Device and method for determining a knowledge graph

A training device for determining a knowledge graph, the device comprising at least one processor configured to:

• determine, using a function, a respective second embedding for each of a plurality of respective first embeddings for a word, wherein each of the plurality of respective first embeddings of the word is represented by a respective vector and each of the plurality of respective first embeddings for the word being a different type of embedding, each type of embedding being selected from:
• a character embedding, a Byte Pair Encoding embedding, a FastText embedding, a FLAIR embedding, a BERT embedding, a domain-specific embedding, and wherein each of the respective second embeddings is determined by the function mapping the respective first embedding to the respective second embedding using a respective image;
• determine based on a weighted combination of the respective second embeddings, using a first classifier, a first classification for the word, the first classification defining whether or not the word is an entity for the knowledge graph, and when the word is an entity, the first classification defining to which type of entity for the knowledge graph the word in the knowledge graph is to be assigned;
• determine for each of the respective second embeddings, using a discriminator, a second classification which defines to which type of embedding from the different types of embeddings of the respective first embeddings, the second embedding is to be assigned;
• train at least one parameter for the function in a training as a function of at least one gradient for a training of the first classifier and as a function of at least one gradient for a training of the discriminator.

A non-transitory computer-readable medium on which is stored a computer program including computer-readable instructions for training, the instructions, when executed by a computer, causing the computer to perform the following steps:

• determining, using a function, a respective second embedding for each of a plurality of respective first embeddings for a word, wherein each of the plurality of respective first embeddings of the word is represented by a respective vector and each of the plurality of respective first embeddings for the word being a different type of embedding, each type of embedding being selected from:
• a character embedding, a Byte Pair Encoding embedding, a FastText embedding, a FLAIR embedding, a BERT embedding, a domain-specific embedding, and wherein each of the respective second embeddings is determined by the function mapping the respective first embedding to the respective second embedding using a respective image;
• determining based on a weighted combination of the respective second embeddings, using a first classifier, a first classification for the word, the first classification defining whether or not the word is an entity for the knowledge graph, and when the word is an entity, the first classification defining to which type of entity for the knowledge graph the word in the knowledge graph is to be assigned;
• determining for each of the respective second embeddings, using a discriminator, a second classification which defines to which type of embedding from the different types of embeddings of the respective first embeddings, the second embedding is to be assigned;
• training at least one parameter for the function in a training as a function of at least one gradient for a training of the first classifier and as a function of at least one gradient for a training of the discriminator.