Research on multi-dimensional end-to-end phrase recognition algorithm based on background knowledge

Currently, the deep end-to-end methods based on supervised learning are widely used in entity recognition and dependency analysis. However, this method has two problems: first of all, it cannot incorporate background knowledge; secondly, it cannot recognize the nested features of multiple granularities in natural language. To solve the problems, we proposed the phrase-window-based annotation rules, and designed the Multi-Dimensional End-to-End Model (MDM). The annotation rules use the phrase as the minimal unit, divide the sentences into seven categories of nestable phrases, and then tag the syntactic dependencies between the phrases. MDM not only supports incorporating background knowledge, and can also recognize both the categories of the nestable phrases in the sentences and the syntactic dependencies between the phrases. The experiment results show that the annotation rules can be used easily without ambiguity; and the MDM is more suitable for processing the multi-granular and diverse features of sentences than the traditional end-to-end algorithms. After incorporating the background knowledge in the experiment on Chinese Phrase Window Dataset (CPWD), the algorithm improved the accuracy of the phrase recognition by more than one percent over the tradition end-to-end methods with the POS dimensions. Also, we applied this algorithm in the CCL 2018 competition and won the first place in the Chinese humor category recognition.