Multilingual Summarization: Dimensionality Reduction and a Step Towards Optimal Term Coverage

In this paper we present three term weighting approaches for multi-lingual document summarization and give results on the DUC 2002 data as well as on the 2013 Multilingual Wikipedia feature articles data set. We introduce a new intervalbounded nonnegative matrix factorization. We use this new method, latent semantic analysis (LSA), and latent Dirichlet allocation (LDA) to give three term-weighting methods for multi-document multi-lingual summarization. Results on DUC and TAC data, as well as on the MultiLing 2013 data, demonstrate that these methods are very promising, since they achieve oracle coverage scores in the range of humans for 6 of the 10 test languages. Finally, we present three term weighting approaches for the MultiLing13 single document summarization task on the Wikipedia featured articles. Our submissions significantly outperformed the baseline in 19 out of 41 languages. 1 Our Approach to Single and Multi-Document Summarization The past 20 years of research have yielded a bounty of successful methods for single document summarization (SDS) and multi-document summarization (MDS). Techniques from statistics, machine learning, numerical optimization, graph theory, and combinatorics are generally languageindependent and have been applied both to single and multi-document extractive summarization of multi-lingual data. In this paper we extend the work of our research group, most recently discussed in Davis et al. (2012) for multi-document summarization, and apply it to both single and multi-document multilingual document summarization. Our extractive multi-document summarization performs the following steps: 1. Sentence boundary detection; 2. Tokenization and term identification; 3. Term-sentence matrix generation; 4. Term weight determination; 5. Sentence selection; 6. Sentence ordering. Sentence boundary detection and tokenization are language dependent, while steps (3)-(6) are language independent. We briefly discuss each of these steps. We use a rule based sentence splitter FASSTE (very Fast, very Accurate Sentence Splitter for Text – English) (Conroy et al., 2009) and its multilingual extensions (Conroy et al., 2011) for determining the boundary of individual sentences. Proper tokenization improves the quality of the summary and may include stemming and also morphological analysis to disambiguate compound words in languages such as Arabic. Tokenization may also include stop word removal. The result of this step is that each sentence is represented as a sequence of terms, where a term can be a single word, a sequence of words, or character n-grams. The specifics of tokenization are discussed in Section 2. Matrix generation (the vector space model) was pioneered by Salton (1991). Later Dumais (1994) introduced dimensionality reduction in document retrieval systems, and this approach has also been