Rule Induction for Sentence Reduction

The field of Automatic Sentence Reduction has been an active research topic, with several relevant approaches being recently proposed. However, in our view many milestones still need to be reached in order to approach human-like quality sentence simplification. In this work, we propose a new framework, which processes huge sets of web news stories and learns sentence reduction rules in a fully automated and unsupervised way. This is our main contribution. Our system is conceptually composed of several modules. In the first one, the system automatically extracts paraphrases from on-line news stories, using new lexically based functions that we have proposed. In our system's second module, the extracted paraphrases are transformed into aligned paraphrases, meaning that the two paraphrasic sentences get their words aligned through DNA-like sequence alignment algorithms, that has been conveniently adapted for aligning sequences of words. These alignments are then explored and specific text structures called bubbles are selected. Afterwards, these structures are transformed into learning instances and used in the last learning module that exploits techniques of Inductive Logic Programming. This module learns the rules for sentence reduction. Results show that this is a good approach for learning automatic sentence reduction, while some pertinent issues still need future investigation.