| Automatic summarization is a very useful technique to condense information and make it easy for readers to digest. In this dissertation, we propose several methods to improve performance for several different summarization tasks.;First, for extractive summarization we propose two improvements for the concept-based integer linear programming (ILP) summarization method. One improvement is that we introduce a new measurement for the importance of a language concept and a regression model to estimate it. The other is that we leverage a number of external resources to extract indicative features to help better estimate a concept's weight, and propose to use a joint concept weighting and sentence selection process to train the concepts' feature weights.;Second, we aim to improve abstractive summarization by better sentence compression. We adopt a pipeline abstractive summarization framework where sentence compression is followed by a summary generation component. For sentence compression, we first propose a summary guided sentence compression model, where a sentence is compressed not only considering information in the sentence itself, but also explicitly guided by the summarization goal. We create a new sentence compression corpus for this purpose. Then we propose a discriminative sentence compression model based on expanded constituent parse trees and implement it in the ILP framework where linguistic constraints are incorporated to improve the linguistic quality of the compressed sentences, and thus the final summary.;Third, we adopt the supervised ILP summarization method for the update summarization problem. We investigate different linguistic features for a concept's novelty and salience estimation at both concept and sentence level, and further propose to output more sentences and utilize a sentence reranking component to choose the final summary sentences.;Finally, we apply extractive and abstractive summarization methods to the social media domain. We focus on single news article summarization for a trending topic with the help of Facebook posts that are closely related to that trending topic. We leverage information from the relevant posts at both word and sentence level. Furthermore, we propose a joint summarization and sentence compression model to generate abstractive summaries for the news articles. |
