Distribution of Multi-Words in Chinese and English Documents

As a hybrid of N-gram in natural language processing and collocation in statistical linguistics, multi-word is becoming a hot topic in area of text mining and information retrieval. In this paper, a study concerning distribution of multi-words is carried out to explore a theoretical basis for probabilistic term-weighting scheme. Specifically, the Poisson distribution, zero-inflated binomial distribution, and G-distribution are comparatively studied on a task of predicting probabilities of multi-words’ occurrences using these distributions, for both technical multi-words and nontechnical multi-words. In addition, a rule-based multi-word extraction algorithm is proposed to extract multiwords from texts based on words’ occurring patterns and syntactical structures. Our experimental results demonstrate that G-distribution has the best capability to predict probabilities of frequency of multi-words’ occurrence and the Poisson distribution is comparable to zero-inflated binomial distribution in estimation of multi-word distribution. The outcome of this study validates that burstiness is a universal phenomenon in linguistic count data, which is applicable not only for individual content words but also for multi-words.