Spelling Correction in Healthcare Query-Answer Systems: Methods, Retrieval Impact, and Empirical Evaluation

Saurabh K Singh

cs.CLFeb 25, 2026

Spelling Correction in Healthcare Query-Answer Systems: Methods, Retrieval Impact, and Empirical Evaluation

Authors: Saurabh K Singh

Organizations: Oracle Corporation

Abstract

Healthcare question-answering (QA) systems face a persistent challenge: users submit queries with spelling errors at rates substantially higher than those found in the professional documents they search. This paper presents the first controlled study of spelling correction as a retrieval preprocessing step in healthcare QA using real consumer queries. We conduct an error census across two public datasets -- the TREC 2017 LiveQA Medical track (104 consumer health questions) and HealthSearchQA (4,436 health queries from Google autocomplete) -- finding that 61.5% of real medical queries contain at least one spelling error, with a token-level error rate of 11.0%. We evaluate four correction methods -- conservative edit distance, standard edit distance (Levenshtein), context-aware candidate ranking, and SymSpell -- across three experimental conditions: uncorrected queries against an uncorrected corpus (baseline), uncorrected queries against a corrected corpus, and fully corrected queries against a corrected corpus. Using BM25 and TF-IDF cosine retrieval over 1,935 MedQuAD answer passages with TREC relevance judgments, we find that query correction substantially improves retrieval -- edit distance and context-aware correction achieve MRR improvements of +9.2% and NDCG@10 improvements of +8.3% over the uncorrected baseline. Critically, correcting only the corpus without correcting queries yields minimal improvement (+0.5% MRR), confirming that query-side correction is the key intervention. We complement these results with a 100-sample error analysis categorising correction outcomes per method and provide evidence-based recommendations for practitioners.

Figures & tables

Metric	TREC LiveQA Medical	HealthSearchQA
Total queries	104	4,436
Queries with $\geq$ 1 error	64 (61.5%)	$\sim$ 88 (est. 1.97%) ∗
Token error rate	11.0%	$\sim$ 2.0% (est.)
Avg errors per query (all)	1.60	—
Avg errors per affected query	2.59	—
Query source	Real user submissions (NLM)	Google autocomplete

Table 1: Spelling error census results. ∗ HealthSearchQA OOV rate (10.9%) calibrated using the TREC LiveQA correction yield ratio (18.1%) to estimate true spelling error rate.

Original	Correction	Error Type	Query Context	Edit Dist.
tabkets	tablets	substitution	Zolmitriptan tabkets 5mg	1
hydrslazine	hydralazine	deletion	blood pressure medication	1
citrobactor	citrobacter	substitution	UTI bacterium	1
osella	ocella	substitution	oral contraceptive	1
precaution	precautions	deletion	disease management	1
hydrslazine	hydralazine	double_deletion	antihypertensive	2

Table 2: Representative real spelling errors in TREC LiveQA Medical queries with ground-truth corrections from NIST paraphrases.

Condition	Method	MRR	R@1	R@5	R@10	NDCG@10
Exp 2: Baseline	—	0.633	0.079	0.253	0.424	0.436
Exp 3a: Orig Q $\times$ Corr C	conservative	0.636	0.079	0.258	0.428	—
Exp 3a: Orig Q $\times$ Corr C	edit distance	0.636	0.079	0.256	0.427	—
Exp 3b: Corr Q $\times$ Orig C	conservative	0.678	0.086	0.276	0.457	0.472
Exp 3b: Corr Q $\times$ Orig C	edit distance	0.687	0.086	0.276	0.453	0.474
Exp 4: Both Corrected	conservative	0.682	0.088	0.278	0.456	0.473

Table 3: BM25 retrieval results across all experimental conditions (103 queries, 1,935 passages, TREC relevance judgements). Bold indicates best result per metric. NDCG@10 values for Experiment 3a are omitted as they were not computed for the corpus-only condition.

Condition	Method	MRR	R@1	R@5	R@10	NDCG@10
Exp 2: Baseline	—	0.667	0.076	0.316	0.512	0.461
Exp 3a: Orig Q $\times$ Corr C	conservative	0.656	0.077	0.311	0.515	—
Exp 3b: Corr Q $\times$ Orig C	conservative	0.728	0.082	0.331	0.537	0.492
Exp 3b: Corr Q $\times$ Orig C	edit distance	0.730	0.082	0.332	0.539	0.496
Exp 4: Both Corrected	conservative	0.717	0.080	0.330	0.540	0.491
Exp 4: Both Corrected	edit distance	0.715	0.080	0.331	0.544	0.495

Table 4: TF-IDF cosine similarity retrieval results (proxy for dense embedding retrieval). Note: TF-IDF is a vocabulary-based method, not a neural embedding approach; results should not be extrapolated to BERT-based dense retrieval. NDCG@10 values for Experiment 3a are omitted as they were not computed for the corpus-only condition.

Category	Conserv.	Edit Dist.	Context	SymSpell	Definition
Total corrections	90	180	180	66	—
correct_fix (%)	18.9	12.0	9.0	12.1	Corrected to paraphrase’s word
partial_improv. (%)	70.0	70.0	77.0	86.4	Changed but not to paraphrase word
unnecessary_change (%)	8.9	16.0	12.0	1.5	Original word was in paraphrase
harmless_synonym (%)	2.2	2.0	2.0	0	Both original and correction in paraphrase

Table 5: Categorisation of spelling corrections per method (sample of

\leq

100 corrections per method; conservative sample = full 90). Ground truth = NIST paraphrase vocabulary.

Medical QA

Medical question answering develops and evaluates AI systems that answer clinical, patient-facing, and health-information questions.

See all 84 papers

Healthier LLMs: Retrieval-Augmented Generation for Public Health Question Answering
Jul 7, 2026Felix Feldman, Joshua Harris, Timothy Laurence +7Retrieval-Augmented Generation Question Answering
GUIDE: Generative Unsupervised Chinese Query Correction via Phonetic and Visual Shared-ID Encoding
Aug 26, 2026Lei Yang, Binbin Huang, Jiwei Tan +4Query Reformulation Information Retrieval
Kuaishou Technology · School of Data Science, Fudan University
MedRECT: A Bilingual Medical Reasoning Benchmark for Error Correction in Clinical Texts
Nov 1, 2025Naoto Iwase, Hiroki Okuyama, Junichiro IwasawaMultilingual Language Model Evaluation Healthcare
Preferred Networks, Inc., Tokyo, Japan · School of Medicine, Nagoya University, Nagoya, Japan

Jul 7, 2026J/K move · Enter open · S save

Healthier LLMs: Retrieval-Augmented Generation for Public Health Question Answering

Felix Feldman, Joshua Harris, Timothy Laurence, Leo Loman+6

Retrieval-Augmented Generation Question Answering Hybrid IR RAG Evaluation

Open paper AskPDF arXiv

Large language models (LLMs) achieve promising results on medical question answering benchmarks, yet their use in public health is constrained by hallucinations and the rapid evolution of official guidance. Retrieval-Augmented Generation (RAG) mitigates these risks by grounding responses in an explicitly maintained corpus, but end-to-end performance depends critically on retrieval configuration and on evaluation beyond multiple-choice formats. We extend PubHealthBench, a question answering (QA) benchmark of 7,929 questions derived from UK Government public health guidance, into a retrieval-augmented setting and systematically evaluate retrieval and generation choices. We compare dense, sparse, and hybrid retrieval across multiple embedding models and corpus variants, and show that hybrid retrieval consistently improves recall and ranking quality, with chunk length and topic interacting with ranking performance. Providing retrieved context substantially increases multiple-choice accuracy across a diverse set of LLMs, enabling smaller open-weight models to match or outperform larger models used without retrieval, with gains primarily driven by retrieval quality and careful context selection. To assess realistic free-form answering, we introduce a rubric-based LLM-as-a-judge covering faithfulness, completeness, clarity, and factual consistency, and validate it against dual human annotations. Judge-human agreement is strongest for faithfulness and completeness, while factual consistency and clarity are less reliably reproduced, motivating caution when interpreting those dimensions at scale. Overall, our results highlight retrieval as a primary lever for reliable public health QA and provide practical guidance for building and evaluating RAG systems grounded in official guidance.

Abstract

Medical QA

Explore similar work

Healthier LLMs: Retrieval-Augmented Generation for Public Health Question Answering

GUIDE: Generative Unsupervised Chinese Query Correction via Phonetic and Visual Shared-ID Encoding

MedRECT: A Bilingual Medical Reasoning Benchmark for Error Correction in Clinical Texts