Natural language research is often equated with attempts to derive structure and meaning from unstructured data. Given this focus, computational treatment of style has largely remained unexplored. In this talk I will argue that elements of style such as the redundancy in text, the level of specificity or its entertaining effect, affect the performance of standard systems and that good approaches to computational style analysis will be beneficial for such systems. 

I will first briefly present our findings from a corpus study of local coherence motivating the need for style analysis. Theories of discourse relations and entity coherence fail to explain local coherence for a large portion of newspaper text, with stylistic factors often involved in the cases where standard theories do not apply. 

Next I will present our work on developing measures of one particular element of style, text specificity. I will discuss how we successfully developed a classifier for sentence specificity. The classifier allows us to analyze specificity in summaries produced by people and machines and reveals that machine summaries are overly specific. Furthermore analysis of sentence compression data shows that when summarizing people often edit a specific sentence in the source into general one for the summary, indicating that specificity is a suitable objective for compression systems that will naturally lead to the need for compression. 

I will conclude with a brief overview of other style-related tasks which affect the performance of summarization systems.