The objective of this research is to generate a coherent and understandable text in Chinese. We extract commonsense knowledge from ConceptNet automatically and select concepts by Monte-Carlo Tree Search (MCTS) algorithm.
Combine text by emplates and use a constructed word embedding model and a Deep Neural Network (DNN) of discourse coherence model as a reward function in MCTS to evaluate the coherent score of generated text. Evaluate generated text by human rating, and the result shows that it is more coherent when using the discourse coherence model.
Employee 30 web users to rate the paragraphs from 1 to 6 according to coherence, fluency and correctness. The text with lower score is less coherent, and the one with higher score is more coherent. We take the human-written texts as gold standards , and normalize the scores from 1 to 5 (the second row of table).
Here are some examples of generated texts in the human evaluation dataset.
We can see that text without discourse coherence model (example C) is less coherent than using it (example B).
Code and detailed information are in my Github
Comments