Classic textual question answering (QA) approaches that rely on statistical keyword relevance scoring without exploiting semantic content are useful to a certain extent, but are limited to questions answered by a small text excerpt. With the maturation of Wikipedia and with upcoming projects like DBpedia, we feel that nowadays QA can adopt a deeper, semantic approach to the task, where answers can be inferred using knowledge bases to overcome the limitations of textual QA approaches. In GikiCLEF, a QA- avoured evaluation task, the best performing systems followed a semantic approach. In this paper, we present our motivations for preferring semantic approaches to QA over textual approaches, with Wikipedia serving as a raw knowledge source.
Question answering (QA) is a challenging task, requiring a good expertise in several natural language processing elds to properly understand information needs in the questions, to obtain a list of answer candidates from documents, and to lter them based on solid evidence that justi es each answer's correctness.
Most state-of-the-art QA approaches are textual QA systems built around a passage retrieval core, where questions (or their a rmative re-phrasings) are treated as bag-of-words or n-grams, ignoring their semantic content. These simpli ed question representations are fed into the information retrieval engine to obtain the paragraphs most likely to contain answers. The candidates are then extracted based on their type and ranked based on their frequency.
Passage-retrieval QA systems have their share of success in QA evaluation tracks such as QA@CLEF, which include a considerable amount of concrete questions (such as factoid or de nition questions). Answers to these questions can usually be found in the collection within a paragraph containing a sentence similar to the question (e.g. \Where is X located ? " ! \X is located in Y "), or by exploiting redundant information in large collections such as the Web.
As textual QA systems focus on selecting text excerpts from the collection, they cannot address structurally complex questions that require advanced reasoning over key concepts, nor questions whose answers are not explicitly represented in a text excerpt, but must be inferred from knowledge that may (or may not) be automatically derived from those text excerpts.
GikiCLEF [
Textual QA Semantic QA
[web of] documents [web of] data
document retrieval (IR core) search for and derive facts (IE core)
query expansion on word level question expansion on entity level
keywords & co-occurrence concepts & relations
ambiguous words (or even word forms) disambiguated concepts
textual snippets graph patterns
gazetteers RDF data
lexical semantics (thesaurus oriented) formal semantics
text with entities linked entities with text
One of the obstacles for semantic QA approaches is the need for a large
repository of \condensed world knowledge" derived from validated, well-structured
and machine-accessible data to enable inferring information that is implicit in
the text. While compiling such a repository is highly unfeasible for any single
researcher or research group (the famous knowledge bottleneck), it is feasible for
community-driven projects such as Wikipedia or Freebase. Along with upcoming
projects like DBpedia [
Our semantic QA systems dedicated special attention to these three tasks: Grounding expected answer types to a workable category/classi cation. The expected answer type (EAT) for a question must be mapped to entities from the world knowledge, so that its meaning becomes unambiguously grounded. Take for instance the question \Which Romanian writers were born in Bucharest in the 20th century? "; the EAT can be grounded to DBpedia resource URL http://dbpedia.org/resource/Category:Romanian_writers which makes it easier to search and validate candidate answers.
Parse constraints from the question. A semantic question interpretation can be
seen as an EAT and a composition of constraints. In the example above, the
question has two restrictions: one geographical (born `in Bucharest' ) and one
temporal (born `in the 20th century' ). By handling each constraint separately,
systems can use its meaning to employ the right strategies for validating
candidate answers [
Semantic analysis. The natural language question is transformed into a
machinereadable representation, which can be rewritten or expanded in some way. One
technique is logical question expansion by employing an inference engine with
rules for paraphrases, entailments, and logical equivalence, as described in [
The QA community nally has access to a considerable amount of world knowledge, created and curated by a dedicated community, encoded in machinereadable formats such as RDF/OWL and freely available through Linked Data. We feel that the time has come to focus on semantic approaches for QA. The core strategy of semantic QA systems is to understand key entities, concepts and relations between them in the question (the information need), and solve it by exploiting multiple knowledge sources and selecting the best strategies to determine and validate the right answers. Classic textual QA approaches are still useful for simple questions, but are quite limited for elaborated questions, and depending on a collection limits their question range. We foresee that the performance of semantic QA systems will improve as the resources and services from associated projects evolve over time, and will be able to answer more complex, open list questions that have more resemblance to real-user information needs.
This work is partially supported by FCT for its LASIGE Multi-annual support, GREASEII project (grant PTDC/EIA/73614/2006) and Nuno Cardoso's scholarship (grant SFRH/BD/45480/2008), partly supported by the EU-funded project QALL-ME (FP6 IST-033860), and supported by the Science Foundation Ireland (Grant 07/CE/I1142) as part of the Centre for Next Generation Localisation (CNGL).