Aligning Interrogative Intent with Answer Semantic Types

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 237 papers38 conceptsfrozen 2026-10-04

In neural text generation, fluency is often mistaken for intelligence. A sequence-to-sequence model can easily generate a grammatically perfect question like "Why was the treaty signed?" for a source text. However, if the intended answer is "1945," the model has fundamentally failed. This "interrogative mismatch"—where the question word (Why) does not align with the semantic type of the answer (Date)—is a primary hurdle in creating reliable Question Generation (QG) and Question Answering (QA) systems.

Recent research has shifted from simple sequence modeling toward explicit semantic alignment, ensuring that the intent of a query and the nature of its answer are logically coupled.

The Interrogative Mismatch Problem

Standard sequence-to-sequence models often treat the target answer as a simple location marker in the source text rather than a semantic constraint. As Samm Sun and Xingwu Sun identify in Answer-focused and Position-aware Neural Question Generation, this leads to "hallucinations" where the model generates interrogative words that are semantically incompatible with the answer span.

This problem persists because the decoder's objective is to maximize the likelihood of the next word based on the previous words, often ignoring the specific category of the answer entity. This "semantic drift" is particularly visible in Knowledge Base Question Generation (KBQG), where Sheng Bi notes that models might ask "when" for a location entity because they fail to condition the generation process on the answer's metadata in Knowledge-enriched, Type-constrained and Grammar-guided Question Generation over Knowledge Bases.

Explicit Type Conditioning and Constraints

To fix this, researchers have introduced mechanisms that force the model to "look" at the answer's type before deciding how to start a question. Qingbao Huang and colleagues developed an entity-guided question-type predictor in Entity Guided Question Generation with Contextual Structure and Sequence Information Capturing. By explicitly mapping answer entities to question types (e.g., mapping a "Person" entity to "Who"), the model constrains the decoder's search space, ensuring the generated question is actually answerable by the target span.

This constraint isn't just for questions. In the reverse task—generating natural language answers from entity "lonely" answers—Fen Zhao uses a "question-aware loss" to ensure the generated sentence reflects the structural intent of the query, as described in Attention-based RNN with question-aware loss and multi-level copying mechanism for natural answer generation.

Moving from Sequences to Semantic Structures

Intuition suggests that a question isn't just a string of words; it's a bridge between a predicate and an argument. Treating text as a linear sequence loses this structural relationship. Yu Chen addresses this in Reinforcement Learning Based Graph-to-Sequence Model for Natural Question Generation by using Graph Neural Networks (GNNs). By encoding the source text as a graph of semantic relationships, the model understands the answer's role in the sentence (e.g., as a Subject or Object), which naturally informs the correct interrogative word.

Verification through Cycle Consistency

One of the most effective ways to ensure alignment is to "close the loop." If a model generates a question from an answer, a separate QA model should be able to find that exact answer using the generated question. Qifan Wang implements this "primal-dual" approach in Learning to Generate Question by Asking Question. By "asking the question back," the system verifies that the generated text uniquely and specifically leads to the intended target. If the QA model fails to find the original answer, the generator is penalized, forcing it to produce more semantically precise queries.

Reversing the Synthesis Pipeline

When forward generation (Question → Answer) is too noisy, researchers have found success in reversing the pipeline. In the context of Text-to-SQL, Yu Guo introduces SQLForge, which generates the structured SQL query first and then "forges" the natural language question around it. This ensures that the synthesized ground truth is always semantically aligned with the logic of the query, a technique that significantly improves the reasoning capabilities of open-source models.

Similarly, for educational testing, Elaf Alhazmi uses contrastive learning in Fine-Tuning Encoder-Decoder Models with Contrastive Learning for In-Context Distractor Generation to ensure that "wrong" answers (distractors) are contextually relevant but logically distinct from the correct answer, preventing the generation of irrelevant noise that students could easily ignore.

Papers in this version

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 22 papers
  2. Version 1 · 2026-10-04 · 37 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_answer_question_semantic_alignment_6c7bcb69,
  title = {answer-question semantic alignment},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/answer-question-semantic-alignment/rdr_6c7bcb6985134c67b3efeb948b6f997f},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_aligning_interrogative_intent_with_answe_6c7bcb69_v1,
  title = {Aligning Interrogative Intent with Answer Semantic Types},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/aligning-interrogative-intent-with-answer-semantic-types/rdr_6c7bcb6985134c67b3efeb948b6f997f/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}