Page 91 - My FlipBook
P. 91
Brochure 2020
Our research focuses on the following issues: (1) How to guide the user to quickly reach a
speci c answer through pragmatic reasoning; (2) How to identify related passages scattered
over various documents via multi-hop searching; (3) How to add intermediate clues (in addition
to supporting evidence) extracted from different external knowledge bases (e.g., Freebase,
Wikidata); (4) How to conduct multi-hop inference enhanced with external knowledge (e.g.,
commonsense, domain-knowledge, etc.) extracted from WordNet/ConceptNet; and (5) How to
handle natural speech and integrate both the NLP and SR modules.
In the first year of the project (2020), we will complete a domain-specific baseline system
based on pre-extracted Wikipedia pages. To do this, we will: (1) design a uni ed framework to
integrate various answer-seeking mechanisms; (2) design various answer-generation modules
to mimic human reasoning; and (3) utilize speech and text clues to closely integrate the NLP
and SR modules.
In the second year (2021), we will complete a multi-domain system with multi-hop inference
capability by: (1) designing a multi-hop searching algorithm to obtain all associated supporting
evidence scattered across different documents; (2) decompose a complex question into
a sequence of simpler questions; (3) conduct multi-hop inference based on the extracted
supporting evidence; (4) conduct domain adaptation; and (5) detect and remove dis uency to
improve spontaneous speech recognition.
In the third year (2022), we will complete an open-domain system with enhanced
commonsense reasoning capability over both Wikipages and other on-line documents by: (1)
conducting commonsense reasoning (introducing various intermediate goals); (2) pragmatic
reasoning in dialog management; (3) enhanced interpretability (i.e. generate associated
rationales); and (4) enhanced robustness (via adversarial learning) to recognize natural speech.
To demonstrate the power of our developed methodologies, we will continuously participate
in the "Formosa Grand Challenge -- Talk to AI" contest (organized by the Ministry of Science and
Technology) and the international Machine Reading for QA (MRQA) shared task.
89
Our research focuses on the following issues: (1) How to guide the user to quickly reach a
speci c answer through pragmatic reasoning; (2) How to identify related passages scattered
over various documents via multi-hop searching; (3) How to add intermediate clues (in addition
to supporting evidence) extracted from different external knowledge bases (e.g., Freebase,
Wikidata); (4) How to conduct multi-hop inference enhanced with external knowledge (e.g.,
commonsense, domain-knowledge, etc.) extracted from WordNet/ConceptNet; and (5) How to
handle natural speech and integrate both the NLP and SR modules.
In the first year of the project (2020), we will complete a domain-specific baseline system
based on pre-extracted Wikipedia pages. To do this, we will: (1) design a uni ed framework to
integrate various answer-seeking mechanisms; (2) design various answer-generation modules
to mimic human reasoning; and (3) utilize speech and text clues to closely integrate the NLP
and SR modules.
In the second year (2021), we will complete a multi-domain system with multi-hop inference
capability by: (1) designing a multi-hop searching algorithm to obtain all associated supporting
evidence scattered across different documents; (2) decompose a complex question into
a sequence of simpler questions; (3) conduct multi-hop inference based on the extracted
supporting evidence; (4) conduct domain adaptation; and (5) detect and remove dis uency to
improve spontaneous speech recognition.
In the third year (2022), we will complete an open-domain system with enhanced
commonsense reasoning capability over both Wikipages and other on-line documents by: (1)
conducting commonsense reasoning (introducing various intermediate goals); (2) pragmatic
reasoning in dialog management; (3) enhanced interpretability (i.e. generate associated
rationales); and (4) enhanced robustness (via adversarial learning) to recognize natural speech.
To demonstrate the power of our developed methodologies, we will continuously participate
in the "Formosa Grand Challenge -- Talk to AI" contest (organized by the Ministry of Science and
Technology) and the international Machine Reading for QA (MRQA) shared task.
89
