# US 10,503,833 B2 · Device and method for natural language processing > Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt - Number: US 10,503,833 B2 - Document ID: US10503833B2 - Status: Granted - Kind: B2 - Application number: 15/498,517 - Filed: 2017-04-27 - Published: 2018-06-07 - Granted: 2019-12-10 - Earliest claimed: 2016-12-06 - Assignees: Siemens Aktiengesellschaft - Applicants: Siemens Aktiengesellschaft - Inventors: Bernt Andrassy; Pankaj Gupta - Web page: https://patentgrep.com/patent/US10503833B2 ## CPC - G06F 40/30: Semantic analysis - G06F 40/295: Named entity recognition - G06N 3/044: Recurrent networks, e.g. Hopfield networks - G06N 3/0442: characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU] - G06N 3/045: Combinations of networks - G06N 3/08: Learning methods - G06N 3/09: Supervised learning - G06N 5/022: Knowledge engineering; Knowledge acquisition - G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N) - G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS ## Abstract A device for relation extraction in a natural language sentence having n words is suggested, the device comprising: a recurrent neural network for joint entity and relation extractions of entities and relations between the entities in the sentence, and an entity-relation table for storing entity labels for the entities and relation labels for the relations, wherein both the entity labels and the relation labels are defined as instances of binary relationships between certain words wi and wj in the sentence, with i∈[1, . . . , n], and j∈[1, . . . , n], wherein each of the entity labels is a first binary relationship for i=j, and wherein each of the relation labels is a second binary relationship for i≠j, wherein the recurrent neural network is configured to fill the cells of the entity-relation table and includes a forward neural network and a backward neural network. ## Claims ### Claim 1 A computing system relation extraction in a natural language sentence having n words, the computing system comprising: a processor for processing the natural language sentence, wherein the processor uses a recurrent neural network for joint entity and relation extractions of entities and relations between the entities in the sentence, and an entity-relation table for storing entity labels for the entities and relation labels for the relations, wherein both the entity labels and the relation labels are defined as instances of binary relationships between certain words wi and wj in the sentence, with i∈[1, . . . , n], and j∈[1, . . . , n], wherein each of the entity labels is a first binary relationship for i=j, and wherein each of the relation labels is a second binary relationship for i≠j, wherein the recurrent neural network is configured to fill the cells of the entity-relation table and includes a forward neural network and a backward neural network, wherein the forward neural network is configured to provide a representation of a current history for a certain word of the sentence, and wherein the backward neural network (120) is configured to provide a representation for a following context of the certain word of the sentence. ### Claim 2 (depends on 1) The computing system of claim 1, wherein the recurrent neural network is a context-aware recurrent neural network framework configured for word-level relation learning via word-pair compositions utilizing their contexts in the sentence. ### Claim 3 (depends on 1) The computing system of claim 1, wherein the recurrent neural network is configured to model interdependencies in entity and relation extraction tasks for the joint entity and relation extractions of the entities and the relations in the sentence. ### Claim 4 (depends on 3) The computing system of claim 3, wherein the recurrent neural network is configured to use a piggybacking approach for modeling relation to entity dependencies as a part of the interdependencies in the entity and relation extraction tasks. ### Claim 5 (depends on 4) The computing system of claim 4, wherein the piggybacking approach includes that, if a relation label for a relation between two entities of the sentence is increased, each entity label of said two entities is also increased. ### Claim 6 (depends on 4) The computing system of claim 4, wherein the piggybacking approach is configured to use a deep-learning approach for modeling the interdependencies in the entity and relation extraction tasks. ### Claim 7 (depends on 3) The computing system of claim 3, further comprising a deriving entity for deriving a list of possible candidate entities for each word of the sentence participating in one of the relations of the sentence, wherein the list includes a count vector for each of the candidate entities. ### Claim 8 (depends on 7) The computing system of claim 7, further comprising a learning entity for piggybacking the count vectors of the candidate entities to enable learning relation to entity dependencies in order to improve the extractions of the entities in the sentence. ### Claim 9 (depends on 1) The computing system of claim 1, wherein the forward neural network and the backward neural network form a bidirectional architecture. ### Claim 10 (depends on 1) The computing system of claim 1, wherein the recurrent neural network is configured to fill the cells of the entity-relation table according to a certain table-filling rule. ### Claim 11 (depends on 1) The computing system of claim 1, wherein the recurrent neural network includes a number of hidden layers and a third direction neural network for providing a representation of a middle context of the sentence between a word wi provided to the hidden layers by the forward neural network and a word wj provided to the hidden layers by the backward neural network. ### Claim 12 A method for relation extraction in a natural language sentence having n words, the method comprising: providing a recurrent neural network for joint entity and relation extractions of entities and relations between the entities in the sentence, wherein the recurrent neural network is configured to fill the cells of an entity-relation table and includes a forward neural network and a backward neural network, wherein the forward neural network is configured to provide a representation of a current history (hfi) for a certain word of the sentence, and wherein the backward neural network is configured to provide a representation for a following context of the certain word of the sentence, providing an entity-relation table for storing entity labels for the entities and relation labels for the relations, wherein both the entity labels and the relation labels are defined as instances of binary relationships between certain words wi and wj in the sentence, with i∈[1, . . . , n], and j∈[1, . . . , n], wherein each of the entity labels is a first binary relationship for i=j, and wherein each of the relation labels is a second binary relationship for i≠j, and filling the cells of the provided entity-relation table using the provided recurrent neural network. ### Claim 13 (depends on 12) The method of claim 12, wherein the recurrent neural network is a context-aware recurrent neural network framework configured for word-level relation learning via word-pair compositions utilizing their contexts in the sentence. ### Claim 14 (depends on 12) The method of claim 12, wherein the recurrent neural network is configured to model interdependencies in entity and relation extraction tasks for the joint entity and relation extractions of the entities and the relations in the sentence. ### Claim 15 (depends on 14) The method of claim 14, wherein the recurrent neural network is configured to use a piggybacking approach for modeling relation to entity dependencies as a part of the interdependencies in the entity and relation extraction tasks. ### Claim 16 (depends on 15) The method of claim 15, wherein the piggybacking approach includes that, if a relation label for a relation between two entities of the sentence is increased, each entity label of said two entities is also increased. ### Claim 17 (depends on 15) The method of claim 15, wherein the piggybacking approach is configured to use a deep-learning approach for modeling the interdependencies in the entity and relation extraction tasks. ### Claim 18 (depends on 12) The method of claim 12, wherein the forward neural network and the backward neural network form a bidirectional architecture. ### Claim 19 (depends on 12) The method of claim 12, wherein the recurrent neural network includes a number of hidden layers and a third direction neural network for providing a representation of a middle context of the sentence between a word wi provided to the hidden layers by the forward neural network and a word wj provided to the hidden layers by the backward neural network. ### Claim 20 A computer program product, comprising a computer readable hardware storage device storing a computer readable program code, the computer readable program code comprising an algorithm that when executed by a computer processor of a computing system implements a method for relation extraction in a natural language sentence having n words, the method comprising: providing a recurrent neural network for joint entity and relation extractions of entities and relations between the entities in the sentence, wherein the recurrent neural network is configured to fill the cells of a entity-relation table and includes a forward neural network and a backward neural network, wherein the forward neural network is configured to provide a representation of a current history (hfi.) for a certain word of the sentence, and wherein the backward neural network is configured to provide a representation for a following context of the certain word of the sentence, providing an entity-relation table for storing entity labels for the entities and relation labels for the relations, wherein both the entity labels and the relation labels are defined as instances of binary relationships between certain words wi and wj in the sentence, with i∈[1, . . . , n], and j∈[1, . . . , n], wherein each of the entity labels is a first binary relationship for i=j, and wherein each of the relation labels is a second binary relationship for i≠j, and filling the cells of the provided entity-relation table using the provided recurrent neural network.