Interpretable Complex Question Answering

We will review cross-community co-evolution of question answering (QA) with the advent of large-scale knowledge graphs (KGs), continuous representations of text and graphs, and deep sequence analysis. Early QA systems were information retrieval (IR) systems enhanced to extract named entity spans from high-scoring passages. Starting with WordNet, a series of structured curations of language and world knowledge, called KGs, enabled further improvements. Corpus is unstructured and messy to exploit for QA. If a question can be answered using the KG alone, it is attractive to ‘interpret’ the free-form question into a structured query, which is then executed on the structured KG. This process is called KGQA. Answers can be high-quality and explainable if the KG has an answer, but manual curation results in low coverage. KGs were soon found useful to harness corpus information. Named entity mention spans could be tagged with fine-grained types (e.g., scientist), or even specific entities (e.g., Einstein). The QA system can learn to decompose a query into functional parts, e.g., “which scientist” and “played the violin”. With increasing success of such systems, ambition grew to address multi-hop or multi-clause queries, e.g., “the father of the director of La La Land teaches at which university?” or “who directed an award-winning movie and is the son of a Princeton University professor?” Questions limited to simple path traversals in KGs have been encoded to a vector representation, which a decoder then uses to guide the KG traversal. Recently the corpus counterpart of such strategies has also been proposed. However, for general multi-clause queries that do not necessarily translate to paths, and seek to bind multiple variables to satisfy multiple clauses, or involve logic, comparison, aggregation and other arithmetic, neural programmer-interpreter systems have seen some success. Our key focus will be on identifying situations where manual introduction of structural bias is essential for accuracy, as against cases where sufficient data can get around distant or no supervision.