Text-based Question Answering Sample Clauses

Text-based Question Answering. Tasks‌ Knowledge-based approaches are often common choices when the query asks for a specific relation of entities. In the previous subsection, the question asked for the height of XxXxxx Xxxxx. In such case, it is relatively easy to extract the entity and its relation, which can be immediately mapped with a knowledge base. Unfortunately, despite the fact that the largest knowl- edge bases contain millions of nodes and billions of edges, several limitations exist. Often, these approaches fail on more difficult queries or when the query requires more advanced linguistic analysis [69]. Also, knowledge bases sometimes lack the actual entity links which cause problems [174]. Text-based approaches, on the other side, use lexical, syntactic and xxxxx- tic analyses to perform a set of tasks based on the natural language text. At first, passage retrieval is performed, where for a given query and a set of passages (documents), the most relevant ones to the query are located. The foundation of early approaches in this task is built on the lexicon-based ap- proaches [120, 12, 37]. Xxxxxx and Xxxxxxx provide a comprehensive overview of term-weighting methods for passage retrieval [120]. More recently, ma- chine learning has been successfully applied in this task by learning to rank the potential candidates [82]. Statistical learning methods such as SVM [18], ListNet [19], decision trees [22], neural networks and genetic algorithms [24] pushed the state of the art forward. Answer sentence selection is a core task in open-domain question answering. Given a question and a set of candidate sentences, the task is to rank the sentences from the most to least likely containing the answer or supporting it. Early approaches to this task were based on syntactic tree manipulations such as tree edit distances and semantic kernels [104, 38, 149]. These approaches are described further in Subsection 2.1.3. Unfortunately, the tree-edit based approaches have a major limitation of being computationally expensive and are often based on the outputs from specific parsers. Quasi-synchronous grammars [149] and deeper semantic analysis such as semantic roles [125, 58] have also been shown successful in the sentence selection task. Recent growth of neural network architectures and distributional semantics contributed to an increased number of approaches using these methods [171, 68, 43, 168, 166, 13]. Currently, researchers are investigating di↵erent attention approaches using neural archite...

Text-based Question Answering. A traditional approach to factoid question answering over text documents, popularized by the TREC QA task, starts by querying a collection with a possibly transformed question and retrieving a set of potentially relevant doc- uments, which are then used to identify the answer. Information retrieval for question answering has certain differences from traditional IR methods [110], which are usually based on keyword matches. A natural language question contains certain information, that is not expected to be present in the answer (e.g., question words who, what, when etc.), and the answer statement might use language that is different from the question (lexical gap problem) [25]. On the other side, there is a certain additional information about expected answer statement, that a QA system might infer from the question (e.g., we expect to see a number in response to the “how many” question). One way to deal with this problem is to transform the question in certain ways before querying a collection [5, 40]. Another idea is to extend the raw text data with certain semantic annotations, e.g., part of speech tags, semantic role labels, named entity tags, etc. By indexing these annotations a question answering system gets an opportunity to query collection with additional attributes, inferred from the question [33, 52, 238]. The next stage in TextQA is to select sentences, which might contain the answer. One of the most frequently used benchmark datasets for the task, designed by Xxxxxxx Xxxx et al. [211], is based on the questions from the TREC QA tasks and sentences retrieved by participating systems2. The early approaches for the task used mostly keyword matching xxxxxx- xxxx [101, 181]. However, in many cases, keywords does not capture the similarity in meaning of the sentences very well and researchers started look- ing on syntactic information. Syntactic and dependency tree edit distances and kernels help us measure the similarity between the structures of the sentences [90, 151, 172, 210, 237]. Recent improvements on the answer sen- tence selection task are associated with deep learning, e.g., recursive neural networks using sentence dependency tree [102], convolutional neural net- works [242, 161], recurrent neural networks [189, 207], and some techniques for term matching in the embeddings space [88, 216, 231]. Another dataset, called WikiQA [233], raises a problem of answer triggering, i.e., detecting cases when the retrieved set of sentences does not con...