{{'TITLE' | translate}}

The NLP module generally takes on two tasks. For one, every question is analyzed. During this analysis, question characteristics such as the type of the expected answer are determined. Additionally, the question is categorized in regards to its general topic or field to enable the use of specific knowledge bases. The second main task consists of providing certain natural language processing methods. These methods are used in the question analysis, but there are also several answer finding modules that make use of them.

In the answerfinding process we aim to find answers to the posed question using the structures generated by the natural language preprocessing. To accomplish this, we scan both offline and online resources for documents that might contain answers to the question. By using methods like Named Entity Tagging we generate answer candidates from these documents, which we subsequently filter considering different criteria. In a final step, we generate more specific documents for each candidate that our scorer modules process in order to estimate a confidence score for each answer candidate.

The ranking module is the last step in answering a question. It takes all answer candidates and their scores from the answer finding modules and uses these scores to rank the candidates by their probability of being the correct answer to the question. This ranking, or the candidate at the top of it, is the output of the system. The module combines all scores of an answer candidate into a single score and ranks the candidates by their combined scores. Because scores come from many answer finding modules of different quality, combining scores without weighting them differently is not optimal and setting the weights manually is not feasible. Thus a machine learning algorithm (logistic regression) is used.