Method, system and storage medium for automatically identifying related compounds in patent documents
By training a chemical entity recognition system to extract and identify related compounds from patent databases, the automated extraction and identification problems in the prior art are solved, and efficient compound extraction and correlation recognition is achieved, reducing maintenance costs and improving retrieval accuracy.
Patent Information
- Application Number
- CN201980029275.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Priority Date
- 2018-03-07
- Filing Date
- 2019-03-06
- Publication Date
- 2024-11-05
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The prior art is difficult to extract and identify related compounds from patent documents automatically and effectively, and manual excerpting methods are time-consuming and labor-intensive. The automatic methods have limitations and maintenance difficulties, and it is impossible to effectively utilize the chemical information in the patent.
By training a chemical entity recognition system, patent documents are obtained from patent databases, standardized and generated chemical patent corpus, and the compound is extracted and its correlation with patent documents is marked by machine learning algorithms, reducing database maintenance costs and improving retrieval efficiency.
Automatic compound extraction and correlation identification are realized, reducing the amount of data, improving the effectiveness and retrieval accuracy of patent resources, and reducing maintenance costs.
Smart Images

Figure CN112543931B_ABST
Abstract
Description
[0001] CROSS-REFERENCE TO RELATED APPLICATIONS
[0002] This application claims priority to U.S. Provisional Patent Application No. 62 / 639,656, filed on March 7, 2018, entitled “Automatic Identification of Related Compounds in Patents,” the entire contents of which are incorporated herein by reference. Technical Field
[0003] The present disclosure relates to methods, systems and storage media for automatically identifying compounds in patent documents, and more specifically, to methods, systems and storage media for training a chemical entity recognition system to automatically extract compounds from patent documents and classify the compounds relative to the corresponding patent document relevance. Background Art
[0004] Chemistry-related publications may include patent applications and scientific journal articles. In a commercial R&D project, the first public disclosure of a new compound may occur during the patent application process. Sometimes, it may take another 1-3 years for these compounds to appear in journal publications. Therefore, these compounds may only be available through patent literature for a period of time. In addition, chemical patent literature may contain unique information such as reactions, experimental conditions, modes of action, biological activity data, and catalysts. Analyzing such information may be necessary because it can help understand the prior art of the compound, provide a means of novelty examination and verification, and point out the starting point for chemical research in academia and industry. Summary of the invention
[0005] One aspect of the present disclosure relates to a method for training a chemical entity recognition system to extract one or more compounds from a patent document and determine the relevance of the one or more compounds to the patent document. The method includes a processing device obtaining a plurality of patent documents from one or more patent databases. The method also includes the processing device normalizing each of the plurality of patent documents into a unified format to achieve a plurality of unified patent documents. The method also includes the processing device generating a chemical patent corpus from the plurality of unified patent documents, the chemical patent corpus including one or more chemical entities extracted from the plurality of unified patent documents. Each of the one or more chemical entities includes one or more relevance annotations indicating the relevance to the patent document from which the chemical entity was extracted. The method also includes the processing device providing the chemical patent corpus to the chemical entity recognition system. In response to receiving the chemical patent corpus, the chemical entity identification system marks the one or more chemical entities in corresponding normalized patent documents of the plurality of unified patent documents, extracts one or more additional chemical entities from the plurality of unified patent documents, assigns a confidence score to each of the one or more additional chemical entities, and marks each of the one or more additional chemical entities as being relevant or irrelevant to the relevant patent document based on information contained in the chemical patent corpus.
[0006] Another aspect of the present disclosure relates to a system configured to train a chemical entity recognition system to extract one or more compounds from a patent document and determine the relevance of the one or more compounds to the patent document. The system includes one or more hardware processors; and a non-transitory processor-readable storage medium containing one or more programming instructions. When the programming instructions are executed, the one or more hardware processors obtain multiple patent documents from one or more patent databases. When the programming instructions are executed, the one or more hardware processors normalize each of the multiple patent documents into a unified format to achieve multiple unified patent documents. When the programming instructions are executed, the one or more hardware processors generate a chemical patent corpus from the multiple unified patent documents. The chemical patent corpus includes one or more chemical entities extracted from the multiple unified patent documents. Each of the one or more chemical entities includes one or more relevance annotations, which indicate the relevance to the patent document from which the chemical entity is extracted. When the programming instructions are executed, the one or more hardware processors provide the chemical patent corpus to the chemical entity recognition system. The chemical entity identification system marks the one or more chemical entities in corresponding normalized patent documents of the multiple unified patent documents, extracts one or more additional chemical entities from the multiple unified patent documents, assigns a confidence score to each of the one or more additional chemical entities, and marks each of the one or more additional chemical entities as being related or unrelated to the relevant patent document based on information contained in the chemical patent corpus.
[0007] Another aspect of the present disclosure relates to a non-transitory storage medium, which contains executable instructions for causing a processing device to perform the following operations: obtain multiple patent documents from one or more patent databases; normalize each of the multiple patent documents into a unified format to achieve multiple unified patent documents, and generate a chemical patent corpus from the multiple unified patent documents. The chemical patent corpus includes one or more chemical entities extracted from the multiple unified patent documents. Each of the one or more chemical entities includes one or more relevance annotations. The one or more relevance annotations indicate the relevance to the patent document from which the chemical entity was extracted. The executable instructions also cause the processing device to provide the chemical patent corpus to the chemical entity recognition system. The chemical entity recognition system marks the one or more chemical entities in the corresponding normalized patent documents of the multiple unified patent documents, extracts one or more additional chemical entities from the multiple unified patent documents, assigns a confidence score to each of the one or more additional chemical entities, and identifies each of the one or more additional chemical entities as being related to or unrelated to the relevant patent document based on the information contained in the chemical patent corpus.
[0008] These and other features, characteristics, and methods of operation and function of the related elements of the present technology, as well as the incorporation of components and the economy of manufacture, will become apparent from consideration of the following description and the appended claims in conjunction with the accompanying drawings, all of which form a part of this specification, wherein like reference numerals indicate corresponding components in the various figures. However, it should be expressly understood that the drawings are for illustration and description purposes only and are not intended as limitations on the scope of the invention. As used in the specification and claims, the singular forms of "a", "an", and "the" include plural referents unless the context clearly indicates otherwise.
[0009] BRIEF DESCRIPTION OF THE DRAWINGS
[0010] Figure 1 schematically depicts a schematic network of a system for training a chemical entity recognition system to automatically extract compounds from patent documents and determine the relevance of the compounds to the patent documents according to one or more embodiments shown and described herein;
[0011] Figure 2A depicts a block diagram of schematic internal components of a training apparatus configured to train a chemical entity recognition system to automatically extract compounds from patent documents and determine the relevance of the compounds to the patent documents according to one or more embodiments shown and described herein;
[0012] Figure 2Bdepicts a block diagram of schematic logic modules within a memory of a training device configured to train a chemical entity recognition system to automatically extract compounds from patent documents and determine the relevance of the compounds to the patent documents according to one or more embodiments shown and described herein;
[0013] Figure 2C A block diagram depicting schematic data components within a storage device of a training apparatus configured to train a chemical entity recognition system to automatically extract compounds from patent documents and determine the relevance of the compounds to the patent documents according to one or more embodiments shown and described herein;
[0014] Figure 3A depicts a block diagram of schematic internal components of a chemical entity recognition system trained to automatically extract compounds from patent documents and determine the relevance of the compounds to the patent documents according to one or more embodiments shown and described herein;
[0015] Figure 3B depicts a block diagram of schematic logic modules within a memory of a chemical entity recognition system trained to automatically extract compounds from patent documents and determine the relevance of the compounds to the patent documents according to one or more embodiments shown and described herein;
[0016] Figure 3C depicts a block diagram of schematic data components within a storage device of a chemical entity recognition system trained to automatically extract compounds from patent documents and determine the relevance of the compounds to the patent documents according to one or more embodiments shown and described herein;
[0017] Figure 4 A flow chart depicting an illustrative general method of training a chemical entity recognition system to automatically extract compounds from patent documents and determine the relevance of the compounds to the patent documents according to one or more embodiments shown and described herein;
[0018] Figure 5 depicts a flow chart of an illustrative method of classifying dependencies according to one or more embodiments shown and described herein;
[0019] Figure 6 depicts a flow chart of an illustrative method for developing a patent corpus according to one or more embodiments shown and described herein;
[0020] Figure 7depicts an illustrative user interface for describing annotations in a patent document using an annotation tool according to one or more embodiments shown and described herein;
[0021] Figure 8 A graph depicting the performance of a chemical entity recognition system based on precision, recall, and F-score according to one or more embodiments shown and described herein; and
[0022] Fig. 9 Depicted is a graph showing the performance of a relevance classification system as a function of a relevance score threshold when relevance features are removed according to one or more embodiments shown and described herein. DETAILED DESCRIPTION
[0023] The present disclosure generally relates to a system that automatically extracts compounds from patent documents and determines the relevance of the compounds to the patent documents. The method described herein relates to a training device, in particular, the training device is configured to extract patent documents from a database, normalize the patent documents and feed the patent documents to a machine learning system (referred to as a chemical entity recognition system in the text), so that the machine learning system can be trained to automatically identify compounds in the normalized patent documents and determine whether these compounds are related to related patent documents.
[0024] Patent data contained in patent documents can be obtained from various patent databases, including but not limited to databases maintained by various patent offices, such as the European Patent Office (EPO), the United States Patent and Trademark Office (USPTO), the World Intellectual Property Organization (WIPO), the Japan Patent Office (JPO), the State Intellectual Property Office of China (SIPO), and the African Regional Intellectual Property Organization (ARIPO). In some embodiments, the patent database can be maintained by a non-governmental entity such as Google. The information contained in the database maintained by the non-governmental entity can be a copy of the information contained in the database of each patent office. Therefore, the term "patent database" used herein generally refers to any database containing patent documents or patent data, including (but not limited to) the database described above in this article.
[0025] Depending on the patent office, the available data may be in one or more formats, including but not limited to XML, HTML, PDF text, optical character recognition (OCR) PDF, image PDF, etc. Patent documents may follow a system structure consisting of the invention title, bibliographic information (e.g., patent number, date, inventor name, assignee, applicant, and International Patent Classification (IPC) category), abstract, specification, and claims. In some embodiments, the chemical data contained in the patent document may be obtained in the experimental section of the specification, and the claimed (i.e., patented) compounds may be obtained in the claims section. The specification drawings, sequence listings, or other information containing chemical data may be obtained at the end of the patent document (e.g., after the claims and abstract).
[0026] Patent offices provide patent documents but do not provide systematic continuous chemical annotation and full-text search capabilities, so manual or automatic extraction methods can be considered. Manual extraction methods are expensive and time-consuming, so they may be limited to commercial content providers such as Elsevier Reaxys (Elsevier BV, Amsterdam NL). Automatic methods for extracting information from patents can extract images and attached files, but the extracted information can only be obtained through text mining and image mining and may only be applicable to certain patent documents published after a certain date (for example, information on digital chemical structures of documents provided by the USPTO for its patent subset (patents granted from 2001 to 2011)). However, it has proven difficult to maintain public databases, so many automatic methods are outdated. In addition, these automatic methods have limitations in interpreting individual image features (such as chemical bonds) found in the structure diagrams of some images. In addition, automatic methods using text mining focus on the identification of compounds in patents, which is limited to the compounds included in the dictionary. It is impossible to add all systematic compound identifiers to the dictionary because these are generated algorithmically based on the structure of the compound and a set of rules. Furthermore, in the field of chemistry, the correctness of chemical structures associated with recognized compounds is of paramount importance. Typically, a combination of the above methods is used in the form of an integrated system for compound identification, which requires a gold standard corpus for training, development, and testing performance. Creating such a corpus is laborious and expensive. It involves developing well-defined annotation criteria, selecting and training domain experts for annotation, selecting data, annotating the data by multiple annotators, and coordinating the annotations.
[0027] Automatically extracting information from patents is fast, but has limitations. Most patent text mining systems have been developed, trained, and tested using invention titles and abstracts of patent documents. Therefore, their use is not evaluated on full-text documents. More importantly, automatic extraction mainly consists in extracting all mentioned compounds. In manually extracted databases, the focus is on relevant compounds. When a compound plays a major role in a patent application (for example, a starting material or product in a reaction specified in the claim section), the compound is relevant to the patent. Related compounds account for only a small proportion of all compounds mentioned in patent documents. Automatic identification of related compounds will greatly reduce the amount of data extracted from patents and can improve the effectiveness of patent resources. In addition, these compounds can be used for predictive analysis to identify key compounds in patents (key compounds are the main compounds protected by patent applications and are usually well hidden in the context).
[0028] Thus, the systems, methods, and media of the present application use an automated method to identify related compounds in patent documents that determines whether a chemical entity is related to a patent document containing the chemical entity, thereby minimizing a database that is maintained to categorize the growing number of available patent documents, which makes the database searched more efficient, retrieves more relevant results, and is less expensive to maintain. Other advantages may also be achieved.
[0029] As used herein, the term "patent document" generally refers to any publication related to patents, including but not limited to published patents (including utility patents, design patents, and plant patents), published patent applications, published utility models, published innovation patents, published utility certificates, published petty patents, published short-term patents, published utility innovations, published functional designs, published utility certificates, etc. In some embodiments, the patent document may be a chemistry-related patent document containing chemical information. That is, the chemistry-related patent document may include but is not limited to one or more chemical symbols, one or more functional groups, one or more chemical class identifiers, one or more chemical formula identifiers, one or more chemical structural formula identifiers, one or more chemical prefix identifiers, one or more chemical suffix identifiers, one or more chemical property identifiers, any chemical nomenclature and / or terminology promulgated by the International Union of Pure and Applied Chemistry (IUPAC), etc.
[0030] Now referring to the accompanying drawings, Figure 1Schematically depicts a schematic network of a system for training a chemical entity recognition system to automatically extract compounds from patent documents and determine the relevance of the compounds to the patent documents according to various embodiments. Figure 1 As shown, the computer network 100 may include a wide area network (WAN), such as the Internet, a local area network (LAN), a mobile communication network, a public service telephone network (PSTN), a personal area network (PAN), a metropolitan area network (MAN), a virtual private network (VPN), and / or other networks. The computer network 100 may generally be configured to be electrically connected to one or more computing devices and / or components thereof. The exemplary computing device may include, but is not limited to, a training device 110, a chemical entity recognition system 120, one or more data repositories 130, and / or a user computing device 140.
[0031] The training device 110 may generally be configured to train the chemical entity recognition system 120, and may further be configured to send and / or receive electronic data from one or more sources (e.g., the chemical entity recognition system 120; one or more data repositories 130; and / or a user computing device 140), etc., directly operate one or more other devices (e.g., the chemical entity recognition system 120; one or more data repositories 130; and / or a user computing device 140), collect data from one or more sources (e.g., patent document data, in particular chemical patent document data from one or more data repositories 130, etc.), store data related to chemical entities in patent documents, related patent documents, data related to the correlation of chemical entities in patent documents, etc. Additional details about the training device 110 are described herein. In some embodiments, the training device 110 may be capable of communicating with one or more other devices according to a client / server architecture and / or other architectures.
[0032] The chemical entity recognition system 120 is typically a machine learning (ML) server that is specifically configured to receive data related to chemical patent documents, analyze the data and extract chemical entities therefrom, and determine whether the extracted chemical entities are related to the chemical patent documents from which the chemical entities were extracted. The chemical entity recognition system 120 may continuously receive data and / or instructions from one or more other devices of the computer network 100, including but not limited to the training device 110, one or more data repositories 130, and / or user computing devices 140. Additional details regarding the chemical entity recognition system 120 are described herein.
[0033] As described herein, one or more data repositories 130 can generally store data for extracting chemical entities and determining their relevance. That is, one or more data repositories 130 can include patent documents, particularly chemical patent documents. In some embodiments, the data contained in one or more data repositories 130 can be a third-party server, which contains information that can be used to provide a dynamic ranking recommendation list, which can be obtained by training device 110 via application programming interface (API), chemical entity recognition system 120 and / or user computing device 140, etc. For example, one or more data repositories 130 may include one or more repositories maintained by patent offices such as USPTO, EPO, SIPO, JPO, WIPO and ARIPO. In some embodiments, in order to perform the process described herein, data can be directly obtained from one or more data repositories 130 automatically and continuously. In other embodiments, in order to perform the process described herein, data can be copied from one or more data repositories 130 to training device 110 and / or chemical entity recognition system 120.
[0034] Regardless of whether or not specifically described herein, the user computing device 140 can generally be used as an interface between a user and other components connected to the computer network 100, and / or various other components communicatively coupled to the user computing device 140 (e.g., components of the user computing device 140 communicatively coupled to the user computing device 140 via one or more networks), respectively. Thus, the user computing device 140 can be used to perform one or more user-oriented functions, such as receiving one or more inputs from a user or providing information to a user. For example, the user computing device 140 can receive user input corresponding to researching patent documents (including chemical patent documents), researching chemical information, researching chemical entities, providing information, performing various searches, etc. In addition, in the event that the training device 110 and / or the chemical entity recognition system 120 require supervision, updates, or corrections, the user computing device 140 can be configured to provide the required supervision, updates, or corrections. The user computing device 140 can also be used to input other data into the data storage portion of the training device 110, the chemical entity recognition system 120, and / or one or more data repositories 130. For example, a user can use the user computing device 140 to upload a patent publication to one or more components connected via the computer network 100. In some embodiments, the user computing device 140 may be configured to communicate with other platforms via a server and / or according to a peer-to-peer architecture and / or other architectures.
[0035] It should be understood that although the user computing device 140 is described as a personal computer and the training device 110, and the chemical entity recognition system 120 and the one or more data repositories 130 are described as servers, these are non-limiting examples. More specifically, in some embodiments, any type of computing device (e.g., a mobile computing device, a personal computer, a server, etc.) or any dedicated device with a computing component can be used for any of these components. In addition, although Figure 1 Each device shown in the figure is a separate piece of hardware, which is also merely an example. More specifically, each of the training device 110, the chemical entity recognition system 120, the one or more data repositories 130, and the user computing device 140 can represent multiple computers, servers, databases, mobile devices, components, special-purpose devices, etc. Similarly, the one or more data repositories 130 can be a single computer, server, database, mobile device, component, special-purpose device, etc.
[0036] Figure 2A Schematic hardware components of the training device 110 are shown. Bus 200 can interconnect various components, including (but not limited to) processing device 210, user interface hardware 220, communication interface hardware 230, memory 240 and / or storage device 260. Processing device 210, such as a computer processing unit (CPU), can be the central processing unit of training device 110 and perform calculations and logical operations required to execute programs. Processing device 210, alone or in combination Figure 2A In combination with one or more other elements disclosed herein, is an exemplary processing device, computing device, processor, or combination thereof, as these terms are used in this disclosure. Memory 240, such as read-only memory (ROM) and random access memory (RAM) may constitute exemplary storage devices (i.e., non-transitory processor-readable storage media). Such memory 240 may include one or more programming instructions thereon, which, when executed by processing device 210, cause processing device 210 to perform various processes, such as the processes described herein. In some embodiments, the program instructions may be stored on a tangible computer-readable medium, such as an optical disk, digital disk, flash memory, memory card, USB drive, optical storage media, such as Blu-ray discs, etc. TM Optical disks and / or other non-transitory processor-readable storage media.
[0037] In some embodiments, the program instructions contained on the memory 240 may be embodied as a plurality of software logic modules, wherein each logic module provides programming instructions for performing one or more tasks. For example, certain software logic modules may be used for the following purposes: collecting information (e.g., information contained in patent documents, particularly information in chemical patent documents); extracting information (e.g., chemical entities in chemical patent documents); providing information (e.g., to the chemical entity identification system 120 ( Figure 1 ) transmission information), etc. Other details about the logic module will be referred to in this article Figure 2B Have a discussion.
[0038] Still refer to Figure 2A The storage device 260 may be a storage medium that is usually separate from the memory 240 and may include one or more data repositories for storing data related to patent documents, especially chemical patent documents, data related to chemical entities, data on whether chemical entities are related to related patent documents, and data for transmission to the chemical entity identification system 120 ( Figure 1 ) to train the chemical entity recognition system 120, data related to annotations, etc. Figure 2A , the storage device 260 may be any physical storage medium, including but not limited to a hard disk drive (HDD), memory, removable storage, etc. Although the storage device 260 is depicted as a local device, it should be understood that the storage device 260 may be a remote storage device, such as a server computing device, one or more data repositories 130 ( Figure 1 ) etc. For further details on the type of data stored in the storage device 260, please refer to Figure 2C Give a description.
[0039] Still refer to Figure 2A , the user interface hardware 220 may allow information from the bus 200 to be provided to a user, whether the user is a local user of the training device 110 or a remote user of the training device 110 (e.g., the user computing device 140 ( Figure 1 )). Still refer to Figure 2A The user interface hardware 220 may be combined with a display and / or one or more input devices so that information is displayed on the display and / or input is received in an audio, video, graphic, or alphanumeric format. Exemplary input devices include, but are not limited to, a keyboard, a mouse, a joystick, a touch screen, a remote control, a pointing device, a video input device, an audio input device, a tactile feedback device, and the like.
[0040] Reference Figure 1 and Figure 2A, the communication interface hardware 230 can generally enable the training device 110 to have the ability to interface with one or more components of the computer network 100. For example, the training device can communicate with components of the computer network 100 via the communication interface hardware 230. The communication interface hardware 230 includes, but is not limited to, the chemical entity recognition system 120, one or more data repositories 130, and / or the user computing device 140. Communication with external devices can be carried out using various communication ports (not shown). The exemplary communication port can be attached to a communication network, such as the Internet, an intranet, a local area network, a direct connection, etc.
[0041] It should be understood that Figure 2A The components shown in the drawings are for illustration only and are not intended to limit the scope of the present disclosure. Figure 2A The components in FIG. 1 are shown as existing within training device 110, but this is a non-limiting example. In some embodiments, one or more components may exist outside of training device 110, or in Figure 1 The components described herein may be included in one or more components, other components, or as independent components. Similarly, one or more components may be included in other computing devices not specifically described herein. Figure 2A The components in are particularly related to the training device 110, which is also a non-limiting example. That is, similar components can be located in other components without departing from the scope of the present disclosure.
[0042] Now refer to Figure 2B , showing a method that can be included in the training device 110 ( Figure 2A ) is an exemplary logic module in the memory 240. Still referring to Figure 2B The logic module may include, but is not limited to, patent document acquisition logic 242 , patent document normalization logic 244 , patent corpus generation logic 246 , patent document provision logic 248 , scoring logic 250 and / or communication logic 252 .
[0043] Patent document acquisition logic 242 typically includes programming instructions for acquiring patent documents. That is, patent document acquisition logic 242 may include a program for causing processing device 210 ( Figure 2A ) accesses one or more data storage components (e.g., storage device 260 ( Figure 2A ), one or more data repositories 130 ( Figure 1) etc.), and obtain patent documents, particularly chemical patent documents, therefrom. Thus, patent document acquisition logic 242 may include programming instructions that allow a connection to be established between devices, protocols for requesting data storage containing data, instructions that cause data to be copied, moved, or read, and / or similar instructions. Thus, as a result of operating in accordance with patent document acquisition logic 242, data and information related to patent documents, particularly chemical patent documents, may be used to perform various other processes, which will be described in greater detail herein.
[0044] The patent document normalization logic 244 typically includes programming instructions for normalizing patent documents that have been obtained from multiple sources. That is, the patent document normalization logic 244 includes such programming instructions that information from patent documents, particularly chemical patent documents, is written in a unified format for later access, thereby generating multiple unified patent documents. Generally, this unified format is understood to be a format common to all patent documents, similar to the unidiff format commonly used in computational data comparison. Therefore, the multiple unified patent documents refer to multiple patent documents that have been modified to conform to a unified format. As a non-limiting example, normalizing each patent document may include: using one or more predefined XML tags corresponding to the heuristic information in the multiple patent documents, converting the multiple patent documents into a unified XML representation format. It should be understood that predefined XML tags typically refer to custom tags that define specific parts of patent documents, and are called differently in different countries and even in different patents in the same database, so that any object or part marked by the custom tag will be read according to the custom tag. For example, a specific text body is called "detailed description" in one patent document, "detailed disclosure of implementation scheme" in another patent document, and "disclosure" in a third patent document. As described herein, predefined XML tags may be set so that these three of these text bodies are identified as the same thing when read later. As used herein, the term "heuristic information" refers to a statistical value associated with a particular portion of a patent document that indicates the relative suitability of the portion among similar portions based on intuition, prior experience, common sense, etc., and may be developed based on machine learning, for example.
[0045] Patent corpus generation logic 246 generally includes programming instructions for generating a corpus from normalized documents generated in accordance with the operation of patent document normalization logic 244. That is, the generated normalized documents are collected into a corpus in accordance with patent corpus generation logic 246. In some embodiments, the corpus is further stored in a data repository in accordance with programming instructions provided by patent corpus generation logic 246. In still other embodiments, the data may be stored separately from the data contained in the patent document and / or the data contained in the normalized document.
[0046] In some embodiments, the patent corpus generation logic 246 may also include programming instructions for generating a chemical patent corpus from a plurality of unified / normalized patent documents. A chemical patent corpus is typically a corpus of unified / normalized documents (or data extracted from documents that have been unified / normalized) containing one or more chemical entities. In some embodiments, all unified / normalized documents may contain chemical entities, so all documents may be included in the chemical patent corpus. Generating a chemical patent corpus may include, for example, identifying compounds within the text contained in each of the plurality of normalized / unified patent documents. Generating a chemical patent corpus may also include accessing a physical property database and obtaining one or more physical properties of the identified compounds. It should be understood that a physical property database is typically a database containing data matching a specific compound with a specific physical property. For example, compound H 2 O can be included in the physical property database together with corresponding data related to the physical properties of water. Generating a chemical patent corpus can also include generating chemical structures corresponding to the compounds based on one or more physical properties. Identifying the compounds can include identifying the compounds using a dictionary-based method and / or a morphology-based method.
[0047] The morphology-based method may include identifying one or more elements in the compound, and if the compound is verified based on the structural chemistry of the compound, combining the one or more elements to produce the compound. As a non-limiting example, generating a chemical patent corpus from the plurality of normalized / unified patent documents includes annotating each patent document in the plurality of unified patent documents with one or more of a compound, a class of compounds, a suffix of a compound, and a prefix of a compound.
[0048] It should be understood that a compound is a chemical substance composed of chemical elements held together by chemical bonds, including molecules (or molecular entities) held together by chemical bonds. A compound can be a molecule held together by covalent bonds, an ionic compound held together by ionic bonds, an intermetallic compound held together by metallic bonds, or a complex held together by coordinated covalent bonds. A compound can be represented by a chemical formula. As a non-limiting example, the compound is selected from a single component compound, a portion of a compound mixture, or a predicted compound. A single component compound can include a pure compound, such as a system identifier, a common name, an element, and a chemical formula. A portion of a compound mixture can be a portion of a compound with a specific percentage of components (e.g., "Magnesiaflux", scientifically 30% MgF 2and 70% MgO). The predicted compound is a specific compound that is not characterized in the text of the patent document and is only mentioned in the claim part of the patent document or the description part of the patent document in the case of protecting intellectual property rights.
[0049] Compound classes can generally be any grouping of compounds based on specific criteria. For example, compounds can be classified based on the elements present in the compound (e.g., the oxide class can include any compound with one or more oxygen atoms; the hydride class can include any compound with one or more hydrogen atoms; the halide class can include any compound with one or more halogen atoms; the organic compound class can include any compound with a backbone of carbon atoms). In another embodiment, compounds can be classified based on the type of bonds they contain (e.g., the ionic compound class includes compounds formed by attractive forces between ions of opposite charge (e.g., salts); the molecular compound class includes compounds formed by covalent bonds). In yet another embodiment, compounds can be classified based on specific compounds (e.g., the acidic compound class includes compounds that produce hydrogen ions (protons or H) when dissolved in water). + ions); basic compounds include compounds that accept hydrogen ions when formed). The suffix of a compound refers to the end of the name of the compound. As a non-limiting example, the compound class can be selected from a chemical class, a biomolecule, a polymer, a mixture class, a mixture part class or a Markush class. It should be understood that biomolecules are generally molecules and ions present in an organism, such as but not limited to proteins, carbohydrates, lipids, nucleic acids, metabolites, etc. It should also be understood that polymers are generally substances with such a molecular structure that is mainly or entirely composed of a large number of similar units combined together, for example, synthetic organic materials used as plastics and resins. It should also be understood that the mixture class is a general class of mixtures of materials, such as solutions, suspensions, colloids, etc. Similarly, the mixture part class refers to the part class that constitutes the mixture (e.g., a compound that constitutes a part of the mixture). Markush class generally refers to a class of compounds accepted with the same Markush group, such as compounds with a single structural similarity, use, etc.
[0050] In some embodiments, the patent corpus generation logic may include programming instructions for grouping one or more chemical entities extracted from multiple normalized / unified patent documents into a specific corpus. It should be understood that the term "chemical entity" generally refers to a physical entity of interest in chemistry, including but not limited to molecular entities, parts of molecular entities, and chemical substances. Each of one or more chemical entities may include one or more relevance annotations. As described in more detail herein, the relevance annotation is a generated annotation of whether a specific chemical entity is related to the patent document from which the specific chemical entity is extracted. The one or more relevance annotations may include related compounds indicated as predicted compounds or Markush classes. As a non-limiting example, the one or more relevance annotations may include unrelated compounds indicated as compound mixture parts, mixture part classes, mixture classes, polymers, or biomolecules. The one or more relevance annotations of a single component compound or chemical class may be assigned based on the context of the corresponding unified patent document. The one or more relevance annotations may indicate the relevance of the patent document from which the chemical entity is extracted.
[0051] Reference Figures 1 to 2B , the patent corpus providing logic 248 generally includes programming instructions for providing a patent corpus to another device in the computer network 100. For example, the patent corpus providing logic 248 may include programming instructions that allow data related to the patent corpus to be transmitted to the chemical entity identification system 120, one or more data stores 130, and / or the user computing device 140.
[0052] Scoring logic 250 generally includes programming instructions for scoring each chemical entity contained in the patent corpus. As described in detail herein, scoring logic 250 includes programming instructions for assigning a relevance score, a confidence score, etc. to each chemical entity within the patent corpus in response to the score received from chemical entity recognition system 120.
[0053] The communication logic 252 typically includes programming instructions for communicating with one or more devices in a computer network. For example, the communication logic 252 may include a communication protocol for establishing a communication connection with the chemical entity recognition system 120, one or more data repositories 130, and / or a user computing device 140 so that data and / or signals can be transmitted between them.
[0054] Figure 2B The described logic modules are only exemplary. Therefore, it should be understood that more or fewer logic modules may be included in the memory 240 without departing from the scope of the present disclosure. In addition, in some embodiments, some logic modules may be combined into a single logic module and / or some logic modules may be divided into separate logic modules.
[0055] Reference Figure 2C , illustrates exemplary types of data that may be contained within storage device 260. The types of data may include, but are not limited to, patent document data 262, unified patent document data 264, patent corpus data 266, chemical entity data 268, and / or annotation data 270.
[0056] Patent document data 262 is generally data related to patent documents, particularly chemical patent documents. In some embodiments, the data contained in patent document data 262 may include full text received from one or more patent databases (eg, the patent databases described herein).
[0057] Unified patent document data 264 is generally data related to a unified patent document that has been standardized as described herein. As described in more detail herein, in some embodiments, the data contained in unified patent document data 264 may include full text with annotations, associated XML files, etc. providing standardized information.
[0058] As described herein, patent corpus data 266 is typically data generated as a result of creating a patent corpus. In some embodiments, patent corpus data 266 may be chemical patent corpus data.
[0059] Chemical entity data 268 may include data related to one or more chemical entities extracted from a plurality of unified patent documents. That is, chemical entity data 268 may identify each chemical entity within each patent document in the patent corpus, and may provide related structures, related related names, related categories, etc.
[0060] Annotation data 270 generally includes data related to annotations made to various chemical entities and / or patent documents in the patent corpus. For example, in some embodiments, each of the chemical entities may include one or more relevance annotations indicating relevance to the patent document from which the chemical entity was extracted.
[0061] Figure 3A Schematic hardware components of the chemical entity recognition system 120 are shown. Bus 300 can interconnect various components, including (but not limited to) processing device 310, user interface hardware 320, communication interface hardware 330, memory 340 and / or storage device 360. Processing device 310, such as a computer processing unit (CPU), can be the central processing unit of the chemical entity recognition system 120, performing calculations and logical operations required to execute programs. Processing device 310, alone or in combination Figure 3AIn combination with one or more other elements disclosed in, is an exemplary processing device, computing device, processor, or combination thereof, as the terms are used in this disclosure. Memory 340, such as read-only memory (ROM) and random access memory (RAM), may constitute an exemplary storage device (i.e., a non-transitory processor-readable storage medium). Such memory 340 may include one or more programming instructions thereon, which, when executed by the processing device 310, causes the processing device 310 to perform various processes, such as the processes described herein. In some embodiments, the program instructions may be stored on a tangible computer-readable medium, such as an optical disk, a digital disk, a flash memory, a memory card, a USB drive, an optical storage medium such as a Blu-ray disc, or a computer program. TM Optical disks and / or other non-transitory processor-readable storage media.
[0062] In some embodiments, the program instructions contained on the memory 340 may be embodied as a plurality of software logic modules, wherein each logic module provides programming instructions for performing one or more tasks. For example, certain software logic modules may be used for the following purposes: collecting information (e.g., information contained in patent documents, particularly information contained in chemical patent documents); extracting information (e.g., chemical entities in chemical patent documents); providing information (e.g., transmitting information to the training device 110 ( Figure 1 )); understand the meaning of specific types of information, etc. For other details about the logic module, refer to Figure 3B Have a discussion.
[0063] Still refer to Figure 3A The storage device 360 may be a storage medium separate from the memory 340, and may include one or more data repositories for storing data related to patent documents, especially chemical patent documents, data related to chemical entities, and data related to whether chemical entities are related to related patent documents, and transmitting the data to the training device 110 ( Figure 1 ), data related to annotations, data related to confidence scores, etc. Figure 3A , the storage device 360 may be any physical storage medium, including but not limited to a hard disk drive (HDD), memory, removable storage medium, etc. Although the storage device 360 is described as a local device, it should be understood that the storage device 360 may be a remote storage device, such as a server computing device, one or more data repositories 130 ( Figure 1 ) etc. For further details on the types of data stored in the storage device 360, please refer to Figure 3C Have a discussion.
[0064] Still refer to Figure 3A, the user interface hardware 320 may allow information from the bus 300 to be provided to a user, whether the user is a local user of the chemical entity identification system 120 or a remote user of the chemical entity identification system 120 (e.g., a user computing device 140 ( Figure 1 ) user). Still referring to 3A, user interface hardware 320 can be combined with a display and / or one or more input devices so that information is displayed on the display and / or input is received in audio, video, graphic or alphanumeric format. Exemplary input devices include but are not limited to keyboards, mice, joysticks, touch screens, remote controls, pointing devices, video input devices, audio input devices, tactile feedback devices, etc.
[0065] Reference Figure 1 and Figure 3A , the communication interface hardware 330 can generally provide the chemical entity recognition system 120 with the ability to interface with one or more components of the computer network 100. For example, the chemical entity recognition system 120 can communicate with components of the computer network 100 via the communication interface hardware 330, and the communication interface hardware 330 includes but is not limited to the training device 110, one or more data repositories 130, and / or the user computing device 140. Communication with external devices can be carried out using various communication ports (not shown). The exemplary communication port can be attached to a communication network, such as the Internet, an intranet, a local area network, a direct connection, etc.
[0066] It should be understood that Figure 3A The components shown are exemplary only and are not intended to limit the scope of the present disclosure. Figure 3A The components shown are located within the chemical entity recognition system 120, but this is a non-limiting example. In some embodiments, one or more components may be located outside of the chemical entity recognition system 120, or located in a Figure 1 The components shown may be located within one or more components, within other components, or as separate components. Similarly, one or more components may be located in other computing devices not specifically described herein. Figure 3A The components of are particularly related to the chemical entity recognition system 120, which is also a non-limiting example. That is, similar components can be located in other components without departing from the scope of the present disclosure.
[0067] Now refer to Figure 3B , showing a method that can be included in the chemical entity recognition system 120 ( Figure 3A ) is an exemplary logic module in the memory 340. Still referring to Figure 3B, the logic module can generally be a module of the machine learning logic 341 module. Exemplary logic modules include, but are not limited to, chemical entity extraction logic 342, chemical entity labeling logic 344, confidence assignment logic 346, labeling logic 348, and / or scoring logic 250.
[0068] The machine learning logic 341 may generally be a logic module in which one or more machine learning algorithms are incorporated. The machine learning logic 341 may be included in the machine learning logic 341 and may be used by the chemical entity recognition system 120 ( Figure 3A ) The machine learning algorithm used is not limited by the present disclosure, and can generally be any algorithm now known or developed later, particularly those specifically adapted to generate a predictive model that can be used to determine the relevance of a particular chemical entity to related chemical patent literature. That is, the machine learning algorithm can be a supervised learning algorithm, an unsupervised learning algorithm, a semi-supervised learning algorithm, and a reinforcement learning algorithm. Specific examples of machine learning algorithms can include, but are not limited to, nearest neighbor algorithms, naive Bayes algorithms, decision tree algorithms, linear regression algorithms, supervised vector machines, neural networks, clustering algorithms, association rule learning algorithms, Q learning algorithms, time difference algorithms, and deep adversarial networks. It should generally be understood that other specific examples of machine learning algorithms within the machine learning logic 341 are included within the scope of the present disclosure.
[0069] The predictive model generated as a result of the operation of the machine learning logic 341 is generally any machine learning model now known or later developed, particularly a model that provides result information that can be used to determine the relevance of a chemical entity to a related chemical patent document. Illustrative examples of machine learning models include, but are not limited to, convolutional neural network (CNN) models, long short-term memory (LSTM) models, neural network (NN) models, dynamic time warping (DTW) models, and the like.
[0070] The chemical entity extraction logic 342 included in the machine learning logic 341 generally includes programming instructions for extracting chemical entities from chemical patent documents. That is, the chemical entity extraction logic 342 may include programming instructions for receiving normalized / unified patent documents from a corpus of patent documents, analyzing the documents, and determining chemical entities contained in the documents, as will be described in more detail herein.
[0071] The chemical entity tagging logic 344 included in the machine learning logic 341 may generally include programmed instructions for tagging, annotating, or otherwise labeling the normalized / unified patent documents with data related to chemical entities extracted from the patent documents, as described in more detail herein.
[0072] Confidence score assignment logic 346 included in machine learning logic 341 generally includes programming instructions for assigning a confidence score to each of the one or more chemical entities. The confidence score generally represents a confidence level of whether the chemical entity is associated with a particular document based on various factors, as described in more detail herein.
[0073] The labeling logic 348 included in the machine learning logic 341 generally includes functions for labeling, marking, or otherwise indicating information in the patent document that may not have been learned by the training device 110 ( Figure 1 ) as described in more detail herein.
[0074] Still refer to Figure 3B , the relevance scoring logic 350 contained in the machine learning logic 341 generally includes programmed instructions for determining the relevance of a chemical entity to the document from which the chemical entity was extracted, as described in more detail herein.
[0075] Figure 3B The logic modules shown are only exemplary. Therefore, it should be understood that more or fewer logic modules may be included in the memory 340 without departing from the scope of the present disclosure. In addition, in some embodiments, some logic modules may be combined into a single logic module and / or some logic modules may be divided into separate logic modules.
[0076] Now refer to Figure 3C , illustrates exemplary types of data that may be contained within storage device 360. The types of data may include, but are not limited to, patent corpus data 362, chemical entity data 364, confidence score data 366, and / or relevance data 368.
[0077] As described herein, patent corpus data 362 is typically data generated as a result of creating a patent corpus. In some embodiments, patent corpus data 362 may be chemical patent corpus data.
[0078] Chemical entity data 364 may include data related to one or more chemical entities extracted from a plurality of unified patent documents, particularly as identified by chemical entity identification system 120 ( Figure 3A ) data related to other entities extracted. That is, chemical entity data 364 can be data identifying each chemical entity within each patent document in the patent corpus, and can provide associated structures, associated related names, associated categories, etc.
[0079] Confidence score data 366 generally includes information related to the chemical entity recognition system 120 ( Figure 3AThat is, confidence score data 366 includes data related to the confidence determined that a chemical entity is or is not associated with the patent document from which the chemical entity was extracted, as described in more detail herein.
[0080] The relevance data 368 generally includes data indicating the relevance of each chemical entity to the patent document from which the chemical entity was extracted. For example, the relevance data 368 can be in the form of a table or other similar data that lists each chemical entity extracted in a particular patent document and an associated relevance indicator, as described in more detail herein.
[0081] Figure 4 A block diagram of an illustrative method 400 for training a chemical entity recognition system to automatically extract compounds from patent documents and determine the relevance of the compounds to the patent documents according to one or more embodiments is shown. The operations of method 400 presented below are for illustrative purposes only. In some embodiments, method 400 may be performed without one or more additional operations described and / or without one or more operations discussed. In addition, in Figure 4 The order in which the operations of method 400 are illustrated and described below is not intended to be limiting.
[0082] In some embodiments, method 400 can be performed by one or more processing devices (e.g., digital processors, analog processors, digital circuits designed to process information, analog circuits designed to process information, state machines, and / or other mechanisms for electronically processing information), such as those described herein. Figure 2A The processing device 210 shown and described herein and / or the processing device 310 shown and described herein with respect to 3A. Still referring to Figure 4 , one or more processing devices may include one or more responses electronically stored in an electronic storage medium (e.g., Figure 2A-2B The memory 240 and / or Figure 3A-3B The one or more processing devices may include one or more devices configured by hardware, firmware, and / or software to be specifically designed to perform one or more operations of method 400.
[0083] See also Figures 1 to 4 At block 402, a plurality of patent documents may be obtained. In some embodiments, the plurality of patent documents may be obtained from one or more patent databases (e.g., one or more data repositories 130). According to one or more implementations, the operations of block 402 may be performed by one or more hardware processors configured by machine-readable instructions, the machine-readable instructions including logic the same or similar to patent document obtaining logic 242.
[0084] At block 404, each of the plurality of patent documents may be normalized into a unified format to achieve a plurality of unified patent documents. According to one or more implementations, the operations of block 404 may be performed by one or more hardware processors configured by machine-readable instructions that include logic that is the same or similar to patent document normalization logic 244.
[0085] At block 406, a one-to-one mapping between each character in the original text of each patent document and the corresponding character in the normalized patent document may be stored. According to one or more implementations, the operations of block 406 may be performed by one or more hardware processors configured by machine-readable instructions that include logic that is the same or similar to patent document normalization logic 244 and / or scoring logic 250.
[0086] At square block 408, a chemical patent corpus can be generated. In some embodiments, the chemical patent corpus can be generated from a plurality of unified patent documents. The chemical patent corpus can include one or more chemical entities extracted from a plurality of unified patent documents. Each chemical entity in the one or more chemical entities can include one or more correlation annotations. One or more correlation annotations can indicate the correlation with the patent document from which the chemical entity is extracted. According to one or more embodiments, the operation of square block 408 can be performed by one or more hardware processors configured by machine-readable instructions, and the machine-readable instructions include logic identical or similar to patent corpus generation logic 246.
[0087] At block 410, the chemical patent corpus may be provided to the chemical entity identification system 120. Thus, as described in more detail herein, the chemical entity identification system 120 may mark one or more chemical entities in a corresponding normalized patent document of a plurality of unified patent documents; extract one or more additional chemical entities from the plurality of unified patent documents; assign a confidence score to each of the one or more additional chemical entities; and mark each of the one or more additional chemical entities as being relevant or irrelevant to the relevant patent document based on information contained in the chemical patent corpus. According to one or more embodiments, the operations of block 410 may be performed by one or more hardware processors configured by machine-readable instructions, the machine-readable instructions including logic that is the same or similar to the patent corpus providing logic 248.
[0088] Now refer to Figure 5, a schematic method for classifying correlation is shown. In box 510, chemical patents are extracted by the patent office. In box 520, the patent source document is normalized into a unified format. Then it is supplied to a chemical entity recognition system 530, which is composed of two different named entity extraction systems, i.e., a chemical entity recognizer (CER) 532 (Elsevier, Frankfurt, Germany) and a mining program 534, such as OCMiner (OntoChem, Halle, Germany). CER 532 extracts chemical entities and marks them in the normalized input documents. OCMiner 534 further enriches the output of CER 532 by extracting additional chemical entities and assigning confidence scores to all extracted entities of the two systems. The related structures of the compounds extracted by CER 532 or OCMiner 534 are generated, verified and standardized using a naming service 536 (e.g., Reaxys naming service (Elsevier, BV, Amsterdam, the Netherlands)). Chemical annotations 542 in patent corpus 540 are used to train and test chemical entity recognition system 530. Relevance annotations 544 in the corpus are used to train and test relevance classifier 550, which labels chemical entities extracted by the chemical entity recognition system as relevant or irrelevant at block 560. Each component is described in more detail below.
[0089] Normalization
[0090] It may be necessary to normalize the various input sources and files into a unified text representation. The normalization step is performed by converting all input files (e.g., XML, HTML, and PDF) into a unified XML representation. Predefined XML tags corresponding to heuristic information such as document parts (invention title, abstract, claims, description, and metadata) are used in this unified representation. Normalization also converts all character encodings into a specific format, such as UTF-8 (8-bit Unicode Transformation Format).
[0091] During normalization, a one-to-one mapping between each character in the original text and the corresponding character in the normalized document is stored. This may provide the possibility to go back from the normalized text to the original document and vice versa. This may also minimize the effort to update the annotations in the patent corpus in case of changes in the normalization method (note that the documents in the corpus have also been normalized).
[0092] Patent corpus development
[0093] The development of a chemical patent corpus with annotations of chemical entities and relevance can be done in two stages. Figure 6An illustrative corpus creation process 600 is shown. The first stage 610 focuses on building a corpus with chemical entity annotations. The second stage 630 can include assigning relevance annotations to the entities annotated in the first stage 610 using the corpus obtained from the first stage 610. In the second stage 630, the annotator can also mark any misspelled compounds. For each stage, a set of clear guidelines can be developed to help achieve annotation consistency.
[0094] Chemical Entity Annotation Guidelines
[0095] Criteria can be developed based on patent corpora, such as the criteria mentioned in "Annotated Chemical Patent Corpus: The Gold Standard for Text Mining" compiled by Akhondi, SA, Klenner, AG, Tyrchan, C, et al. (2014), which is disclosed in PLoS One, 9, e107477 and incorporated herein by reference in its entirety, and chemical entity annotation criteria are developed according to boxes 610 and 612. The criteria define the entities to be annotated. For each entity, positive and negative examples are provided. In addition, any exceptions are defined and explained by examples. The criteria also define how to annotate within the brat quick annotation tool (available from http: / / brat.nlplab.org / ). The brat annotation tool allows text to be annotated online using predefined entity types. Annotators are required to annotate compounds (e.g., tetrahydrofuran), chemical classes (e.g., zirconium alkoxides), and suffixes or prefixes of these compounds (e.g., "stable" as a prefix of "stable zirconium oxide" and "nanoparticles" as a suffix of "silver nanoparticles").
[0096] Compounds can be divided into three categories: single-component compounds (pure compounds, such as system identifiers, common names, elements, and chemical formulas), compound mixtures (such as "Magnesiaflux", which is scientifically 30% MgF 2 and 70% MgO) or predicted compounds (specific compounds mentioned in the claims or specification only for the purpose of intellectual property protection and not described in the text).
[0097] Compound classes can be annotated into six categories: chemical class (natural product or substructure name, e.g. heterocycle); biomolecule (e.g. insulin); polymer (e.g. polyethylene); mixture class (e.g. opium); mixture part class (e.g. quinupristin) or Markush class (a textual description of the Markush formula, e.g. H a X b CCH).
[0098] Dependency Annotation Guidelines
[0099] For relevance annotation according to box 630, a new set of criteria were developed that define how relevant compounds are identified. When defining the criteria, the legal status of a compound (e.g., a predicted or claimed compound) and its characterization (e.g., NMR or MS measurements), properties (e.g., superconductivity), utility (e.g., toxicity), and transformations (e.g., reactions) are considered. The relevance annotation does not include suffixes and prefixes of the compound. In short, relevance is assigned as follows: Predicted compounds and Markush classes are related. Compound mixture parts, mixture part classes, mixture classes, polymers, and biomolecules are not related. Single component compounds and chemical classes are assigned relevance based on the context of the full patent text. Entities are considered relevant to a patent if (a) they are present in the invention title or abstract section of the patent, (b) the entity is part of the reaction environment (e.g., a product, intermediate, catalyst, or starting material used in a synthesis), or (c) an entity or its measured property belongs to the invention in the claim section and is related to the core invention of the patent document. Single component compounds and chemical classes are irrelevant if (a) the entity is introduced solely for the purpose of further explanation and is described outside the scope of the invention, (b) the entity is described as a reference or comparison, or (c) the entity is involved in a chemical reaction but is not a starting material, product, or catalyst.
Claims
1. A method for training and using a chemical entity recognition system to extract and identify one or more compounds from a patent document and determine the relevance of the one or more compounds to the patent document, the method comprising: The processing device obtains a plurality of patent documents from one or more patent databases; The processing device normalizes each of the plurality of patent documents into a unified format to achieve a plurality of unified patent documents; The processing device generates a chemical patent corpus from the plurality of unified patent documents, the chemical patent corpus includes one or more chemical entities extracted from the plurality of unified patent documents, each of the one or more chemical entities includes one or more relevance annotations, the one or more relevance annotations indicate relevance to the patent document from which the chemical entity is extracted, wherein the one or more relevance annotations indicate related compounds for each of the predicted compounds and Markush classes, the one or more relevance annotations are irrelevant compounds that are compound mixture parts, mixture part classes, mixture classes, polymers, and biomolecules, and when the chemical entity is present in the invention title or abstract part of the patent document, and the chemical entity is part of a reaction environment, and the chemical entity or a measured property of the chemical entity is in the claim part of the patent document, the one or more relevance annotations are single-component compounds and chemical classes that are each related to the patent; The processing device provides the chemical patent corpus to the chemical entity identification system, wherein the chemical entity identification system, in response to receiving the chemical patent corpus, marks the one or more chemical entities in corresponding normalized patent documents of the plurality of unified patent documents, extracts one or more additional chemical entities from the plurality of unified patent documents, The chemical entity identification system calculates a confidence score for each of the one or more additional chemical entities extracted from the associated patent document; The chemical entity identification system determines, based at least in part on the confidence score, a label for each of the one or more additional chemical entities, wherein each label indicates that the one or more additional chemical entities are related or unrelated to the associated patent document; and the chemical entity identification system assigning the confidence score and the label for each of the one or more additional chemical entities to the associated patent document, wherein the confidence score for each additional chemical entity extracted from the associated patent document is a linear combination of features of the compound representing the additional chemical entity, the features comprising frequency of the compound in the patent document, occurrence of the compound in a predefined portion of the patent document, length of terms, occurrence of the compound in special characters, occurrence of a single compound in a portion of the patent document, compounds free of solvents or laboratory chemicals, and presence of the compound in one or more predefined groups representing the frequency of the compound in a large number of chemical patent documents.
2. The method according to claim 1, wherein: Obtaining the plurality of patent documents from the one or more patent databases includes obtaining patent documents classified as chemistry-related patent documents.
3. The method according to claim 1, wherein: Normalizing each of the multiple patent documents includes converting the multiple patent documents into a unified XML representation format, utilizing one or more predefined XML tags corresponding to heuristic information in the multiple patent documents, and storing a one-to-one mapping between each character in the original text of each patent document and the corresponding character in the normalized patent document.
4. The method according to claim 1, wherein: Generating the chemical patent corpus includes: identifying a chemical compound in text contained in each of the plurality of unified patent documents; accessing a physical property database and obtaining one or more physical properties of the identified compound; and A chemical structure corresponding to the compound is generated based on the one or more physical properties.
5. The method according to claim 4, wherein: Identifying the compound includes: using one or more of a dictionary-based method and a morphology-based method to identify the compound, wherein the morphology-based method includes identifying one or more elements in the compound, and if the compound is verified based on the structural chemistry of the compound, the one or more elements are combined to produce the compound.
6. The method according to claim 1, wherein: Generating a chemical patent corpus from the plurality of unified patent documents includes annotating each patent document in the plurality of unified patent documents with one or more of a compound, a compound class, a compound suffix, and a compound prefix.
7. The method according to claim 6, wherein: The compound is selected from the single component compound, the compound mixture part or the predicted compound.
8. The method according to claim 6, wherein: The compound class is selected from the chemical class, the biomolecule class, the polymer class, the mixture class, the mixture portion class, or the Markush class.
9. A system configured for training a chemical entity recognition system to extract one or more compounds from a patent document and determine the relevance of the one or more compounds to the patent document, the system comprising: one or more hardware processors; and A non-transitory processor-readable storage medium having one or more programming instructions contained thereon, wherein the one or more programming instructions, when executed, cause the one or more hardware processors to perform the method of any one of claims 1-8.
10. A non-transitory storage medium comprising executable instructions for causing a processing device to perform the method according to any one of claims 1 to 8.
Citation Information
Patent Citations
Method for automatically extracting key phrases of patent documents
CN103885934A
Academic-literature semantic restructuring method based on image processing and sequence labeling
CN105260727A
Chemistry resource database
US20020049548A1
System and Method for Identifying Similar Molecules
US20080004810A1
Method and software for extracting chemical data
US20110202331A1
