PaperlayerView exact revision on arXiv

Source-linked AI summary

OLAPH: Improving Factuality in Biomedical Long-form Question Answering

Minbyul Jeong, Hyeon Hwang, Chanwoong Yoon, Taewhoo Lee, Jaewoo Kang

arXiv:2405.12701v3cs.CLcs.AI

TL;DR

Patient-facing medical long-form answers need factual claims, yet their automatic evaluation is difficult. The paper introduces MedLFQA and OLAPH, and reports improved factuality, including 7B answers comparable to medical experts. These findings support the potential of appropriately trained 7B models to produce long-text medical answers at a desirable level.

  • Problem

    Patient-facing medical long-form answers need crucial factual claims without hallucinations, while automatic factuality evaluation remains difficult and human review is costly.

    Source: §1 · pp. 1–2

  • Method

    The paper reconstructs MedLFQA with answers and crucial statements, then uses multifaceted automatic evaluation to build synthetic preference sets for iterative OLAPH training.

    Source: §1 · p. 2

  • Results

    OLAPH-trained 7B models improve factuality, including on metrics not used in training, and produce long answers comparable to medical experts’ answers in factuality.

    Source: §1 · p. 2, §7 · p. 11

  • Takeaways & Limitations

    With appropriate data and methods, 7B models can produce long-text medical answers at a desirable level, while personalized physician-assistance conversations remain a future possibility.

    Source: §7 · p. 11

    The analysis compares 7B-parameter models; the method’s effectiveness remains to be assessed with smaller and larger models.

    Source: §7 · p. 11

Abstract

from arXiv · show

In the medical domain, numerous scenarios necessitate the long-form generation ability of large language models (LLMs). Specifically, when addressing patients' questions, it is essential that the model's response conveys factual claims, highlighting the need for an automated method to evaluate those claims. Thus, we introduce MedLFQA, a benchmark dataset reconstructed using long-form question-answering datasets related to the biomedical domain. We use MedLFQA to facilitate a cost-effective automatic evaluations of factuality. We also propose OLAPH, a simple and novel framework that utilizes cost-effective and multifaceted automatic evaluation to construct a synthetic preference set and answers questions in our preferred manner. Our framework leads us to train LLMs step-by-step to reduce hallucinations and include crucial medical claims. We highlight that, even on evaluation metrics not used during training, LLMs trained with our OLAPH framework demonstrate significant performance improvement in factuality. Our findings reveal that a 7B LLM trained with our OLAPH framework can provide long answers comparable to the medical experts' answers in terms of factuality. We believe that our work could shed light on gauging the long-text generation ability of LLMs in the medical domain. Our code and datasets are available.

1 INTRODUCTION

Medical long-form answers must convey crucial factual claims without hallucinations, but evaluating them automatically is difficult. The paper introduces MedLFQA and OLAPH, reporting improved factuality and 7B answers comparable to medical experts.

1 INTRODUCTION

  • Patient-facing medical LFQA requires factual, crucial claims while avoiding hallucinations, and existing automatic measures or human review leave evaluation difficult or costly.The paper identifies human annotation as labor-intensive and cost-ineffective.Source: §1 · pp. 1–2
  • MedLFQA restructures biomedical LFQA datasets to evaluate factuality, semantic similarity, and word composition, using answers and statements to assess crucial claims.The benchmark includes GPT-4-generated answers and statements where needed, and medical experts evaluated them pairwise.Source: §1 · p. 2
  • OLAPH constructs synthetic preference sets through cost-effective, multifaceted automatic evaluation and iteratively trains models on preferred responses.The process starts with supervised fine-tuning, then samples model predictions and repeats preference construction and alignment tuning.Source: §1 · p. 2
  • 7B models trained with OLAPH generate long-form answers comparable to medical experts’ answers, with improvements also reported on factuality metrics not used during training.The paper also reports gains in semantic similarity and word composition.Source: §1 · p. 2

2 PRELIMINARIES

Biomedical LFQA addresses patient and consumer questions with elaborate answers, but limited patient-centered data and difficult automatic evaluation constrain research. The section reviews evaluation approaches and MedLFQA’s multifaceted assessment.

2.1 LONG-FORM QUESTION ANSWERING

  • Biomedical LFQA uses elaborate, in-depth answers to open-ended questions, while patient-centered datasets remain relatively limited because usable data are restricted by privacy concerns.The paper studies existing benchmarks built around questions consumers or patients typically ask.Source: §2.1 · p. 3

2.2 EVALUATING LONG-TEXT GENERATION

  • Long-text evaluation spans word overlap, semantic similarity, and factuality, but comprehensive LFQA evaluation remains difficult.Prior work also relies on human verification, while Figure 1 presents MedLFQA answers and statements for multifaceted assessment of factuality, semantic similarity, and word composition.Source: §2.2 · p. 3, Figure 1 · p. 4
  • HALLUCINATION measures clinical claims contradicted by a response, while COMPREHENSIVENESS measures crucial claims included in it.The paper uses BioBERT trained on NLI datasets to predict response entailment instead of GPT-3.5-turbo because of API costs.Source: §2.2 · p. 3

3 MEDLFQA: RECONSTRUCTION AND QUALIFICATION

MedLFQA reconstructs biomedical long-form question-answering datasets to support automatic factuality evaluation. It also qualifies GPT-4-generated answers as a source for datasets lacking expert answers.

3 MEDLFQA: RECONSTRUCTION AND QUALIFICATION

  • MedLFQA reconstructs existing biomedical LFQA datasets to evaluate whether responses convey factual claims, using two statements for automatic assessment.The section describes how the datasets and statements are constructed and how generated answers and statements are qualified.Source: §3 · p. 4, §3.1 · p. 4

3.1 RECONSTRUCTION OF BIOMEDICAL LONG-FORM QUESTION-ANSWERING DATASETS

  • The K Health source contains 1,212 patient questions, but only 202 question-answer pairs with expert-curated long-form answers.Its Must Have statements mark content required for medical accuracy, while Nice to Have statements provide supplemental information.Source: §3.1 · p. 4
  • MedLFQA combines patient questions, long-form answers, Must Have statements, and Nice to Have statements from four biomedical LFQA datasets.For datasets without answers, the authors generate answers and statements with GPT-4; where expert answers exist, they decompose those answers.Source: §3.1 · pp. 4–5

3.2 QUALIFICATION OF GENERATED ANSWERS AND STATEMENTS

  • Medical experts judged GPT-4 answers better than expert annotations on reasoning, irrelevant content, omissions, and possible harm likelihood.The authors use the generated answers to derive statements for automatic evaluation of long-form predictions.Source: §3.2 · p. 5, Figure 2 · p. 5

4 HOW TO TRAIN OLAPH?

OLAPH trains LLMs through SFT followed by iterative preference optimization using self-generated answers and multifaceted automatic evaluation. Its process selects preferred responses and trains models to favor them over dispreferred responses.

4 HOW TO TRAIN OLAPH?

  • OLAPH iteratively samples answers, scores them, forms preferred and dispreferred pairs, and applies preference optimization such as DPO.The framework begins with SFT and repeats preference-set construction and training until convergence.Source: §4 · p. 6, Figure 3 · p. 7

4.1 SUPERVISED FINE-TUNING

  • SFT uses a top-scoring self-generated response as the labeled answer for the next training step.The question-answer collection includes Must Have and Nice to Have statements, and the SFT stage familiarizes the model with the task.Source: §4.1 · p. 6

4.2 COST-EFFECTIVE AND MULTIFACETED AUTOMATIC EVALUATION

  • Sampled predictions are ranked using word-composition, semantic-similarity, and factuality criteria, with the hallucination score deducted.A predetermined threshold separates the preferred high-score set from the dispreferred low-score set.Source: §4.2 · p. 6

4.3 DIRECT PREFERENCE OPTIMIZATION

  • DPO trains the student model to favor preferred responses over dispreferred responses from the collected preference sets.The student starts from the SFT model, while β controls deviation from that model.Source: §4.3 · p. 7

4.4 ITERATIVE LEARNING WITH SELF-GENERATED PREFERENCE SET

  • OLAPH iteratively builds preference sets from self-generated responses and trains until convergence, avoiding dependence on human-annotated preference data.The authors state that multifaceted evaluation improves long-form answer quality, including on evaluation metrics not used during training.Source: §4.4 · p. 7

5 EXPERIMENTAL SETTINGS

Because MedLFQA has no separate training sets, the experiments hold out one dataset for testing and train on the remaining datasets.

5 EXPERIMENTAL SETTINGS

  • Each of the five MedLFQA datasets serves as a test set in turn, with the remaining datasets used for training.The authors describe this as the same concept as cross-validation.Source: §5 · p. 8

6 EXPERIMENTS & ANALYSIS

The experiments compare clinical-answering performance across open and proprietary LLMs, then test whether iterative OLAPH training improves 7B models. BioMistral’s factuality improves through iterative learning, including on FACTSCORE, which was not used during training.

RQ 1

  • Base LLaMA2 and Mistral answer properly on some datasets but inconsistently, with lower factuality despite preserving word-composition and semantic-similarity scores.The evaluation is zero-shot and uses MedLFQA; Table 3 covers three proprietary models, but its caption supplies no result values.Source: §6 · pp. 8–9, Table 3 · p. 9
  • Meditron and BioMistral show very low factuality, while Self-BioRAG consistently performs well on medical questions.The authors attribute the differing pattern to instruction data: Self-BioRAG was trained on diverse instructions including long-form question answering.Source: §6 · p. 9
  • Proprietary LLMs perform remarkably well on clinical long-form responses relative to open-foundation models, but cannot be reached to elicit or update their knowledge through training.The authors therefore focus subsequent OLAPH experiments on open-source models under 7B parameters.Source: §6 · p. 9

RQ 2

  • After iterative DPO training at Step 3, BioMistral 7B reaches the GPT-4 upper-bound performance on the K-QA Golden evaluation.Performance tends to saturate after DPO training at Step 2.Source: §6 · pp. 9–10, Figure 4 · p. 9

RQ 3

  • FACTSCORE factuality rises substantially from Step 0 to Step 3, with OLAPH-trained 7B models reaching the medical-expert upper-bound performance.FACTSCORE is not used during training; Figure 5 reports it without a length penalty and includes GPT-4 as an upper bound.Source: §6 · p. 10, Figure 5 · p. 10

7 CONCLUSION, LIMITATIONS, AND FUTURE WORK

The paper presents MedLFQA for cost-effective evaluation and OLAPH for training models to produce more factual, semantically similar, and coherent medical answers. Its analysis focuses on 7B models and notes several boundaries for extending the findings.

Conclusion

  • OLAPH combines SFT, preference-set construction from multifaceted automatic evaluation, and iterative alignment tuning to improve 7B answers’ factual accuracy, semantic similarity, and word composition.MedLFQA is reconstructed in a unified format with long-form answers and crucial statements for cost-effective evaluation.Source: §7 · pp. 10–11

Limitations

  • The analysis compares 7B-parameter models; the authors say smaller and larger models must be considered to determine whether the findings generalize.They note smaller models may be less likely to generate correct predictions, making sampled responses less suitable.Source: §7 · p. 11
  • MedLFQA’s biomedical knowledge is fixed to a timestamp, which could make it outdated in the future.Source: §7 · p. 11
  • The authors identify possible evaluation error propagation from using trained NLI models.They frame the approach as aiming to replace substantial API costs with cost-effective evaluation metrics.Source: §7 · p. 11

Future Work

  • Future work could test whether 7B or larger models can use patient history and multi-turn conversations to assist physicians in personalized settings.Source: §7 · p. 11

A.1 CRITERIA OF GENERATED ANSWER EVALUATION

The appendix describes a fine-grained, pairwise evaluation of GPT-4 and medical-expert answers using nine criteria. Medical experts’ agreement is assessed across the criteria.

A.1 Criteria of Generated Answer Evaluation

  • The evaluation uses nine criteria, with the first four preferred when selected and the other five preferred when not selected.The criteria follow prior work, and their details are provided in Table 4.Source: §A.1 · p. 15, Table 4 · p. 16
  • Agreement is counted when at least two of three medical experts choose the same answer; the authors report high agreement that GPT-4 responses are highly usable.Source: §A.1 · p. 15, Table 4 · p. 16

A.2 SENSITIVITY ANALYSIS OF HYPERPARAMETERS

The sensitivity analysis examines how factuality weighting and preference thresholds affect OLAPH’s training setup. The reported pattern links higher factuality weight to improved factuality scores, while iterative DPO is associated with reduced score variability.

Sensitivity Analysis of Hyperparameters

  • The sensitivity experiments use BioMistral 7B and Self-BioRAG 7B, training on four datasets and evaluating on KQA-Golden.The training datasets are LiveQA, MedicationQA, HealthSearchQA, and KQA-Silver.Source: §A.2 · p. 16
  • Increasing α3 correlates with improved factuality scores, while setting α3 to 0 yields minimal factuality changes despite gains in word composition and semantic similarity.The analysis varies α3 from 0 to 1 in increments of 0.2, with α1 and α2 fixed at 1.Source: §A.2 · p. 16
  • Iterative DPO training reduces the standard deviation in overall scores, indicating more reliable answers as training progresses.Source: §A.2 · p. 16
  • The Equation 4 threshold separates preferred from dispreferred responses using automatic-evaluation scores, and the authors test thresholds from 0 to 200.The threshold was set at 200 after manual evaluation of responses according to the authors’ preferences.Source: §A.2 · pp. 16–17
  • The authors normalize metric scales because factuality scores range from -100 to 100 while ROUGE scores range from 0 to 1.The passage also describes scaling ROUGE, BLEURT, and BERTScore to align their ranges.Source: §A.2 · p. 16

A.3 EXPERIMENTAL DETAILS

OLAPH training used multi-GPU infrastructure, specified optimization settings, and sampled predictions for training and inference.

A.3 EXPERIMENTAL DETAILS

  • Training used 8 Nvidia A100 GPUs with 80GB memory, PyTorch, HuggingFace, DeepSpeed stage 3, and FlashAttention.The implementation used DeepSpeed for multi-GPU settings and FlashAttention for efficient training.Source: §A.3 · p. 17
  • The initial step used a 5e-7 learning rate and 0.1 warmup ratio; after Step-2 DPO training, the learning rate was 1e-7 with β=0.01.The β value was used for DPO learning.Source: §A.3 · p. 17
  • Prediction generation used k=6 outputs—one deterministic at τ = 0 and five sampled at τ = 1.0—with a 200 threshold for long-form response quality.Inference used vLLM to speed up inference time.Source: §A.3 · p. 17

A.4 DETAIL PERFORMANCE OF OLAPH FRAMEWORK

The appendix details OLAPH’s evaluation metrics and iterative performance patterns, its biomedical evidence retrieval and entailment model, and the rationale and results for sampling-based predictions.

A.4 DETAIL PERFORMANCE OF OLAPH FRAMEWORK

  • Factuality and semantic-similarity scores generally increase across OLAPH steps, while LLaMA2’s initially higher scores decline after SFT and may reflect repetitive answers.The authors caution that one automatic metric cannot capture every aspect of responses and should be checked against another metric or a human evaluator.Source: §A.4 · pp. 17–19, Figure 6 · p. 19

A.5 BIOMEDICAL KNOWLEDGE SOURCE & DOMAIN-SPECIFIC RETRIEVER

  • For biomedical factuality assessment, the authors supplement FACTSCORE’s Wikipedia-based atomic-fact support with biomedical documents retrieved using Self-BioRAG; the corpus includes PubMed abstracts, PMC full text, clinical guidelines, and English medical textbooks.The appendix describes using the off-the-shelf MedCPT retriever and reranking retrieved evidence.Source: §A.5 · pp. 19–20

A.6 DETAILS OF ENTAILMENT MODEL

  • The entailment model uses BioBERT fine-tuned on MultiNLI, SNLI, and MedNLI, with the authors aiming to avoid substantial API costs while using models accessible to medical researchers.The model assesses how much of two statements is included in a model response.Source: §A.6 · p. 20

A.7 WHY DO WE USE SAMPLING-BASED PREDICTIONS?

  • In Figure 7, sampled responses with the highest scores surpass deterministic responses across evaluation metrics, and the selected responses supply labels and preference sets for iterative training.The figure groups metrics by category: pink for word composition, yellow for semantic similarity, and blue for factuality; its left and right sides show deterministic and best-scoring sampled responses, respectively.Source: §A.7 · p. 20, Figure 7 · p. 20
Loading 2405.12701v3…