Processing devices, methods, apparatuses, and media for performing natural language processing tasks

Abstract

This application discloses a processing apparatus, method, device, and medium for performing natural language processing tasks, relating to the field of computer technology. The processing apparatus is integrated into a neural network accelerator and includes: a memory for storing multiple integer input data; a lookup circuit for searching for reference data from the multiple integer input data; the integer input data being attention scores generated based on an attention mechanism in the natural language processing task; a shift circuit for generating offset differences between each integer input data and the reference data, and performing shift operations on each offset difference based on shift control parameters; and an output circuit for outputting attention weights based on the shift results, so as to perform weighted processing on semantic features in the natural language processing task based on the attention weights. It is evident that this application effectively reduces computational complexity, reduces hardware resource consumption, reduces memory bandwidth pressure and access latency, and can flexibly adapt to accuracy and efficiency requirements through preset shift control parameters.

Images