Structured QA Annotation Scheme

• Structured QA Annotation Schemes are formal frameworks that specify taxonomies, mapping rules, and data structures to reliably pair questions with answers in diverse domains.
• They use systematic annotation workflows that transform utterances into structured (type, slot) pairs via deterministic mapping functions for both questions and commands.
• The schemes are adaptable across languages and domains, enhancing neural model training and providing consistent intent extraction for dialogue systems.

A structured question–answer annotation scheme specifies formal taxonomies, annotation workflows, and data representations for pairing questions with corresponding answers in a way that enables downstream computational treatment. Such schemes appear throughout question answering (QA), argument mining, dialog modeling, semantic parsing, and educational assessment; they are foundational for creating reliable, scalable training corpora and interpretable model outputs. Contemporary structured QA annotation frameworks define not only the permissible types of questions and answers but also the granularity and content of argument extraction, the layer(s) of semantic or pragmatic interpretation (illocutionary force, intent, slot filling), explicit mapping rules, and precise data structures. Methods may differ in their alignment with conversational vs. reading comprehension goals, their focus on dialog acts or predicate-argument relations, and the degree of cross-linguistic adaptation.

1. Taxonomic Foundations: Types, Roles, and Layering

Structured annotation schemes specify a discrete typology covering both question and answer forms. In dialog-centric schemes for typologically complex languages such as Korean, a primary distinction is drawn between questions and commands, with illocutionary labeling at the utterance level: questions (Q) seek information, while commands (C) request action. Each is characterized by a fine-grained set of types: for questions, the annotation scheme in (Cho et al., 2018) prescribes T_Q = {YN, ALT, WH} corresponding to yes/no (polar), alternative, and wh-questions; for commands, T_C = {PH, REQ, SR} designating prohibition, requirement, and strong requirement. Utterances with purely rhetorical, exclamatory, or idiomatic illocution are explicitly excluded.

Layered mapping is central: each utterance u is mapped to a structured tuple

$$(\text{type}, \text{argument}, [\text{negativeness}]),$$

where argument is a nominalized phrase (“slot value”) capturing the core requested content or action, and “negativeness” (negation marker) is only instantiated for commands.

No multi-level syntactic decomposition of argument content is performed; all mappings are utterance→slot, maintaining a strict two-layer structure. This ensures both conceptual clarity and process efficiency in annotation.

2. Annotation Workflow and Formalization

Annotation proceeds via systematic assignment of (type, argument) pairs using transformation rules derived from linguistic analysis:

• Questions: For each u∈U that is a question, $\varphi_Q$ applies a deterministic mapping based on surface structure. Yes/no questions are nominalized via suffixes marking “whether or not” (e.g., “-yepwu” in Korean), alternative questions express “what to do/choose among A and B,” and wh-questions abstract away from the wh-word to a fixed noun (e.g., who→person, what→meaning, where→place).
• Commands: For u∈U that is a command, $\varphi_C$ converts the predicate to a nominalized action, marked for negation in prohibitions (“-ha-ki an-ki”) but not in neutral requirements (“-ha-ki”). For strong requirements, only the positive imperative is annotated.

Formally:

$$\varphi_Q(u) = (\text{YN}, P\text{-yepwu}) \quad \text{for yes/no question with predicate } P$$

$$\varphi_C(u) = (\text{PH}, P\text{-ha-ki an-ki}) \quad \text{for prohibition with predicate } P$$

and for wh-questions, the mapping is

$$\varphi_Q(u) = (\text{WH}_{\text{wh}}, a_{\text{wh}}(P))$$

where $a_{\text{wh}}$ maps the wh-word to its semantic class noun.

Examples:

• “Did you apply for medical service?” → (YN, “의료 서비스 신청 여부”)
• “Don’t go outside.” → (PH, “밖에 나가기 안키”)

Congruence of this formalism is critical for subsequent systematic extraction and machine learning.

3. Integration with Neural Paraphrasing and Machine Learning

Structured annotation is directly leveraged for supervised training of neural paraphrasing systems. The primary task is to learn a function mapping non-canonical utterances to canonical argument slots:

• Model architecture: Bidirectional LSTM or Transformer encoder over utterances, classification head for type prediction ($T_Q \cup T_C$), and sequence decoder for slot phrase (argument) generation.
• Data: 30,837 supervised pairs of utterances and their canonical slot arguments, with additional paraphrase augmentation (e.g., synonym replacement, slot order shuffling for ALT questions).
• Evaluation metrics: Exact-match accuracy (slot generation), BLEU/ROUGE for similarity with gold slots, and classification accuracy for type labels.

This joint structured annotation plus modeling pipeline enables machines to identify intent and extract actionable arguments from linguistically variable dialog (Cho et al., 2018).

4. Scheme Generalizability and Typological Adaptation

The framework’s central abstractions allow adaptation to other languages and domains:

• Typological transfer: To agglutinative languages (Japanese, Turkish), one substitutes equivalent nominalizers and semantic class mappings for wh-words. For SVO languages such as English, slots like “whether P” for YN questions, “doing X”/“not doing X” for commands, and role-like nouns (“location,” “reason”) for wh-questions maintain correspondence.
• Domain extension: For application in areas like e-commerce, domain-specific slots (e.g., product type, delivery time) are appended, but the underlying $\varphi_Q/\varphi_C$ mapping machinery is unchanged.

This transferability demonstrates the inductive bias provided by the two-layer (type, slot) organization and semantic abstraction.

5. Empirical Properties, Use Cases, and Illustrative Mappings

Structured annotation schemes have direct utility in building robust argument mining pipelines, slot-filling in task-oriented dialog systems, and as foundation for neural sequence-to-sequence models. Representative mappings include:

Utterance	Type	Argument
“Did you apply for medical service?”	YN	의료 서비스 신청 여부
“Will you come by bus or taxi?”	ALT	버스 택시 중 타고 올 kes
“Who came today?”	WH_who	오늘 온 salam
“Don’t go outside, a typhoon’s coming.”	PH	밖에 나가기 안키
“Please check personal info.”	REQ	개인 정보 확인하기
“Don’t be greedy, just sell now!”	SR	지금 팔기

The mapping enforces both illocution (type) and a nominalized actionable or informational slot, which downstream neural systems can target for intent extraction, paraphrase generation, or argument summarization.

6. Implications, Applications, and Limitations

By systematizing the representation of conversational intent and argument, structured annotation schemes enable typologically-informed corpus creation and consistent training targets for neural intent extraction modules. Their two-layer structure ensures both processable simplicity and theoretical coverage of dialog phenomena relevant to intention recognition in spoken and written language.

However, the scheme focuses strictly on mapping utterances to single nominalized slots, deliberately eschewing deeper argument structure decomposition or multi-slot syntactic parsing. Rhetorical, exclamatory, and idiomatic forms are filtered out, which could limit coverage in open-domain dialog but enforces high precision for task-oriented applications. The method's adaptation relies on redefining nominalizers and argument labels, which may require expert linguistic intervention in less typologically similar languages (Cho et al., 2018).

Through explicit formalism, representative mappings, and compatibility with modern neural models, structured question–answer annotation schemes provide a theoretically sound, empirically validated methodology for dialog, argument mining, and robust intent extraction in computational linguistics.