Recent Advances in Computational Linguistics

ive summarization approaches use information extraction, ontological information, information fusion, and compression. Automatically generated abstracts (abstractive summaries) 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. An abstract contains at least some sentences (or phrases) that do not exist in the original document. Of course, true abstraction involves taking the process one step further. Abstraction involves recognizing that a set of extracted passages together constitute something new, something that is not explicitly mentioned in the source, and then replacing them in the summary with the new concepts. The requirement that the new material not be in the text explicitly means that the system must have access to external information of some kind, such as an ontology or a knowledge base, and be able to perform combinatory inference. Recently, Ledeneva et al. [Led08a, Led08b, Led08c] and Garcia et al. [Gar08a, Gar08b, Gar09] have successfully employed the word sequences from the selftext for detecting the candidate text fragments for composing the summary. Ledeneva et al. [Led08a] suggest a typical automatic extractive summarization approach composed by term selection, term weighting, sentence weighting and sentence selection steps. One of the ways to select the appropriate sentences is to assign some numerical measure of usefulness of a sentence for the summary and then select the best ones; the process of assigning these usefulness weights is called sentence weighting. One of the ways to estimate the usefulness of a sentence is to sum up usefulness weights of individual terms of which the sentence consists; the process of estimating the individual terms is called term weighting. For this, one should decide what the terms are: for example, they can be words; deciding what objects will count as terms is the task of term selection. Different extractive summarization methods can be characterized by how they perform these tasks. Ledeneva et al. [Led08a, Led08b, Led08c] has proposed to extract all the frequent grams from the selftext, but she only considers those that are not contained (as subsequence) in other frequent grams (maximal frequent word sequences). In comparison with n-grams, the Maximal Frequent Sequences (MFS) are attractive for extractive text summarization since it is not necessary to define the gram size (n), it means, the length of each RECENT ADVANCES IN CL Informatica 34 (2010) 501–517 511 MFS is determined by the self-text. Moreover, the set of all extracted MFSs is a compact representation all frequent word sequences, reducing in this way the dimensionality in a vector space model. Garcia et al. [Gar08b, Gar09] have extracted all the sequences of n words (n-grams) from the self-text as features of its model. In this work, we evaluate the ngrams and maximal frequent sequences as domainand languageindependent models for automatic text summarization. In this work, sentences were extracted using unsupervised learning approach. Some other methods are also developed for abstractive summarization. For example, techniques of sentence fusion [Dau04, Bar03, Bar05], information fusion [Bar99], sentence compression [Van04, Mad07], headline summarization [Sar05], etc. 4.3.3 Recent applications of text