Building enterprise-specific knowledge graphs

Abstract

A framework for constructing an enterprise-specific knowledge base based on enterprise-specific data including structured data and unstructured data is provided. For each of a plurality of tuples included in the structured data, a relationship between entities that matches a known relationship is identified. Where possible, relationships between entities that match a known relationship are also identified for tuples included in the unstructured data. If a matching relationship between entities cannot be identified for a tuple in the unstructured data, the extracted relationships are sequentially clustered into similar relationships and the relationships are assigned to the clustered tuples. An enterprise-specific knowledge graph is constructed based on structured data tuples and their identified relationships, unstructured data tuples where relationships can be mapped to known relationships and their identified relationships, and unstructured data tuples that cannot be mapped to known relationships and their assigned relationships. The knowledge graph is enriched with any information determined to be missing from the knowledge graph.

Description

Background Technology

[0001] Organizing and constructing online information is an important and challenging task, as it greatly supports information retrieval and question answering. Given its importance, the number of large-scale, publicly available knowledge bases has increased significantly in recent years. Examples include, but are not limited to, Google Knowledge Graph, Yago Knowledge Base, and DBpedia. The most readily accessible large-scale knowledge bases are built on a combination of structured data (i.e., data residing in fixed fields within documents, files, or records, such as infoboxes, tables, etc.) and unstructured data (i.e., data without a predefined data model or not organized in a predefined way).

[0002] While the success of these knowledge bases only proves their usefulness, most existing knowledge bases are built using general-purpose data corpora. Therefore, the information existing in these knowledge bases often cannot be extended to the needs of specific domains or businesses. For example, these knowledge bases are often proven to be useful for enterprise clients (such as, for instance, It is almost useless for retrieving information because, at least in part, the data that meets the needs of an enterprise's knowledge organization is not available in general sources. Furthermore, the types of data available within a specific enterprise or field vary in quality. Summary of the Invention

[0003] Embodiments of the present invention particularly relate to a framework for constructing enterprise-specific knowledge graphs. It receives enterprise-specific data, including both structured and unstructured data. Structured data is data residing in fixed fields within documents, files, or records, such as infoboxes, tables, etc. Structured data in the form of text field / value pairs naturally occurs as multiple tuples or triples with a <subject><verb><object> format. Each distinct text field represents an edge between an entity (referenced by <subject> and <object> elements) and its corresponding text field / value. Relationships between entities in the structured data (referenced by <verb> elements) are identified by comparison with predefined relation classifications.

[0004] Unstructured data is data that does not have a predefined data model or is not organized in a predefined manner. Exemplary unstructured data is free text, such as alphanumeric text presented in sentence format. In this embodiment, tuples (or triples) in the format <subject><verb><object> are extracted from the unstructured data. Entities (referenced by <subject> and <object> elements) and relations (referenced by <verb> objects) are normalized to standardize the tuples to predefined categories. Entity normalization maps entities to predefined entity categories that include entity classes (e.g., people, locations, etc.). Where possible, relation normalization maps relations to predefined relation categories. Relations that cannot be mapped to predefined relation categories are normalized via sequential clustering to group similar relations together and assign relations to clustered tuples.

[0005] Enterprise-specific knowledge graphs are constructed based on the relationships between structured data tuples and their corresponding identifiers, the relationships that can be mapped to predefined relation categories of unstructured data tuples and their corresponding identifiers, and the relationships that cannot be mapped to predefined relation categories of unstructured data tuples and their corresponding assignments. Any information identified as missing from the enterprise-specific knowledge graph using probabilistic methods is then used to enrich the graph.

[0006] This "Summary" is provided to introduce some concepts in a simplified form, which will be further described in the "Detailed Description" below. This "Summary" is not intended to identify key or essential features of the claimed subject matter, nor is it intended to help determine the scope of the claimed subject matter. Attached Figure Description

[0007] Some embodiments or implementations of this disclosure are illustrated in the accompanying drawings by way of example rather than limitation, in which the same reference numerals denote similar elements, and in the accompanying drawings:

[0008] Figure 1 It is a block diagram of an exemplary computing system that can employ some implementations of this disclosure;

[0009] Figure 2 This is a schematic diagram illustrating exemplary enterprise-specific data that can be used in some implementations of this disclosure;

[0010] Figure 3 This is an exemplary schematic diagram of cluster-based normalization that can be adopted in some implementations of this disclosure;

[0011] Figure 4 This is a flowchart illustrating an exemplary method for constructing an enterprise-specific knowledge graph according to some implementations of this disclosure;

[0012] Figure 5This is a flowchart illustrating another exemplary method for constructing an enterprise-specific knowledge graph according to some implementations of this disclosure;

[0013] Figure 6 This is a flowchart illustrating yet another exemplary method for constructing an enterprise-specific knowledge graph according to some implementations of this disclosure; and

[0014] Figure 7 This is a block diagram of an exemplary computing environment suitable for use according to some implementations of this disclosure. Detailed Implementation

[0015] The subject matter of this invention has been specifically described herein to satisfy legal requirements. However, the description itself is not intended to limit the scope of this patent. Rather, the inventors have anticipated that the claimed subject matter can also be implemented in other ways, in combination with other existing or future techniques, including combinations of steps different from or similar to those described herein. Furthermore, while the terms “step” and / or “block” may be used herein to denote different elements of the method employed, these terms should not be construed as implying any particular order among or between the various steps disclosed herein, unless the order of individual steps is explicitly described.

[0016] For the purposes of this disclosure, the word "including" has the same broad meaning as the word "comprising." Furthermore, unless otherwise indicated, words such as "a" and "an" include both plural and singular forms. Thus, for example, the requirement of "feature" is satisfied when one or more features are present. Additionally, the term "or" includes conjunctions, adversative conjunctions, and both (therefore, a or b includes a or b as well as a and b).

[0017] As mentioned earlier, organizing and constructing online information is an important and challenging task, as it greatly supports information retrieval and question answering. Given its importance, the number of large-scale, publicly available knowledge bases has increased significantly in recent years. Examples include, but are not limited to, Google Knowledge Graph, Yago Knowledge Base, and DBpedia. The most readily accessible large-scale knowledge bases are built on a combination of structured data (i.e., data residing in fixed fields within documents, files, or records, such as infoboxes, tables, etc.) and unstructured data (i.e., data without a predefined data model or not organized in a predefined way, such as free text).

[0018] The success of these knowledge bases alone proves their usefulness. While not typically built for a specific domain or industry, knowledge graphs can be used even in enterprise environments to retrieve relevant results or answer questions that require knowledge from more than one source / document. For example, by simultaneously viewing all facts about "mobile phones from Samsung" and "mobile phones with dual SIM cards," queries such as "Which mobile phone from Samsung has dual SIM cards?" can be solved.

[0019] When considering enterprise-specific domains, most existing knowledge bases have significant limitations. First, most existing knowledge bases are built using general-purpose databases. Therefore, the information existing in these knowledge bases often cannot be extended to the needs of a specific domain or enterprise, at least in part because the data required to meet the needs of enterprise knowledge organization is not available from general sources. Furthermore, the types of data available within a specific enterprise or domain vary in quality. Therefore, the limited availability and non-standardized nature of available data make extracting meaningful facts from the mixture of structured and unstructured data particularly challenging in an enterprise environment.

[0020] In enterprise environments, linking various entities across documents / sources to provide a more comprehensive view is also challenging. This involves leveraging ontological constraints and entity disambiguation to identify multiple representations of the same entity in different locations / contexts. The lack of a standard ontology for enterprise environments makes this even more challenging, as it requires considering multiple representations of facts from the extracted knowledge and deduplicating these representations to achieve accurate knowledge representation.

[0021] Enriching enterprise-specific knowledge graphs by inferring additional relationships from existing facts is particularly challenging due to the limited amount of data in enterprise environments. Therefore, it is important to learn relationships from existing data to infer new facts and thus enrich the knowledge graph with information not explicitly present in the dataset.

[0022] As mentioned earlier, knowledge bases such as the Google Knowledge Graph, Yago Knowledge Base, and DBpedia have been built by processing data in the public domain. These knowledge bases extract meaningful relations and construct tuples (triples) in the format <entity, relation, entity>. Each of these triples constitutes a fact. However, all these knowledge bases cover a wide range of real-world concepts and are not limited to any particular domain.

[0023] Building a knowledge base for a specific industry / enterprise requires a fundamentally different process. Unlike facts extracted from large corpora such as WIKIPEDIA, facts from much smaller enterprise data corpora are often noisy and require a mechanism to filter out relevant and irrelevant information. Standard classifications are not entirely suitable for classifying relations within domain-specific ontologies because these relations in the facts vary across industries / enterprises. Furthermore, classifications derived from standard sources (such as DBpedia) have been manually curated and are not feasible for each individual domain or enterprise.

[0024] Given a content repository with both structured and unstructured data, this disclosure provides an implementation of a method for pruning facts from content by processing ontology constraints through all possible tuples (triples) in the data. Each processed fact is represented as a tuple in the format <subject><verb><object> (or <entity><relation><entity>). Further pruning is performed by mapping tuples to standard classifications to link entities within documents and across documents. Based on content derived from multiple enterprise-specific sources / documents, this provides a comprehensive view of all facts associated with a single entity.

[0025] The implementation of this disclosure also provides a mechanism for identifying and clustering semantically similar relations expressed differently in data. This is accomplished by combining semantic and syntactic pairings of relations that cannot be mapped to a predefined relation classification. The implementation further provides a method for enriching facts by probabilistically inferring new facts from previously extracted facts.

[0026] Therefore, an exemplary aspect of the present invention relates to a computer-implemented method for constructing an enterprise-specific knowledge graph. The method includes receiving enterprise data from at least one enterprise-specific source. A first portion of the enterprise data is structured data, and a second portion of the enterprise data is unstructured data. The structured data includes a plurality of text fields, each text field having a corresponding value, such that the structured data includes a first plurality of text field / value pairs. The method further includes identifying a relation corresponding to each text field / value pair in the first plurality of text field / value pairs, which matches one of a plurality of predefined relations included in a predefined relation classification. Furthermore, the method includes extracting a second plurality of text field / value pairs from the unstructured data, and determining that each text field / value pair in the first portion of the second plurality of text field / value pairs references a relation that matches one of the plurality of predefined relations included in the predefined relation classification, and that each text field / value pair in the second portion of the second plurality of text field / value pairs references a relation that does not match one of the plurality of predefined relations. The method further includes identifying a predefined relation among the plurality of predefined relations included in the predefined relation classification that matches each text field / value pair in the first portion of the second plurality of text field / value pairs. Furthermore, the method includes sequentially clustering relations referenced by a second portion of a second plurality of text field / value pairs, such that similar relations are grouped together, and assigning assigned relations to at least a portion of the text field / value pairs in the second plurality of text field / value pairs based on sequential clustering. Additionally, the method includes constructing an enterprise-specific knowledge graph at least in part based on relations between a first plurality of text field / value pairs and their corresponding identifiers, relations between a first portion of the second plurality of text field / value pairs and their corresponding identifiers, and relations between a portion of the second portion of the second plurality of text field / value pairs and their corresponding assignments. In embodiments, the method may further include enriching the enterprise-specific knowledge graph by adding third plurality of text field / value pairs that are determined to be missing based on previously known existing patterns.

[0027] Another exemplary embodiment relates to one or more computer storage media storing computer-usable instructions that, when executed by one or more processors, cause one or more processors to perform a method for constructing an enterprise-specific knowledge graph. The method includes receiving enterprise data from an enterprise knowledge base, a first portion of which is structured data and a second portion of which is unstructured data. The structured data comprises a first plurality of tuples, each structured data tuple including a <subject> element, a <predicate> element, and a <object> element arranged in a <subject><predicate><object> format. The method also includes identifying a relation corresponding to each of the plurality of structured data tuples, the relation being matched against one of a plurality of predefined relations included in a predefined relation classification. Furthermore, the method includes extracting a second plurality of tuples from the unstructured data, each unstructured data tuple including an <subject> element, a <predicate> element, and a <object> element arranged in a <subject><predicate><object> format. Furthermore, the method includes identifying relationships where each unstructured data tuple reference in the first part of the second plurality of tuples matches one of a plurality of predefined relationships included in a predefined relationship classification, and relationships where each unstructured data tuple reference in the second part of the second plurality of tuples does not match one of a plurality of predefined relationships included in the predefined relationship classification. Additionally, the method includes identifying predefined relationships among the plurality of predefined relationships included in the predefined relationship classification that match each unstructured data tuple in the first part of the second plurality of tuples. The method also includes sequentially clustering relationships referenced by the second part of the second plurality of tuples such that similar relationships are grouped together, and assigning assigned relationships to at least a portion of the unstructured data tuples of the second plurality of tuples based on sequential clustering. The method also includes constructing an enterprise-specific knowledge graph at least partially based on the structured data tuples of the first plurality of tuples and their corresponding identified relationships, the unstructured data tuples of the first part of the second plurality of tuples and their corresponding identified relationships, and at least a portion of the unstructured data tuples of the second part of the second plurality of tuples and their corresponding assigned relationships. Furthermore, the method includes enriching the enterprise-specific knowledge graph by adding third plurality of tuples identified as missing from the enterprise-specific knowledge graph based on existing patterns.

[0028] Another exemplary embodiment relates to a computer system including one or more processors and one or more computer storage media storing computer-usable instructions that, when used by the one or more processors, cause the one or more processors to perform a specific function. These functions may include, but are not limited to: (a) receiving enterprise data from at least one source, the at least one source having enterprise-specific information associated with the at least one source, a first portion of the received enterprise data being structured data and a second portion of the received enterprise data being unstructured data, the structured data having a first plurality of text fields, each of the first plurality of text fields corresponding to a structured data tuple; (b) normalizing the structured data by: (1) identifying the relation corresponding to each structured data tuple, and (2) mapping the relation to a relation included in a predefined relation classification; (c) normalizing the unstructured data by: (3) extracting a plurality of unstructured data tuples from the unstructured data, each of the plurality of unstructured data tuples having a <subject> element, a <predicate> element, and an <object> element, the <subject> element and the <object> element representing entities and the <predicate> element representing the relationship between the <subject> element and the <object> element. (4) Map the <subject> and <object> elements of each unstructured data tuple to the entity categories included in the predefined entity classification; (5) Determine which <predicate> elements can be mapped to the first part of the predefined relation classification and which <predicate> elements cannot be mapped to the second part of the predefined relation classification; (6) Map each <predicate> element of the first part of the <predicate> elements to the relation included in the predefined relation classification; (7) Sequentially cluster the second part of the <predicate> elements to group similar <predicate> elements together; (8) Assign the assigned relation to at least a part of the second part of the <predicate> elements based on the sequential clustering; (d) Construct an enterprise-specific knowledge graph at least in part based on the relation of structured data tuples and their corresponding identifiers, the relation of unstructured data tuples containing the first part of the <predicate> elements and their corresponding identifiers, and the relation of at least a part of the unstructured data tuples containing the second part of the <predicate> elements and their corresponding assigned relations.

[0029] Now refer to the attached diagram, Figure 1This is a block diagram illustrating an exemplary system 100 for constructing an enterprise-specific knowledge graph according to an implementation of this disclosure. It should be understood that this and other arrangements described herein are merely illustrative examples. Other arrangements and elements (e.g., machines, interfaces, functions, sequences, and functional groupings, etc.) may be used in addition to or instead of those shown, and some elements may be omitted entirely. Furthermore, many of the elements described herein are functional entities that can be implemented as discrete or distributed components or in combination with other components, and can be implemented in any suitable combination and location. The various functions described herein as being performed by one or more entities can be performed by hardware, firmware, and / or software. For example, various functions can be performed by a processor that executes instructions stored in memory.

[0030] System 100 is an example of a suitable architecture for implementing certain aspects of this disclosure. In addition to other components not shown, System 100 also includes a user device 110 that interacts with a knowledge graph infrastructure 112 to construct an enterprise-specific knowledge graph based on enterprise-specific data. Figure 1 Each component shown may be one or more computer devices discussed below (such as...) Figure 7 The present disclosure is provided on a computing device 700. It should be noted that implementations of this disclosure are equally applicable to mobile computing devices and devices that accept gesture, touch, and / or voice input. Any and all such variations and any combinations thereof are contemplated within the scope of implementations of this disclosure.

[0031] like Figure 1 As shown, user equipment 110 and knowledge graph infrastructure 112 can communicate via network 114, which may include, but is not limited to, one or more local area networks (LANs) and / or wide area networks (WANs). Such network environments are common in offices, enterprise-wide computer networks, intranets, and the Internet. Therefore, network 114 will not be further described herein. It should be understood that any number of user equipment and knowledge graph infrastructure can be employed within system 100 within the scope of this disclosure. Each may include a single device or multiple devices collaborating in a distributed environment. For example, knowledge graph infrastructure 112 may be provided by multiple server devices that collectively provide the functionality of knowledge graph infrastructure 112 as described herein. Additionally, other components not shown may also be included in the network environment.

[0032] The knowledge graph infrastructure 112 can access at least one data source or repository 116, including enterprise-specific data. In implementations of this disclosure, the data repository 116 is configured to search one or more items stored therein. It should be understood that the information stored in association with the data repository 116 can be configurable and can include any information related to data generated by and / or belonging to a specific enterprise or domain, as exemplified only. The content and amount of such information are not intended to limit the scope of any aspect of this technology. Furthermore, the data repository 116 can be a single, independent component (as shown) or multiple storage devices, such as a database cluster (parts of which may reside in association with the knowledge graph infrastructure 112), another external computing device (not shown), and / or any combination thereof. Additionally, the data repository 116 can include multiple unrelated data repositories or sources within the scope of embodiments of this technology.

[0033] The knowledge graph building infrastructure is typically configured to receive enterprise-specific data from a data repository (such as data repository 116) and construct an enterprise-specific knowledge graph based on the received data. Instructions for constructing or modifying the enterprise-specific knowledge graph can be received from user device 110, which can access and communicate with the knowledge graph building infrastructure 112 via a web browser or other applications running on user device 110. The knowledge graph building infrastructure 112 includes a data receiving component 118, a structured data processing component 120, an unstructured data processing component 122, a knowledge graph construction component 124, and a knowledge graph enrichment component 126.

[0034] Data receiving component 118 is configured to receive enterprise-specific data (e.g., from data repository 116) as a mixture of structured data (i.e., data residing in fixed fields in documents, files, or records, such as infoboxes, tables, etc.) and unstructured data (i.e., data without a predefined data model or not organized in a predefined manner). Structured data in the form of text field / value pairs naturally appears in tuples with a <entity><relation><entity> format and more specifically, a <subject><verb><object> format. Each distinct field represents an edge between an entity (referenced by <subject> and <object> elements) and its corresponding text field / value.

[0035] The structured data processing component 120 is configured to identify the relation corresponding to each text field / value pair contained in the structured data. This relation corresponding to a text field / value pair can be identified by comparing the <predicate> element or <relation> element of each tuple with relations appearing in one or more known relation classifications (e.g., YAGO relation classifications). This is merely an example and for reference only. Figure 2Tuples can be extracted from the structured text field "Author". <product62392> <iswrittenby> <tlau>(<Product 62392><Written by... <tlau>The text field is mapped to the relation "isWrittenBy" of the <predicate> element, which is further mapped to a predefined known relation category. In an embodiment, this normalization of the text field to the relation from the predefined known relation category can be performed in a non-automatic manner.

[0036] As described more fully below regarding the knowledge graph building component 124, tuples extracted from structured data, along with the identified corresponding relations that match predefined known relation categories, can be used to begin building enterprise-specific knowledge graphs.

[0037] Unstructured data processing component 122 is configured to process unstructured data into tuples with relational relationships. Initially, the unstructured data can be preprocessed using a co-reference parser to resolve any co-reference pronouns. Various co-reference parsers are known to those skilled in the art and therefore will not be described further herein. From the preprocessed text, tuples with the format <subject><verb><object> are extracted based on an OpenIE framework (known to those skilled in the art), which identifies relational phrases in the unstructured text based on syntactic and lexical constraints and assigns a pair of noun parameters (based on the <verb> element of the extracted tuple) to each extracted relation. For each tuple, the entity types of <subject> and <object> are identified, for example, using a Stanford Named Entity Recognizer (known to those skilled in the art).

[0038] Because tuple extraction is largely based on the structure of unstructured data (or lacks it) (e.g., based on sentence structure), it can produce noisy tuples. To reduce noisy tuples and address redundancy and ambiguity, entity sets are normalized, and relations extracted from unstructured data are also normalized. A two-pronged approach is used for this purpose.

[0039] Classification-based normalization. The entities (<subject> and <object>) and relations (<predicate> elements) included in each tuple are mapped to a predefined, known relation classification (e.g., YAGO relation classification). For entities, edit distances are calculated from matching entities in the relation classification, and a threshold (σ) is used to determine the edit distance. 实体 The mapping of ) is used as a normalized entity. Ontology constraints (such as the same entity should have the same type) are also incorporated to maintain consistency during normalization.

[0040] For these relationships, a set of three techniques can be used to identify appropriate relationships that can be mapped to a predefined relationship category, which is constrained by the properties of the entities associated with the tuple. These constraints contain ontological knowledge of the relationships between entity types and relationships to eliminate noisy triples. For example, the domain and scope constraints DOMAIN(isWrittenBy,book) and RANGE(isWrittenBy,person) specify that the relationship "isWrittenBy" is a mapping from entities of type book to entities of type person.

[0041] The first technique uses a vector space model to capture the semantic similarity of words in a phrase and classification-based relations. It involves computing the cosine similarity between relations derived from component tuples and Word2Vec embeddings from classification relations. Word2Vec captures the semantic space of words, and thus such a metric maps relations to relations in the classification based on their semantic similarity.

[0042] The second technique is syntactic mapping. In this technique, the similarity between two phrases is determined by the similarity of their main verbs. To capture the syntax of the component descriptions, a dependency parser is used, and a network of "cognitive synonyms" of the root verb of the dependency tree is extracted. This synonym network is compared with the root verbs of the categorized relations to establish a correspondence between the relations in a syntactic sense.

[0043] The third technique is pattern-based mapping. In this technique, two verb phrases are considered similar if they share some common word patterns, and some words, such as auxiliary verbs and adjectives, may differ. Based on this intuition, this technique acquires textual patterns of binary relations and transforms them into syntactic ontology lexical patterns using frequent itemset mining, and then constructs classifications for these patterns. These classifications are used to identify the closest matching relation in a predefined classification corresponding to the current pattern tuple by checking whether the two relations contain the same patterns.

[0044] Votes from all three techniques mentioned above are used to determine the best normalization relation for the current tuple.

[0045] Cluster-based normalization. Since enterprise-specific data is expected to be unique, normalization based solely on known predefined classifications is often insufficient to eliminate all extracted tuples. To normalize new relations not present in the predefined classifications, a relation clustering step is performed to group similar phrases together and represent them as a single relation. As mentioned above, the cosine similarity between Word2vec relation vectors is a similarity measure to which clustering can be performed. In the implementation, density-based spatial clustering (DBSCAN) can be used because such clustering recognizes clusters of arbitrary size and does not require a predefined number of clusters as input. An additional constraint could be maintaining the same NER label for <subject> and <object> elements throughout the cluster. In the implementation, clusters are further partitioned based on NER labels.

[0046] From the acquired clusters, candidate relations can be selected to represent the clusters based on the frequency of their occurrence in the dataset. (Reference) Figure 3 The diagram illustrates an illustrative example of the clustering phase. Relationships can first be clustered using DBSCAN, which clusters them based on either "savings" or "bids". Then, the clusters are further partitioned (if possible) using NER labels for the relationships. For example, the relationship "typical charges" has the NER label "money", and "bidding up to" has the NER label "percentage". Therefore, they can be partitioned. Candidates can then be selected based on the criteria described above.

[0047] Refer again Figure 1 The knowledge graph building component 124 of the knowledge graph infrastructure 112 is configured to build a knowledge graph based at least in part on the relationship between structured data tuples and their corresponding identifiers, the relationship between unstructured data tuples that can be mapped to predefined categories and their corresponding identifiers, and the relationship between at least a portion of the unstructured data tuples and their corresponding assignments.

[0048] Because knowledge graphs are built upon both structured and unstructured data, they are subject to the open-world assumption, which states that any missing tuple is not necessarily false, but simply unknown. Therefore, for completeness, knowledge graphs are enriched using new tuples by inferring any missing tuples based on existing patterns. Thus, knowledge graph enrichment component 126 is configured to enrich the constructed knowledge graph. Enrichment addresses the following issues: extracting missing information, enhancing the confidence values ​​of correct facts, and resolving conflicts in the data.

[0049] Association rule mining is used to extract confidence logic rules based on their support in the knowledge graph. Logic rules of the following form...

[0050]

[0051] Encode frequent correlations in the data. Assign normalized confidence scores to rules that take into account their frequency in the data. Based on these rules, a probabilistic soft logic model can be used with a Markov hinge loss function, which uses the input tuples and the weights of the resulting rules to compute the weights of the inferred tuples. Furthermore, higher confidence scores are assigned to facts supported by multiple rules, thereby reinforcing the correct tuples in the knowledge graph.

[0052] Turn now Figure 4 The diagram illustrates some implementations of this disclosure, showing an exemplary method 400 for constructing an enterprise-specific knowledge graph. As shown in box 410, from at least one enterprise-specific source (e.g., Figure 1 Data repository 116) receives enterprise data (e.g., through... Figure 1 (Data receiving component 118). The first part of the enterprise data is structured data, and the second part of the enterprise data is unstructured data.

[0053] The structured data includes multiple text fields, each with a corresponding value, such that the structured data includes a first plurality of text field / value pairs. As shown in box 412, a relationship is identified corresponding to each text field / value pair in the first plurality of text field / value pairs (e.g., through...). Figure 1 The structured data processing component 120 matches one of several predefined relations included in the predefined relation classification.

[0054] As shown in box 414, a second set of text field / value pairs is extracted from unstructured data (e.g., via...). Figure 1 The unstructured data processing component 122). As shown in box 416, (e.g., via...) Figure 1 The unstructured data processing component 122) determines that each text field / value pair in the first part of the second plurality of text field / value pairs references a relationship that matches one of a plurality of predefined relationships included in a predefined relationship classification. It further determines relationships that each text field / value pair in the second part of the second plurality of text field / value pairs references a relationship that does not match one of a plurality of predefined relationships included in the predefined relationship classification. As shown in box 418, predefined relationships among the plurality of predefined relationships included in the predefined relationship classification that match each text field / value pair in the first part of the second plurality of text field / value pairs are identified. With respect to the second part of the second plurality of text field / value pairs, the referenced relationships are sequentially clustered such that similar relationships are grouped together. This is indicated at box 420. As shown in box 422, assigned relationships are assigned to at least a portion of the text field / value pairs in the second plurality of text field / value pairs based on sequential clustering.

[0055] As shown in box 424, an enterprise-specific knowledge graph is constructed at least in part based on the relationships between a first plurality of text field / value pairs and their corresponding identifiers, the relationships between the first portion of a second plurality of text field / value pairs and their corresponding identifiers, and the relationships between a portion of the second portion of a second plurality of text field / value pairs and their corresponding assignments (e.g., through knowledge graph construction component 124).

[0056] Turn now Figure 5 The diagram illustrates some implementations of this disclosure, showing an exemplary method 500 for constructing an enterprise-specific knowledge graph. As shown in box 510, from at least one enterprise knowledge repository (e.g., Figure 1 Data repository 116) receives (e.g., via) Figure 1 The data receiving component 118) receives enterprise data. The first part of the enterprise data is structured data, and the second part is unstructured data. The structured data includes a first plurality of tuples, each of which includes a <subject> element, a <predicate> element, and a <object> element arranged in the format <subject><predicate><object>. As shown in box 512, a relation corresponding to each of the plurality of structured data tuples is identified (e.g., using structured data processing component 120), which matches one of a plurality of predefined relations included in a predefined relation classification.

[0057] As shown in box 514, a second set of tuples is extracted from unstructured data (e.g., through...). Figure 1 The unstructured data processing component 122). Each of the second plurality of tuples includes an unstructured data tuple <subject> element, an unstructured data tuple <predicate> element, and an unstructured data tuple <object> element arranged in the format <subject>><predicate>><object>.

[0058] As shown in box 516, determine (for example, by...) Figure 1 The unstructured data processing component 122) identifies each unstructured data tuple in the first part of the second plurality of tuples that matches one of the plurality of predefined relations included in the predefined relation classification, and each unstructured data tuple in the second part of the second plurality of tuples that does not match one of the plurality of predefined relations included in the predefined relation classification. As shown in box 518, predefined relations that match each unstructured data tuple in the first part of the second plurality of tuples are identified among the plurality of predefined relations included in the predefined relation classification. As shown in box 520, relations referenced by the second part of the second plurality of tuples are sequentially clustered such that similar relations are grouped together, and as shown in box 522, assigned relations are assigned to at least a portion of the unstructured data tuples of the second plurality of tuples based on sequential clustering.

[0059] As shown in box 524, the structure is constructed at least in part based on the relationship between the structured data tuples of the first plurality of tuples and their corresponding identifiers, the relationship between the unstructured data tuples of the first part of the second plurality of tuples and their corresponding identifiers, and the relationship between at least a portion of the unstructured data tuples of the second part of the second plurality of tuples and their corresponding assignments (e.g., through...). Figure 1 Knowledge graph building component 124) Enterprise-specific knowledge graph. As shown in box 526, it is enriched by adding a third or more tuples that are identified as missing from the enterprise-specific knowledge graph based on existing patterns (e.g., through...). Figure 1 Knowledge graph enrichment component 126) Enterprise-specific knowledge graph.

[0060] refer to Figure 6 This illustrates another exemplary method 600 for constructing an enterprise-specific knowledge graph according to an implementation of this disclosure. As shown in box 610, from at least one source having associated enterprise-specific information (e.g., Figure 1 Data repository 116) receives (e.g., via) Figure 1 The data receiving component 118) receives enterprise data. The first part of the received enterprise data is structured data, and the second part of the received enterprise data is unstructured data.

[0061] Structured data comprises multiple text fields, each corresponding to a structured data tuple. As shown in box 612, structured data is normalized in the following ways (e.g., through...). Figure 1 The structured data processing component 120): (1) identifies the relation corresponding to each structured data tuple; and (2) maps the relation to the relations included in a predefined relation classification.

[0062] As shown in box 614, unstructured data is normalized in the following ways (e.g., by...). Figure 1 The unstructured data processing component 122): (1) extracts multiple unstructured data tuples from unstructured data, each unstructured data tuple having a <subject> element, a <predicate> element and an <object> element, the <subject> element and the <object> element representing entities, and the <predicate> element representing the correlation between the <subject> element and the <object> element; (2) maps the <subject> element and the <object> element of each unstructured data tuple to entity categories included in a predefined entity classification; (3) determines the first part of the <predicate> elements that can be mapped to a predefined relation classification and the second part of the <predicate> elements that cannot be mapped to a predefined relation classification; (4) maps each <predicate> element of the first part of the <predicate> elements to a relation included in the predefined relation classification; (5) sequentially clusters the second part of the <predicate> elements to group similar <predicate> elements together; (6) assigns the assigned relation to at least a part of the second part of the <predicate> elements based on sequential clustering.

[0063] As shown in box 616, an enterprise-specific knowledge graph is constructed at least in part based on the relationships between structured data tuples and their corresponding identifiers, the relationships between unstructured data tuples containing the first part of the <predicate> element and their corresponding identifiers, and the relationships between at least a portion of unstructured data tuples containing the second part of the <predicate> element and their corresponding assignments. (For example, using...) Figure 1 Knowledge graph building components 124).

[0064] Having briefly described an overview of embodiments of the invention, the following describes exemplary operating environments in which at least exemplary embodiments may be implemented, in order to provide a general context for various aspects of the invention. Reference is generally made to the accompanying drawings and initially specifically to... Figure 7 The illustration shows an exemplary operating environment for implementing embodiments of the described techniques, and is generally designated as computing device 700. Computing device 700 is merely one example of a suitable computing environment and is not intended to impose any limitation on the purpose or scope of the embodiments of the invention. Computing device 700 should also not be construed as having any dependency or requirement relating to any of the components or any combination of the components shown.

[0065] This invention can be described in the general context of computer code or machine-usable instructions, including computer-usable or computer-executable instructions, such as program modules, that are executed by a computer or other machine (such as a personal data assistant or other handheld device). Typically, program modules include routines, programs, objects, components, data structures, etc., and / or refer to code that performs a specific task or implements a specific abstract data type. This invention can be implemented in various system configurations, including but not limited to handheld devices, consumer electronics, general-purpose computers, and more specialized computing devices. This invention can also be implemented in distributed computing environments, where tasks are performed by remote processing devices linked via a communication network.

[0066] Continue to refer Figure 7 The computing device 700 includes a bus 710 that directly or indirectly couples to the following devices: memory 712, one or more processors 714, one or more presentation components 716, one or more input / output (I / O) ports 718, one or more I / O components 720, and an illustrative power supply 722. Bus 710 represents something that can be one or more buses (such as an address bus, a data bus, or a combination thereof). Although Figure 7 The various boxes are shown with lines for clarity, but in reality, these boxes represent logical, not necessarily physical, components. For example, a presentation component such as a display device can be considered an I / O component. Furthermore, the processor has memory. The inventors recognize here that this is the nature of the art and reiterate... Figure 7 The figures are merely illustrations of exemplary computing devices that can be used in conjunction with one or more exemplary embodiments of the present invention. No distinction is made between categories such as "workstation," "server," "laptop," and "handheld device," as all of these fall under this category. Figure 7 Within the scope and refer to "Computing Devices".

[0067] Computing device 700 typically includes a variety of computer-readable media. Computer-readable media can be any available medium accessible to computing device 700, and includes volatile and non-volatile media, removable and non-removable media. Computer-readable media includes computer storage media and communication media; computer storage media does not include the signal itself. Computer storage media includes volatile and non-volatile, removable and non-removable media implemented using any method or technology for storing information such as computer-readable instructions, data structures, program modules, or other data. Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, Digital Universal Disc (DVD) or other optical disc storage devices, magnetic tape cassettes, magnetic tape, disk storage devices or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible by computing device 700. On the other hand, communication media contains computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and includes any information transmission medium. The term "modulated data signal" refers to a signal whose one or more characteristics are set or altered in a manner that allows information to be encoded in the signal. By way of example and not limitation, communication media include wired media such as wired networks or direct wired connections, and wireless media such as acoustic, RF, infrared, and other wireless media. Any combination of the above should also be included within the scope of computer-readable media.

[0068] Memory 712 includes computer storage media in the form of volatile and / or non-volatile memory. Memory can be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state memory, hard disk drives, optical disk drives, etc. Computing device 700 includes one or more processors that read data from various entities such as memory 712 or I / O components 720. Multiple presentation components 716 present data indications to a user or other device. Exemplary presentation components include display devices, speakers, printing components, vibration components, etc.

[0069] I / O port 718 allows computing device 700 to be logically coupled to other devices including I / O components 720, some of which may be built-in. Exemplary I / O components include microphones, joysticks, game controllers, satellite antennas, scanners, printers, wireless devices, controllers (such as styluses, keyboards, and mice), natural user interfaces (NUIs), hard / soft buttons, touchscreen displays, etc.

[0070] NUI processes air gestures (i.e., gestures made in the air by one or more parts of the user's body or by a device controlled by the user's body), voice, or other physiological input generated by the user. These inputs can be interpreted as package creation requests, generic package selections, download requests, etc., presented by computing device 700. These requests can be transmitted to appropriate network elements for further processing. NUI implements any combination of voice recognition, touch and stylus recognition, face recognition, biometrics, on-screen and near-screen gesture recognition, air gestures, head and eye tracking, and touch recognition associated with the display on computing device 700. Computing device 700 may be equipped with depth cameras, such as stereo camera systems, infrared camera systems, RGB camera systems, and combinations thereof, for gesture detection and recognition. Additionally, computing device 700 may be equipped with accelerometers or gyroscopes capable of detecting motion. The output of the accelerometer or gyroscope is provided to the display of computing device 700 to present immersive augmented reality or virtual reality.

[0071] It is understood that aspects of the present invention provide systems, methods, and computer storage media for constructing enterprise-specific knowledge bases based on enterprise-specific data including structured and unstructured data. For each of the plurality of tuples included in the structured data, relationships between entities that match relationships in a known relation classification are identified. Where possible, relationships between entities that match relationships in a known relation classification are also identified for tuples included in the unstructured data. If relationships between entities that match relationships in a known relation classification cannot be identified for tuples in the unstructured data, then the extracted relationships are sequentially clustered with similar relationships in other unstructured tuples, and the relationships are assigned to the clustered tuples. An enterprise-specific knowledge graph is constructed based on structured data tuples and their corresponding identified relationships, unstructured data tuples where relationships can be mapped to relationships in a known relation classification and their corresponding identified relationships, and unstructured data tuples that cannot be mapped to relationships in a known relation classification and their corresponding assigned relationships. The knowledge graph is enriched using any information determined to be missing according to a probabilistic method.

[0072] Some specific embodiments of the invention have been described, and these embodiments are illustrative in all respects and not limiting. Alternative embodiments will become apparent to those skilled in the art without departing from the scope of the invention.

[0073] Certain embodiments of the invention are illustrated in the accompanying drawings and have been described in detail above. However, it should be understood that the invention is not intended to be limited to the specific forms disclosed, but rather, it is intended to cover all modifications, alternative constructions, and equivalents that fall within the spirit and scope of the invention.

[0074] Those skilled in the art will understand that Figure 4 Method 400 Figure 5 Method 500 and Figure 6 The order of steps shown in method 600 does not imply any limitation on the scope of the invention, and in fact, these steps may occur in various different orders in embodiments of the invention. Any and all such variations and any combinations thereof are contemplated within the scope of embodiments of the invention.< / tlau> < / tlau> < / iswrittenby> < / product62392>

Claims

1. A computer-implemented method for constructing a knowledge graph, the method comprising: Receive enterprise data from the repository, wherein the first part of the data is structured data and the second part of the enterprise data is unstructured data; Multiple structured data tuples are generated from the structured data. Each structured data tuple includes a structured data tuple <subject> element, a structured data tuple <predicate> element, and a structured data tuple <object> element arranged in the format <subject>><predicate>><object>. Identify the relation corresponding to each of the plurality of structured data tuples, wherein the relation matches one of a plurality of predefined relations included in a predefined relation classification; Multiple unstructured data tuples are extracted from the multiple unstructured data, each unstructured data tuple including a <subject> element, an <predicate> element, and an <object> element arranged in the format of <subject>><predicate>><object>. Determine a relationship in which each unstructured data tuple reference in the first part of the plurality of unstructured data tuples matches one of the plurality of predefined relations included in the predefined relation classification, and a relationship in which each unstructured data tuple reference in the second part of the plurality of unstructured data tuples does not match one of the plurality of predefined relations included in the predefined relation classification; Identify the predefined relation among the plurality of predefined relations included in the predefined relation classification that matches each unstructured data tuple of the first part of the plurality of unstructured data tuples; Clustering the relations referenced by the second part of the plurality of unstructured data tuples, such that similar relations are grouped together; At least a portion of the unstructured data tuples in the second part of the unstructured data tuples are assigned the distribution relationship based on clustering; The knowledge graph is constructed at least in part based on the relationship between the structured data tuples and their corresponding identifiers, the relationship between the unstructured data tuples and their corresponding identifiers in the first part of the plurality of unstructured data tuples, and the relationship between at least a portion of the unstructured data tuples and their corresponding allocations in the second part of the plurality of unstructured data tuples. as well as The knowledge graph is enriched by adding multiple tuples that are identified as missing from the knowledge graph based on existing patterns.

2. The computer-implemented method according to claim 1, wherein within a given structured data tuple, the structured data tuple <subject> element and the structured data tuple <object> element reference entities.

3. The computer-implemented method according to claim 2, wherein the structured data tuple <predicate> element references the correlation between the entities referenced by the structured data tuple <subject> element and the structured data tuple <object> element.

4. The computer-implemented method of claim 3, wherein identifying the relation corresponding to each of the plurality of structured data tuples, the relation matching one of the plurality of predefined relations included in the predefined relation classification, includes: Identify the specific predefined relation among the plurality of predefined relations included in the predefined relation classification that matches the <predicate> element of the structured data tuple for each structured data tuple.

5. The computer-implemented method according to claim 1, wherein within a given unstructured data tuple, the unstructured data tuple <subject> element and the unstructured data tuple <object> element reference entities.

6. The computer-implemented method of claim 5, wherein the unstructured data tuple <predicate> element references the correlation between the entities referenced by the unstructured data tuple <subject> element and the unstructured data tuple <object> element.

7. The computer-implemented method of claim 6, wherein the predefined relation that identifies each unstructured data tuple of the first portion of the plurality of unstructured data tuples among the plurality of predefined relations included in the predefined relation classification includes a specific predefined relation that identifies the <predicate> element of each unstructured data tuple among the plurality of predefined relations included in the predefined relation classification.

8. The computer-implemented method of claim 1, wherein identifying a predefined relation among the plurality of predefined relations included in the predefined relation classification that matches each unstructured data tuple of the first portion of the plurality of unstructured data tuples comprises identifying the predefined relation among the plurality of predefined relations using one or more of semantic mapping, syntactic mapping, and pattern-based mapping.

9. The computer-implemented method of claim 1, wherein at least a portion of the unstructured data tuples for assigning the assigned relation to the second portion of the plurality of unstructured data tuples based on the clustering comprises: Based on the clustering, the assigned relations are attributed to at least a portion of the second part of the unstructured data tuple <predicate> element.

10. One or more computer storage media storing computer-usable instructions, which, when executed by one or more processors, cause the one or more processors to perform a method for constructing a knowledge graph, the method comprising: Receive enterprise data from the repository, wherein the first part of the data is structured data and the second part of the enterprise data is unstructured data; Multiple structured data tuples are generated from the structured data. Each structured data tuple includes a structured data tuple <subject> element, a structured data tuple <predicate> element, and a structured data tuple <object> element arranged in the format <subject>><predicate>><object>. Identify the relation corresponding to each of the plurality of structured data tuples, wherein the relation matches one of a plurality of predefined relations included in a predefined relation classification; Multiple unstructured data tuples are extracted from the multiple unstructured data, each unstructured data tuple including a <subject> element, an <predicate> element, and an <object> element arranged in the format of <subject>><predicate>><object>. Determine a relationship in which each unstructured data tuple reference in the first part of the plurality of unstructured data tuples matches one of the plurality of predefined relations included in the predefined relation classification, and a relationship in which each unstructured data tuple reference in the second part of the plurality of unstructured data tuples does not match one of the plurality of predefined relations included in the predefined relation classification; Identify the predefined relation among the plurality of predefined relations included in the predefined relation classification that matches each unstructured data tuple of the first part of the plurality of unstructured data tuples; Clustering the relations referenced by the second part of the plurality of unstructured data tuples, such that similar relations are grouped together; At least a portion of the unstructured data tuples in the second part of the unstructured data tuples are assigned the distribution relationship based on clustering; The knowledge graph is constructed at least in part based on the relationship between the structured data tuples and their corresponding identifiers, the relationship between the unstructured data tuples and their corresponding identifiers in the first part of the plurality of unstructured data tuples, and the relationship between at least a portion of the unstructured data tuples and their corresponding allocations in the second part of the plurality of unstructured data tuples. as well as The knowledge graph is enriched by adding multiple tuples that are identified as missing from the knowledge graph based on existing patterns.

11. One or more computer storage media according to claim 10, wherein within a given structured data tuple, the structured data tuple <subject> element and the structured data tuple <object> element reference entities, and the structured data tuple <predicate> element references the correlation between the entities referenced by the structured data tuple <subject> element and the structured data tuple <object> element.

12. The one or more computer storage media of claim 11, wherein identifying the relation corresponding to each of the plurality of structured data tuples, the relation matching one of the plurality of predefined relations included in the predefined relation classification, includes identifying a specific predefined relation among the plurality of predefined relations included in the predefined relation classification that matches the structured data tuple <predicate> element of each structured data tuple.

13. One or more computer storage media according to claim 10, wherein within a given unstructured data tuple, the unstructured data tuple <subject> element and the unstructured data tuple <object> element reference entities, and the unstructured data tuple <predicate> element references the correlation between the entities referenced by the unstructured data tuple <subject> element and the unstructured data tuple <object> element.

14. The computer storage medium of claim 13, wherein the predefined relation that identifies each unstructured data tuple of the first portion of the plurality of unstructured data tuples among the plurality of predefined relations included in the predefined relation classification includes a specific predefined relation that identifies the <predicate> element of each unstructured data tuple among the plurality of predefined relations included in the predefined relation classification.

15. The computer storage medium of claim 10, wherein identifying a predefined relation among the plurality of predefined relations included in the predefined relation classification that matches each unstructured data tuple of the first portion of the plurality of unstructured data tuples comprises identifying the predefined relation among the plurality of predefined relations using one or more of semantic mapping, syntactic mapping, and pattern-based mapping.

16. A computer system, comprising: One or more processors; as well as One or more computer storage media storing computer-usable instructions that, when used by one or more processors, cause the one or more processors to: Receive enterprise data from the repository, wherein the first part of the data is structured data and the second part of the enterprise data is unstructured data; Multiple structured data tuples are generated from the structured data. Each structured data tuple includes a structured data tuple <subject> element, a structured data tuple <predicate> element, and a structured data tuple <object> element arranged in the format <subject>><predicate>><object>. Identify the relation corresponding to each of the plurality of structured data tuples, wherein the relation matches one of a plurality of predefined relations included in a predefined relation classification; Multiple unstructured data tuples are extracted from the multiple unstructured data, each unstructured data tuple including a <subject> element, an <predicate> element, and an <object> element arranged in the format of <subject>><predicate>><object>. Determine a relationship in which each unstructured data tuple reference in the first part of the plurality of unstructured data tuples matches one of the plurality of predefined relations included in the predefined relation classification, and a relationship in which each unstructured data tuple reference in the second part of the plurality of unstructured data tuples does not match one of the plurality of predefined relations included in the predefined relation classification; Identify the predefined relation among the plurality of predefined relations included in the predefined relation classification that matches each unstructured data tuple of the first part of the plurality of unstructured data tuples; Clustering the relations referenced by the second part of the plurality of unstructured data tuples, such that similar relations are grouped together; At least a portion of the unstructured data tuples in the second part of the unstructured data tuples are assigned the distribution relationship based on clustering; A knowledge graph is constructed at least in part based on the relationship between the structured data tuples and their corresponding identifiers, the relationship between the unstructured data tuples and their corresponding identifiers in the first part of the plurality of unstructured data tuples, and the relationship between at least a portion of the unstructured data tuples and their corresponding allocations in the second part of the plurality of unstructured data tuples. as well as The knowledge graph is enriched by adding multiple tuples that are identified as missing from the knowledge graph based on existing patterns.

17. The computer system of claim 16, wherein within a given structured data tuple, the structured data tuple <subject> element and the structured data tuple <object> element reference entities, and the structured data tuple <predicate> element references the correlation between the entities referenced by the structured data tuple <subject> element and the structured data tuple <object> element.

18. The computer system of claim 17, wherein identifying the relation corresponding to each of the plurality of structured data tuples, the relation matching one of the plurality of predefined relations included in the predefined relation classification, includes identifying a specific predefined relation among the plurality of predefined relations included in the predefined relation classification that matches the structured data tuple <predicate> element of each structured data tuple.

19. The computer system of claim 16, wherein within a given unstructured data tuple, the unstructured data tuple <subject> element and the unstructured data tuple <object> element reference entities, and the unstructured data tuple <predicate> element references the correlation between the entities referenced by the unstructured data tuple <subject> element and the unstructured data tuple <object> element.

20. The computer system of claim 19, wherein the predefined relation that identifies each unstructured data tuple of the first portion of the plurality of unstructured data tuples among the plurality of predefined relations included in the predefined relation classification includes a specific predefined relation that identifies the <predicate> element of each unstructured data tuple among the plurality of predefined relations included in the predefined relation classification.

Images