Conversational QA: Techniques & Advances

• Conversational Question Answering (ConvQA) is a task where systems combine dialogue history with evidence (passages, corpora, or knowledge graphs) to answer context-dependent, follow-up queries.
• It employs techniques like History Answer Embedding, positional encoding, and attention mechanisms to effectively integrate prior dialogue and resolve ambiguous queries.
• Recent research extends ConvQA to open-retrieval and heterogeneous-source settings, emphasizing adaptive context management and robust multi-turn evidence integration.

Conversational Question Answering (ConvQA) is the task of answering a current question by combining it with dialogue history and an evidence source, rather than treating each question as an isolated query. In the passage-grounded setting, this evidence is typically a passage; in open-retrieval settings it is a large corpus; in other variants it may be a knowledge graph or a heterogeneous collection of knowledge bases, text, and tables. A common premise across these formulations is that follow-up questions are frequently elliptical, anaphoric, or otherwise underspecified on their own, so the system must recover an evolving information need over multiple turns. For that reason, ConvQA is often treated as a simplified but concrete setting of conversational search rather than merely a multi-turn extension of single-turn QA (Qu et al., 2019, Qu et al., 2019).

1. Task formulations and scope

A standard passage-grounded formulation represents each training instance as $(p, q_k, \mathbf{H}_k, a_k)$, where $p$ is the passage, $q_k$ the current question, $\mathbf{H}_k$ the conversation history, and $a_k$ the ground-truth answer. One influential decomposition separates the problem into three components: a ConvQA model that predicts answers, a history selection module that chooses which past turns to use, and a history modeling module that integrates the selected history into the answering model. In this view, the central technical issue is not only answer extraction, but also how to represent conversational context so that the current information request becomes interpretable (Qu et al., 2019).

Open-retrieval ConvQA adds a retrieval stage. Instead of receiving a gold passage, the system is given the current question, conversational history, and a large corpus $C$, and must first retrieve evidence and then answer. This makes context modeling doubly important: the history must support both interpretation of the current question and retrieval of the right evidence. Weak supervision becomes a major issue in this setting, because gold evidence passages are often unavailable and exact answer-string matching can fail when answers are freeform rather than literal spans (Qu et al., 2021).

The scope of ConvQA has also expanded beyond single passages and homogeneous corpora. One line of work treats ConvQA over knowledge graphs as path reasoning under conversational context, while another treats it as heterogeneous evidence integration over knowledge bases, text, tables, and infoboxes. In the latter case, the same conversational turn may be best answered from different source types depending on whether the needed evidence is relational, textual, or tabular. This broadens ConvQA from a passage-reading problem into a general framework for multi-turn grounded information access (Kacupaj et al., 2022, Christmann et al., 2022).

2. Passage-grounded history modeling

A foundational question in passage-grounded ConvQA is how to inject prior dialogue into a reader without overloading the encoder with concatenated history. A particularly influential answer is History Answer Embedding (HAE), which marks passage tokens according to whether they belong to selected previous answer spans and adds this signal as an extra embedding channel: $$\mathbf{x}_i = \mathbf{e}^{tok}_i + \mathbf{e}^{seg}_i + \mathbf{e}^{pos}_i + \mathbf{e}^{HAE}_i.$$ This design reflects the empirical claim that, in QuAC-style information-seeking conversations, previous answers matter more than previous questions. On QuAC, plain BERT reached 54.4 validation F1, while BERT + PHQA and BERT + PHA reached 62.0 and 61.8; BERT + HAE reached 63.1 validation and 62.4 test F1. HAE was also robust to longer dialogue context, with best performance at about 5–6 history turns, whereas prepending baselines peaked at only 1–2 turns (Qu et al., 2019).

Subsequent work refined this idea in two directions. First, Positional History Answer Embedding (PosHAE) augments answer-span marking with relative turn position. Second, a History Attention Mechanism (HAM) performs soft selection over history turns rather than relying on fixed recency. In HAM, the model encodes one history turn at a time, then learns attention weights over the resulting representations, including a token-level fine-grained variant. On QuAC, HAM achieved 65.7 validation and 64.4 test F1, and HAM with BERT-Large reached 65.4 test F1. Ablations are revealing: removing history attention reduced F1 to 61.1, and replacing PosHAE with the older HAE reduced F1 to 64.2. These results support two claims: relative position matters, and selecting which turn matters now is at least as important as encoding all turns indiscriminately (Qu et al., 2019).

A recurring misconception is that more raw history is automatically beneficial. The passage-grounded literature instead shows a sharper pattern: conversation history is essential, but only when represented in a way that preserves passage budget, emphasizes prior answers, and allows dynamic turn weighting. This suggests that ConvQA history is better treated as structured evidence about the passage than as unbounded appended text (Qu et al., 2019, Qu et al., 2019).

3. Dependency resolution, structured cues, and adaptive context management

Another major line of work addresses conversational dependency by modifying the question representation itself. A prominent critique of full question rewriting is that it can produce verbose standalone questions, remove the conversational character of the input, and create pipeline dependence on rewrite quality. An alternative is to preserve the original question and generate a compact structured cue: $$SR = (\text{context entity} \mid \text{question entity}).$$ In CONVSR, relevant history turns are selected with a soft cosine threshold of 0.75, then the model generates context and question entities as intermediate representations. On QuAC/CANARD-based experiments, CONVSR improved over a standard rewrite-then-answer pipeline for all tested backbones; with RoBERTa, the pipeline reached 66.1 F1 while CONVSR reached 67.9, and removing the question entity slot reduced F1 to 62.8, indicating that explicit question-entity recovery is especially important for resolving ellipsis and anaphora (Zaib et al., 2023).

A more explicitly history-centric alternative is DHS-ConvQA, which performs hard pruning of history turns, attention-based reranking of the retained turns, and binary term classification to keep only useful tokens. Its complete system reached 67.5 F1, 65.3 HEQ-Q, and 7.5 HEQ-D on QuAC, while removing pruning reduced F1 to 64.3. The paper also injected irrelevant negative turns and showed a monotonic decline from 67.5 F1 with no negatives to 52.4 with 11 negatives. This directly supports the claim that irrelevant conversational context is not merely redundant but actively harmful (Zaib et al., 2023).

Recent work has extended this concern from turn relevance to token-budget management. In document-grounded ConvQA with long histories and LLM backbones, Adaptive Context Management (ACM) organizes memory into unmodified recent context $UNC$, summarized middle context $SMC$, and entity-only oldest context $p$0, dynamically reallocating the context window under a token limit. The framework’s Context Manager, Summarization, and Entity Extraction modules were evaluated on a 10% sample of coqa_chat across six backbones. Reported gains were consistent; for example, phi-2 improved from 58.57 to 69.35 F1, and Mistral-7B from 56.31 to 66.71. This suggests that, under modern context-window constraints, ConvQA increasingly depends on memory compression and selective retention rather than raw history concatenation (Perera et al., 22 Sep 2025).

4. Open retrieval, realistic inference, and preference-based RAG

In open-retrieval ConvQA, weak supervision for answer extraction becomes a central problem. A widely used strategy is to search retrieved passages for an exact string match of the known answer, but this fails on freeform conversational answers. A learned weak supervisor instead predicts a paraphrastic answer span in a passage. On OR-QuAC, span-match weak supervision was already sufficient, yielding 23.6 test F1 against 23.1 for learned weak supervision. On OR-CoQA, however, the learned approach became important: span-match yielded 24.3 F1, while combining span-match with learned weak supervision yielded 28.8, reflecting much better coverage for freeform answers (Qu et al., 2021).

Retrieval itself can also benefit from answer history. ConvADR-QA extends dense conversational retrieval by encoding previous question-answer pairs, not just previous questions, in the query representation. On OR-QuAC with an extractive reader, ConvDR $p$1 Reader achieved 36.2 F1, whereas ConvADR-QA reached 38.4; with FiD, ConvDR $p$2 FiD achieved 31.5 and ConvADR-QA 33.6. These gains support the claim that historical answers often supply the lexical anchors needed for later retrieval (Fang et al., 2022).

A related problem appears at inference time even in passage-given ConvQA: prior answers are usually unavailable, so realistic systems must feed back their own predictions. AS-ConvQA shows that using all predicted answers in history is often barely better than omitting answers entirely, because wrong predictions contaminate later turns. The proposed remedy is to keep only prior predicted answers with sufficiently high confidence or sufficiently low uncertainty, optionally after temperature scaling. On QuAC with BERT, the All Pred. baseline reached 55.76 F1, while uncertainty-based AS-ConvQA reached 57.35; with RoBERTa, All Pred. reached 61.53 and confidence-based AS-ConvQA reached 62.47. The underlying point is that realistic ConvQA is partly a dialogue-state filtering problem, not just an answering problem (Jeong et al., 2023).

More recent RAG-style work has turned these stages into an explicit pipeline. PRAISE decomposes ConvQA into Question Understanding, Evidence Retrieval and Filtering, and Answer Generation, all implemented as LLM adapters trained from self-generated preference data. Instead of requiring gold rewrites or gold evidence labels, it treats successful and unsuccessful intermediate generations as preference pairs and applies Direct Preference Optimization. On ConvMix, the full system reached $p$3, $p$4, and $p$5, compared with $p$6, $p$7, and $p$8 for UniHGKR-7B. The gain of 15.5 percentage points in precision indicates that ConvQA pipelines can be improved subtask by subtask even when only final answer supervision is readily available (Kaiser et al., 28 Mar 2025).

5. Knowledge-graph and heterogeneous-source ConvQA

Knowledge-graph ConvQA has often been cast as a reasoning problem under conversational incompleteness. One influential direction uses reinforcement learning from conversational behavior rather than explicit answer labels. CONQUER treats question reformulations as noisy implicit feedback: a reformulation suggests the previous answer was unsatisfactory, whereas a new intent suggests the previous answer was acceptable. It models answering as multiple agents walking in parallel on a KG from context entities and trains with rewards derived from reformulation behavior. On ConvRef, CONQUER achieved up to 0.353 P@1, 0.428 Hit@5, and 0.387 MRR, compared with 0.225, 0.257, and 0.241 for Convex (Kaiser et al., 2021).

A different weakly supervised perspective is to rank candidate KG paths directly. PRALINE formulates ConvQA over KGs as contrastive path ranking, jointly embedding conversation context and KG paths while incorporating full dialog history, fluent responses, and domain information. On ConvQuestions it reached 0.292 P@1, 0.529 H@5, and 0.398 MRR; on ConvRef it reached 0.335 P@1, 0.599 H@5, and 0.441 MRR. These results indicate that rich conversational context can compensate substantially for the absence of gold logical forms, although the method still struggles when gold positive paths are missing (Kacupaj et al., 2022).

Another strand combines reformulation with KG reasoning more directly. CornNet uses a teacher-student architecture in which a teacher learns from human reformulations and a student mimics it using LLM-generated reformulations, with the resulting representation driving RL-based KG traversal. On ConvQuestions, CornNet achieved Hit@5 $p$9 and MRR $q_k$0; on ConvRef, Hit@5 $q_k$1 and MRR $q_k$2. The paper explicitly notes that its strongest gains are on Hit@5 rather than uniformly on MRR, which suggests broader answer coverage rather than consistently sharper top-rank precision (Liu et al., 2023).

ConvQA has also expanded beyond homogeneous KG settings. CONVINSE addresses heterogeneous evidence by learning a source-agnostic structured representation with slots for context entities, question entities, question predicates, and expected answer types, then retrieving from Wikidata, Wikipedia text, tables, and infoboxes. On ConvMix, the All-sources setting reached answer presence 0.542 and P@1 0.342, compared with 0.495 and 0.312 for the best prepend-history baseline. This suggests that heterogeneous-source ConvQA benefits from making the conversational turn explicit in a compact semantic frame before retrieval and answer generation (Christmann et al., 2022).

6. Synthetic data, robust training, and fluent responses

Because ConvQA datasets are expensive to annotate, a substantial body of work focuses on synthetic data and robustness training. CQAG-AR generates conversational QA data from passages with a BERT-based contextual answer extractor and a T5-based question generator that revises the answer after generating the question. On CoQA$q_k$3, ConvQA models trained on CQAG-AR synthetic data reached 83.1/73.8 F1/EM on Wikipedia, compared with 71.3/59.9 for Vanilla CQAG; on unseen-domain DoQA cooking, synthetic data reached 51.5 F1 with 3.7k examples, compared with 45.1 for human-annotated data of the same size. This implies that answer-question consistency is a major determinant of synthetic ConvQA usefulness (Hwang et al., 2022).

Dialogizer extends synthetic generation by training a T5-based dialog generator with dialog reconstruction, question-answer matching, and topic-aware dialog generation, then reranking beam candidates with R$q_k$4UGE. It produces four datasets—WikiDialog2, PubmedDialog, CC-newsDialog, and ElsevierDialog—and reports consistent gains over a dialog-inpainting baseline across automatic and human evaluations. On WikiDialog2, for example, R$q_k$5UGE improved from 2.7579 to 3.8303, and the full QAM + TDG + reranking configuration outperformed all ablations (Hwang et al., 2023).

A closely related development targets open-retrieval ConvQA directly. Synthetic dialogs can be generated from domain documents by first extracting ask-worthy propositions, then generating decontextualized and contextualized question-answer pairs grounded in those propositions. This enables training of lightweight question rewriters or retrievers without any manually annotated in-domain dialogs. In that setting, proposition-based retrieval is markedly better than sentence-based retrieval, and domain-specific synthetic rewrites can beat a T5 rewriter trained on much larger QReCC data; fine-tuned MiniLM retrieval sometimes matches the performance of gold decontextualized queries on synthetic tests and transfers usefully to doc2dial and multidoc2dial test sets (Vlachos et al., 7 Jul 2025).

Robustness to surface variation can also be trained directly. REIGN defines 15 reformulation categories based on insertion, deletion, and substitution over entities, relations, types, and answer-type cues, then uses deep Q-learning to select the reformulations that most improve a target KG-ConvQA model. On ConvMix, Conquer improved from 0.218 to 0.245 P@1 and Explaignn from 0.370 to 0.384 under Reign; on GPT-expanded robustness tests, the framework also increased the number of correctly answered formulations per intent. This suggests that conversational robustness is not just a matter of inference-time rewriting, but of training-time exposure to multiple intent-preserving variants (Kaiser et al., 2023).

Finally, ConvQA is not only about retrieving or extracting correct answers; it is also about responding naturally. Earlier work on fluent response generation treated the input as a question $q_k$6 plus an expert answer phrase $q_k$7, generated candidate responses through syntactic transformations, ranked them with a BERT-based classifier, and then used the resulting data to train end-to-end seq2seq models. The explicit goal was to move from bare spans such as “2011” toward dialogue-appropriate responses such as “It was approved in 2011,” thereby adding an NLG layer to otherwise extractive ConvQA pipelines (Baheti et al., 2020).

Taken together, these strands show that ConvQA has evolved from a narrow extractive benchmark into a family of problems about contextual interpretation, evidence access, and grounded response generation. A plausible synthesis is that future ConvQA systems will continue to combine compact question reformulation, selective and adaptive memory, retrieval over increasingly heterogeneous evidence, and training signals derived from downstream success rather than dense intermediate annotation.