Page 90 - My FlipBook
P. 90
Collaborative Projects所
內
合
作
計
畫
Figure 2 : Proposed Q&A operation ow.
Given a Question, Wikipedia and some external Resources (such as WordNet, ConceptNet or Freebase), the
task is to establish the most likely answer from related Wikipages and on-line retrieved documents. To reduce
computation load, we rst extract related Wikipages using an o -the-shelf information retrieval tool (e.g., the
Lucene search engine) as a pre-processing stage. Then, di erent kinds of questions (e.g., Who, What, When,
Where, How, and Why) might be posed to those results, generating different answer types (e.g., person/
location/organization/product/event name, date/time, duration, distance, procedure, reason, etc.), in di erent
answer forms (e.g., Yes/No, specific answer, free text). Importantly, different mechanisms (termed "answer-
modes" in Figure 2) would be required to obtain various desired answers. For example, acquiring an answer
might require conducting the following different operations: locating one or more spans from the given
text, conducting entailment judgment (for yes/no questions), and performing logic (e.g., intersection, union,
complement), arithmetic (e.g., sum, di erence), and aggregative (e.g., comparison, Max/Min, counting, etc.)
operations on the extracted information.
Obviously, it is exceedingly di cult to design a general-purpose module to handle all of the various answer-
seeking mechanisms mentioned above. Therefore, we propose to utilize a Divide-and-Conquer framework
to convert this complex task into a set of simple sub-tasks. Accordingly, each speci c answer-mode will be
handled by a di erent answer-generation module/model. Moreover, one answer-type might be handled by
several answer-generation modules if an ensemble approach is adopted. Thus, we have designed a dispatcher
module to rst identify possible answer-types and answer-modes associated with a given question, allowing
us to activate various corresponding answer-generation modules to get a set of results. Then, an aggregator
module can generate the nal answer by merging all of the results.
88
內
合
作
計
畫
Figure 2 : Proposed Q&A operation ow.
Given a Question, Wikipedia and some external Resources (such as WordNet, ConceptNet or Freebase), the
task is to establish the most likely answer from related Wikipages and on-line retrieved documents. To reduce
computation load, we rst extract related Wikipages using an o -the-shelf information retrieval tool (e.g., the
Lucene search engine) as a pre-processing stage. Then, di erent kinds of questions (e.g., Who, What, When,
Where, How, and Why) might be posed to those results, generating different answer types (e.g., person/
location/organization/product/event name, date/time, duration, distance, procedure, reason, etc.), in di erent
answer forms (e.g., Yes/No, specific answer, free text). Importantly, different mechanisms (termed "answer-
modes" in Figure 2) would be required to obtain various desired answers. For example, acquiring an answer
might require conducting the following different operations: locating one or more spans from the given
text, conducting entailment judgment (for yes/no questions), and performing logic (e.g., intersection, union,
complement), arithmetic (e.g., sum, di erence), and aggregative (e.g., comparison, Max/Min, counting, etc.)
operations on the extracted information.
Obviously, it is exceedingly di cult to design a general-purpose module to handle all of the various answer-
seeking mechanisms mentioned above. Therefore, we propose to utilize a Divide-and-Conquer framework
to convert this complex task into a set of simple sub-tasks. Accordingly, each speci c answer-mode will be
handled by a di erent answer-generation module/model. Moreover, one answer-type might be handled by
several answer-generation modules if an ensemble approach is adopted. Thus, we have designed a dispatcher
module to rst identify possible answer-types and answer-modes associated with a given question, allowing
us to activate various corresponding answer-generation modules to get a set of results. Then, an aggregator
module can generate the nal answer by merging all of the results.
88
