23.4.1 Content Selection in Multi-Document Summarization

In single document summarization we used both supervised and unsupervised methods for content selection. For multiple document summarization supervised training sets are less available, and we focus more on unsupervised methods.

The major difference between the tasks of single document and multiple document summarization is the greater amount of redundancy when we start with multiple documents. A group of documents can have significant overlap in words, phrases, and concepts, in addition to information that might be unique to each article. While we

want each sentence in the summary to be about the topic, we don't want the summary to consist of a set of identical sentences.

For this reason, algorithms for multi-document summarization focus on ways to avoid redundancy when selected sentences for the summary. When adding a new sentence to a list of extracted sentences we need some way to make sure the sentence doesn't overlap too much with the already-extracted sentences.

A simple method of avoiding redundancy is to explicitly include a redundancy factor in the scoring for choosing a sentence to extract. The redundancy factor is based on the similarity between a candidate sentence and the sentences that have already been extracted into the summary; a sentence is penalized if it is too similar to the summary. For example the MMR or Maximal Marginal Relevance scoring system Carbonell and Goldstein (1998), Goldstein et al. (2000) includes the following penalization term for representing the similarity between a sentence s and the set of sentences already extracted for the summary Summary, where $ \lambda $ is a weight that can be tuned and Sim is some similarity function:

$$ \mathrm{M M R~penalization~factor}(s)=\lambda m a x_{s_{i}\in\mathrm{S u m m a r y}}\mathrm{S i m}(s,s_{i}) $$

An alternative to MMR-based method is to instead apply a clustering algorithm to all the sentences in the documents to be summarized to produce a number of clusters of related sentences and then to select a single (centroid) sentence from each cluster into the summary.

By adding MMR or clustering methods for avoiding redundancy, we can also do sentence simplification or compression at the content selection stage rather than at the sentence realization stage. A common way to fit simplification into the architecture is to run various sentence simplification rules (Sec. 23.3.1) on each sentence in the input corpus. The result will be multiple versions of the input sentence, each version with different amounts of simplification. For example, the following sentence:

Former Democratic National Committee finance director Richard Sullivan faced more pointed questioning from Republicans during his second day on the witness stand in the Senate's fund-raising investigation.

might produce different shortened versions:

• Richard Sullivan faced pointed questioning.

• Richard Sullivan faced pointed questioning from Republicans

Richard Sullivan faced pointed questioning from Republicans during day on stand in Senate fundraising investigation

• Richard Sullivan faced pointed questioning from Republicans in Senate fundraising investigation

This expanded corpus is now used as the input to content extraction. Redundancy methods such as clustering or MMR will choose only the (optimally long) single version of each original sentence.