Visualization for Coreference Annotation

Proceedings of Recent Advances in Natural Language Processing, pages 692–697, Hissar, Bulgaria, 12-14 September 2011. Visualization for Coreference Annotation Andre Burkovski and Gunther Heidemann Intelligent Systems Group Institute for Visualization and Interactive Systems University of Stuttgart {firstname.lastname}@vis.uni-stuttgart.de Abstract The annotation of documents with linguis- tic information requires time-consuming and therefore expensive manual annota- tion. Especially, a complex task, like coreference resolution, needs large data sets for the training of supervised machine learning methods. We present a tool which combines visualization techniques and un- supervised machine learning to support the annotation of documents with coref- erence information. Self-organizing Maps are used to cluster similar data and visu- alize the feature space. For link visualiza- tion, precise annotation, and error correc- tion a matrix-based coreference visualiza- tion is used which exploits the transitive property of the coreference relation. 1 Introduction The task of finding noun phrases which refer to the same discourse entity in a plain text is called coreference resolution. Many applications in in- formation retrieval (Nicolov et al., 2008), machine translation, and text summarization (Mitkov et al., 2007) use coreference resolution to improve the results. Currently, the popular Jeopardy! winner machine “Watson” uses a coreference resolution module among other modules for question answer- ing (Ferruci et al., 2010). In many cases, coreference is resolved using su- pervised machine learning methods. These meth- ods need large amounts of training samples. Yet, particularly for languages other than English, such data sets are rare and they mostly contain rela- tively few samples. Another problem is that super- vised methods which are trained on a specific do- main, such as news articles, may degrade on texts from other domains, like books or reports. E.g. Bakkenson and Soroka (2010) report on how the genre influences pronominal anaphora resolution. The manual annotation of a document with coreference information is time consuming, be- cause it depends not only on the background knowledge of the annotator, but also requires a high level of concentration to avoid annotation er- rors. Most annotation tools use text-based visual- izations to show and highlight the noun phrases for coreference annotation. However, using only text- based visualizations tools has a major drawback. Annotators are prone to annotation errors because they often need to repeatedly read passages in a text for every new unlabeled noun phrase. Mitkov et al. (2000) introduce some strategies for annota- tors to reduce errors and to accelerate the annota- tion speed, but the annotation of large data sets still is a problem. Rule-based or unsupervised machine learning methods may be used to highlight prob- able coreference pairs and chains and to support the annotators in identifying coreferences. Such an approach further reduces the time spent on the search of suitable coreference candidates. In our approach we address the annotation effi- ciency from the visualization and interaction point of view. We propose a combination of unsu- pervised machine learning method and visualiza- tions for annotation purposes. We use the Self- organizing Map (SOM) to visualize gropus of sim- ilar links …

Content truncated to 500 words