Automatic Question Answering using Domain Knowledge

We everybody knows question answering is a fastly growing field of NLP and IR research. There are two kinds of question answering systems, first one is domain specific question answering and other one is open domain question answering. In this blog post we will mainly focus on domain specific question answering. For automatic question answering the domain knowledge base or ontology can be used to a great extent to answer the user query most efficiently and accurately. For this blog I mainly used music domain knowledge to answer the questions related to song, singer, album, company etc.

ARCHITECTURE OF THE SYSTEM:

Image taken from Reference: 1

The basic architecture of the system consists of four main building blocks:

A. Question classification

B. Matching with FAQ question

C. Question Analyzer

D. Answer Extractor

Question classification:

This is the first module of the question answering system. In this stage, the system classifies the user
query into 8 main classes like the song, album, person, location, digital, company etc. and 46
subclasses. Which helps the system to process the question accurately.  In this stage, the query is
initially compared using ontology and then further processed in the bayesian classifier. Using ontology
the concepts present in the user questions extracted so that it can be used for ontology knowledge
base. For further discrimination, bayesian based method is used for the allocation of class.  

B. Matching with FAQ question

In this module, a set of frequently asked questions get compared with the asked query. For this
purpose cosine similarity between the queries can be calculated by converting the queries into
word2vec representation. If the asked query did not find any matching pair with the available FAQ
set, it is sent to question analyzer stage else answer is returned directly to the user.

C. QUESTION ANALYZER:

In this stage, the semantic representation of the question is extracted and analyzed by the system.
For the domain-specific analysis, the query is divided into divided into different domains. In some
places questions, semantic is identified and analyzed for getting the finer details. Two levels can be
used to the construction of best analyzer: level 1 question

block pair and level 2 is sequential semantic block.

D. Answer Extractor:

Mainly two strategies are adopted for finding out the appropriate answers i.e. by matching with the
frequently asked questions and the second one is by the checking with ontology but the first one is
more preferable.

Above experiment shows 77.25% precision in the result, but In future steps,

an author has assured to improve the precision.

Reference:

[1] International Conference on Computer Engineering and Technology Domain, 2009, Domain
Ontology-Based Automatic Question Answering Jibin Jibin Fu, Jinzhong Xu, Keliang Jia