Research On The Alignment Of Video Features And The Descriptive Vocabulary

Video based Meaning Acquisition of Chinese Verb (ViMac-V) is a supervised model which could generate natural languages. ViMac-V takes video-text paired inputs as training samples. Then pre-process for both videos and text is implemented following requests of verb semantic definition based on Frame Semantics. After these steps, verb semantic representation based on video information is constructed.In order to make video-text paired, we need to achieve the alignment of words and visual features. on condition of the correspondence between 4-dimensional characteristics of moving objects and three types of words which mean direction of movement, movement position or velocity correspondence, we need to classify the descriptive vocabulary into three types of words mean direction of movement, movement position or velocity correspondence. And achieve the alignment of words and visual features.Classification of words needs to calculate the similarity between the words. And The current word similarity algorithms are dependent on external knowledge (Hownet, TongYiCi CiLin, etc.), leading to only words within external knowledge can be resolved, or rely solely on one feature of words (minimum edit distance, google distance, etc), the accuracy is insufficient (75%). This paper presents a similarity algorithm based on minimum edit distance in combination of parts of speech, it breaks the limitation of external knowledge, and increased the accuracy of which relies solely on minimum edit distance.in order to achieve a natural language description, We design a language production model based on Bigram. It is able to generate natural languages about the input videos.