Paraphrases and other semantically related sentences present a challenge to NLP and IR applications such as multi-document summarization and question answering systems. While it is generally agreed that paraphrases contain approximately equivalent ideas, they often differ from one another in subtle, yet non-trivial, ways. In this paper, we examine semantic differences in cases of paraphrase and subsumption, in an effort to understand what makes one sentence significantly more informative than another. Using manually annotated data from the news domain, we concentrate on developing a framework for analyzing and comparing pairs of related sentences.