Method for training natural language retrieval system, retrieval system and corresponding uses

Abstract

The present invention provides a method and system for training a machine learning-based patent search or novelty assessment system. The method includes providing a plurality of patent documents, each of which has a computer-readable claim block and a full specification block, wherein the full specification block includes at least a portion of the specification of the patent document. The method also includes providing a machine learning model and training the machine learning model using a training data set, wherein the training data set includes data from the patent documents used to form the trained machine learning model. According to the present invention, the training includes using claim block and full specification block pairs from the same patent document as training cases in the training data set.

Description

Technical Field

[0001] The present invention relates to natural language processing. In particular, the present invention relates to systems and methods based on machine learning, such as neural networks, for searching, comparing, or analyzing documents containing natural language. The documents may be technical or scientific documents. In particular, the documents may be patent documents. Background Art

[0002] The comparison of written technical concepts is required in many areas of business, industry, economics, and culture. A specific example is the examination of patent applications, where one objective is to determine whether a technical concept defined in a claim of a patent application semantically encompasses another technical concept defined in another document.

[0003] Currently, there are a growing number of retrieval tools available for finding individual documents, but the analysis and comparison of concepts exposed by documents is still largely a manual effort involving human inference of the meaning of words, sentences, and larger linguistic entities.

[0004] Scientific research around natural language processing has produced tools for automatically parsing language using computers. These tools can be used, for example, to tokenize text, perform part-of-speech tagging, identify entities, and identify relationships between words or entities.

[0005] Scientific work has also been conducted to automatically analyze patents by extracting key concepts from the documents, for example, for the purposes of text summarization and technology trend analysis.

[0006] Recently, word embedding, which uses multidimensional word vectors, has become an important tool for mapping the meaning of words into a form that can be processed by digital computers. This method can be used by neural networks, such as recurrent neural networks, to provide computers with a deeper understanding of the content of documents.

[0007] Traditionally, patent searches have been conducted using keyword searches, which involve defining the correct keywords, their synonyms, inflections, etc., and creating Boolean search strategies. This is time-consuming and requires specialized knowledge. More recently, semantic searches have also been developed, which are more fuzzy and may involve the use of artificial intelligence techniques. They help quickly find a large number of documents that are related in some way to a concept discussed in another document. However, they are relatively limited in applications such as patent novelty searches because, in practice, their ability to assess novelty—that is, to find documents that disclose specific content that falls under the general concept defined in the patent claims—is limited. These methods have proven powerful, for example, in machine translation applications.

[0008] In summary, there are available techniques that are well suited for general searching and extracting core concepts, for example, from texts and text summaries. However, they are not well suited for performing detailed comparisons between concepts disclosed in different documents in large amounts of data, which is crucial, for example, for patent novelty search purposes or other technology comparison purposes.

[0009] In particular, improved techniques for text analysis and comparison are needed to enable more efficient retrieval and novelty assessment tools. Summary of the Invention

[0010] An object of the present invention is to address at least some of the above problems and to provide a novel system and method for improving the accuracy of technology searches. A specific object is to provide a solution that helps an automated system better assess the novelty of concepts disclosed in a document relative to each other and to better account for the technical relationships between them.

[0011] A specific object is to provide an improved machine learning based retrieval system and a method of training such a system.

[0012] Specific objects include providing a patent search or novelty assessment system with improved accuracy, and providing new uses for publicly available patent data.

[0013] According to one aspect, the present invention provides a method for training a machine learning-based patent search or novelty assessment system, the method comprising providing a plurality of patent documents, each of the patent documents having a computer-recognizable claim block and a full specification block, the full specification block comprising at least a portion of the specification of the patent document. The method further comprises providing a machine learning model and training the machine learning model using a training data set, the training data set comprising data from the patent documents used to form the trained machine learning model. According to the present invention, the training comprises using claim block and full specification block pairs from a single, i.e., the same, patent document as training cases for the training data set.

[0014] The machine learning model is preferably capable of embedding the claim block and the full specification block into a vector. The above-mentioned training cases (i.e., training samples) are positive training samples, so the learning goal of the model can be to minimize the vector angle between the claim block and the full specification block. Other positive training samples can be claim blocks and full specification blocks that are not derived from the same document but are associated with each other via database references. Another learning goal can be to maximize the vector angle between the claim block and the full specification block that are derived from at least some different documents and are not associated with each other in this way, or to provide a non-zero vector angle, thereby forming a negative training sample.

[0015] According to one aspect, a machine learning-based patent search or novelty assessment system is provided, comprising: a machine learning training subsystem adapted to read patent claim blocks and full specification blocks of patent documents and utilize them as training data; and a machine learning search engine utilizing a trained machine learning model to search for subsets of patent documents within a larger collection of patent documents. In the present invention, the machine learning training subsystem is configured to use pairs of claim blocks and full specification blocks from the same patent document as training cases in the training data set.

[0016] In one aspect, a natural language retrieval system is provided, comprising a digital data storage device for storing a plurality of natural language chunks and a data graph corresponding to the chunks. A first data processing device is also provided for converting the chunks into the graph, the graph being stored in the storage device. The graph comprises a plurality of nodes, preferably consecutive nodes, each node comprising a natural language unit extracted from the chunk as a node value or a portion thereof. A second data processing device is also provided for executing a machine learning algorithm capable of traversing the graph and reading the node values ​​to form a trained machine learning model based on the node structure and node values ​​of the graph. A third data processing device is adapted to read a fresh graph or a fresh natural language chunk converted into a fresh graph, and utilize the machine learning model to determine a subset of the natural language chunks based on the fresh graph. The first and second data processing devices are part of the machine learning trainer subsystem described above. The third data processing device is a machine learning search engine as described above.

[0017] The graph may in particular be a tree-like recursive graph having partial-word relationships between node values ​​of consecutive nodes.

[0018] The method and system are preferably neural network based, whereby the machine learning model is a neural network model.

[0019] More specifically, the invention is characterized by what is stated in the independent claims.

[0020] The present invention provides significant benefits. While patent novelty search data and citation data provided by patent offices and patent applicants can be used to train neural networks, they have the disadvantage that the quality of the data varies. In particular, only some of the novelty barriers raised by patent offices are actually novelty barriers. Despite this, all citations cited by patent examiners end up in public records and patent databases, and without manual evaluation, it is impossible to discern from the public records and patent databases which citations are truly relevant. This reduces the reliability of publicly available training data. The present invention allows for at least one truly relevant training case for each claim, and in particular for one or more independent claims. As a result, neural networks can be trained more accurately to find relevant prior art documents.

[0021] The presently disclosed same document training cases may be the only positive (indicative of relevant prior art) training cases used, or novelty search data and / or citation data may also be used to form additional training cases.

[0022] The present method is also compatible with advanced training schemes such as data augmentation, as will be discussed in detail later. The combination of these methods provides particularly good training results.

[0023] All of this contributes to more targeted searches and more accurate automatic novelty assessment, requiring less manual work.

[0024] Tree-shaped graphs with partial word edges are particularly beneficial because they can be modified quickly and safely while preserving coherent technical and semantic logic within the graph.

[0025] The dependent claims relate to selected embodiments of the invention.

[0026] Next, selected embodiments of the present invention and their advantages are discussed in more detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1A A block diagram of an exemplary retrieval system is shown at a general level.

[0028] Figure 1B A block diagram showing a more detailed embodiment of the retrieval system comprising a series of neural network based search engines and their trainers is shown.

[0029] Figure 1C A block diagram of a patent search system according to one embodiment is shown.

[0030] Figure 2A Block diagram showing an exemplary nested graph with only part-word relationships / whole-word relationships.

[0031] Figure 2B A block diagram showing an exemplary nested graph with part-word / whole-word relationships and hyponym / hypernym relationships is shown.

[0032] Figure 3 A flowchart of an exemplary graph parsing algorithm is shown.

[0033] Figure 4A Block diagram showing patent search neural network training using patent search / citation data as training data.

[0034] Figure 4B A block diagram showing the training of a neural network using claim-description graphs from the same patent document as training data.

[0035] Figure 4CA block diagram showing neural network training using an augmented set of claim graphs as training data.

[0036] Figure 5 The functionality of an exemplary graphics feed user interface according to one embodiment is illustrated. DETAILED DESCRIPTION

[0037] definition

[0038] In this article, a "natural language unit" refers to a text chunk, or after embedding, a vector representation of a text chunk. The chunk can be a single word or a multi-word sub-concept that appears once or multiple times in the original text stored in a computer-readable form. The natural language unit can be represented as a set of character values ​​(commonly called a "string" in computer science), or digitally represented as a multidimensional vector value, or a reference to such a value.

[0039] A "natural language chunk" is a data instance containing a linguistically meaningful combination of natural language units, such as one or more complete or incomplete sentences in a language (such as English). The natural language chunk can be represented as a single string and stored in a file in a file system and / or displayed to a user via a user interface.

[0040] "Document" means a machine-readable entity containing natural language content and associated with a machine-readable document identifier that is unique with respect to other documents within the system.

[0041] "Patent document" refers to the natural language content of a patent application or granted patent. A patent document is associated in this system with a publication number and / or another machine-readable unique document identifier assigned by a recognized patent authority, such as the EPO, WIPO, or USPTO, or another national or regional patent office of another country or region. The term "claims" refers to the essential content of the claims of a patent document, particularly independent claims. The term "full specification" refers to the content of a patent document that covers at least a portion of the specification of the patent document. The full specification may also cover other parts of the patent document, such as the abstract or claims. Claims and full specification are examples of natural language blocks.

[0042] A "claim" is defined herein as a block of natural language that would be considered a claim by the European Patent Office on the date of entry into force of this patent application. In particular, a "claim" is a computer-readable block of a natural language document identified by a machine-readable integer number therein, for example, in a string format preceding the block and / or as (part of) the associated information in a markup file format (such as XML or HTML).

[0043] "Full specification" is defined herein as a computer-recognizable block of natural language that is computer-recognizable within a patent document that also includes at least one claim and includes at least one other portion of the document other than the claims. Alternatively, the "full specification" may be recognizable by relevant information in a markup document format (such as XML or HTML).

[0044] In this context, an "edge relation" may be a technical relation extracted from a block and / or a semantic relation derived from the semantics of the relevant natural language units. In particular, the edge relation may be

[0045] - Part-word relationship (also: part-word / whole-word relationship); Part-word: X is a part of Y; Whole-word: Y has X as its own part; for example: "wheel" is a part-word of "car",

[0046] -Hyponym relationships (also: hyponym / hypernym relationships); hyponym: X is a hyponym of Y; hypernym: X is a hypernym of Y; for example: "electric car" is a hyponym of "car", or

[0047] - Synonym relationship: X is the same as Y.

[0048] In some embodiments, edge relationships are defined between consecutively nested nodes of a recursive graph, with each node containing a natural language unit as a node value.

[0049] Other possible technical relationships include thematic relationships, which refer to the role played by a subconcept of the text relative to one or more other subconcepts in addition to the above relationships. At least some thematic relationships can be defined between consecutively nested units. In one example, the thematic relationship of a parent unit is defined in a child unit. An example of a thematic relationship is the role class "function". For example, the function of a "handle" can be "allow manipulation of an object". Such a thematic relationship can be stored as a child unit of the "handle" unit, and the "function" role is associated with the child unit. Thematic relationships can also be general relationships that do not have a predefined classification (or have a general classification such as "relationship"), but the user can freely define the relationship. For example, the general relationship between a handle and a cup can be "[handle] is attached to [cup] with adhesive". Such a thematic relationship can be stored as a child unit of the "handle" unit or the "cup" unit or both, preferably referencing each other.

[0050] If a relation unit is linked to computer executable code that, when executed by a data processor, produces a natural language block comprising a relation in a particular relational category or subcategory, the relation unit is considered to define a relation in that category or subcategory.

[0051] "Graph" or "data graph" refers to a data instance that follows a general non-linear recursive and / or network data pattern. The system can simultaneously contain several different graphs that follow the same data pattern and whose data comes from different sources and / or is related to different sources. In practice, the graph can be stored in any suitable text or binary format, which allows data items to be stored recursively and / or as a network. The graph is in particular a semantic graph and / or a technical graph (describing semantic and / or technical relationships between node values), rather than a syntactic graph (which only describes linguistic relationships between node values). The graph can be a tree graph. A forest-shaped graph comprising a plurality of trees is considered a tree graph in this document. In particular, the graph can be a technical tree graph.

[0052] "Data pattern" refers to the rules according to which data, particularly natural language units and data associated therewith, such as information about the technical relationships between the units, is organized.

[0053] "Nesting" of a natural language unit refers to the unit's ability to have one or more children and one or more parents, as determined by the data schema. In one example, a unit can have one or more children and only a single parent. A root unit has no parent, and a leaf unit has no children. Sibling units have the same parent. "Continuous nesting" refers to nesting between a parent unit and its immediate children.

[0054] A "recursive" nesting or data pattern is one that allows for the nesting of natural language units containing data items.

[0055] A "(natural language) token" is a word or phrase within a larger natural language block. A token may also contain metadata associated with the word or phrase, such as a part-of-speech (POS) tag or syntactic dependency annotations. A "set" of natural language tokens specifically refers to tokens that can be grouped based on their textual value, POS tags, dependency annotations, or any combination thereof, according to predetermined rules or fuzzy logic.

[0056] The terms "data storage means", "processing means" and "user interface means" primarily refer to software means, i.e., computer executable code (instructions) that can be stored on a non-transitory computer-readable medium and adapted to perform designated functions when executed by a processor, in other words, storing digital data, allowing a user to interact with said data, and processing said data. All of these components of the system can be hosted in a software that is run by a local computer or by a web server via a locally installed web browser, for example, supported by suitable hardware for running the software components. The methods described herein are computer-implemented methods.

[0057] Description of selected embodiments

[0058] The following describes a natural language search system comprising a digital data storage device for storing a plurality of natural language chunks and data graphs corresponding to the chunks. The storage device may include one or more local or cloud data stores. The storage device may be file-based or query language-based.

[0059] The first data processing device is a converter unit adapted to convert the block into the graph. Each graph comprises a plurality of nodes, each node containing a natural language unit extracted from the block as a node value. Edges are defined between pairs of nodes, defining a technical relationship between the nodes. For example, an edge or edges may define a partial-word relationship between two nodes.

[0060] In some embodiments, the number of at least some nodes in the graph that contain specific natural language unit values ​​is less than the number of occurrences of the specific natural language unit in the corresponding natural language block. In other words, the graph is a condensed representation of the original text, which can be implemented using, for example, the symbol recognition and matching methods described later. By allowing each node to have multiple subnodes, the basic technical content (and optionally semantic content) of the text can still be maintained in the graphical representation. It is also efficient to process the condensed graph through a graph-based neural network algorithm, whereby they can learn the basic content of the text better and faster than learning from direct text representations. This method has been shown to be particularly powerful in the comparison of technical texts and in particular in the automatic evaluation of the novelty of the claims based on the retrieval of the full patent specification and the claims.

[0061] In some embodiments, the number of all nodes containing a particular natural language unit is one. In other words, there are no duplicate nodes. While this may result in a simplification of the original content of the text, at least when using a tree graph, it results in a graph that is highly efficient to process and still relatively expressive, suitable for patent searching and novelty assessment.

[0062] In some embodiments, the graph is a condensed graph of at least the nouns and noun chunks found in the original text. In particular, the graph can be a condensed graph of noun-valued nodes arranged according to their part-word relationships. In an average patent document, many noun terms appear dozens or even hundreds of times throughout the text. Through this approach, the content of such documents can be compressed into a fraction of the original space, making them more accessible for machine learning.

[0063] In some embodiments, a plurality of terms that appear multiple times in at least one original natural language chunk appear exactly once in the corresponding graph.

[0064] A streamlined graph representation is also beneficial because synonyms and coreferences (expressions that mean the same thing in a particular context) can be taken into account when constructing the graph. This results in an even more streamlined graph. In some embodiments, multiple terms that appear in at least two different written forms in at least one original natural language block appear exactly once in the corresponding graph.

[0065] The second data processing device is a neural network trainer for executing a neural network algorithm that iteratively traverses the graph structure and learns both from the internal structure of the graph and from its node values, as defined by a loss function defining a learning objective and a training data case. The trainer typically receives a training data set for the graph or an augmented graph derived therefrom, as specified by the training algorithm. The trainer outputs a trained neural network model.

[0066] Supervised machine learning methods using graph-form data, such as those described herein, have been found to be extremely powerful for finding technically relevant documents among patent and scientific documents.

[0067] In some embodiments, the storage device is further configured to store reference data linking at least some of the blocks to each other. The reference data is used by the trainer to derive training data, i.e., to define combinations of graphs to be used as positive training cases or negative training cases (i.e., training samples) during training. The learning objectives of the trainer depend on this information.

[0068] The third data processing device is a search engine adapted to read a fresh graph or a fresh natural language block, typically via a user interface or a web interface. If necessary, the block is converted into a graph in the converter unit. The search engine uses the trained neural network model to determine a subset of natural language blocks (or graphs derived therefrom) based on the fresh graph.

[0069] Figure 1A An embodiment of the present system is shown, which is particularly suitable for searching technical documents, such as patent documents or scientific documents. The system includes a document storage 10A containing a plurality of natural language documents. A graph parser 12 is adapted to read the documents from the document storage 10A and convert them into a graph format, which will be discussed in more detail later. The converted graph is stored in a graph storage 10B.

[0070] The system includes a neural network trainer unit 14 that receives as training data a set of parsed graphs from a graph memory and some information about their relationships to each other. In this case, a document citation data memory 10C is provided, including, for example, citation data about documents and / or novelty search results. Trainer unit 14 runs a graph-based neural network algorithm that generates a neural network model for a neural network-based search engine 16. Engine 16 uses the graphs from graph memory 10B as the target search set and uses user data—typically text or graphs—obtained from user interface 18 as a reference.

[0071] The search engine 16 may be, for example, a graph-to-vector search engine that is trained to find the vector corresponding to the graph in the graph memory 10B that is closest to the vector formed by the user data. The search engine 16 may also be a classifier search engine, such as a binary classifier search engine, that compares the user graph or a vector derived therefrom with a graph obtained from the graph memory 10B or a vector derived therefrom in pairs.

[0072] Figure 1B An embodiment of the system is shown, which also includes a text embedding unit 13 that converts the natural language units of the graphics into a multi-dimensional vector format. This is done for graphics that have been converted from the graphics memory 10B and for graphics that have been input through the user interface 18. Typically, the vectors have at least 100 dimensions, such as 300 dimensions or more.

[0073] Also shown in Figure 1B In one embodiment, the neural network search engine 16 is divided into two parts forming a series. For example, the engine 16 includes a graph embedding engine that converts a graph into a multidimensional vector format using a model trained by the graph embedding trainer 14A of the neural network trainer 14 using reference data from the document reference data storage 10C. A vector comparison engine 16B compares the user graph with a graph pre-generated by the graph embedding engine 16A. As a result, a narrowed subset of graphs that are closest to the user graph is found. This subset of graphs is further compared with the user graph by the graph classifier engine 16C to further narrow the group of related graphs. The graph classifier engine 16C is trained by the graph classifier trainer 14C using data from the document reference data storage 10C, for example, as training data. This embodiment is advantageous because vector comparison of pre-generated vectors by the vector comparison engine 16B is very fast, while the graph classifier engine has access to the detailed data content and structure of the graphs and can accurately compare the graphs to identify differences. The pattern embedding engine 16A and the vector comparison engine 16B act as efficient pre-filters for the pattern classifier engine 16C, thereby reducing the amount of data that needs to be processed by the pattern classifier engine 16C.

[0074] The graph embedding engine may convert the graph into a vector having at least 100 dimensions, preferably 200 dimensions or more, and even 300 dimensions or more.

[0075] The neural network trainer 14 is divided into two parts: a graph embedding part and a graph classifier part. The graph embedding part and the graph classifier part are trained using a graph embedding trainer 14A and a graph classifier trainer 16C, respectively. The graph embedding trainer 14A forms a neural network-based graph-to-vector model, the purpose of which is to form nearby vectors for graphs with similar text content and internal structure. The graph classifier trainer 14B forms a classifier model that can rank graph pairs based on the similarity of their text content and internal structure.

[0076] After being embedded in the embedding unit 13, the user data obtained from the user interface 18 is fed to the graphic embedding engine for vectorization. After that, the vector comparison engine 16B finds a set of closest vectors corresponding to the graphic in the graphic memory 10B. This set of closest graphics is fed to the graphic classifier engine 16C, which uses a trained graphic classifier model to compare them one by one with the user graphic to obtain an accurate match.

[0077] In some embodiments, the graph embedding engine 16A, as trained by the graph embedding trainer 14A, outputs vectors whose angles are closer to each other, the more similar the graphs are in terms of node content and node structure, as learned from the reference data using the learning objectives dependent thereon. Through training, the angles of the vectors of positive training cases (graphs describing the same concept) derived from the reference data can be minimized, while the angles of the vectors of negative training cases (graphs describing different concepts) are maximized, or at least significantly deviate from zero.

[0078] The graphics vector may be selected to have, for example, 200-1000 dimensions, such as 250-600 dimensions.

[0079] It has been found that such a supervised machine learning model is able to efficiently evaluate the similarity of technical concepts disclosed by a graph and further by the natural language chunks from which the graph is derived.

[0080] In some embodiments, graph classifier engine 16C, as trained by graph classifier trainer 14C, outputs a similarity score, the higher the similarity score, the more similar the compared graphs are in terms of node content and node structure, as learned from the reference data using the learning objectives dependent thereon. Through training, the similarity scores of positive training cases (graphs describing the same concept) derived from the reference data can be maximized, while the similarity scores of negative training cases (graphs describing different concepts) are maximized.

[0081] Cosine similarity is one possible criterion for similarity of graphs or vectors derived from graphs.

[0082] It should be noted that the graph classifier trainer 14C or engine 16C is not mandatory, but the graph similarity can be evaluated directly based on the angle between the vectors embedded by the graph embedding engine. For this purpose, a fast vector index known per se can be used to find one or more neighboring graph vectors of a given fresh graph vector.

[0083] The neural network used by the trainer 14 and the search engine 16 or any one or both of their sub-trainers 14A, 14C or sub-engines 16A, 16C can be a recurrent neural network, in particular a recurrent neural network using long short-term memory (LSTM) units. In the case of a tree-structured graph, the network can be a Tree-LSTM network, such as a Child-Sum-Tree-LSTM network. The network can have one or more LSTM layers and one or more network layers. The network can use an attention mechanism that internally or externally relates parts of the graph to each other when training and / or running the model.

[0084] Some further embodiments of the invention are described below in the context of a patent search system, whereby the documents processed are patent documents.The general embodiments and principles described above are applicable to the patent search system.

[0085] In a certain embodiment, the system is configured to store natural language documents in a storage device, each natural language document comprising a first natural language block and a second natural language block different from the first natural language block. The trainer may use a plurality of first graphs corresponding to the first block of the first document, and for each first graph, use one or more second graphs based at least in part on the second block of a second document different from the first document, as defined by reference data. In this way, the neural network model learns from the interrelationships between different parts of different documents. On the other hand, the trainer may use a plurality of first graphs corresponding to the first block of the first document, and for each first graph, use a second graph based at least in part on the second block of the first document. In this way, the neural network model can learn from the internal relationships of the data within a single document. These two learning schemes can be used separately or simultaneously by the patent search system described in detail below.

[0086] The simplified graphic representation discussed above is particularly suitable for patent search systems, namely for claim graphics and full specification graphics, and in particular for full specification graphics.

[0087] Figure 1CA system is shown that includes a patent document storage 10A containing patent documents, each of which includes at least a computer-readable specification portion and a claims portion. A graph parser 12 is configured to parse claims via a claims graph parser 12A and to parse the full specification via a full specification graph parser 12B. The parsed graphs are stored separately in the claims and full specification graph storage 10B. A text embedding unit 13 prepares the graphs for processing in a neural network.

[0088] Citation data may include search and / or examination data for public patent applications and patents and / or citation data between patent documents. In one embodiment, the citation data includes prior patent search results, i.e., information on earlier patent documents considered as obstacles to novelty and / or inventive step for later-filed patent applications. The citation data is stored in a prior patent search and / or citation data storage 10C.

[0089] The neural network trainer 14 uses the parsed and embedded graph to form a neural network model that is trained specifically for patent search purposes. This is achieved by using patent search and / or citation data as input to the trainer 14. The goal is, for example, to minimize the vector angle or maximize the similarity score between the claim graph of a patent application and the full specification graph of the patent document that serves as its novelty barrier. In this way, applied to a plurality of (typically hundreds of thousands or millions) of claims, the model learns to evaluate the novelty of the claims relative to the prior art. The model is used by the search engine 16 for the user graph obtained through the user interface 18A to find the most likely novelty barriers. The results can be shown in the search results view interface 18B.

[0090] Figure 1C The system can utilize a range of search engines. The engines can be trained using the same or different subsets of training data obtained from previous patent searches and / or citation data stores 10C. For example, one can use a graph embedding engine trained on a large or complete set of citation data (i.e., positive claim / full description pairs and negative claim / full description pairs) to filter a set of graphs from a complete prior art data set. The filtered set of graphs is then classified against the user graph in a classification engine, which can be trained on a smaller, for example, patent classification-specific, citation data set (i.e., positive claim / full description pairs and negative claim / full description pairs) to find graph similarities.

[0091] Next, refer to Figure 2A and Figure 2B Describes a tree-like graphical structure particularly suitable for patent search systems.

[0092] Figure 2A A tree graph is shown with only part-word relationships as edge relationships. Text units AD are arranged within the graph as linear recursive nodes 10, 12, 14, and 16, originating from root node 10, with text unit E as a child of node 12, as child node 18, as derived from the natural language block shown. In this document, part-word relationships are detected from the part-word / whole-word expressions "include," "have," "are contained within," and "contain."

[0093] Figure 2B Another tree graph is shown with two different edge relationships, which in this example are a part-word relationship (first relationship) and a hyponym relationship (second relationship). Text units AC are arranged as linear recursive nodes 10, 12, 14 with part-word relationships. Text unit D is arranged as a child node 26 of the parent node 14 with a hyponym relationship. Text unit E is arranged as a child node 24 of the parent node 12 with a hyponym relationship. Text unit F is arranged as a child node 28 of the node 24 with a part-word relationship. In this article, part-word relationships and hyponym relationships are detected from the part-word / whole-word expressions "including", "having", "such as" and "is for example".

[0094] According to one embodiment, the first data processing device is adapted to convert the block into a graph by first identifying a first set of natural language symbols (e.g., nouns and noun chunks) and a second set of natural language symbols (e.g., part words and whole word expressions) different from the first set of natural language symbols from the block. Then, a matcher is executed using the first set of symbols and the second set of symbols to form matching pairs of the first set of symbols (e.g., "subject" and "component" from "subject includes component"). Finally, the matching pairs are used to arrange the first set of symbols into nodes of the graph (e.g., "subject" - (part word edge) - "component").

[0095] In one embodiment, at least partial-word edges are used in the graph, whereby individual nodes contain natural language units having partial-word relationships with each other, as derived from the chunks.

[0096] In one embodiment, hyponymy edges are used in the graph, whereby each node contains natural language units that have a hyponymy relationship with each other, as derived from a natural language chunk.

[0097] In one embodiment, edges are used in the graph, with at least one of the nodes of the graph containing a reference to one or more nodes in the same graph and additionally at least one natural language unit derived from the corresponding natural language block (e.g., "below" [node id: X]). In this way, graph space is conserved and a simple, e.g., tree-like, graph structure can be maintained, still allowing for expressive data content in the graph.

[0098] In some embodiments, the graph is a tree graph whose node values ​​contain words or multi-word chunks, which are typically derived from the natural language chunks or from their vectorized forms by a graph conversion unit using the parts of speech and syntactic dependencies of the words.

[0099] Figure 3 An example of how text to graph conversion is implemented in a first data processing device is shown in detail. First, text is read in step 31 and a first set of natural language symbols (such as nouns) and a second set of natural language symbols (such as symbols indicating part-of-speech (meronymity) or whole-of-speech (holonymity) (e.g., "include") are detected from the text. This can be achieved by tokenizing the text in step 32, tagging the symbols with parts of speech (POS) in step 33, and deriving their syntactic dependencies in step 34. Using this data, noun chunks can be determined in step 35 and part-of-word and whole-of-word expressions can be determined in step 36. In step 37, matching pairs of noun chunks are formed using the part-of-word and whole-of-word expressions. The noun chunk pairs form or can be used to infer part-of-word relationship edges of a graph.

[0100] In one embodiment, the noun chunk pairs are arranged into a tree graph, where the part words are children of the corresponding whole words, as shown in step 38. The graph is saved in a graph memory in step 39 for further use, as discussed above.

[0101] In one embodiment, the graph formation step involves using a probabilistic graphical model (PGM) such as a Bayesian network for reasoning about a preferred graph structure. For example, different edge probabilities of the graph can be calculated based on a Bayesian model, after which the edge probabilities are used to calculate the most likely graph form.

[0102] In one embodiment, the graph formation step includes feeding the text - typically in a tokenized, POS-tagged and dependency-parsed form - into a neural network-based technical parser that finds related chunks from the text block and extracts their desired edge relationships, such as part-word relationships and / or hyponym relationships.

[0103] In one embodiment, the graph is a tree graph comprising edge relationships recursively arranged according to a tree data pattern, the tree data pattern being acyclic. This allows for the use of efficient tree-based neural network models of either cyclic or non-recurrent type. An example is the Tree-LSTM model.

[0104] In another embodiment, the graph is a network graph that allows cycles, ie edges between branches. This has the benefit of allowing complex edge relationships to be expressed.

[0105] In yet another embodiment, the graph is a forest of linear and / or nonlinear branches having one or more edge lengths. Linear branching has the benefit of avoiding or significantly simplifying the tree or network construction step, and maximizing the amount of source data available for the neural network.

[0106] In each model, the marginal likelihood—if obtained through the PGM model—can be stored and used by the neural network.

[0107] It should be noted that, as mentioned above Figure 3 The graphic formation methods described elsewhere in this document can be implemented independently of the other methods and system portions described herein to form and store a technically concise representation of the technical content of a document, particularly a full patent specification and claims.

[0108] Figures 4A-4C Different, but not mutually exclusive, approaches to training neural networks, particularly for patent search purposes, are shown.

[0109] For general use, the term "patent document" may be replaced with a "document" (having a unique computer-readable identifier among other documents in the system), "claims" may be replaced with a "first computer-identifiable block," and "full specification" may be replaced with a "second computer-identifiable block that is at least partially different from the first block."

[0110] exist Figure 4AIn one embodiment, a plurality of claim graphs 41A and a corresponding prior art full specification graph 42A for each claim graph, such as those related by reference data, are used as training data by a neural network trainer 44A. They form positive training cases, indicating that a low vector angle or high similarity score between such graphs will be achieved. Additionally, for each claim graph, negative training cases—i.e., one or more distant prior art graphs—can be used as part of the training data. A high vector angle or low similarity score between such graphs will be achieved. Negative training cases can be randomized, for example, from the complete set of graphs.

[0111] According to one embodiment, in at least one phase of training, such as that performed by neural network trainer 44A, a plurality of negative training cases are selected from a subset of all possible training cases that are harder than the average of all possible negative training cases. For example, the hard negative training cases may be selected such that the claim graphics and the specification graphics are from the same patent classification (up to a predetermined classification level), or such that the neural network was previously unable to correctly classify the specification graphics as a negative case (with a predetermined confidence level).

[0112] According to one embodiment, which may also be implemented independently of the other methods and systems described herein, training of the neural network-based patent search or novelty assessment system is performed by providing a plurality of patent documents, each of which has a computer-readable claim block and a full specification block, wherein the full specification block includes at least a portion of the specification of the patent document. The method further includes providing a neural network model and training the neural network model using a training data set, wherein the training data set includes data from the patent documents used to form the trained neural network model. Training includes using claim block and full specification block pairs from the same patent document as training cases in the training data set.

[0113] Typically, these document-internal positive training cases form a small fraction of all training cases for training, eg 1%-25%, with the remainder comprising eg search report (examiner novelty citation) training cases.

[0114] The machine learning model is generally configured to convert claims and full specifications into vectors, and a learning objective of the model training may be to minimize the vector angle between the claim vectors and the full specification vectors of the same patent document. Another learning objective may be to maximize the vector angle between the claim vectors and the full specification vectors of at least some different patent documents.

[0115] exist Figure 4BIn one embodiment, multiple claim graphs 41A and full specification graphs 42A originating from the same patent document are used as training data by a neural network trainer 44B. A claim's "own" full specification typically forms a perfect positive training case. In other words, a patent document itself is technically an ideal novelty barrier for its claims. Therefore, these graph pairs form positive training cases, indicating that low vector angles or high similarity scores between such graphs will be achieved. Citation data and / or negative training cases can also be used in this scenario.

[0116] Tests have shown that simply adding claim-description pairs from the same document to realistic novelty-based search training data improves prior art classification accuracy by more than 15% when tested with realistic novelty-based search test data pairs.

[0117] Typically, at least 80%, often at least 90%, and in many cases 100% of the machine-readable content (natural language units, in particular words) of a claim is found somewhere in the full specification of the same patent document. Thus, the claims and the full specification of a patent document are linked to each other not only by cognitive content and the same unique identifier (e.g., publication number), but also by their byte-level content.

[0118] According to one embodiment, which may also be implemented independently of the other methods and system portions described herein, training of the present neural network-based patent search or novelty assessment engine includes deriving at least one reduced data instance partially corresponding to the original block from at least some original claims or full specification blocks, and using the reduced data instance together with the original claims or full specification blocks as training cases of the training data set.

[0119] exist Figure 4C In an embodiment, positive training cases are augmented by forming a plurality of reduced claim graphs 41C"-41C"" from an original claim graph 41C'. A reduced claim graph is a graph wherein

[0120] - At least one node is removed (e.g. Phone-Display-Sensor -> Phone-Display)

[0121] - at least one node is moved to another position at a higher (more general) position of the branch (e.g., Phone-Display-Sensor -> Phone-(Display, Sensor), and / or

[0122] - The natural language unit value of at least one node is replaced with a more general natural language unit value (phone-display-sensor -> electronic device-display-sensor).

[0123] This augmentation scheme allows the training set for the neural network to be expanded, thereby producing a more accurate model. It also allows for meaningful searches and evaluations of the novelty of so-called trivial inventions using only a few nodes or in very general terms, which are not often seen in actual patent novelty search data. Figure 4A and Figure 4B Data augmentation can be performed using any one or a combination of the embodiments of [ 0 ]. Also in this scenario, negative training cases can be used.

[0124] Negative training cases can also be augmented by removing, moving, or replacing nodes or their values ​​in the full specification graph.

[0125] Tree-like graph structures, such as those based on part-word relations, are beneficial for augmentation schemes because they can be augmented by deleting nodes or moving them to higher tree positions in a straightforward and robust manner, thus still preserving coherent logic. In this case, both the original data instances and the reduced data instances are graphs.

[0126] In one embodiment, the reduced graph is a graph that has deleted at least one leaf node relative to the original graph or another reduced graph.In one embodiment, all leaf nodes at a certain depth of the graph are deleted.

[0127] This type of augmentation can also be performed directly on the natural language block, in particular by deleting parts of the natural language block or partially changing the content of the natural language block to a more general content.

[0128] The number of reduced data instances per original instance may be, for example, 1-10000, in particular 1-100. Good training results are achieved in claim augmentations with 2-50 augmentation patterns.

[0129] In some embodiments, the search engine reads fresh natural language blocks, such as fresh claims, which are converted to fresh graphics by a converter, or directly reads fresh graphics through a user interface. Next, a user interface suitable for direct graphic input is discussed.

[0130] Figure 5The representation and modification of an exemplary graphic on a display element 50 of a user interface is illustrated. The display element 50 includes a plurality of editable data cells AF, whose values ​​are functionally connected to corresponding natural language units of the underlying graphic (e.g., cells AF, respectively), and are shown in respective user interface (UI) data elements 52, 54, 56, 54', 56', 56". The UI data elements can be, for example, text fields, the values ​​of which can be edited via the keyboard after the element is activated. The UI data elements 52, 54, 55, 54', 56', 56" are positioned horizontally and vertically on the display element 50 according to their position in the graphic. In this document, the horizontal position corresponds to the depth of the cell in the graphic.

[0131] Display element 50 may be, for example, a window, frame, or panel of a web browser running a web application, or a graphical user interface window of a stand-alone program executable in a computer.

[0132] The user interface also includes a shifting engine that allows the natural language unit to be moved horizontally (vertically) on the display element in response to user input and to modify the graphics accordingly. To illustrate this, Figure 5 Data cell F (element 56") is shown shifted one level to the left (arrow 59A). Due to this, the original element 56" nested under element 54' ceases to exist and is formed as element 54" nested under the higher-level element 52 and including data cell F (with its original value). If data element 54' is thereafter shifted two levels to the right (arrow 59B), data element 54' and its children will be shifted to the right and nested under data element 56 as data element 56" and data element 58. Each shift is reflected by a corresponding shift of the nesting level in the underlying graphic. Thus, when the children of a cell are shifted to a different nesting level in the user interface, the children of the cell will be retained in the graphic.

[0133] In some embodiments, the UI data element includes a natural language helper element that is shown in relation to an editable data cell and is used to help the user input natural language data. The content of the helper element can be formed using relationship elements associated with the relevant natural language element and, optionally, the natural language element of its parent element.

[0134] Instead of Figure 5 The graphical user interface illustrated in , which may allow entry of a block of text, such as the independent claim. The text block is then fed to a graphical parser to obtain a graph that can be used in further stages of the search system.

Claims

1. A computer-implemented method for training a machine learning-based patent search or novelty assessment system, comprising: providing a plurality of patent documents, each of the patent documents having a computer-recognizable claim block and a computer-recognizable full specification block, the full specification block including at least a portion of the specification of the patent document, Provide machine learning models, training the machine learning model using a training data set by a machine learning training subsystem, the training data set including data from the patent documents used to form the trained machine learning model, wherein the machine learning training subsystem is adapted to read patent claim blocks and full specification blocks of the plurality of patent documents, The training includes converting the claim block and the specification block into a graph, the graph comprising a plurality of nodes, each of the nodes comprising a natural language unit extracted from the corresponding block, A graph-based neural network model is used, wherein the graph-based neural network model is capable of embedding the graph into a vector, and a first pair of claim graphs and full specification graphs from the same patent document are used as training cases of the training data set, wherein the learning objective of the training is to minimize the vector angle between the first pair of claim graphs and full specification graphs from the same patent document. A second pair of claim blocks and full specification blocks from different patent documents is also used as training cases of the training data set.

2. The method according to claim 1 , comprising using the first pair of claim graphics and full specification graphics derived from the same patent document as positive training cases, wherein the positive training cases indicate positive search hits or negative novelty evaluation results. 3 . The method of claim 1 , wherein the graph format is a recursive tree format comprising nested nodes having natural language data units as node values.

4. The method of claim 1, wherein the converting comprises identifying from the block a first set of natural language symbols and a second set of natural language symbols different from the first set of natural language symbols, executing a matcher using the first set of natural language symbols and the second set of natural language symbols to form matching pairs of the first set of natural language symbols, At least a portion of the first set of natural language tokens is arranged as consecutive nodes of the graph using the matching pairs.

5. The method of claim 1, wherein the graph comprises a plurality of edges, each node of the edges comprising natural language units having part-word or hyponym relationships with each other, as derived from natural language chunks.

6. The method according to claim 1 or 2, wherein the claim block comprises independent claims of the patent document, such as the first independent claim.

7. The method according to claim 1 or 2, wherein the claim block comprises a combination of an independent claim of a patent document and its dependent claims.

8. The method of claim 1, wherein the positive training cases are augmented by forming a plurality of reduced claim graphs from an original claim graph.

9. A natural language document comparison system based on machine learning, including Machine learning training subsystem for reading computer-recognizable claim blocks and computer-recognizable full specification blocks of a collection of patent documents, each patent document having a computer-recognizable claim block and a full specification block, the full specification block including at least a portion of the specification of the patent document, and utilizing the block as training data for forming a trained machine learning model, wherein the full specification block is at least partially distinct from the claims block, a machine learning search engine that uses the trained machine learning model to find a subset of patent documents in the collection of patent documents, The machine learning training subsystem is configured as Converting the claim block and the full specification block into a first graph and a second graph, wherein the first graph and the second graph include a plurality of nodes, each of the nodes includes a natural language unit extracted from the corresponding block, Using a graph-based neural network algorithm and using a first pair of first and second graphs from the same patent document as training cases of the training data, wherein the learning objective of the training is to minimize the vector angle between the first pair of first and second graphs from the same patent document, A second pair of claim blocks and full specification blocks from different patent documents is also used as training cases of the training data set.

10. The system of claim 9, wherein the graph comprises a plurality of edges, each node of the edges comprising natural language units having part-word or hyponym relationships with each other, as derived from natural language chunks.

11. Use of the claims and full specification of the same patent document as training cases for machine learning-based patent search or novelty assessment systems, The machine learning-based patent search or novelty assessment system includes a machine learning model, which is configured to convert claims and full specifications into vectors, and wherein the learning goal of the model training is to minimize the vector angle between the claim vector and the full specification vector of the same patent document.

12. The use according to claim 11 further comprises using claims and full specifications of different patent documents as training cases, wherein the learning objective of the training of the model is to maximize the vector angle between the claim vectors and full specification vectors of different patent documents.

Images