Unsupervised Learning Summarization Templates from Concise Summaries

Proceedings of NAACL-HLT 2013, pages 270–279, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Unsupervised Learning Summarization Templates from Concise Summaries Horacio Saggion∗ Universitat Pompeu Fabra Department of Information and Communication Technologies TALN Group C/Tanger 122 - Campus de la Comunicaci´on Barcelona - 08018 Spain http://www.dtic.upf.edu/˜hsaggion/ Abstract We here present and compare two unsuper- vised approaches for inducing the main con- ceptual information in rather stereotypical summaries in two different languages. We evaluate the two approaches in two differ- ent information extraction settings: mono- lingual and cross-lingual information extrac- tion. The extraction systems are trained on auto-annotated summaries (containing the in- duced concepts) and evaluated on human- annotated documents. Extraction results are promising, being close in performance to those achieved when the system is trained on human-annotated summaries. 1 Introduction Information Extraction (Piskorski and Yangarber, 2013) and Automatic Text Summarization (Saggion and Poibeau, 2013) are two Natural Language Pro- cessing tasks which require domain and language adaptation. For over two decades (Riloff, 1993; Riloff, 1996) the natural language processing com- munity has been interested in automatic or semi- automatic methods which could be used to port sys- tems from one domain or task to another, aiming at reducing at least in part the cost associated with the creation of human annotated datasets. Automatic system adaptation can take different forms: if high ∗This work is partially supported by Ministerio de Econom´ıa y Competitividad, Secretar´ıa de Estado de Investigaci´on, De- sarrollo e Innovaci´on, Spain under project number TIN2012- 38584-C06-03 and Advanced Research Fellowship RYC-2009- 04291. We thank Biljana Drndarevi´c for proofreading the paper. quality human annotated data is available, then rule- based or statistical systems can be trained on this data (Brill, 1994), reducing the efforts of writing rules and handcrafting dictionaries. If high quality human annotated data is unavailable, a large non- annotated corpus and a bootstrapping procedure can be used to produce annotated data (Ciravegna and Wilks, 2003; Yangarber, 2003). Here, we concen- trate on developing and evaluating automatic proce- dures to learn the main concepts of a domain and at the same time auto-annotate texts so that they be- come available for training information extraction or text summarization applications. However, it would be naive to think that in the current state of the art we would be able to learn all knowledge from text au- tomatically (Poon and Domingos, 2010; Biemann, 2005; Buitelaar and Magnini, 2005). We therefore here concentrate on learning template-like represen- tations from concise event summaries which should contain the key information of an event. 18 de julio de 1994DateOfAttack . Un atentado contra la sede de la Asociaci´on Mutual Israelita ArgentinaTarget de Buenos AiresPlaceOfAttack causa la muerte de 86NumberOfVictims personas. (18th July 1994. An attack against the headquarters of the Jewish Mutual Association in Buenos Aires, Ar- gentina, kills 86 people.) Figure 1: Sample of Human Annotated Summary in Spanish An example of the summaries we want to learn from is presented in Figure 1. It is a summary in the terrorist attack …

Content truncated to 500 words