The CIST Summarization System at TAC 2011

In this report, we present our extractive summarization system on both summarization and multiling tracks of TAC 2011. We introduce an extractive multi-document summarization method based on hierarchical topic model of hierarchical Latent Dirichlet Allocation (hLDA) and sentence compression. hLDA is a representative generative probabilistic model, which not only can mine latent topics from a large amount of discrete text data, but also can organize these topics into a hierarchy to achieve a deeper semantic analysis. We try to combine the hLDA model with some traditional features. The evaluation results showed some improvement compared with our own system in TAC 2010, which is based on sentence clustering. But there are still many problems needed to be studied in the future. As to the new multiling task of TAC 2011, we used the frame of the hLDA model again but deleting those knowledge base for English. We tried all the 7 languages, including Arabic, Czech, English, French, Hebrew, Hindi and Greek. The evaluations of human confirmed that our method has better performance than some other ones.