Skip to main content
Rethinking "Questions" for
Context-aware Question-answering
2026. 08. 04 @Agent Memory Study
Won Ik Cho
Why does context matter?
• 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
Key paper
2
Assume information structure is universal
3
Discourse as language game
4
Goals, rules, moves, strategies, and CG/Context set
5
Question as a set of propositions
6
Entailment
7
Question, Information structure,
Question under discussion
Question
9
Question
10
e.g., Who was the 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.
Question
11
Answer
12
Answer
13
Entailment
14
Information structure
15
Information structure
16
Information structure
17
Information structure
18
Information structure
19
QUD
20
QUD
21
QUD
22
QUD
23
QUD
24
QUD
25
QUD
26
QUD
27
QUD
28
Questions in information structure
29
Questions in information structure
30
Questions in information structure
31
Multi-hop QA and Multi-turn dialogue
Multi-hop QA
33
Multi-hop QA
34
Multi-hop QA
35
Multi-hop QA
36
Multi-hop QA
37
MHQA as QUD in 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
QA in multi-turn dialogue 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
Takeaways
• Multi-turn QA intuitively 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
Thank you!
Questions & Discussion welcome.