Loading...
Publication
Rapport : a fact-based question answering system for portuguese
Authors
Abstract(s)
Question answering is one of the longest-standing problems in natural language processing. Although natural language interfaces for computer systems can be considered
more common these days, the same still does not happen regarding access to specific
textual information. Any full text search engine can easily retrieve documents containing user specified or closely related terms, however it is typically unable to answer user
questions with small passages or short answers.
The problem with question answering is that text is hard to process, due to its syntactic structure and, to a higher degree, to its semantic contents. At the sentence level,
although the syntactic aspects of natural language have well known rules, the size and
complexity of a sentence may make it difficult to analyze its structure. Furthermore, semantic aspects are still arduous to address, with text ambiguity being one of the hardest
tasks to handle. There is also the need to correctly process the question in order to define its target, and then select and process the answers found in a text. Additionally, the
selected text that may yield the answer to a given question must be further processed
in order to present just a passage instead of the full text. These issues take also longer
to address in languages other than English, as is the case of Portuguese, that have a lot
less people working on them.
This work focuses on question answering for Portuguese. In other words, our field
of interest is in the presentation of short answers, passages, and possibly full sentences,
but not whole documents, to questions formulated using natural language. For that purpose, we have developed a system, RAPPORT, built upon the use of open information
extraction techniques for extracting triples, so called facts, characterizing information
on text files, and then storing and using them for answering user queries done in natural language. These facts, in the form of subject, predicate and object, alongside other
metadata, constitute the basis of the answers presented by the system. Facts work both
by storing short and direct information found in a text, typically entity related information, and by containing in themselves the answers to the questions already in the
form of small passages. As for the results, although there is margin for improvement,
they are a tangible proof of the adequacy of our approach and its different modules for
storing information and retrieving answers in question answering systems.
In the process, in addition to contributing with a new approach to question answering for Portuguese, and validating the application of open information extraction to
question answering, we have developed a set of tools that has been used in other natural language processing related works, such as is the case of a lemmatizer, LEMPORT,
which was built from scratch, and has a high accuracy. Many of these tools result from
the improvement of those found in the Apache OpenNLP toolkit, by pre-processing their
input, post-processing their output, or both, and by training models for use in those
tools or other, such as MaltParser. Other tools include the creation of interfaces for
other resources containing, for example, synonyms, hypernyms, hyponyms, or the creation of lists of, for instance, relations between verbs and agents, using rules.