Semantic analysis and NLP
Semantic analysis and NLP
Despite the significant progress in machine translation, modern solutions still do not always cope with the translation of stable phrases, word play, as well as the choice of suitable cases and the correct construction of sentences.
Contents
A semantic memory module, similar to a knowledge base, consists of pre-prepared GloVe vectors that are used to create sequences of vector representations of words from incoming sentences. These vectors will be used as input to the model.
Data processing
The input module processes question-related input vectors into vector sets called facts. This module is implemented using the Gated Recurrent Unit (hereinafter - GRU), which allows the network to find out the relevance of the proposal in doctranslator.
The question module processes the question word by word and outputs a vector using the same GRU as in the input module, with the same weights.
Data analysis
The episodic memory module stores the fact and question vectors extracted at the input, encoded as attachments. This is similar to the process that takes place in the hippocampus of the brain to extract temporary states in response to sound or sight.
The response module generates a suitable response. In the last step, the episodic memory contains the information necessary for a response. This module uses another GRU, trained with the cross-entropy error classification of correct sequence, which is converted back to natural language.