Why does contextmatter?
• The problem
Long-context QA needs evidence spread across a large corpus
Sometimes this context is not necessarily in document format
‘Stuff everything in the prompt' affects accuracy and cost
The agent must choose what to retrieve, keep, and discard
• This relates to the pragmatics of “Question”
How is question defined?
How interaction regarding question yields the information structure?
What is ‘question under discussion’?
How can we relate QUDs with multi-hop question-answering?
How should the clues for question-answering be managed while interaction?
1
Question
10
e.g., Who wasthe president of United States in the year in which World War II began?
Mavi et al. (2022)
p1: World War II began in 1939.
p2: Franklin D. Roosevelt was the president of United States in 1939.
MHQA as QUDin information structure
• Finding answer for q which can be iteratively answered by
decomposing q to multiple (wh-)subquestions
Solving MHQA: QUD reaching the end (getting answer / unanswerable)
Set of contexts: Common ground
Set of related contexts: Strategy of inquiry (step by step manner)
: Goal of the discourse (among interlocutors)
: Assertion for each subquestion
Hop: Adding an accepted assertion
Reasoning chain: An ordered set of assertion for each subquestion
Ordered permutation: <
38
40.
QA in multi-turndialogue as MHQA
• Common ground
Dialogue history (and that they talked) becomes context
• Strategy of inquiry
“Relevant” contents of dialogue history
• Goal of the discourse
Continuing the conversation based on the premise that all participants
comprehends the utterances of each other
• < (order)
Logical + Temporal factors both apply
39
41.
Takeaways
• Multi-turn QAintuitively becomes multi-hop QA so while
managing knowledge for solving questions in multi-turn manner, it
is important to
formalize questions for atomic accumulation (hopefully multiple wh-
manner)
accummulate common ground as context
assign timestamps to moves for ordering
remove (or omit) unaccepted moves
remove (or edit) negated moves
manage QUD based on relevance
distinguish resolved (answered, unanswerable) from QUDs (while
maintainig that those are answered/unanswerable)
use appropriate strategy depending on the type of MHQA the multi-turn
QA belongs to
40