Typed-RAG: Type-aware Multi-Aspect Decomposition for Non-Factoid Question Answering
Non-factoid question-answering (NFQA) poses a significant challenge due to its open-ended nature, diverse intents, and the need for multi-aspect reasoning, which renders conventional factoid QA approaches, including retrieval-augmented generation (RAG), inadequate. Unlike factoid questions, non-factoid questions (NFQs) lack definitive answers and require synthesizing information from multiple sources across various reasoning dimensions. To address these limitations, we introduce Typed-RAG, a type-aware multi-aspect decomposition framework within the RAG paradigm for NFQA. Typed-RAG classifies NFQs into distinct types—such as debate, experience, and comparison—and applies aspect-based decomposition to refine retrieval and generation strategies. By decomposing multi-aspect NFQs into single-aspect sub-queries and aggregating the results, Typed-RAG generates more informative and contextually relevant responses. To evaluate Typed-RAG, we introduce Wiki-NFQA, a benchmark dataset covering diverse NFQ types.
3 Method
This section introduces the RAG pipeline specifically designed for NFQ types. The overall methodology is illustrated in Figure 1. The types of NFQ differ in their intent, the directionality of their aspects, and the degree of contrast between them. NFQs often manifest as multi-aspect queries, where answers span multiple perspectives. These aspects can be categorized into two types: contrasting aspects, which exhibit high contrast and opposing directions, and related aspects, which have lower contrast and align in the same direction. Based on these characteristics, we designed methods that effectively align with user expectations.12
In the following sections, we discuss these question types and their processing rationale.
Evidence-based Evidence-based type questions seek to understand the characteristics or definitions of specific concepts, objects, or events, typically requiring fact-based explanations. As single-aspect queries, they do not undergo multi-aspect decomposition. Instead, we apply a straightforward RAG approach, retrieving relevant passages and generating accurate answers while preserving clarity and simplicity. The retriever first retrieves passages using the question as a query, and the generator produces responses based on these passages.
Comparison Comparison-type questions are used to examine differences, similarities, or superiority between keywords, with answers tailored to the comparison’s purpose and targets. 1Using the prompts described in Appendix A.4, our approach involves either applying a multi-aspect decomposer, tailored to the question type, or aggregating answers generated for multiple single-aspect queries.
As multi-aspect queries, they require decomposition. First, a keyword extractor identifies the comparison purpose (compare_type) and targets (keywords_list). Subsequently, the retriever is employed to search for passages that are related to each keyword. Following the deduplication of passages, the remaining results are reranked according to their relevance. Finally, the generator combines the information to create a response that aligns with the comparison purpose, thereby effectively addressing the question.
Experience Experience-type questions seek advice or recommendations, providing explanations based on personal experiences. As multi-aspect queries, they require decomposition. The retriever first retrieves relevant passages, followed by similarity-based reranking using extracted keywords by the keyword extractor. In the end, the generator produces an optimized response aligned with the intent of the question, effectively meeting its requirements.
Reason/Instruction Reason-type questions aim to explore the causes of specific phenomena, while instruction-type questions focus on understanding procedures or methods. The intent behind these question types is to receive clear and comprehensive answers that present reasons or steps in a well-structured manner. These questions are treated as multi-aspect queries and require a multi-aspect decomposer. First, the query is decomposed into single-aspect queries using a single-aspect query generator. Each generated query is then processed individually by the retriever and generator to produce separate answers. Finally, an answer aggregator combines these individual responses into a concise and accurate final answer.
Debate Debate-type questions are hypothetical questions designed to explore diverse perspectives, including opposing viewpoints. The purpose of this question type is to generate a balanced response that reflects multiple perspectives. Being a multi-aspect question, it naturally requires the multi-aspect decomposer. The question is first processed by a single-aspect query generator, which extracts the discussion topic and various opinions. Generated queries for each opinion are then handled by the retriever and generator to produce individual responses. Finally, a debate mediator combines the topic and perspectives to produce a well-balanced final response.