Document retrieval device

Abstract

The invention relates to a file retrieval device. High-precision file retrieval is achieved through a simple input method, and particularly high-precision file retrieval related to intellectual property is achieved. A file retrieval apparatus for retrieving a file related to or similar to an input file includes a processing section configured to: derive a first distribution representation vector and a first weight of the first distribution representation vector from the input file; extracting a second distribution representation vector based on a similarity with the first distribution representation vector; obtaining a second weight of the second distribution representation vector, wherein the second weight comprises a product of the first weight and the similarity; outputting the first weight and the second weight; receiving a variation value of one or more of the first weight and the second weight; and performing a retrieval of the file by using the edited first weight and the edited second weight.

Description

[0001] This application is a divisional application of an application filed on March 13, 2019, with PCT application number PCT / IB2019 / 052022, which entered the Chinese national phase on November 18, 2020, with national application number 201980033402.2, and invention name “Document Retrieval System, Document Retrieval Method, Program, and Non-transitory Computer-readable Storage Medium”. Technical Field

[0002] One embodiment of the present invention relates to a document search system, a document search method, a program, and a non-transitory computer-readable storage medium. One embodiment of the present invention relates to a document search system and a document search method related to intellectual property.

[0003] Note that one embodiment of the present invention is not limited to the above-mentioned technical field. As an example of the technical field of one embodiment of the present invention, a semiconductor device, a display device, a light-emitting device, a power storage device, a storage device, an electronic device, a lighting device, an input device (e.g., a touch sensor, etc.), an input-output device (e.g., a touch panel, etc.), and a driving method or a manufacturing method of the above-mentioned device can be cited. Background Art

[0004] By conducting a prior art investigation on the invention before the application, it is possible to investigate whether there is any related intellectual property. The domestic and foreign patent documents and papers obtained through the prior art investigation can be used to confirm the novelty and creativity of the invention and to determine whether to apply for a patent. In addition, by conducting an invalidation document investigation on patent documents, it is possible to investigate whether one's own patent rights have been invalidated or whether it is possible to invalidate the patent rights of others.

[0005] For example, a user inputs a keyword into a patent document search system, thereby outputting patent documents containing the keyword.

[0006] In order to conduct a prior art search with high accuracy using such a system, the user is required to have high skills such as searching with appropriate keywords and extracting necessary patent documents from a large number of patent documents output by the system.

[0007] In addition, the application of artificial intelligence is being studied for various purposes. In particular, it is expected that computers with higher performance than existing Neumann computers can be realized by using artificial neural networks, and in recent years, various research works on building artificial neural networks on electronic circuits have been carried out.

[0008] For example, Patent Document 1 discloses an invention in which a memory device using a transistor including an oxide semiconductor in a channel formation region holds weight data required for calculation using an artificial neural network.

[0009] [Prior technical literature]

[0010] [Patent Document]

[0011] [Patent Document 1] U.S. Patent Application Publication No. 2016 / 0343452 Summary of the invention

[0012] Technical problem to be solved by the invention

[0013] Therefore, one of the purposes of one embodiment of the present invention is to provide a document retrieval system capable of performing document retrieval with high accuracy. Another purpose of one embodiment of the present invention is to provide a document retrieval method capable of performing document retrieval with high accuracy. Another purpose of one embodiment of the present invention is to achieve high-accuracy document retrieval through a simple input method, in particular, to achieve high-accuracy document retrieval related to intellectual property rights.

[0014] Note that the description of these objectives does not prevent the existence of other objectives. One embodiment of the present invention does not need to achieve all of the above objectives. Objectives other than the above objectives can be extracted from the description of the specification, drawings, and claims.

[0015] Solutions to technical problems

[0016] One embodiment of the present invention is a document retrieval system including a processing unit. The processing unit has: a function of extracting a keyword included in text data; a function of extracting a keyword-related word from words included in a plurality of first reference text analysis data; a function of assigning a weight to each of the keyword and the related word; a function of assigning a score to each of a plurality of second reference text analysis data based on the weight; a function of generating sorted data by sorting the plurality of second reference text analysis data based on the score; and a function of outputting the sorted data.

[0017] One embodiment of the present invention is a document retrieval method. The method includes the following steps: extracting keywords included in text data; extracting keywords-related words from words included in a plurality of first reference text analysis data; assigning weights to each of the keywords and related words; scoring each of a plurality of second reference text analysis data based on the weights; ranking the plurality of second reference text analysis data based on the scores to generate ranking data; and outputting the ranking data.

[0018] One embodiment of the present invention is a program that causes a processor to execute the following steps: a first step of extracting keywords included in text data; a second step of extracting related words of the keywords from words included in a plurality of first reference text analysis data; a third step of assigning weights to each of the keywords and related words; a fourth step of scoring each of a plurality of second reference text analysis data based on the weights of keywords or related words that match words included in the second reference text analysis data; and a fifth step of ranking the plurality of second reference text analysis data based on the scores to generate ranking data. One embodiment of the present invention is a non-transitory computer-readable storage medium storing the above-mentioned program.

[0019] The weight of a keyword is preferably a value of the keyword based on the inverse document frequency in a plurality of first reference text analysis data or a plurality of second reference text analysis data. The weight of a related word is preferably a product of a value based on the similarity or distance between a distribution representation vector of the related word and a distribution representation vector of the keyword and the weight of the keyword.

[0020] The second reference text analysis data containing words matching the keyword or the related words is preferably scored.

[0021] The plurality of first reference text analysis data may also be the same as the plurality of second reference text analysis data.

[0022] The related words are preferably extracted using a distribution representation vector obtained through machine learning of the distribution representation of words included in the plurality of first reference text analysis data.

[0023] The related words are preferably extracted from the words included in the plurality of first reference text analysis data based on the similarity or proximity of the word distribution representation vector and the keyword distribution representation vector. The word distribution representation vector is preferably a vector generated using a neural network.

[0024] As a function of extracting keywords included in text data, it is preferred to have a function of performing morphological analysis of text data to generate analysis data and a function of extracting keywords from the analysis data. Keywords are preferably extracted from words included in the analysis data based on the degree of inverse document frequency in a plurality of first reference text analysis data or a plurality of second reference text analysis data.

[0025] Preferably, the weights can be modified by the user.

[0026] Preferably, the first reference text analysis data is data generated by performing a morphological analysis on the first reference text data, and the second reference text analysis data is data generated by performing a morphological analysis on the second reference text data.

[0027] A document retrieval system of one embodiment of the present invention preferably includes an electronic device and a server. The electronic device includes a first communication unit. The server includes the processing unit and a second communication unit. The first communication unit has a function of providing text data to the server through one or both of wired communication and wireless communication. The processing unit has a function of providing sorting data to the second communication unit. The second communication unit has a function of providing sorting data to the electronic device through one or both of wired communication and wireless communication.

[0028] The processing unit may include either a transistor including a metal oxide in a channel formation region or a transistor including silicon in a channel formation region.

[0029] Effects of the Invention

[0030] According to one embodiment of the present invention, a document retrieval system capable of performing document retrieval with high accuracy can be provided. In addition, according to one embodiment of the present invention, a document retrieval method capable of performing document retrieval with high accuracy can be provided. In addition, according to one embodiment of the present invention, a high-accuracy retrieval system can be implemented by a simple input method, and in particular, a high-accuracy document retrieval related to intellectual property rights can be implemented.

[0031] Note that the description of these effects does not prevent the existence of other effects. One embodiment of the present invention does not necessarily have all of the above effects. Effects other than the above effects can be extracted from the description of the specification, drawings, and claims. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] [ Figure 1 ]A block diagram showing an example of a document retrieval system.

[0033] [ Figure 2 ]A flowchart showing an example of a method for generating retrieval data.

[0034] [FIG. 3] (A) to (C) are diagrams showing an example of a method of generating search data.

[0035] [ Figure 4 ]A flowchart showing an example of a file retrieval method.

[0036] [ Figure 5 ]A flowchart showing an example of a file retrieval method.

[0037] [ Fig. 6 ] (A) to (C) are diagrams showing an example of a file retrieval method.

[0038] [Fig. 7] (A) and (B) are diagrams showing examples of a file retrieval method.

[0039] [ Figure 8 ]A flowchart showing an example of a file retrieval method.

[0040] [ Fig. 9 ]A flowchart showing an example of a file retrieval method.

[0041] [Fig. 10] (A) and (B) are diagrams showing examples of a file retrieval method.

[0042] [ Fig.11 ]A block diagram showing an example of a document retrieval system.

[0043] [Figure 12] (A) and (B) are diagrams showing examples of the structure of a neural network.

[0044] [ Fig.13 ]A diagram showing a structural example of a semiconductor device.

[0045] [ Fig.14 ]A diagram showing a structural example of a storage unit.

[0046] [ Fig.15 ]A diagram showing a structural example of a bias circuit.

[0047] [ Fig.16 ]Timing diagram. DETAILED DESCRIPTION

[0048] The embodiments are described in detail with reference to the accompanying drawings. Note that the present invention is not limited to the following description, and a person skilled in the art can easily understand that the method and details can be transformed into various forms without departing from the purpose and scope of the present invention. Therefore, the present invention should not be interpreted as being limited to the contents described in the embodiments shown below.

[0049] Note that in the structure of the invention described below, the same reference numerals are used in common in different drawings to represent the same parts or parts with the same function, and their repeated description is sometimes omitted. In addition, when representing parts with the same function, the same hatching is sometimes used without adding a reference numeral in particular.

[0050] In addition, for ease of understanding, the position, size, range, etc. of each component shown in the drawings may not represent its actual position, size, range, etc. Therefore, the disclosed invention is not necessarily limited to the position, size, range, etc. disclosed in the drawings.

[0051] Note that the terms "film" and "layer" may be used interchangeably depending on the situation or condition. For example, "conductive layer" may be replaced with "conductive film" in some cases. Also, "insulating film" may be replaced with "insulating layer" in some cases.

[0052] (Implementation Method 1)

[0053] In this embodiment, using Figures 1 to 11 A document search system and a document search method according to one embodiment of the present invention will be described.

[0054] One embodiment of the present invention is a document retrieval system including a processing unit. The processing unit has: a function of extracting a keyword included in text data; a function of extracting a keyword-related word from a word included in a plurality of first reference text analysis data; a function of assigning a weight to each of the keyword and the related word; a function of assigning a score to each of a plurality of second reference text analysis data based on the weight; a function of generating ranking data by ranking the plurality of second reference text analysis data based on the score; and a function of outputting the ranking data.

[0055] In the document search system of one embodiment of the present invention, the first reference text analysis data may be used to extract related words of a keyword, and the second reference text analysis data may be used as a search target to search for data related to or similar to the text data.

[0056] The first reference text analysis data may be the same as the second reference text analysis data. In this case, in a document retrieval system of one embodiment of the present invention, the reference text analysis data for the retrieval object may be used to extract related words of the keyword. In addition, the first reference text analysis data may include part or all of the second reference text analysis data.

[0057] That is, the processing unit included in the file retrieval system of one embodiment of the present invention may also have: a function of extracting keywords included in text data; a function of extracting related words of the keywords from words included in multiple reference text analysis data; a function of assigning weights to each of the keywords and related words; a function of scoring each of the multiple reference text analysis data based on the weight; a function of sorting the multiple reference text analysis data based on the scores to generate sorted data; and a function of outputting the sorted data.

[0058] When a user selects a keyword for retrieval in a document retrieval system, the user needs to select the keyword for retrieval by considering not only the keyword but also the synonyms, near synonyms, and inconsistencies of the keyword. Therefore, keyword selection is a heavy task for the user and is prone to differences due to different skills. In addition, it is also a heavy task for the user to find a necessary document from a large number of documents output by the document retrieval system.

[0059] Here, a document retrieval system of one embodiment of the present invention has a function of extracting keywords and related words of the keywords included in text data. Therefore, the user of the document retrieval system of one embodiment of the present invention does not need to personally select keywords for retrieval. The user can directly input text data whose weight is larger than the keywords into the document retrieval system. In addition, when the user wants to personally select keywords and related words, it is not necessary to select them from the beginning, and the keywords and related words can be added, modified and deleted with reference to the keywords and related words extracted by the document retrieval system. Thus, the burden of the user in the document retrieval can be reduced, and the difference in the retrieval results based on the skills of the user can be made less likely to occur.

[0060] In particular, a document retrieval system in one embodiment of the present invention has a function of extracting related words of a keyword from words included in a plurality of reference text analysis data. When extracting related words of a keyword from words included in an existing concept dictionary, it is sometimes difficult to extract unique records included in the data to be searched as related words. On the other hand, in one embodiment of the present invention, related words of a keyword are extracted from words included in data prepared for extracting related words of a keyword (first reference text analysis data) or data to be searched (second reference text analysis data). This makes it easy to extract the unique record as a related word, which can reduce search omissions, and is therefore preferred.

[0061] Furthermore, a document retrieval system of one embodiment of the present invention has a function of assigning a weight to each of the extracted keywords and related words. Since each of the keywords and related words has a weight, a score based on the weight can be assigned to the reference text analysis data in which the keywords or related words are found. Furthermore, a document retrieval system of one embodiment of the present invention has the following function: generating and outputting sorted data by sorting the reference text analysis data as the search object based on the score. The search results are sorted and output according to relevance or similarity, so that the user can easily find the necessary documents from the search results, the work efficiency is improved, and it is not easy to cause negligence. In this way, a document retrieval system of one embodiment of the present invention can perform document retrieval simply and with high precision.

[0062] The weight of a keyword is preferably a value of the inverse document frequency (IDF) of the keyword based on a plurality of first or second reference text analysis data. IDF indicates that a certain word rarely appears in a document. A word that appears in many documents has a small IDF, and a word that appears in only a portion of documents has a high IDF. Therefore, it can be said that a word with a high IDF is a unique word in the first or second reference text analysis data.

[0063] Keyword extraction from text data may also include the IDF of words in the text data based on the first or second reference text analysis data. For example, words with an IDF value above a certain value may be extracted as keywords, or any number of words may be extracted in order of high IDF as keywords.

[0064] Keywords can be extracted based on the IDF in the first reference text analysis data or the IDF in the second reference text analysis data. It is preferable to extract keywords from text data based on the IDF in the second reference text analysis data that is the object of retrieval because it is easy to extract unique words in the document of the object of retrieval. However, in the case where the number of documents to be retrieved is small, it is sometimes easier to extract keywords from text data by extracting keywords from text data based on the IDF in the first reference text analysis data.

[0065] Alternatively, keywords may be extracted based on word class information obtained by morphological analysis of text data input by the user. For example, when performing morphological analysis of Japanese text, nouns are preferably extracted. Also, when performing morphological analysis of English text, adjectives, nouns, and verbs are preferably extracted.

[0066] As related words, synonyms, near-synonyms, antonyms, hypernyms, hyponyms, etc. can be cited. The related words are preferably extracted from words included in a plurality of reference text analysis data based on the similarity or proximity of the distribution representation vector of the word and the distribution representation vector of the keyword. With respect to the keywords included in the text data input by the user, synonyms, near-synonyms, etc. included in the reference text analysis data can be extracted as related words. This can improve the search accuracy.

[0067] The weight of the related word is preferably the product of the value of the similarity or distance between the distribution representation vector of the related word and the distribution representation vector of the keyword and the weight of the keyword. By setting the weight of the related word based on both the relevance between the related word and the keyword and the weight of the keyword itself, the sorting accuracy can be further improved.

[0068] The document search system preferably utilizes artificial intelligence (AI) in at least a part of its processing.

[0069] In particular, the document search system preferably uses an artificial neural network (ANN: Artificial Neural Network, hereinafter sometimes referred to as a neural network). The neural network can be implemented by a circuit (hardware) or a program (software).

[0070] For example, when generating a distribution representation vector of a word, it is preferred to use machine learning, and more preferably a neural network. Specifically, the related words are preferably extracted using a distribution representation vector obtained by machine learning of the distribution representation of words included in a plurality of reference text analysis data. Thus, the accuracy of the extraction of the related words and the weight of the related words can be improved.

[0071] In this specification, etc., a neural network refers to any model that simulates a biological neural circuit network and determines the connection strength between neurons through learning, thereby obtaining problem-solving capabilities. A neural network includes an input layer, an intermediate layer (hidden layer), and an output layer.

[0072] In this specification and other documents, when describing a neural network, determining the connection strength (also called weight coefficient) between neurons based on existing information may be referred to as "learning."

[0073] In this specification and the like, constructing a neural network using the combination strength obtained through learning and deriving a new conclusion from the structure is sometimes referred to as "inference".

[0074] <1. Configuration Example 1 of Document Retrieval System>

[0075] In this embodiment, a document retrieval system that can be used for searching intellectual property is described as an example of a document retrieval system. Note that the application of the document retrieval system of one embodiment of the present invention is not limited to searching for intellectual property, and can also be used for searches other than intellectual property.

[0076] Figure 1 1 is a block diagram showing the document retrieval system 100. In the drawings of this specification, the block diagram shows the components classified according to the functions in the independent blocks, but it is difficult to clearly divide the actual components according to the functions, and one component may have multiple functions. In addition, one function may involve multiple components, for example, two processes performed in the processing unit 103 may be performed in different servers.

[0077] The document search system 100 includes at least a processing unit 103 . Figure 1 The document retrieval system 100 shown further includes an input unit 101 , a transmission channel 102 , a storage unit 105 , a database 107 and an output unit 109 .

[0078] [Input unit 101]

[0079] The input unit 101 is supplied with data from outside the document search system 100. The data supplied to the input unit 101 is supplied to the processing unit 103, the storage unit 105, or the database 107 through the transmission channel 102.

[0080] [Transmission channel 102]

[0081] The transmission channel 102 has a function of transmitting data. The transmission channel 102 can be used to transmit and receive data between the input unit 101 , the processing unit 103 , the storage unit 105 , the database 107 , and the output unit 109 .

[0082] [Processing unit 103]

[0083] The processing unit 103 has a function of performing calculations and inferences using data supplied from the input unit 101, the storage unit 105, the database 107, etc. The processing unit 103 can supply calculation results, inference results, etc. to the storage unit 105, the database 107, the output unit 109, etc.

[0084] It is preferable to use a transistor containing a metal oxide in the channel formation region in the processing unit 103. Since the off-state current of the transistor is extremely small, by using the transistor as a switch to hold the charge (data) flowing into the capacitor used as a storage element, a long data retention period can be ensured. By applying this characteristic to at least one of the register and the cache memory included in the processing unit 103, the processing unit 103 can be operated only when necessary, and in other cases, the previous processing information is stored in the storage element, and the processing unit 103 can be turned off. That is, normally off computing is realized, thereby achieving low power consumption of the file retrieval system.

[0085] In this specification and the like, a transistor using an oxide semiconductor or a metal oxide in a channel formation region is referred to as an oxide semiconductor transistor or an OS transistor. The channel formation region of the OS transistor preferably includes a metal oxide.

[0086] In this specification, etc., metal oxide refers to an oxide of a metal in a broad sense. Metal oxides are classified into oxide insulators, oxide conductors (including transparent oxide conductors) and oxide semiconductors (Oxide Semiconductor, also referred to as OS), etc. For example, when a metal oxide is used in a semiconductor layer of a transistor, the metal oxide is sometimes referred to as an oxide semiconductor. In other words, when a metal oxide has at least one of an amplification effect, a rectification effect and a switching effect, the metal oxide can be referred to as a metal oxide semiconductor (metal oxide semiconductor), or it can be abbreviated as OS.

[0087] The metal oxide possessed by the channel formation region preferably contains indium (In). In the case where the metal oxide possessed by the channel formation region contains indium, the carrier mobility (electron mobility) of the OS transistor is improved. In addition, the metal oxide possessed by the channel formation region is preferably an oxide semiconductor containing element M. Element M is preferably aluminum (Al), gallium (Ga) or tin (Sn) and the like. As other elements that can be used as element M, there are boron (B), silicon (Si), titanium (Ti), iron (Fe), nickel (Ni), germanium (Ge), yttrium (Y), zirconium (Zr), molybdenum (Mo), lanthanum (La), cerium (Ce), neodymium (Nd), hafnium (Hf), tantalum (Ta), tungsten (W) and the like. Note that as element M, multiple of the above elements may also be combined sometimes. Element M is, for example, an element with a high bond energy with oxygen. Element M is, for example, an element with a higher bond energy with oxygen than indium. In addition, the metal oxide possessed by the channel formation region preferably contains zinc (Zn). Metal oxides containing zinc are sometimes easy to crystallize.

[0088] The metal oxide contained in the channel formation region is not limited to the metal oxide containing indium. The semiconductor layer may be a metal oxide containing zinc, gallium or tin without containing indium, such as zinc tin oxide or gallium tin oxide.

[0089] Alternatively, a transistor whose channel formation region includes silicon may be used for the processing unit 103 .

[0090] It is preferable to use a combination of a transistor including an oxide semiconductor in a channel formation region and a transistor including silicon in a channel formation region and use them for the processing unit 103 .

[0091] The processing unit 103 includes, for example, a calculation circuit or a central processing unit (CPU).

[0092] The processing unit 103 may also include a microprocessor such as a DSP (Digital Signal Processor) or a GPU (Graphics Processing Unit). The microprocessor may also be implemented by a PLD (Programmable Logic Device) such as an FPGA (Field Programmable Gate Array) or an FPAA (Field Programmable Analog Array). The processing unit 103 may perform various data processing and program control by interpreting and executing instructions from various programs by the processor. The program executable by the processor is stored in at least one of the memory area included in the processor and the storage unit 105.

[0093] The processing unit 103 may include a main memory. The main memory includes at least one of a volatile memory such as a RAM (Random Access Memory) and a nonvolatile memory such as a ROM (Read Only Memory).

[0094] As the RAM, for example, DRAM (Dynamic Random Access Memory) or SRAM (Static Random Access Memory) is used, and the RAM is allocated with virtual storage space as a work space of the processing unit 103 and is used for the processing unit 103. The operating system, application program, program module, program data, and lookup table stored in the storage unit 105 are loaded into the RAM when executed. The processing unit 103 directly accesses and operates these data, programs, and program modules loaded into the RAM.

[0095] ROM can store BIOS (Basic Input / Output System) and firmware that do not need to be rewritten. Examples of ROM include mask ROM, OTPROM (One Time Programmable Read Only Memory), EPROM (Erasable Programmable Read Only Memory), etc. Examples of EPROM include UV-EPROM (Ultra-Violet Erasable Programmable Read Only Memory) that can erase stored data by ultraviolet irradiation, EEPROM (Electrically Erasable Programmable Read Only Memory), and flash memory.

[0096] [Storage unit 105]

[0097] The storage unit 105 has a function of storing programs executed by the processing unit 103. The storage unit 105 may also have a function of storing calculation results and inference results generated by the processing unit 103, data input to the input unit 101, and the like.

[0098] The storage unit 105 includes at least one of a volatile memory and a non-volatile memory. The storage unit 105 may include, for example, a volatile memory such as a DRAM or an SRAM. The storage unit 105 may include, for example, a non-volatile memory such as a ReRAM (Resistive Random Access Memory), a PRAM (Phase change Random Access Memory), a FeRAM (Ferroelectric Random Access Memory), a MRAM (Magnetoresistive Random Access Memory), or a flash memory. In addition, the storage unit 105 may also include a recording medium drive such as a hard disk drive (HDD) and a solid state drive (SSD).

[0099] [Database 107]

[0100] The database 107 has a function of storing at least reference text analysis data, IDF data, and vector data that are search targets. In addition, the database 107 may also have a function of storing calculation results and inference results generated by the processing unit 103, and data input to the input unit 101. The storage unit 105 and the database 107 may not be separated. For example, the document search system may include a storage unit having the functions of both the storage unit 105 and the database 107.

[0101] [Output unit 109]

[0102] The output unit 109 has a function of supplying data to the outside of the document search system 100. For example, the calculation results and inference results in the processing unit 103 can be supplied to the outside.

[0103] <2. File retrieval method>

[0104] First, use Figure 2 3 illustrates the processing performed in advance for searching using the document search system 100. Figure 2 FIG3 shows a flow chart showing Figure 2 Note that the diagram of each data shown in FIG3 is only an example and is not limited to this.

[0105] In the following, the reference text data used to extract the related words of the keyword is the same as the reference text data for the search object (referred to as reference text data TD ref) is used as an example to illustrate. As mentioned above, the two reference text data may also be different data. For example, the first reference text data may be used to generate vector data VD (described later) and the second reference text data may be used as the search object. In this case, the IDF data ID described later may be generated using both the first reference text data and the second reference text data.

[0106] [Step S1]

[0107] First, multiple reference text data TD ref Input to the input unit 101 ( Figure 2 and Figure 3A ).

[0108] In the document search method of this embodiment, a plurality of pre-prepared documents can be used as search targets to search for documents related to or similar to the input document. ref The reference text data TD is text data of a document to be searched in the document search system 100. The reference text data TD may be appropriately selected according to the purpose of the document search system 100. ref data.

[0109] Reference text data TD ref The document retrieval system 100 is inputted to the input unit 101. Then, the reference text data TD ref The input unit 101 supplies the data to the processing unit 103 through the transmission channel 102. Alternatively, the reference text data TD ref The data may be stored in the storage unit 105 or the database 107 from the input unit 101 through the transmission channel 102 , and supplied from the storage unit 105 or the database 107 to the processing unit 103 through the transmission channel 102 .

[0110] exist Figure 3A In the figure, n (n is an integer greater than or equal to 2) reference text data TD are shown. ref , which are recorded as data TD ref (x) (x is an integer greater than or equal to 1 and less than or equal to n).

[0111] In this embodiment, an example is shown in which the search target is a document related to intellectual property. ref It is text data of documents related to intellectual property.

[0112] Note that, as described above, the reference text data input in step S1 is sometimes different from the data to be searched. The reference text data input in step S1 is preferably text data of documents related to intellectual property rights, but is not limited thereto. In addition, a part of the reference text data input in step S1 is sometimes data to be searched. For example, the reference text data input in step S1 may be text data of patent documents and papers, while the search object may be only text data of patent documents.

[0113] Here, as documents related to intellectual property rights, specifically, publications such as patent documents (open patent publications, patent gazettes, etc.), utility model gazettes, design gazettes, and papers can be cited. Not only publications published domestically, but publications published in various countries around the world can also be used as documents related to intellectual property rights.

[0114] A part or all of each of the specification, claims, and abstract included in the patent document can be used as the reference text data TD ref For example, an embodiment, an example, or a claim of a specified invention may also be used as reference text data TD ref Similarly, part or all of the text in other publications such as papers can be used as reference text data TD ref .

[0115] Documents related to intellectual property are not limited to publications. For example, text documents owned by users or user groups of the document retrieval system can also be used as reference text data TD. ref .

[0116] In addition, documents related to intellectual property rights include texts describing inventions, plans, designs, or industrial products.

[0117] Reference text data TD ref For example, it may include patent documents of a specific applicant or patent documents in a specific technical field.

[0118] Reference text data TD ref In addition to the description of the intellectual property itself (e.g., specification, etc.), various information related to the intellectual property (e.g., literature information, etc.) can also be included. Examples of such information include the patent applicant, technical field, application number, publication number, status (pending, authorized, abandoned, etc.), etc.

[0119] Reference text data TD ref It is preferred to include date information on the intellectual property. When the intellectual property is a patent document, the date information may include, for example, the application date, publication date, and authorization announcement date, and when the intellectual property is technical information of an industrial product, the date information may include, for example, the release date.

[0120] Thus, in the reference text data TD ref Including various information related to intellectual property, you can use the document search system to select various search scopes.

[0121] For example, the document search system of this embodiment can be used to search for patent documents, papers, or industrial products related to or similar to the invention before the application. This allows a prior art investigation of the invention before the application. By grasping the relevant prior art and reconsidering it, the invention can be strengthened and made into a powerful patent that other companies cannot circumvent.

[0122] In addition, for example, the document retrieval system of this embodiment can be used to search for patent documents, papers, or industrial products related to or similar to the industrial product before it is sold. ref When a company's own patent documents are included, it is possible to confirm whether the patent application for the technology related to the company's pre-sale industrial products is sufficient. ref When including information on other companies' intellectual property rights, you can check whether the industrial products before they are sold infringe on other companies' intellectual property rights. By understanding the relevant prior art and reconsidering the technology related to the industrial products before they are sold, you can find new inventions that can become powerful patents that contribute to your company's business. Note that the search is not limited to industrial products before they are sold, but can also be performed on industrial products after they are sold.

[0123] For example, by using the document search system of this embodiment, it is possible to search for patent documents, papers, or industrial products related to or similar to a specific patent. In particular, by conducting an investigation based on the application date of the patent, it is possible to easily and accurately investigate whether the patent has grounds for invalidation.

[0124] [Step S2]

[0125] Next, multiple reference text data TD are performed ref A plurality of reference text analysis data AD is generated by morphological analysis of each ref ( Figure 2 and Figure 3A ).

[0126] The processing unit 103 performs n reference text data TD ref The morphological analysis of each of the n reference text analysis data AD is generated ref (respectively recorded as reference text analysis data AD ref (x) (x is an integer greater than or equal to 1 and less than or equal to n)). For example, by referring to the text data TD ref Morphological analysis of (n) generates reference text analysis data ADref (n).

[0127] In morphological analysis, a text written in a natural language is divided into morphemes (the smallest unit of language meaning), and the word class of the morpheme can be identified. ref Extract only nouns.

[0128] exist Figure 3A In the input reference text data TD ref (1) Including long sentences, in the output reference text analysis data AD ref In (1), the sentence is divided into multiple words.

[0129] exist Figure 3A The reference text analysis data AD generated is shown in ref For example, the processing unit 103 may output the plurality of reference text analysis data AD ref Provided to the database 107. Note that the processing unit 103 may also generate a plurality of reference text analysis data AD ref Collect into a corpus of data and output it.

[0130] Note that, in addition to Japanese text, text analysis of various languages ​​(eg, English, Chinese, Korean, etc.) can be performed. As a text analysis method, various methods can be applied according to the language.

[0131] [Step S3]

[0132] After step S2, the analysis data AD included in the plurality of reference texts is calculated. ref IDF of the words in, generate IDF data ID ( Figure 2 and Figure 3B ). Here, it is preferred to normalize the IDF.

[0133] Reference text analysis data AD ref The data is provided from the database 107 to the processing unit 103 through the transmission channel 102 .

[0134] The processing unit 103 calculates the number of reference text analysis data AD included in n ref The IDF of the words in is normalized to generate the IDF data ID.

[0135] like Figure 3B As shown, the IDF data ID includes a word (Word) and a normalized IDF.

[0136] The IDF(t) of a word t is obtained by normalizing the idf(t) of equation (1). There is no particular restriction on the normalization method. For example, the idf(t) can be normalized using equation (2). In equation (1), N represents the number of all documents (refer to the text analysis data AD ref The number of words is equal to the above n), df(t) represents the number of documents in which a word t appears (refer to text analysis data AD ref In formula (2), idf MAX Indicates that the reference text analysis data AD is included ref The maximum value of the idf(t) of the word in MIN Indicates that the reference text analysis data AD is included ref The minimum value of idf(t) of the words in .

[0137] [Formula 1]

[0138]

[0139] For example, the normalized IDF of word A is 0.868, the normalized IDF of word B is 0.115, and the normalized IDF of word C is 0.642. Therefore, it can be said that word A rarely appears in the reference text analysis data AD because its IDF is higher than that of word B and word C. ref Note that in Figure 3B In FIG. 1 , an example of sorting by name is shown, but the present invention is not limited to this, and words may be sorted in the order of IDF or the like.

[0140] exist Figure 3B , an example is shown in which the generated IDF data ID is output to the outside of the processing unit 103. For example, the processing unit 103 may provide the IDF data ID to the database 107 via the transmission channel 102.

[0141] [Step S4]

[0142] After step S2, a plurality of reference text analysis data AD is generated. ref The distribution of words in represents the vector, generating vector data VD( Figure 2 and Figure 3C ).

[0143] Note that step S3 or step S4 may be performed first, or step S3 and step S4 may be performed simultaneously.

[0144] The processing unit 103 generates the n reference text analysis data AD ref The distribution of words in is represented by vectors, generating vector data VD.

[0145] The word distribution representation is also called word embedding, and the word distribution representation vector is a vector that represents the word using a continuous value quantified for each feature element (dimension). The vectors of words with similar meanings are close to each other.

[0146] like Figure 3C As shown, the processing unit 103 preferably uses a neural network NN to generate a distribution representation vector of a word.

[0147] Here, an example of a method for generating a distribution representation vector of a word using a neural network NN is described. The learning of the neural network NN is performed using supervised learning. Specifically, a certain word is provided to the input layer, and the surrounding words of the word are provided to the output layer, and the neural network NN learns the probability of the surrounding words relative to the certain word. The middle layer (hidden layer) preferably has a lower dimensional vector of more than 10 dimensions and less than 1000 dimensions. The learned vector is a distribution representation vector of the word.

[0148] The word distribution representation can be performed using, for example, the open-code algorithm Word2vec. In Word2vec, words (including word features and semantic structures) are vectorized under the assumption that words used in the same context have the same meaning.

[0149] In word vectorization, by generating a distribution representation vector of a word, the similarity or distance between words can be calculated using operations between vectors. When the similarity between two vectors is high, it can be said that the two vectors are highly correlated. In addition, when the distance between two vectors is close, it can be said that the two vectors are highly correlated.

[0150] In addition, in one-hot representation, one dimension is assigned to one word, while in distributed representation, a word can be represented by a low-dimensional real-valued vector, so even if the vocabulary increases, it can be represented with a smaller number of dimensions. Therefore, even if the number of words included in the corpus is large, the amount of calculation is not likely to increase and a large amount of data can be processed in a short time.

[0151] like Figure 3C As shown, the vector data VD includes words and vectors. For example, the vector of word A is (0.12, 0.90, 0.32, ...), the vector of word B is (0.88, 0.10, 0.29, ...), and the vector of word C is (0.23, 0.56, 0.47, ...).

[0152] exist Figure 3C 2 shows an example of outputting the generated vector data VD to the outside of the processing unit 103. For example, the processing unit 103 may provide the vector data VD to the database 107.

[0153] By performing the above steps in advance, reference text analysis data AD is generated ref , IDF data ID and vector data VD, and the file retrieval system 100 can be used for file retrieval.

[0154] Note that in this embodiment, the reference text analysis data AD is generated using the document retrieval system 100. ref , IDF data ID and vector data VD, however, the reference text analysis data AD may be generated outside the document retrieval system 100. ref , at least one of IDF data ID and vector data VD. At this time, the data generated externally is input to the input unit 101 of the document search system 100 and stored in the database 107, thereby enabling retrieval using the document search system 100.

[0155] Next, a search method using the document search system 100 will be described. Figure 4 , Figure 5 , Figure 8 and Fig. 9 Shown is a flow chart, FIG6, Fig. 7A FIG. 10 is a schematic diagram showing each step in the flowchart. Fig. 7A The diagrams of the various data shown in FIG. 10 are merely examples, and the present invention is not limited to these.

[0156] [Step S11]

[0157] First, text data TD is input to the input unit 101 ( Figure 4 , Figure 5 and Fig. 6A ).

[0158] In the file retrieval method of this embodiment, the reference text data TD ref The text data TD is searched for text data related to or similar to the text data TD.

[0159] The text data TD is inputted from the outside of the document retrieval system 100 to the input unit 101. Then, the text data TD is supplied from the input unit 101 to the processing unit 103 through the transmission channel 102. Alternatively, the text data TD may be stored in the storage unit 105 or the database 107 through the transmission channel 102, and supplied from the storage unit 105 or the database 107 to the processing unit 103 through the transmission channel 102.

[0160] In this embodiment, an example is shown in which the search target is a document related to intellectual property. The text data TD is text data of a document related to intellectual property. The example of the document related to intellectual property is as described above.

[0161] The text data TD may include, for example, texts describing inventions, plans, designs, industrial products, technical information, or technical ideas before filing an application.

[0162] In particular, as the text data TD, the claims, the abstract of the specification, or the text describing the content of the invention can be appropriately used. Such text data TD with a small amount of text (less text than the full text of the specification) is preferred because it is easy to extract characteristic keywords included in the text data TD. Since the file retrieval system of this embodiment can extract related words of characteristic keywords, a high-precision search can be performed even if the vocabulary of the text data TD is small.

[0163] [Step S12]

[0164] Next, the text data TD is subjected to morphological analysis to generate text analysis data AD ( Figure 4 , Figure 5 and Fig. 6A ).

[0165] The processing unit 103 performs morphological analysis on the text data TD to generate text analysis data AD.

[0166] exist Fig. 6A In the example, the input text data TD includes a long sentence, and in the output text analysis data AD, the sentence is divided into a plurality of words.

[0167] exist Fig. 6A 2 shows an example of outputting the generated text analysis data AD to the outside of the processing unit 103. For example, the processing unit 103 may provide the text analysis data AD to the storage unit 105 or the database 107.

[0168] [Steps S13, S33]

[0169] Next, by checking the text analysis data AD and the IDF data ID, the keyword data KD is generated ( Figure 4 , Figure 5 and Figure 6B ).

[0170] The text analysis data AD is supplied from the storage unit 105 or the database 107 to the processing unit 103 through the transmission channel 102. The IDF data ID is supplied from the database 107 to the processing unit 103 through the transmission channel 102.

[0171] The processing unit 103 collates the text analysis data AD and the IDF data ID, sorts the words included in the text analysis data AD in order of high IDF, and generates the keyword data KD.

[0172] The keyword data KD includes the keyword KW and its IDF. Here, an example using normalized IDF is shown.

[0173] It can be said that the keyword KW is a characteristic word included in the text analysis data AD. The keyword KW may be, for example, all words included in both the text analysis data AD and the IDF data ID, words whose IDF is greater than a specified value, or a specified number of words whose IDF is relatively high and in a high position. It is preferred to set the extraction criteria of the keyword KW according to the text volume of the text data TD. For example, the number of keywords KW is preferably more than two and less than one hundred, and more preferably more than five and less than thirty.

[0174] Here, the normalized IDF of the keyword KW is equivalent to the reference text analysis data AD later. ref The weight of the keyword KW used when scoring.

[0175] exist Figure 6B , an example is shown in which the weight of Word D is 0.873, the weight of Word A is 0.868, and the weight of Word E is 0.867.

[0176] exist Figure 6B , an example of outputting the keyword data KD to the outside of the processing unit 103 is shown. For example, the processing unit 103 can provide the keyword data KD to the storage unit 105 or the database 107 through the transmission channel 102.

[0177] [Steps S14, S34]

[0178] Next, the text analysis data AD or keyword data KD and vector data VD are used to generate related word data RD ( Figure 4 , Figure 5 and Figure 6C ).

[0179] The text analysis data AD or the keyword data KD is supplied from the storage unit 105 or the database 107 to the processing unit 103 through the transmission channel 102. The vector data VD is supplied from the database 107 to the processing unit 103 through the transmission channel 102.

[0180] In step S14, the processing unit 103 analyzes the distribution representation vector of the keyword KW based on the distribution representation vector of the keyword KW and the distribution representation vector of the keyword KW included in the reference text analysis data AD. refThe distribution of words in the representation vector is used to extract related words RW of the keyword KW based on the similarity or proximity of the distance. And, by sorting the related words RW in the order of high similarity or close distance, the related word data RD is generated. Specifically, it is preferred to extract more than one and less than ten related words RW for one keyword KW, and it is more preferred to extract more than two and less than five related words RW. The related words RW can be, for example, words with a similarity greater than a specified value, words with a distance less than a specified value, words with a high similarity and a specified number of upper positions, or words with a close distance and a specified number of upper positions. The number of synonyms, near-synonyms, antonyms, hypernyms, hyponyms, etc. varies depending on the keyword KW, so the number of related words RW can also vary depending on the keyword KW. By analyzing the related words RW from the reference text data AD included in the reference text analysis data AD, the number of related words RW can be different depending on the keyword KW. ref Extract the related words RW of the keyword KW from the words in the text analysis data AD ref By expressing the keyword KW with a unique record, the record can be extracted as the related word RW. Therefore, it is possible to reduce search omissions due to inconsistent records, which is preferable.

[0181] In step S34, the processing unit 103 compares the distribution representation vectors of the words included in the text analysis data AD with the distribution representation vectors of the words included in the reference text analysis data AD. ref The distribution of words in represents the similarity of the vectors or the proximity of the distances, and the related words RW are extracted. The rest is the same as step S14.

[0182] The similarity between two vectors can be calculated using cosine similarity, covariance, unbiased covariance, Pearson correlation coefficient, etc. In particular, cosine similarity is preferably used.

[0183] The distance between two vectors can be calculated using Euclidean distance, standard (normalized, average) Euclidean distance, Mahalanobis distance, Manhattan distance, Chebyshev distance, Minkowski distance, etc.

[0184] The related word data RD includes related words RW and their relevance RS.

[0185] It can be said that the related word RW is a word related to the word or keyword KW included in the text analysis data AD.

[0186] The relevance RS is a value indicating the proximity of the above-mentioned similarity or distance or a value obtained by normalizing the similarity or distance. The relevance RS is used to calculate the weight of the related word, which is later given to the reference text analysis data AD. ref Used for scoring. Specifically, the product of the normalized IDF of the keyword KW and the relevance RS of the related word RW is equivalent to the weight of the related word.

[0187] exist Figure 6C2 shows an example in which Word X (correlation RS: 0.999), Word Y (correlation RS: 0.901), and Word Z (correlation RS: 0.712) are extracted in the order of highest correlation RS as the related words RW of Word D.

[0188] exist Figure 6C 2 shows an example of outputting the related word data RD to the outside of the processing unit 103. For example, the processing unit 103 may provide the related word data RD to the storage unit 105 or the database 107 through the transmission channel 102.

[0189] Note that when using keyword data KD to extract related words, such as Figure 4 As shown, step S14 is performed after step S13. On the other hand, when the text analysis data AD is used to extract related words, as shown in FIG. Figure 5 As shown, step S33 or step S34 may be performed first, or step S33 and step S34 may be performed simultaneously.

[0190] Furthermore, the following steps may also be included: determining whether the related word RW is a word included in the concept dictionary, and determining the weight of the related word RW. In the case where the related word RW is included in the concept dictionary, it can be said that the related word RW has a higher correlation with the keyword KW. Therefore, the weight of the related word RW when included in the concept dictionary may also be set to be greater than the weight of the related word RW when not included in the concept dictionary. For example, according to the judgment result, a value obtained by adding a specified value to a value representing the proximity of similarity or distance, or a value obtained by subtracting a specified value from a value representing the proximity of similarity or distance may be used as the weight of the related word RW. Alternatively, when the related word RW is included in the concept dictionary or when the related word RW is not included in the concept dictionary, the specified value may also be used as the weight of the related word RW regardless of the proximity of similarity or distance. For example, in the case where the related word RW is included in the concept dictionary, the weight of the related word RW may also be the same as the weight of the keyword KW.

[0191] [Step S15]

[0192] Next, based on the analysis data AD included in the reference text ref The weight of the keyword KW or related word RW matched by the word in the reference text analysis data AD ref Score( Figure 4 , Figure 5 and Fig. 7A ).

[0193] Reference text analysis data AD refThe keyword data KD and the related word data RD are provided from the storage unit 105 or the database 107 to the processing unit 103 through the transmission channel 102. The processing unit 103 can provide the scoring result to the storage unit 105 or the database 107 through the transmission channel 102.

[0194] First, use Figure 7B Explain specific examples of scoring. Figure 7B 2 shows an example using three keywords KW and four related words RW for one keyword KW.

[0195] The denominator of the fraction is the sum of the weights of the keywords KW. Figure 7B In the above equation, 0.9+0.9+0.8=2.6.

[0196] The numerator of the score is the same as the AD included in the reference text analysis data ref The sum of the weights of the keywords KW or related words RW that match the words in Figure 7B It is 1.95 which is the sum of the weights of Word D, Word e and Word f.

[0197] Therefore, the score can be calculated as 1.95 / 2.6=0.75 (75%).

[0198] use Figure 8 Step S15 is described in detail. Figure 8 As shown, step S15 includes steps S21 to S27. When the number of keywords KW is p and the number of related words RW of each keyword KW is q, x represents an integer greater than 1 and less than p, and y represents an integer greater than 1 and less than q.

[0199] [Step S21]

[0200] First, select an unscored reference text analysis data AD ref .

[0201] [Step S22]

[0202] Next, determine the reference text analysis data AD ref Is the keyword KW found in x If found, proceed to step S25. If not found, proceed to step S23.

[0203] [Step S23]

[0204] Next, determine the reference text analysis data AD ref Is the keyword KW found in xRelated words for RW xy If found, proceed to step S25. If not found, proceed to step S24.

[0205] [Step S24]

[0206] Next, it is determined whether all related words RW of the keyword KWx have been searched. If the search has been completed, the process proceeds to step S26. If the search has not been completed, the process proceeds to step S23. For example, if there are two keywords KW x and determine whether the related word RW is found in the previous step S23. x1 When the relevant word RW is found, the process returns to step S23 to determine whether the relevant word RW is found. x2 .

[0207] [Step S25]

[0208] In step S25, the weight corresponding to the found word is added to the score. x The IDF of the keyword KW is added to the score. x IDF and related words RW xy The product of the correlation RS is added to the score. Note that in the above scoring example, the numerator of the score is added.

[0209] [Step S26]

[0210] Next, it is determined whether all the keywords KW have been searched. If the search has been completed, the process proceeds to step S27. If the search has not been completed, the process proceeds to step S22. For example, if there are two keywords KW x When it is determined in the previous step S22 whether the keyword KW1 is found, the process returns to step S22 to determine whether the keyword KW2 is found.

[0211] [Step S27]

[0212] Next, determine whether to analyze data AD for all reference texts ref If the scoring is completed, proceed to step S16. If the scoring is not completed, proceed to step S21.

[0213] [Step S16]

[0214] And, give reference text analysis data AD ref Sorting, generating and outputting sorted data LD( Figure 4 , 5 and Fig. 7A ).

[0215] The processing unit 103 can provide the sorting data LD to the storage unit 105 or the database 107 through the transmission channel 102. In addition, the processing unit 103 can provide the sorting data LD to the output unit 109 through the transmission channel 102. Thus, the output unit 109 can supply the sorting data LD to the outside of the document retrieval system 100.

[0216] The sorting data LD may include the level (Lank), the reference text data TD ref Information (name or identification number, etc.) (Doc), score (Score), etc. Note that in the reference text data TD ref When stored in the database 107 or the like, the sorting data LD preferably includes reference text data TD ref Thus, the user can easily access the target file from the sorting data LD.

[0217] It can be said that the reference text analysis data AD ref The higher the score, the more corresponding to the reference text analysis data AD ref Reference text data TD ref The higher the correlation or similarity with the text data TD.

[0218] exist Fig. 7A In the example shown, according to the sorting data LD, among the n reference text data TD ref The data with the greatest correlation or similarity to the text data TD is the reference text data TD ref (7) The data with the second largest relevance or similarity is the reference text data TD ref (4) The third most relevant or similar data is the reference text data TD ref (13).

[0219] Through the above steps, the document retrieval system 100 can be used for retrieval.

[0220] Furthermore, the keyword KW, the weight (IDF) of the keyword KW, the related word RW, and the weight (IDF×RS) of the related word RW output in step S14 may be manually edited before proceeding to step S15.

[0221] Fig. 9 FIG. 10 shows a flowchart including editing steps. Fig. 9 Note that the data diagram shown in FIG10 is only an example and is not limited to this.

[0222] [Step S41]

[0223] After step 14, a list of the keyword data KD and related word data RD used in step S15 is output. Fig. 9 and Fig. 10A ).

[0224] according to Fig. 10A It can be seen that Word D, Word A, and Word E are cited as the keyword KW, and their weights (IDF) are 0.9, 0.9, and 0.8, respectively.

[0225] It can be seen that Word X, Word Y, Word Z, and Word a are cited as related words RW of Word D, and their weights (IDF×RS) are 0.9, 0.8, 0.6, and 0.5, respectively.

[0226] It can be seen that Word b, Word c, Word d, and Word e are cited as related words RW of Word A, and their weights (IDF×RS) are 0.5, 0.5, 0.45, and 0.3, respectively.

[0227] It can be seen that Word f, Word g, Word h, and Word i are cited as related words RW of Word E, and their weights (IDF×RS) are 0.75, 0.75, 0.75, and 0.75, respectively.

[0228] [Step S42]

[0229] Next, edit the keyword data KD and related word data RD ( Fig. 9 and Fig. 10B ).

[0230] exist Fig. 10B An example of editing three parts is shown in . Specifically, Word A and its related words RW are deleted, Word a (weight 0.5) is changed to Word x (weight 0.8), and the weight of Word f is changed (from 0.75 to 0.8).

[0231] In this way, the user can also improve the search accuracy by editing at least one of the words and the weights.

[0232] Then, in step S15, the edited keyword data KD and related word data RD are used to analyze the reference text data AD. ref After scoring, in step S16, ranking data LD is generated and output.

[0233] [Step S43]

[0234] Next, confirm whether the sorted data LD is the expected result ( Fig. 9). When the sorting data LD is the expected result, the search ends. When the expected result is not obtained, the process may return to step S41, output the edited keyword data KD and the list of related word data RD, and edit again in step S42.

[0235] Note that editing of words and weights is not limited to manual editing, and can also be automatically edited using dictionary data, analysis data generated by natural language processing, etc. Editing can improve search accuracy.

[0236] <3. Configuration Example 2 of Document Retrieval System>

[0237] Next, Fig.11 The document retrieval system 150 shown is described.

[0238] Fig.11 2 is a block diagram of a document search system 150. The document search system 150 includes a server 151 and a terminal 152 (a personal computer or the like).

[0239] The server 151 includes a communication unit 161a, a transmission channel 162, a processing unit 163a and a database 167. Fig.11 Although not shown in the figure, the server 151 may also include a storage unit, an input / output unit, and the like.

[0240] The terminal 152 includes a communication unit 161b, a transmission channel 168, a processing unit 163b, a storage unit 165, and an input / output unit 169. Fig.11 Although not shown in the figure, the terminal 152 may also include a database, etc.

[0241] The user of the document search system 150 inputs text data TD to the server 151 from the terminal 152. The text data TD is transmitted from the communication unit 161b to the communication unit 161a.

[0242] The text data TD received by the communication unit 161a is stored in the database 167 or a storage unit (not shown) through the transmission channel 162. Alternatively, the text data TD may be directly supplied from the communication unit 161a to the processing unit 163a.

[0243] The various processes described in the above (2. File retrieval method) are performed in the processing unit 163a. ​​Since these processes require high processing capabilities, they are preferably performed in the processing unit 163a included in the server 151.

[0244] Then, the sorting data LD is generated by the processing unit 163a. ​​The sorting data LD is stored in the database 167 or the storage unit (not shown) through the transmission channel 162. Alternatively, the sorting data LD can also be directly supplied from the processing unit 163a to the communication unit 161a. Then, the sorting data LD is output from the server 151 to the terminal 152. The sorting data LD is transmitted from the communication unit 161a to the communication unit 161b.

[0245] [Input / output unit 169]

[0246] Data is supplied from outside the document search system 150 to the input / output unit 169. The input / output unit 169 has a function of supplying data to the outside of the document search system 150. Note that, as in the document search system 100, the input unit and the output unit may be separated from each other.

[0247] [Transmission channel 162 and transmission channel 168]

[0248] The transmission channel 162 and the transmission channel 168 have the function of transmitting data. The transmission and reception of data between the communication unit 161a, the processing unit 163a and the database 167 can be performed through the transmission channel 162. The transmission and reception of data between the communication unit 161b, the processing unit 163b, the storage unit 165 and the input / output unit 169 can be performed through the transmission channel 168.

[0249] [Processing Unit 163a and Processing Unit 163b]

[0250] The processing unit 163a has a function of performing calculations, inferences, etc. by using data supplied from the communication unit 161a and the database 167. The processing unit 163b has a function of performing calculations, etc. by using data supplied from the communication unit 161b, the storage unit 165, the input / output unit 169, etc. For the processing unit 163a and the processing unit 163b, reference can be made to the description of the processing unit 103. In particular, the processing unit 163a can perform various processes described in (2. File retrieval method>). Therefore, the processing capacity of the processing unit 163a is preferably higher than that of the processing unit 163b.

[0251] [Storage unit 165]

[0252] The storage unit 165 has a function of storing programs executed by the processing unit 163b and also has a function of storing calculation results generated by the processing unit 163b, data input to the communication unit 161b, and data input to the input / output unit 169.

[0253] [Database 167]

[0254] The database 167 has a storage function for reference text analysis data AD ref, IDF data ID and vector data VD. In addition, the database 167 may also have a function of storing the calculation results generated by the processing unit 163a and the data input to the communication unit 161a. Alternatively, the server 151 may also include a storage unit other than the database 167, and the storage unit may also have a function of storing the calculation results generated by the processing unit 163a and the data input to the communication unit 161a.

[0255] [Communication Unit 161a and Communication Unit 161b]

[0256] By using the communication unit 161a and the communication unit 161b, data can be sent and received between the server 151 and the terminal 152. As the communication unit 161a and the communication unit 161b, a hub, a router, a modem, etc. can be used. The data can be sent and received by wire or by wireless (for example, radio waves, infrared rays, etc.).

[0257] As described above, in the document retrieval system of the present embodiment, it is possible to retrieve documents related to or similar to the input document using the documents prepared in advance as the retrieval object. The user does not need to select a keyword for retrieval and can use text data with more weight than the keyword for retrieval, so the difference in retrieval accuracy between individuals can be reduced, and the document retrieval system can be performed simply and with high accuracy. In addition, in the document retrieval system of the present embodiment, related words of the keyword are extracted from the documents prepared in advance, so that unique records included in the document can be extracted as related words and retrieval omissions can be reduced. In addition, in the document retrieval system of the present embodiment, the retrieval results can be sorted and output in the order of high relevance or similarity, so that the user can easily find the necessary documents from the retrieval results and it is not easy to cause negligence.

[0258] This embodiment mode can be appropriately combined with other embodiment modes. In this specification, when a plurality of configuration examples are shown in one embodiment mode, the configuration examples can be appropriately combined.

[0259] (Implementation Method 2)

[0260] In this embodiment, a structural example of a semiconductor device that can be used for a neural network is described.

[0261] The semiconductor device of this embodiment can be used for example in a processing unit of a document search system which is one embodiment of the present invention.

[0262] like Fig. 12AAs shown, the neural network NN can be composed of an input layer IL, an output layer OL and an intermediate layer (hidden layer) HL. The input layer IL, the output layer OL and the intermediate layer HL all include one or more neurons (units). Note that the intermediate layer HL can be one layer or more than two layers. A neural network including more than two layers of intermediate layers HL can be called a DNN (deep neural network), and learning using a deep neural network can be called deep learning.

[0263] Each neuron in the input layer IL is input with input data, each neuron in the intermediate layer HL is input with output signals of neurons in the previous or next layer, and each neuron in the output layer OL is input with output signals of neurons in the previous layer. Note that each neuron can be connected to all neurons in the previous and next layers (fully connected) or to some neurons.

[0264] Fig. 12B An example of operation using neurons is shown. Here, neuron N and two neurons in the previous layer that output signals to neuron N are shown. Neuron N is input with the output x1 of the neuron in the previous layer and the output x2 of the neuron in the previous layer. In neuron N, the sum x1w1+x2w2 of the multiplication result (x1w1) of the output x1 and the weight w1 and the multiplication result (x2w2) of the output x2 and the weight w2 is calculated, and then a bias b is added as needed to obtain the value a=x1w1+x2w2+b. The value a is transformed by the activation function h, and the output signal y=h(a) is output from neuron N.

[0265] In this way, the operation using neurons includes the operation of adding the product of the output of the neurons in the previous layer and the weight, that is, the sum-of-products operation (the above x1w1+x2w2). The sum-of-products operation can be performed in software through a program or in hardware. When the sum-of-products operation is performed by hardware, a sum-of-products operation circuit can be used. As the sum-of-products operation circuit, both a digital circuit and an analog circuit can be used. When an analog circuit is used as a sum-of-products operation circuit, the circuit scale of the sum-of-products operation circuit can be reduced or the number of accesses to the memory can be reduced, thereby achieving an increase in processing speed and a reduction in power consumption.

[0266] The multiplication and sum operation circuit can be composed of a transistor containing silicon (single crystal silicon, etc.) in the channel formation region (also called "Si transistor"), and can also be composed of a transistor containing an oxide semiconductor of one of the metal oxides in the channel formation region (also called "OS transistor"). In particular, since the OS transistor has an extremely small off-state current, it is preferably used as a transistor for the memory constituting the multiplication and sum operation circuit. Note that the multiplication and sum operation circuit can also be composed of both Si transistors and OS transistors. Below, a structural example of a semiconductor device having the function of a multiplication and sum operation circuit is described.

[0267] (Structural example of semiconductor device>

[0268] Fig.13 An example of the structure of a semiconductor device MAC having a function of performing a neural network operation is shown. The semiconductor device MAC has a function of performing a product-sum operation of first data corresponding to the connection strength (weight) between neurons and second data corresponding to input data. Note that the first data and the second data may be analog data or multi-valued digital data (dispersed data), respectively. In addition, the semiconductor device MAC has a function of transforming data obtained by the product-sum operation using an activation function.

[0269] The semiconductor device MAC includes a cell array CA, a current source circuit CS, a current mirror circuit CM, a circuit WDD, a circuit WLD, a circuit CLD, a bias circuit OFST, and an activation function circuit ACTV.

[0270] The cell array CA includes a plurality of memory cells MC and a plurality of memory cells MCref. Fig.13 The cell array CA includes m rows and n columns (m and n are integers greater than 1) of memory cells MC (MC[1,1] to MC[m,n]) and m memory cells MCref (MCref[1] to MCref[m]). The memory cell MC has a function of storing first data. In addition, the memory cell MCref has a function of storing reference data for product-sum operations. Note that the reference data may be analog data or multi-valued digital data.

[0271] The memory cell MC[i, j] (i is an integer greater than 1 and less than m, and j is an integer greater than 1 and less than n) is connected to the wiring WL[i], the wiring RW[i], the wiring WD[j], and the wiring BL[j]. In addition, the memory cell MCref[i] is connected to the wiring WL[i], the wiring RW[i], the wiring WDref, and the wiring BLref. Here, the current flowing between the memory cell MC[i, j] and the wiring BL[j] is recorded as I MC[i，j] , the current flowing between the memory cell MCref[i] and the wiring BLref is recorded as I MCref[i] .

[0272] Fig.14 Specific structural examples of the memory cell MC and the memory cell MCref are shown. Fig.141 and 2 are typical examples of memory cells MC[1,1], MC[2,1] and memory cells MCref[1], MCref[2]. However, the same structure can be used in other memory cells MC and memory cells MCref. Both memory cells MC and memory cells MCref include transistors Tr11, Tr12, and capacitor C11. Here, the case where transistors Tr11 and Tr12 are n-channel transistors is described.

[0273] In the memory cell MC, the gate of the transistor Tr11 is connected to the wiring WL, one of the source and the drain is connected to the gate of the transistor Tr12 and the first electrode of the capacitor C11, and the other of the source and the drain is connected to the wiring WD. One of the source and the drain of the transistor Tr12 is connected to the wiring BL, and the other of the source and the drain is connected to the wiring VR. The second electrode of the capacitor C11 is connected to the wiring RW. The wiring VR has a function of supplying a predetermined potential. Here, as an example, a case where a low power supply potential (ground potential, etc.) is supplied from the wiring VR is described.

[0274] The node connected to one of the source and drain of transistor Tr11, the gate of transistor Tr12, and the first electrode of capacitor C11 is referred to as node NM. In addition, the nodes NM of memory cells MC[1,1] and MC[2,1] are referred to as nodes NM[1,1] and NM[2,1], respectively.

[0275] The memory cell MCref also has the same structure as the memory cell MC. However, the memory cell MCref is connected to the wiring WDref instead of the wiring WD and is connected to the wiring BLref instead of the wiring BL. In the memory cells MCref[1] and MCref[2], nodes connected to one of the source and drain of the transistor Tr11, the gate of the transistor Tr12, and the first electrode of the capacitor C11 are respectively referred to as nodes NMref[1] and NMref[2].

[0276] The node NM and the node NMref are used as holding nodes of the memory cell MC and the memory cell MCref, respectively. The node NM holds the first data, and the node NMref holds the reference data. MC[1，1] ,I MC[2，1] The current I flows from the wiring BL[1] to the transistors Tr12 of the memory cells MC[1,1] and MC[2,1]. MCref[1] ,I MCref[2] The power flows from the wiring BLref to the transistors Tr12 of the memory cells MCref[1] and MCref[2], respectively.

[0277] Since the transistor Tr11 has the function of maintaining the potential of the node NM or the node NMref, the off-state current of the transistor Tr11 is preferably small. Therefore, as the transistor Tr11, it is preferable to use an OS transistor with a very small off-state current. As a result, the potential variation of the node NM or the node NMref can be suppressed to improve the calculation accuracy. In addition, the frequency of the operation of refreshing the potential of the node NM or the node NMref can be suppressed to a low level, thereby reducing power consumption.

[0278] There is no particular limitation on the transistor Tr12, and for example, a Si transistor or an OS transistor may be used. When an OS transistor is used as the transistor Tr12, the transistor Tr12 can be manufactured using the same manufacturing apparatus as the transistor Tr11, thereby reducing manufacturing costs. Note that the transistor Tr12 may be an n-channel transistor or a p-channel transistor.

[0279] The current source circuit CS is connected to the wirings BL[1] to BL[n] and the wiring BLref. The current source circuit CS has a function of supplying current to the wirings BL[1] to BL[n] and the wiring BLref. Note that the current value supplied to the wirings BL[1] to BL[n] may be different from the current value supplied to the wiring BLref. Here, the current supplied from the current source circuit CS to the wirings BL[1] to BL[n] is referred to as I C The current supplied from the current source circuit CS to the wiring BLref is recorded as I Cref .

[0280] The current mirror circuit CM includes wirings IL[1] to IL[n] and wiring ILref. The wirings IL[1] to IL[n] are connected to the wirings BL[1] to BL[n], respectively, and the wiring ILref is connected to the wiring BLref. Here, the connection between the wirings IL[1] to IL[n] and the wirings BL[1] to BL[n] is referred to as nodes NP[1] to NP[n]. In addition, the connection between the wiring ILref and the wiring BLref is referred to as a node NPref.

[0281] The current mirror circuit CM has a function of converting a current I corresponding to the potential of the node NPref. CM The function of flowing to the wiring ILref and also transferring the current I CM Flow to the function of wiring IL[1] to IL[n]. Fig.13 Shows the current I CM The current I is discharged from the wiring BLref to the wiring ILref. CM This is an example of a current flowing from the wirings BL[1] to BL[n] to the wirings IL[1] to IL[n]. The current flowing from the current mirror circuit CM to the cell array CA through the wirings BL[1] to BL[n] is expressed as I B[1] to I B [n] In addition, the current flowing from the current mirror circuit CM to the cell array CA through the wiring BLref is recorded as I Bref .

[0282] The circuit WDD is connected to the wirings WD[1] to WD[n] and the wiring WDref. The circuit WDD has a function of supplying a potential corresponding to the first data stored in the memory cell MC to the wirings WD[1] to WD[n]. In addition, the circuit WDD has a function of supplying a potential corresponding to the reference data stored in the memory cell MCref to the wiring WDref. The circuit WLD is connected to the wirings WL[1] to WD[m]. The circuit WLD has a function of supplying a signal selecting the memory cell MC or the memory cell MCref to which data is written to the wirings WL[1] to WD[m]. The circuit CLD is connected to the wirings RW[1] to RW[m]. The circuit CLD has a function of supplying a potential corresponding to the second data to the wirings RW[1] to RW[m].

[0283] The bias circuit OFST is connected to the wirings BL[1] to BL[n] and the wirings OL[1] to OL[n]. The bias circuit OFST has a function of detecting the amount of current flowing from the wirings BL[1] to BL[n] to the bias circuit OFST and / or the amount of change in the current flowing from the wirings BL[1] to BL[n] to the bias circuit OFST. In addition, the bias circuit OFST has a function of outputting the detection result to the wirings OL[1] to OL[n]. Note that the bias circuit OFST can output the current corresponding to the detection result to the wiring OL, and can also convert the current corresponding to the detection result into a voltage and output it to the wiring OL. The current flowing between the cell array CA and the bias circuit OFST is recorded as I α [1] to I α [n].

[0284] Fig.15 A configuration example of the bias circuit OFST is shown. Fig.15 The bias circuit OFST shown includes circuits OC[1] to OC[n]. Circuits OC[1] to OC[n] all include transistors Tr21, Tr22, Tr23, capacitor C21 and resistor R1. The connection relationship of each component is as follows: Fig.15 Note that the node connected to the first electrode of the capacitor C21 and the first terminal of the resistor R1 is referred to as the node Na. In addition, the node connected to the second electrode of the capacitor C21, one of the source and drain of the transistor Tr21, and the gate of the transistor Tr22 is referred to as the node Nb.

[0285] The wiring VrefL has a function of supplying the potential Vref, the wiring VaL has a function of supplying the potential Va, and the wiring VbL has a function of supplying the potential Vb. The wiring VDDL has a function of supplying the potential VDD, and the wiring VSSL has a function of supplying the potential VSS. Here, the case where the potential VDD is a high power supply potential and the potential VSS is a low power supply potential is described. The wiring RST has a function of supplying a potential for controlling the conduction state of the transistor Tr21. The transistor Tr22, the transistor Tr23, the wiring VDDL, the wiring VSSL, and the wiring VbL constitute a source follower circuit.

[0286] Next, an operation example of circuits OC[1] to OC[n] is described. Note that although the operation example of circuit OC[1] is described here as a typical example, circuits OC[2] to OC[n] can also operate in the same manner. First, when a first current flows to wiring BL[1], the potential of node Na becomes a potential corresponding to the first current and the resistance value of resistor R1. At this time, transistor Tr21 is turned on, and potential Va is supplied to node Nb. Then, transistor Tr21 is turned off.

[0287] Next, when the second current flows to the wiring BL[1], the potential of the node Na changes to a potential corresponding to the second current and the resistance value of the resistor R1. At this time, the transistor Tr21 is in the off state and the node Nb is in the floating state, so when the potential of the node Na changes, the potential of the node Nb changes due to capacitive coupling. Here, the potential change at the node Na is △V Na When the capacitance coupling coefficient is 1, the potential of node Nb is Va+△V Na The threshold voltage of transistor Tr22 is V th When the potential Va+△V is output from the wiring OL[1] Na -V th Here, by satisfying Va=V th , the potential △V can be output from the wiring OL[1] Na .

[0288] Potential △V Na The voltage Vref is determined by the change from the first current to the second current, the resistance value of the resistor R1, and the potential Vref. Na The amount of change in the current flowing to the wiring BL.

[0289] As described above, a signal corresponding to the amount of current and / or the amount of change in current detected by the bias circuit OFST is input to the activation function circuit ACTV through the wirings OL[1] to OL[n].

[0290] The activation function circuit ACTV is connected to the wiring OL[1] to OL[n] and the wiring NIL[1] to NIL[n]. The activation function circuit ACTV has a function of performing an operation to transform the signal input from the bias circuit OFST according to a predetermined activation function. As the activation function, for example, a sigmoid function, a tanh function, a softmax function, a ReLU function, and a threshold function can be used. The signal transformed by the activation function circuit ACTV is output as output data to the wiring NIL[1] to NIL[n].

[0291] <Operation Example of Semiconductor Device>

[0292] The semiconductor device MAC described above can perform a product-sum operation on the first data and the second data. Next, an operation example of the semiconductor device MAC when performing a product-sum operation will be described.

[0293] Fig.16 A timing chart showing an operation example of the semiconductor device MAC. Fig.16 Show Fig.14 The potential changes of the wiring WL[1], wiring WL[2], wiring WD[1], wiring WDref, node NM[1,1], node NM[2,1], node NMref[1], node NMref[2], wiring RW[1], and wiring RW[2], and the current I B [1]-I α [1] and current I Bref The value of current I B [1]-I α [1] corresponds to the total of currents flowing from the wiring BL[1] to the memory cells MC[1,1] and MC[2,1].

[0294] Although we focus on Fig.14 The operation is described using the memory cells MC[1,1], MC[2,1] and the memory cells MCref[1], MCref[2] as typical examples, but other memory cells MC and memory cells MCref can also perform the same operation.

[0295] [Storage of First Data]

[0296] First, during the period from time T01 to time T02, the potential of the wiring WL[1] becomes high (High), and the potential of the wiring WD[1] becomes V higher than the ground potential (GND). PR -V W[1，1] The potential of the wiring WDref becomes V greater than the ground potential. PR The potential of wiring RW[1] and wiring RW[2] becomes the standard potential (REFP). Note that the potential VW[1，1] corresponds to the first data stored in the memory cell MC[1,1]. PR Therefore, the transistor Tr11 of the memory cell MC[1,1] and the memory cell MCref[1] is turned on, and the potential of the node NM[1,1] becomes V PR -V W[1，1] , the potential of node NMref[1] becomes V PR .

[0297] At this time, the current I flowing from the wiring BL[1] to the transistor Tr12 of the memory cell MC[1,1] MC[1，1]，0 It can be expressed as follows. Here, k is a constant depending on the channel length, channel width, mobility, and capacitance of the gate insulating film of the transistor Tr12. In addition, V th is the threshold voltage of transistor Tr12.

[0298] I MC[1，1]，0 = k(V PR -V W[1，1] -V th ) 2 (E1)

[0299] Furthermore, the current I flows from the wiring BLref to the transistor Tr12 of the memory cell MCref[1]. MCref[1]，0 It can be expressed as the following formula.

[0300] I MCref[1]，0 = k(V PR -V th ) 2 (E2)

[0301] Next, during the period from time T02 to time T03, the potential of wiring WL[1] becomes low. Therefore, transistor Tr11 included in memory cell MC[1,1] and memory cell MCref[1] is turned off, and the potentials of node NM[1,1] and node NMref[1] are maintained.

[0302] As described above, it is preferable to use an OS transistor as the transistor Tr11. This can suppress the leakage current of the transistor Tr11 and accurately maintain the potentials of the node NM[1, 1] and the node NMref[1].

[0303] Next, during the period from time T03 to time T04, the potential of the wiring WL[2] becomes high level, and the potential of the wiring WD[1] becomes V higher than the ground potential. PR -V W[2，1] The potential of the wiring WDref becomes V greater than the ground potential.PR Note that the potential V W[2，1] Corresponding to the first data stored in the memory cell MC[2,1], the transistor Tr11 of the memory cell MC[2,1] and the memory cell MCref[2] is turned on, and the potential of the node NM[2,1] becomes V PR -V W[2，1] , the potential of node NMref[2] becomes V PR .

[0304] At this time, the current I flowing from the wiring BL[1] to the transistor Tr12 of the memory cell MC[2,1] MC[2，1]，0 It can be expressed as the following formula.

[0305] I MC[2，1]，0 = k(V PR -V W[2，1] -V th ) 2 (E3)

[0306] Furthermore, the current I flows from the wiring BLref to the transistor Tr12 of the memory cell MCref[2]. MCref[2]，0 It can be expressed as the following formula.

[0307] I MCref[2]，0 = k(V PR -V th ) 2 (E4)

[0308] Next, during the period from time T04 to time T05, the potential of wiring WL[2] becomes low level. Therefore, transistor Tr11 included in memory cell MC[2,1] and memory cell MCref[2] is turned off, and the potentials of node NM[2,1] and node NMref[2] are maintained.

[0309] Through the above operation, the first data is stored in the memory cells MC[1, 1] and MC[2, 1], and the reference data is stored in the memory cells MCref[1] and MCref[2].

[0310] Here, during the period from time T04 to time T05, the current flowing to the wiring BL[1] and the wiring BLref is considered. The current is supplied from the current source circuit CS to the wiring BLref. The current flowing through the wiring BLref is discharged to the current mirror circuit CM and the memory cells MCref[1] and MCref[2]. The current supplied from the current source circuit CS to the wiring BLref is referred to as I Cref The current discharged from the wiring BLref to the current mirror circuit CM is called I CM，0 , then the following formula is satisfied.

[0311] I Cref -I CM，0 =I MCref[1]，0 +I MCref[2]，0 (E5)

[0312] Current is supplied from the current source circuit CS to the wiring BL[1]. The current flowing through the wiring BL[1] is discharged to the current mirror circuit CM and the memory cells MC[1,1] and MC[2,1]. In addition, current flows from the wiring BL[1] to the bias circuit OFST. The current supplied from the current source circuit CS to the wiring BL[1] is referred to as I C，0 The current flowing from the wiring BL[1] to the bias circuit OFST is called I α，0 , then the following formula is satisfied.

[0313] I C -I CM，0 =I MC[1，1]，0 +I MC[2，1]，0 +I α，0 (E6)

[0314] [Product-sum operation of first data and second data]

[0315] Next, during the period from time T05 to time T06, the potential of wiring RW[1] is higher than the standard potential by V. X[1] At this time, the potential V X[1] The voltage V is supplied to each capacitor C11 of the memory cell MC[1,1] and the memory cell MCref[1], and the gate potential of the transistor Tr12 rises due to capacitive coupling. X[1] Corresponding to the second data supplied to the memory cell MC[1, 1] and the memory cell MCref[1].

[0316] The change in the potential of the gate of the transistor Tr12 is equivalent to the value obtained by multiplying the change in the potential of the wiring RW by the capacitance coupling coefficient determined by the structure of the memory cell. The capacitance coupling coefficient is calculated based on the capacitance of the capacitor C11, the gate capacitance of the transistor Tr12, and the parasitic capacitance. Below, for the sake of convenience, the case where the change in the potential of the wiring RW is equal to the change in the potential of the gate of the transistor Tr12, that is, the case where the capacitance coupling coefficient is 1 is described. In fact, considering that the capacitance coupling coefficient determines the potential V X , that's it.

[0317] When the potential V X[1] When the voltage is supplied to the capacitor C11 of the memory cell MC[1,1] and the memory cell MCref[1], the potential of the node NM[1,1] and the node NMref[1] both rises by V X[1] .

[0318] Here, during the period from time T05 to time T06, the current I flowing from the wiring BL[1] to the transistor Tr12 of the memory cell MC[1,1] is MC[1，1]，1 It can be expressed as the following formula.

[0319] I MC[1，1]，1 = k(V PR -V W[1，1] +V X[1] -V th ) 2 (E7)

[0320] That is, by supplying the potential V to the wiring RW[1] X[1] , the current flowing from the wiring BL[1] to the transistor Tr12 of the memory cell MC[1,1] increases by ΔI MC[1，1] =I MC[1，1]，1 -I MC[1，1]，0 .

[0321] In addition, during the period from time T05 to time T06, the current I flowing from the wiring BLref to the transistor Tr12 of the memory cell MCref[1] is MCref[1]，1 It can be expressed as the following formula.

[0322] I MCref[1]，1 = k(V PR +V X[1] -V th ) 2 (E8)

[0323] That is, by supplying the potential V to the wiring RW[1] X[1] , the current flowing from the wiring BLref to the transistor Tr12 of the memory cell MCref[1] increases by ΔI MCref[1] =I MCref[1]，1 -I MCref[1]，0 .

[0324] In addition, consider the current flowing to the wiring BL[1] and the wiring BLref. The current I is supplied from the current source circuit CS to the wiring BLref. Cref The current flowing through the wiring BLref is discharged to the current mirror circuit CM and the memory cells MCref[1] and MCref[2]. The current discharged from the wiring BLref to the current mirror circuit CM is referred to as I CM，1 , then the following formula is satisfied.

[0325] I Cref -I CM，1 =I MCref[1]，1 +I MCref[2]，0 (E9)

[0326] The current I is supplied from the current source circuit CS to the wiring BL[1].C The current flowing through the wiring BL[1] is discharged to the current mirror circuit CM and the memory cells MC[1,1] and MC[2,1]. Furthermore, the current flows from the wiring BL[1] to the bias circuit OFST. The current flowing from the wiring BL[1] to the bias circuit OFST is referred to as I α，1 , then the following formula is satisfied.

[0327] I C -I CM，1 =I MC[1，1]，1 +I MC[2，1]，1 +I α，1 (E10)

[0328] According to equations (E1) to (E10), the current I can be expressed as follows: α，0 With current I α，1 The difference (difference current △I α ).

[0329] △I α =I α，1 -I α，0 =2kV W[1，1] V X[1] (E11)

[0330] Thus, the difference current △I α Indicates the potential V W[1，1] With V X[1] The value of the product of .

[0331] Then, during the period from time T06 to time T07 , the potential of wiring RW[ 1 ] becomes the standard potential, and the potentials of node NM[ 1 , 1 ] and node NMref[ 1 ] become the same as during the period from time T04 to time T05 .

[0332] Next, during the period from time T07 to time T08, the potential of wiring RW[1] becomes V higher than the standard potential. X[1] The potential of wiring RW[2] becomes V greater than the standard potential. X[2] Therefore, the potential V X[1] The capacitor C11 supplied to the memory cell MC[1,1] and the memory cell MCref[1] increases the potential of both the node NM[1,1] and the node NMref[1] by V due to capacitive coupling. X[1] In addition, the potential V X[2] The capacitor C11 supplied to the memory cell MC[2,1] and the memory cell MCref[2] increases the potential of both the node NM[2,1] and the node NMref[2] by V due to capacitive coupling. X[2] .

[0333] Here, during the period from time T07 to time T08, the current I flowing from the wiring BL[1] to the transistor Tr12 of the memory cell MC[2,1] is MC[2，1]，1 It can be expressed as the following formula.

[0334] I MC[2，1]，1 = k(V PR -V W[2，1] +V X[2] -V th ) 2 (E12)

[0335] That is, by supplying the potential V to the wiring RW[2] X[2] , the current flowing from the wiring BL[1] to the transistor Tr12 of the memory cell MC[2,1] increases by ΔI MC[2，1] =I MC[2，1]，1 -I MC[2，1]，0 .

[0336] In addition, during the period from time T07 to time T08, the current I flowing from the wiring BLref to the transistor Tr12 of the memory cell MCref[2] is MCref[2]，1 It can be expressed as the following formula.

[0337] I MCref[2]，1 = k(V PR +V X[2] -V th ) 2 (E13)

[0338] That is, by supplying the potential V to the wiring RW[2] X[2] , the current flowing from the wiring BLref to the transistor Tr12 of the memory cell MCref[2] increases by ΔI MCref[2] =I MCref[2]，1 -I MCref[2]，0 .

[0339] In addition, consider the current flowing to the wiring BL[1] and the wiring BLref. The current I is supplied from the current source circuit CS to the wiring BLref. Cref The current flowing through the wiring BLref is discharged to the current mirror circuit CM and the memory cells MCref[1] and MCref[2]. The current discharged from the wiring BLref to the current mirror circuit CM is referred to as I CM，2 , then the following formula is satisfied.

[0340] I Cref -I CM，2 =I MCref[1]，1 +I MCref[2]，1 (E14)

[0341] The current I is supplied from the current source circuit CS to the wiring BL[1]. C The current flowing through the wiring BL[1] is discharged to the current mirror circuit CM and the memory cells MC[1,1] and MC[2,1]. Furthermore, the current flows from the wiring BL[1] to the bias circuit OFST. The current flowing from the wiring BL[1] to the bias circuit OFST is referred to as I α，2 , then the following formula is satisfied.

[0342] I C -I CM，2 =I MC[1，1]，1 +I MC[2，1]，1 +I α，2 (E15)

[0343] According to equations (E1) to (E8) and equations (E12) to (E15), the current I can be expressed as follows: α，0 With current I α，2 The difference (difference current △I α ).

[0344] △I α =I α，2 -I α，0 =2k(V W[1，1] V X[1] +V W[2，1] V X[2] )(E16)

[0345] Thus, the difference current △I α Indicates the potential V W[1，1] With potential V X[1] The product of the potential V W[2，1] With potential V X[2] The value of the result of adding the products.

[0346] Then, during the period from time T08 to time T09, the potentials of wirings RW[1] and [2] become the standard potential, and the potentials of nodes NM[1,1] and NM[2,1] and nodes NMref[1] and NMref[2] are the same as during the period from time T04 to time T05.

[0347] As shown in equations (E11) and (E16), the differential current △I input to the bias circuit OFST is α The potential V corresponding to the first data (weight) can be W and the potential V corresponding to the second data (input data) X That is, by using the bias circuit OFST to calculate the difference current △I α By performing the measurement, a result of a product-sum operation of the first data and the second data can be obtained.

[0348] Note that although the above description focuses on the memory cells MC[1,1], MC[2,1] and the memory cells MCref[1], MCref[2], the number of memory cells MC and memory cells MCref can be set arbitrarily. When the number of rows m of the memory cells MC and memory cells MCref is set to an arbitrary number i, the difference current ΔI can be expressed by the following formula: α .

[0349] △I α =2kΣ i V W[i，1] V X[i] (E17)

[0350] Furthermore, by increasing the number n of columns of memory cells MC and memory cells MCref, the number of parallel sum-of-product operations can be increased.

[0351] As described above, by using the semiconductor device MAC, a product-sum operation can be performed on the first data and the second data. Fig.14 The structures of the memory cells MC and MCref shown above can form a product-sum operation circuit using fewer transistors, thereby reducing the circuit scale of the semiconductor device MAC.

[0352] When the semiconductor device MAC is used for calculations using a neural network, the number m of rows of storage cells MC can correspond to the number of input data supplied to one neuron and the number n of columns of storage cells MC can correspond to the number of neurons. Fig. 12A In the case where a product-sum operation using the semiconductor device MAC is performed in the intermediate layer HL shown in the figure, the number m of rows of the storage cells MC can be set to the number of input data supplied from the input layer IL (the number of neurons in the input layer IL) and the number n of columns of the storage cells MC can be set to the number of neurons in the intermediate layer HL.

[0353] Note that there is no particular limitation on the structure of the neural network using the semiconductor device MAC. For example, the semiconductor device MAC can be used for a convolutional neural network (CNN), a recurrent neural network (RNN), an autoencoder, a Boltzmann machine (including a restricted Boltzmann machine), and the like.

[0354] As described above, by using the semiconductor device MAC, it is possible to perform product-sum operations in a neural network. Fig.14 The memory cell MC and the memory cell MCref shown are used in the cell array CA, and an integrated circuit with high calculation accuracy, low power consumption or small circuit scale can be provided.

[0355] This embodiment mode can be combined with other embodiment modes as appropriate.

[0356] [Explanation of symbols]

[0357] AD: Text Analysis Data, AD ref : Reference text analysis data, C11: Capacitor, C21: Capacitor, ID: IDF data, KD: Keyword data, KW: Keyword, KW1: Keyword, KW2: Keyword, KW x : keyword, LD: sorting data, NN: neural network, R1: resistor, RD: related word data, RS: relevance, RW: related word, RW x1 :Related words, RW x2 :Related words, RW xy :Related words, TD: Text data, TD ref : Reference text data, Tr11: transistor, Tr12: transistor, Tr21: transistor, Tr22: transistor, Tr23: transistor, VD: vector data, 100: document retrieval system, 101: input unit, 102: transmission channel, 103: processing unit, 105: storage unit, 107: database, 109: output unit, 150: document retrieval system, 151: server, 152: terminal, 161a: communication unit, 161b: communication unit, 162: transmission channel, 163a: processing unit, 163b: processing unit, 165: storage unit, 167: database, 168: transmission channel, 169: input and output unit

Claims

1. A file retrieval device for retrieving files related to or similar to an input file, comprising a processing unit, the processing unit being configured to: Obtaining a first distribution representation vector and a first weight of the first distribution representation vector from the input file; extracting a second distribution representation vector based on a similarity to the first distribution representation vector; Obtaining a second weight of the second distribution representation vector, wherein the second weight includes a product of the first weight and the similarity; outputting the first weight and the second weight; receiving a change value of one or more of the first weight and the second weight; as well as The retrieval of the file is performed by using the edited first weight and the second weight.

2. The file retrieval device according to claim 1, The first distribution representation vector and the second distribution representation vector are obtained by machine learning of distribution representations of words included in a plurality of reference text analysis data.

3. The file retrieval device according to claim 1, The first distribution representation vector is obtained from keywords extracted from the input file.

4. The file retrieval device according to claim 2, The plurality of reference text analysis data are generated by performing morphological analysis on reference text data.

5. The file retrieval device according to claim 2, The machine learning described herein uses a neural network.

6. A device comprising: A storage unit stores thereon instructions, which, when executed by the processing unit, cause the processing unit to perform an operation of retrieving a file based on an input file, the operation comprising: Obtaining a first distribution representation vector and a first weight of the first distribution representation vector from the input file; extracting a second distribution representation vector based on a similarity to the first distribution representation vector; Obtaining a second weight of the second distribution representation vector, wherein the second weight includes a product of the first weight and the similarity; outputting the first weight and the second weight; and The retrieval of the document is performed using a variation of one or more of the first weight and the second weight, the variation being provided by editing the keyword data and the related word data.

7. The device according to claim 6, The first distribution representation vector and the second distribution representation vector are obtained by machine learning of distribution representations of words included in a plurality of reference text analysis data.

8. The device according to claim 7, The plurality of reference text analysis data are generated by performing morphological analysis on reference text data.

9. The device according to claim 7, The machine learning described herein uses a neural network.

10. The device according to claim 6, The first distribution representation vector is obtained from keywords extracted from the input file.

Images