Natural Language Generation Using Monte Carlo Tree Search

We propose a method of simulation-based natural language generation that accounts for both building a correct syntactic structure and reflecting the given situational information as input for the generated sentence. We employ the Monte Carlo tree search for this nontrivial search problem in simulation, using context-free grammar rules as search operators. We evaluated numerous generation results from two aspects: the appropriateness of sentence contents for the given input information and the sequence of words in a generated sentence. Furthermore, in order to realize an efficient search in simulation, we introduced procedures to unfold syntactic structures from words strongly related to the given situational information, and increased the probability of selecting those related words. Through a numbers of experiments, we confirmed that our method can effectively generate a sentence with various words and phrasings.