The 2016 Shared Task was concluded with a Workshop held at KONVENS 2016 on Thursday, September 22nd at Ruhr University, Bochum. To learn more about its results, please take a look at the proceedings.
Accurate opinion mining requires the exact identification of the sources and targets of opinions. To evaluate diverse tools, the research community relies on the existence of a gold standard corpus addressing this need.
The 2014 Gestalt Shared Task organized by the Interest Group on German Sentiment Analysis (IGGSA) was a first step towards providing such data for German to the research community in the context of a shared task.
In 2014 IGGSA offered two different shared tasks. The STEPS shared task on Source, Subjective Expression and Target Extraction from Political Speeches will see a second iteration in 2016 with some significant modifications so as to lower the barrier to entry for participating teams.
The second iteration of STEPS maintained continuity with the first iteration in the following key respects:
- It was again concerned with the extraction of opinion roles/entities, that is, Sources and Targets.
- The textual basis continued to be political speeches from the Swiss parliament.
The main differences to the first iteration lay in the following areas:
- The annnotation schema was revised to make the task more accessible. This involved excluding from consideration some linguistic aspects that we felt were too complex to solve as part of this shared task (e.g. sub-lexical analysis in compound words and dealing with so-called polar facts).
- Training data was available, consisting of an adjudicated version of the 2014 task’s test data. Thus, unlike in 2014, participants were able to employ supervised learning for system implementation.