What distinguishes contrasting aspects from related aspects in question structure?
This explores how a question's structure shifts depending on whether it asks you to weigh things against each other (comparison, debate) versus pull apart facets that belong together (decomposition into complementary parts) — and why that distinction changes how a system should retrieve and answer.
This explores how a question's structure shifts depending on whether it pits aspects against each other or breaks a single thing into complementary parts — and the corpus suggests the difference isn't cosmetic, it changes the whole retrieval and aggregation strategy. The cleanest map comes from work showing that non-factoid questions split into five types, where the question's *type* determines both how you retrieve and how you recombine evidence Does question type determine the right retrieval strategy?. Comparison and debate questions need **aspect-specific retrieval** — you go find what each side says about each dimension, then hold them in tension. Experience and reason questions instead need **decomposition** — you split the question into sub-parts that all point at the same answer and filter or aggregate them. So 'contrasting' aspects get retrieved in opposition; 'related' aspects get retrieved as a set you reassemble.
What makes contrast hard is coverage and balance, not just finding material. Research on debatable summarization shows that throwing one uniform query at every source collapses perspectives — the fix is assigning each document its own specialized speaker and a tailored query, which produces large jumps in topic coverage and balance Can tailoring queries per document improve debatable summarization?. The structural signature of a contrasting question is that you have to *deliberately preserve disagreement*; a single averaged pass erases exactly the thing the question is asking about.
Related aspects work the opposite way: the structure is decomposition into attributes that complement rather than compete. The ALFA framework breaks 'question quality' into theory-grounded facets — clarity, relevance, specificity — and trains on each separately, beating a single blended score Can models learn to ask genuinely useful clarifying questions?. These aspects aren't in conflict; they're orthogonal dimensions of one good question. That orthogonality echoes the argument-scheme 'periodic table,' where three independent axes (subject-predicate structure, order of reasoning, proposition pairings) jointly locate any scheme — related coordinates, not rival positions Can three axes organize all possible argument schemes?.
The deeper lesson hiding here is *why contrast is computationally heavier*. Recognizing that two aspects genuinely oppose each other requires reading inferential patterns spread across distributed text spans, not local surface cues — which is exactly why argument-scheme classification plateaus where simpler tagging tasks succeed Why does argument scheme classification stumble where other NLP tasks succeed?. Telling 'these complement' from 'these conflict' is an integrative judgment, and it's the same judgment a reader makes when deciding whether a question wants synthesis or a verdict.
The thing you might not have expected to learn: the contrast-vs-related distinction isn't only in the question's wording — it determines whether a system should *converge* its evidence toward one answer or *hold it apart* to keep the tension visible. Get that wrong, and a comparison question gets flattened into a bland summary, or a decomposable question gets fragmented into a false debate.
Sources 5 notes
Research shows non-factoid questions split into five types, each requiring different retrieval and aggregation methods. Evidence-based questions suit standard RAG, while debate and comparison need aspect-specific retrieval, and experience/reason questions need decomposition or filtering strategies.
MODS achieves 38–58% improvement in topic coverage and balance by assigning each document a specialized speaker LLM that receives tailored queries, rather than applying uniform queries across all documents. This reframes summarization as a retrieval problem solved through source-aware query planning.
The ALFA framework breaks down question quality into theory-grounded attributes (clarity, relevance, specificity) and trains models on 80K attribute-specific preference pairs. Attribute-specific optimization outperforms single-score training, especially in clinical reasoning where asking the right clarifying question directly impacts decision quality.
Wagemans's Periodic Table maps all argument schemes onto coordinates across three axes: subject-predicate structure, first-order versus second-order reasoning, and proposition-type pairings. This combinatorial approach replaces Walton's open-ended list with a closed, systematic space enabling computational analysis and discovery of unstudied scheme types.
Scheme classification requires recognizing inferential patterns across distributed text spans, not local surface features. Models plateau at F1 0.55–0.65 while the same systems exceed 0.80 on component tagging and stance, suggesting the integrative reasoning demand is fundamentally different.