A Comparison of Model Free versus Model Intensive Approaches to Sentence Compression

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 391–399, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP A Comparison of Model Free versus Model Intensive Approaches to Sentence Compression Tadashi Nomoto National Institute of Japanese Literature 10-3 Midori Tachikawa Tokyo 190-0014 Japan nomoto@acm.org Abstract This work introduces a model free approach to sentence compression, which grew out of ideas from Nomoto (2008), and examines how it com- pares to a state-of-art model intensive approach known as Tree-to-Tree Transducer, or T3 (Cohn and Lapata, 2008). It is found that a model free approach significantly outperforms T3 on the par- ticular data we created from the Internet. We also discuss what might have caused T3’s poor perfor- mance. 1 Introduction While there are a few notable exceptions (Hori and Furui, 2004; Yamagata et al., 2006), it would be safe to say that much of prior research on sen- tence compression has been focusing on what we might call ‘model-intensive approaches,’ where the goal is to mimic human created compressions as faithfully as possible, using probabilistic and/or machine learning techniques (Knight and Marcu, 2002; Riezler et al., 2003; Turner and Charniak, 2005; McDonald, 2006; Clarke and Lapata, 2006; Cohn and Lapata, 2007; Cohn and Lapata, 2008; Cohn and Lapata, 2009). Because of this, the question has never been raised as to whether a model free approach − where the goal is not to model what humans would produce as compres- sion, but to provide compressions just as useful as those created by human − will offer a viable alter- native to model intensive approaches. This is the question we take on in this paper.1 1One caveat would be in order. By model free approach, we mean a particular approach which does not furnish any parameters or weights that one can train on human created compressions. An approach is said to be model-intensive if it does. So as far as the present paper is concerned, we might do equally well with a mention of ‘model free’ (‘model- intensive’) replaced with ‘unsupervised’ (‘supervised’), or ‘non-trainable’ (‘trainable’). An immediate benefit of the model-free ap- proach is that we could free ourselves from the drudgery of collecting gold standard data from hu- mans, which is costly and time-consuming. An- other benefit is intellectual; it opens up an alterna- tive avenue to addressing the problem of sentence compression hitherto under-explored. Also breaking from the tradition of previous re- search on sentence compression, we explore the use of naturally occurring data from the Internet as the gold standard. The present work builds on and takes further an approach called ’Generic Sen- tence Trimmer’ (GST) (Nomoto, 2008), demon- strating along the way that it could be adapted for English with relative ease. (GST was originally intended for Japanese.) In addition, to get a per- spective on where we stand with this approach, we will look at how it fares against a state-of-the-art model intensive approach known as ’Tree-to-Tree Transducer’ (T3) (Cohn, 2008), on the corpus …

Content truncated to 500 words