Improvement of a structured language model: arbori-context tree

In this paper we present an extention of a context tree for a structured language model (SLM), which we call an arbori-context tree. The state-of-the-art SLM predicts the next word from a xed partial tree of the history tree, such as two exposed heads, etc. An arbori-context tree allows us to select an optimum partial tree of a history tree for the next word prediction depending on the e ectiveness in the similar way that a context tree selects the length of the history (n of n-gram). The experiment we conducted showed that the test set perplexity of the SLM based on an arbori-context tree (79.98) was lower than that of the SLM with a xed history (101.56).