The session took place at the Indisch Herinneringscentrum, whose staff had been most helpful in putting us into contact with representatives of this user group. We were received with a lot of enthusiasm and already during the introduction round several ideas popped up on how to best share memories with a larger audience. We also got to see a real example of a teacher/artist who travels to schools with a suitcase that she decorated and filled with photos and items from her Dutch-Indonesian family and past. Through this suitcase, she tells her story and hopes to get people interested in the history so they will seek out more information about it. Some of the other persons present have tried to find out as much as they could about their families, but preferred to keep this within the setting of their family.

We also had in-depth discussions with the participants on how they see the interaction between online and offline storytelling for the purpose of increasing historical awareness and how Agora could support this. There were also some critical remarks on the fact that every platform will inevitably suffer from  a bias in presented perspectives and how to find a balance between inputs to the platform from different parties (experts vs. lay people, victims vs. instigators). Another issue that was raised was that professionals who would use a platform to find information and create links would very much like to keep this to themselves, at least until their work is finished and published. Lots of food for thought still.

Overall the participants were positive about the direction of the Agora platform and the proposed ideas.

We are currently working on analysing the interviews, and incorporating our findings into our future work.

There was quite a bit of discussion about the legal perspectives of sharing data, protecting the people who had for example contributed to resources (in the case of research into endangered language for example, one cannot make the sources or transcripts available without giving up anonymity of the interviewees, something that is specific to the domain). Also, some of the projects were well on their way to creating linked data, but were unsure how to make the last few steps. I think therefore it was really good to have a mix of (enthusiastic) computer scientists and linguists present.

Concretely, we have agreed on trying to have follow-up workshops at both linguistics and semantic web venues, and collaborate within a working group.

View more presentations from Agora Project

This presentation is part of a series of student presentations on their MSc thesis design. This is part of the joint VU and UvA Information Systems MSc program.

Towards an Efficient Combination of Similarity Measures for Semantic Relation Extraction ~ Alexander Panchenko
In this talk, the differences in evaluation measures for semantic relations are investigated. There are several measures around and each provides a different type of semantic information. Panchenko tried to combine several measures to provide a more rounded evaluation.

Joint Learning of dependency parsing and semantic role labelling ~ Antal van den Bosch, Roser Morante and Sander Canisius
The problem with a lot of NLP pipeline is that modules work sequentially, and errors from the one module permeate through to the others. In this talk two task that are generally performed sequentially, namely dependency parsing and semantic role labelling are combined in order to prevent the errors cascading. Their findings indicate that dependency parsing works better separately, whereas semantic role labelling actually benefits from combining it. I think this is an interesting way to think about problems, and perhaps it would also be interesting for the Agora setting where we start dealing with roles.

Defining modality categories for NLP ~ Roser Morante
In this talk ongoing work on an annotation scheme for modality was presented based on older theories and recent models. For now this is not really on the Agora radar, but it will become more important as we drill down to the finer grained information about events, to find instances in texts of events that may have happened but didn’t, such as the assassination of Hitler or mythical events. Currently we cannot distinguish between things that people have only talked about and things that actually happened, but we should probably do something with this. Also, mythological ‘events’ did not actually take place and their actors are not real persons, but they are depicted in paintings, so we should not do away with them entirely, but figure out how to detect and represent these properly.

Evaluating DAISY Summarisation with Approximated Summaries ~ Mandy Schiffrin, Fabrice Nauze and Begoña Villada
Summarisation is a difficult task to evaluate, as there is not one correct possible summary. Here an approach was presented on a modified pyramid evaluation and rouge evaluation where automatically generated summaries were evaluated with queries. A similar situation arises in Agora where events descriptions can have similar configurations, sometimes one can refer to an event by mentioning its most important actors, and sometimes more fine grained, we are currently working on figuring out what to represent where, and such evaluation measures may come in handy there.

Mining Cultural Heritage Metadata ~ Kalliopi Zervanou, Ioannis Korkontzelos, Antal van den Bosch and Sophia Anadiadou
From our sister project, HiTime, work was presented about extracting information from object descriptions from archive documents. This is something that one of our master’s students is currently working on, as opening up the treasures from is something we hadn’t gotten round to yet. The work of Zervanou et al. focuses on an analysis of the languages in the archive they use, and they only extract named entities (person, location and organisation) and subjects and other, whereas we are trying to extract some more types of information, but the lessons they learnt from their work (see also Zervanou, K., Korkontzelos, I, Van den Bosch, A., and Ananiadou, S. (2011). Enrichment and structuring of archival description metadata. In K. Zervanou and P. Lendvai (Eds.), Proceedings of the Fifth ACL-HLT Workshop on Language Technology for Cultural Heritage, Social Sciences, and Humanities (LaTeCH-2011), Portland, OR, pp. 44-53. [pdf]) will also come in handy, in particular the way they employed the domain experts’ knowledge to help evaluate their approach.

Final words: topic detection in suicide notes ~ Bart Desmet and Véronique Hoste
Desmet presented work on emotion (topic) detection and classification in a very special text corpus: suicide notes. A corpus of 900 suicide notes was annotated with 13 emotions such as anger, pity and peacefullness. For each emotion, a SVM was trained to detect the best features. Results show that half of the emotion types in the testcorpus was difficult to detect, due to data-sparseness. For the long run emotion detection in documents is interesting for Agora as it adds view and polarity pertaining to the events reported in a document.

Evaluating cross-domain Dutch semantic role annotation ~ Orphée De Clercq, Véronique Hoste and Paola Monachesi
In this talk, work on semantic role labeling (SRL) in the context of the SONAR corpus was presented. A manually verified subset was used to train a labeler that could perform SRL on other parts of the SONAR corpus. Training was performed for in-domain text and combinations of text coming from different domains. Genre specific training proves crucial for a optimal performance.

Event prediction through Social Network Analysis ~ Matje van de Camp
Matje van de Camp, from the sister project HITIME, presented ongoing research on the creation of a time-stamped social network based on biographies of Dutch people related to the socialist movement. Ultimately, the social network will be employed to predict events based on sudden bursts of activity in the network. Hitime and Agora already started a collaboration to share results and techniques.

Computational Linguistics in public safety and security ~
Wauter Bosma discussed three NLP related issues that play an important role in diverse projects the NFI (Dutch Forensic Institute). Specific areas of interest are 1) deviations in processes and events that may point to fraud, 2) network analysis (e.g. constructing criminal networks from various documents and reports) and 3) author identification and disambiguation.

DutchSemCor: from manual annotation to active learning ~
In this talk, DutchSemcor was presented, a semantically annotated corpus based on Cornetto, a lexico-semantic database. For the most frequent and polysemous part of the Dutch language, at least 25 example sentences are annotated manually for each sense. Sentences come from the SONAR corpus and additional webqueries using TextCorp. Next, different WSD systems are trained on these annotations: a KB-based WSD and TiMBL. The DutchSemCor will be available medio 2012.

Unfortunately the speaker presenting on detection and classification of event nominals could not make it, but we will contact her. We keep you posted!

For our session, we have invited colleagues from the Amsterdam Museum, Netherlands Institute for Sound and Vision and the Rijksmuseum to talk about their experiences with opening up their (digital) collections to the public, whilst preserving the context and their collaborations with the public in making their collections better. The session is concluded by a panel.

The session programme:
Chair: Susan Legêne – VU University Amsterdam
14:00 – 14:10: Introduction by chair Susan Legêne (VU University Amsterdam)
14:10 – 14:30: Judith van Gent (Amsterdam Museum) and Victor de Boer (VU University Amsterdam) – Amsterdam Museum Linked Data
14:30 – 14:50: Lotte Belice Baltussen and Johan Oomen (Netherlands Institute for Sound and Vision) – Crowdsourcing
14:50 – 15:05: Geertje Jacobs (Rijksmuseum Amsterdam) – the Rijksmuseum API
15.05 – 15:20 Marieke van Erp and Lora Aroyo (VU University Amsterdam) – the Agora project
15:20 – 15:55: Panel, led by Susan Legêne (VU University Amsterdam) – will include Geertje Jacobs, Judith van Gent, Lora Aroyo and Johan Oomen.
15:55 – 16:00: Closing remarks

You can still register for DISH at http://www.dish2011.nl/tickets