As more information becomes available online, information-overloading results. This problem can be resolved through the application of automatic summarization. Traditional summarization models consider a document as a sequence of sentences. Actually, a large document has a well-defined hierarchical structure. Human abstractors use the hierarchical structure of the document to extract topic sentences. They start searching for topic sentences from the top level of the document structure downwards. Similarly, hierarchical summarization generates a summary for a document based on the hierarchical structure and salient features of the document. User evaluations that have been conducted indicate that hierarchical summarization outperforms traditional summarization.
In general, automatic summarization is represented by a three-stage framework, that is, representation of source document, extraction of information, and generation of summary (Sparck-Jones, 1999). Most of the current research work focuses on the second stage. Traditionally, the summarization system calculates the significance of sentences to the document based on the salient features of the document (Edmundson, 1969; Luhn, 1958). The most significant sentences are then extracted and concatenated as a summary. The compression ratio of the summary can be adjusted to specify the amount of information to be extracted. A lot of extraction features have been proposed.
The extraction approaches are usually classified into three major groups according to the level of processing in the linguistic space (Mani & Maybury, 1999). The surface-level approaches use salient features of a document to extract the important information. The entity-level approaches build an internal representation for text units and their relationships, and use graph theories to determine the significance of units. The discourse-level approaches model the global structure of the text, and the text units are extracted based on the structure. Generally, the deeper approaches are more promising to give more informative summaries. However, the surface-level approaches are proved to be robust and reliable (Goldstein, Kantrowitz, Mittal, & Carbonell 1999). They are still widely adopted at present.
The summarization systems can be evaluated either by intrinsic or extrinsic evaluation (Sparck-Jones & Galliers, 1996). The intrinsic evaluation judges the quality of the summarization by direct analysis of the summary (Kupiec, Pedersen, & Chen, 1995). The extrinsic evaluation judges the quality of the summarization based on how it affects the completion of some other tasks (Morris, Kasper, & Adams, 1992). A number of general-purpose summarization systems have been developed. Experiments have been conducted on these systems. All the systems identify an upper bound for the precision of the summarization system, the performance of the system grows fast with addition of extraction features, and they reach their upper bound after three or four extraction features (Kupiec et al., 1995).
Key Terms in this Chapter
Extrinsic Evaluation: Summarization evaluation methods which judge the quality of the summaries based on how they affect the completion of some other tasks.
Thematic Feature: An extraction feature which calculates the significance of a term to a document based on the properties of the terms.
Extraction Feature: Most summarization models use salient features of a document to extract the important information content. There are many features identified as key features, including thematic feature, location feature, background feature, cue feature, and so forth.
Automatic Summarization: A technique where a computer program summarizes a text. The existing automatic text summarization is mainly the selection of sentences from the source document based on their significance to the document using statistical techniques and linguistic analyses.
Heading Feature: An extraction feature which calculates the significance of a sentence based on the presence of heading or subheading words in the sentence.
Location Feature: An extraction feature which calculates the significance of a sentence based on the position of the sentence within the document.
Intrinsic Evaluation: Summarization evaluation methods which judge the quality of summaries by direct analyses in terms of some set of norms.
Cue Feature: An extraction feature which calculates the significance of a sentence based on the presence of some pragmatic words.
Hierarchical Summarization: A document exhibits a well-defined hierarchical structure. Hierarchical summarization extracts the important information from the source document by exploring the hierarchical structure and salient features of the document.