normalized auxiliary paragraph; The information processing apparatus according to claim 2 , wherein the searched sentence acquisition unit acquires the searched sentence of the normalized auxiliary paragraph.

4. 2. The information processing apparatus according to claim 1, wherein said sentence vector calculation unit calculates sentence vectors of said search sentence and said to-be-searched sentence in advance.

5. The information processing device according to claim 1, wherein the search sentence is one or more sentences among sentences described in the claims, the problem that the invention aims to solve, the means for solving the problem, the gist of the form for implementing the invention, the effects of the invention, or the summary of the invention.

6. The information processing device according to claim 5, wherein the gist of the form for implementing the invention is obtained by concatenating multiple paragraphs of text described in the form for implementing the invention that have a high similarity to the text of the claims so that the number of tokens is within a predetermined number.

7. 6. The information processing apparatus according to claim 5, further comprising: a multiple search unit that performs a search using a plurality of search sentences; and a composite processing unit that acquires highly similar documents based on the results of the search by said multiple search unit.

8. The information processing device according to claim 7, wherein the composite processing unit searches using the multiple search unit with the plurality of search sentences, obtains document numbers of a predetermined number of searched documents in descending order of similarity for each of the plurality of search sentences, assigns evaluation points to the document numbers according to their rankings of similarity, sums up the evaluation points for each document number, and obtains searched documents with high summed evaluation points as highly similar documents.

9. 9. The information processing device according to claim 8, wherein the composite processing unit assigns the evaluation points based on different weights for each of the plurality of search sentences.

10. The information processing device according to claim 1 , wherein the text vector calculation unit uses a text vector calculation model that is fine-tuned using patent document data.

11. An information processing method for searching for documents that are highly similar to a search sentence, Obtaining a search sentence; Obtaining a searched document including a paragraph number and a searched sentence in paragraph units; calculating a sentence vector of the search sentence and the searched sentence; Calculating a similarity between the calculated sentence vector of the search sentence and the calculated sentence vector of the to-be-searched sentence; and acquiring high-similarity documents including highly similar searched sentences based on the similarity. Information processing methods.

12. A computer program that causes a computer to function as each unit of the information processing device according to any one of claims 1 to 10.

Images