JEAPC: A Joint Extraction Model of Action Sequence from Chinese Instructions for Home Service Robot

The language interaction between humans and robots is one of the critical issues in the field of home service robots. In particular, irrelevant information in feature vectors interferes with the extraction task during Chinese instruction parsing. Moreover, the relations between feature vectors of different time steps affect the accuracy of action sequence extraction. In this paper, overlapping action entities in Chinese instructions are labeled through span-based mode, and a Joint Extraction Model of Action Sequences with Partition Coding(JEAPC) is proposed for Chinese instructions. The JEAPC is divided into four modules: BERT, partition encoder, and two decoders. BERT is utilized to obtain the feature vector of each Chinese character in the instructions. The partition encoder is composed of an entity gate, and an action gate, in which the features in the vector are classified, and the irrelevant features are filtered through multi-dimensional vector operations. Furthermore, adversarial training is employed to improve the robustness of JEAPC. Extensive experiments are conducted on a self-built Chinese instructions dataset(FCI) and three entity and relation extraction datasets (CoNLL04, ADE, and SciERC). The experimental results show that the JEAPC can accurately generate action sequences from Chinese instructions and obtain optimal results compared to several competitive approaches.