Aiming at problem that danmaku are short

irregular in nature and mixed with alphanumerical characters

which cause it is hard to handle with traditional new word detection methods. On the basis of previous methods based on internal cohesion and boundary freedom of words

the concept of boundary-boosted context entropy is introduced

making context entropy compatible with words mostly appearing at sentence fragment boundaries. Experimental results on a danmaku corpus collected from a well-known live-streaming service show effectiveness despite model simplicity

adaptability to differently sized input by means of parameter tuning

time and spatial complexity are also good

as is execution performance.