Sentence Fusion for Multidocument News Summarization

Abstract

A system that can produce informative summaries, highlighting common information found in many online documents, will help Web users to pinpoint information that they need without extensive reading. In this article, we introduce sentence fusion, a novel text-to-text generation technique for synthesizing common information across documents. Sentence fusion involves bottom-up local multisequence alignment to identify phrases conveying similar information and statistical generation to combine common phrases into a sentence. Sentence fusion moves the summarization field from the use of purely extractive methods to the generation of abstracts that contain sentences not found in any of the input documents and can synthesize information across sources.

Automatic Summarization from Multiple Documents

“…This is particularly useful for people who constantly need to be aware of latest advances in their respective discipline, such as medicine or engineering. This need drives research towards information summarization in domains ranging from news [Nen06,BM05,ROWBG05] to medical [AKS05a,ENW04] and other areas of interest [SYGH05]. What has become obvious during the last decades is the overwhelming volume of information, either being structured or unstructured, appearing in all domains of one's e-life 1 .…”

“…Syntactic analysis , as in [BM05], and grammatical analysis, as in [RJZ89], both offer additional knowledge about tokens (or chunks), which can prove useful in the attempt to reach for semantics. It is very common to apply Part-OfSpeech (POS) tagging to the words of a text, as applied in [CSS05,ZSN05,HMR05], with the aim to retrieve linguistic and functional information about the words of a text.…”

“…In these graphs, units of information (entities and relations in [MB97], sentences in [Mih05]) are interconnected to form a graph based either on collocation (indicative, for instance, of context [MB97]), or on sentence similarity (indicating for instance presumed semantic similarity [Mih05,ER04a]). Trees, which are actually a subconcept of graphs, like the dependency trees in [BM05], can be used as well to represent a sentence, given the output of a syntactic parser and a set of rules. The representation of structured relationships conveys information that has been utilized either for salience indication and selection of sentences, or for reformulation of a summary.…”

“…Information fusion comes in many shapes, ranging from the selection of a set of most prominent elements [FL03] and their ordering, to the addition of outlying information from other elements [MLDBGH04], to sentence fusion [BM05] or paraphrasing (see [ZLMH06] for paraphrasing detection).…”

“…The domains of application of such methodologies are countless, ranging from news summarization [WL03,BM05,ROWBG05] to scientific article summarization [TM02] and meeting summarization [NPDP05, ELH + 03].…”