Source-linked AI summary
Answerability in Retrieval-Augmented Open-Domain Question Answering
Rustam Abdumalikov, Pasquale Minervini, Yova Kementchedjhieva
TL;DR
ODQA systems can retrieve excerpts that do not contain an answer, while many datasets lack examples for training models to identify such cases. This paper tests random-excerpt training and finds it fails on semantically overlapping irrelevant excerpts, then uses SQuAD 2.0 unanswerable questions to achieve near-perfect performance on challenging generated excerpts.
Problem
Many ODQA datasets lack examples targeting irrelevant retrieved excerpts, although models need to abstain when excerpts do not provide sufficient information.
Source: §1 · p. 1Method
The authors train with SQuAD 2.0 unanswerable questions to recognize semantically related but practically irrelevant excerpts.
Source: §4 · p. 4Results
Near-perfect performance on ChatGPT-generated texts of varying relevance is achieved by training on SQuAD 2.0 unanswerable examples.
Source: §6 · p. 5Takeaways & Limitations
Recognizing when retrieved excerpts lack an answer and abstaining is important for developing trustworthy question-answering systems.
Source: §6 · pp. 4–5Counterfactual data was not used for training because it did not make a difference to answerability in any experimental setting.
Source: §3 · p. 3
Abstract
from arXiv · showhide
The performance of Open-Domain Question Answering (ODQA) retrieval systems can exhibit sub-optimal behavior, providing text excerpts with varying degrees of irrelevance. Unfortunately, many existing ODQA datasets lack examples specifically targeting the identification of irrelevant text excerpts. Previous attempts to address this gap have relied on a simplistic approach of pairing questions with random text excerpts. This paper aims to investigate the effectiveness of models trained using this randomized strategy, uncovering an important limitation in their ability to generalize to irrelevant text excerpts with high semantic overlap. As a result, we observed a substantial decrease in predictive accuracy, from 98% to 1%. To address this limitation, we discovered an efficient approach for training models to recognize such excerpts. By leveraging unanswerable pairs from the SQuAD 2.0 dataset, our models achieve a nearly perfect (~100%) accuracy when confronted with these challenging text excerpts.
1 Introduction
ODQA systems may retrieve excerpts that do not support an answer, yet models trained to always answer may guess rather than abstain. Existing unanswerable data and random-excerpt training leave gaps in studying whether models can recognize irrelevant excerpts that still overlap semantically with the question.
1 Introduction
- ODQA models need to recognize when retrieved excerpts lack answer-relevant information and abstain, rather than guessing and confounding model errors with retrieval deficiencies.Retrieval may be inaccurate or the corpus may lack sufficient knowledge.Source: §1 · p. 1
- Natural Questions labels as unanswerable both questions with no relevant excerpt and questions whose answers cannot be straightforwardly formulated from relevant documents, conflating distinct challenges.The authors focus on the first type: missing relevant excerpts.Source: §1 · p. 1
- Randomly pairing questions with excerpts can yield roughly 99% abstention, but may teach a lack-of-semantic-overlap heuristic unlike realistic excerpts that overlap with questions.The paper investigates training strategies for abstaining on semantically related but practically irrelevant excerpts, and whether correct answers are extracted more often than same-type entities in context.Source: §1 · p. 2
2 Related Work
Prior work has included unanswerable questions in question-answering datasets and studied how annotators identify them. Other work has used randomly paired excerpts to train ODQA models to abstain when relevant answers are absent from context.
2 Related Work
- SQuAD 2.0 added unanswerable questions, which Yatskar (2019) found often involve entity salads, false premises, topic errors, or missing information and are generally easy to detect.Rajpurkar et al. (2018) had annotators write unanswerable questions from presented contexts.Source: §2 · p. 2
- Asai and Choi (2021) studied unanswerability during dataset creation by having annotators mark questions with no answer in the retrieved documents presented to them.Neeman et al. (2022) paired questions with excerpts from the complete pool and labeled them unanswerable to address when models should abstain.Source: §2 · p. 2
3 Methodology
The study uses Neeman et al. (2022)'s augmented Natural Questions data and fine-tunes T5 models, while using ChatGPT to generate excerpts for its research questions. Generated excerpts are for evaluation, and counterfactual training data is excluded.
3 Methodology
- The evaluation set contains 1,365 factual and 1,365 matched counterfactual examples, while the training split has fewer counterfactual than factual examples.The data are based on Neeman et al. (2022)'s answerable NQ questions, augmented with random-context unanswerable examples and same-type counterfactual answers.Source: §3 · p. 2
- Both T5-Large and T5-XL are used, with similar patterns; the main paper reports T5-XL results and Appendix A reports T5-Large results.The models are fine-tuned on the union of factual and answerability training data from Neeman et al. (2022).Source: §3 · pp. 2–3
- ChatGPT generates excerpts as a cheap, scalable alternative to human annotators, and the authors report a 96% success rate for including instructed entities.Prompts request factual-looking paragraphs of one to two sentences; generated excerpts are for evaluation, based on evaluation-split questions and factual answers unless stated otherwise.Source: §3 · p. 3
- Counterfactual data is not used for training because it did not make a difference to answerability in any experimental setting.Source: §3 · p. 3
4 RQ1: Abstaining from Answering
Random-excerpt training does not reliably transfer to semantically related irrelevant excerpts. Training with SQuAD 2.0 unanswerable questions instead achieves flawless performance across the tested types of irrelevant excerpts.
4 RQ1: Abstaining from Answering
- 1.1% is the abstention rate on semantically related excerpts for a model fine-tuned on factual and answerability data, despite near-ceiling performance on random excerpts.With excerpts mentioning the correct answer in another context, abstention decreases by 30% relative to the random-excerpt setting; the authors suggest possible confirmation bias.Source: §4 · p. 3
- For the SQuAD row, including SQuAD 2.0 unanswerable questions yielded flawless performance across irrelevant excerpt types, and the model generalized to highly relevant excerpts designed for this study.Table 2 reports abstention rates on unanswerable data points with random and semantically related excerpts.Source: §4 · p. 4, Table 2 · p. 4
- The authors address the randomization strategy's limited generalization by training on high-semantic-overlap examples using SQuAD 2.0 unanswerable questions.Their prompts generate excerpts that mention relevant question entities in another context or mention the answer without discussing the question or answering it.Source: §4 · pp. 3–4
5 RQ2: Hallucinate or Extract
In the hard-unanswerable setting, models often extracted answers from the excerpt but also produced a substantial number of hallucinations. Extracted answers showed a strong preference for familiar factual answers over counterfactual entities.
5 RQ2: Hallucinate or Extract
- The third excerpt set discussed both the question and answer entity without claiming they were related, using a prompt that requested this separation.The prompt asked for discussion of the question, then a transition to the answer entity without directly answering.Source: §5 · p. 4
- 73.1 (70.7 + 2.4) was the answer-extraction rate, while the hallucination rate was 26.5.Answer origin was assessed by checking for an exact match in the excerpt.Source: Table 3 · p. 4, §5 · p. 4
- Correct factual answers were extracted more than 28 times as often as other context entities, and counterfactual entities were extracted 10 percent less often.The counterfactual comparison replaced the factual answer at the same excerpt position.Source: §5 · p. 4
6 Conclusions
The conclusion identifies recognizing when retrieved text lacks an answer as an ODQA challenge and reports that random excerpt pairing can fail to generalize. It presents unanswerable SQuAD 2.0 examples as an alternative, with near-perfect performance on ChatGPT-generated texts of varying relevance.
6 Conclusions
- Randomized excerpt pairing can fail to generalize effectively when models must recognize that retrieved text does not answer the question.The conclusion frames this recognition and abstention as important when retrieval provides relevant documents without an answer.Source: §6 · p. 4
- Training with unanswerable SQuAD 2.0 questions achieved near-perfect performance on ChatGPT-generated texts of varying relevance.The conclusion describes this as overcoming the limitation of text excerpt randomization.Source: §6 · pp. 4–5
- The conclusion identifies recognizing when retrieved text lacks an answer and abstaining as important to trustworthy question answering.Source: §6 · pp. 4–5
A T5-Large Results
The appendix tables describe T5-Large results for abstention on unanswerable data and for answer extraction, hallucination, or abstention in hard-unanswerable cases.
A T5-Large Results
- Table 4 reports T5-Large abstention rates on unanswerable data points with random versus semantically related excerpts.Source: Table 4 · p. 6
- Table 5 reports T5-Large extraction, hallucination, and abstention rates in a hard-unanswerable setting.Source: Table 5 · p. 6
B Hyperparameters
Fine-tuning used the hyperparameters adopted from Neeman et al. (2022).
B Hyperparameters
- Fine-tuning used batch size 32, learning rate 0.0001, and maximum sequence lengths of 256 for training and 396 for evaluation.Source: B Hyperparameters · p. 6
C Prompting ChatGPT
The prompting instructions standardized answer formatting and sought more varied output, addressing bracket carryover and repetitive openings.
C Prompting ChatGPT
- Prompts explicitly asking ChatGPT to discuss an answer were followed by a fixed sequence of instructions.Source: C Prompting ChatGPT · p. 6
- The instructions requested unquoted, unbracketed answers and more creative discussion when the answer was numeric.Source: C Prompting ChatGPT · p. 6
- Because square brackets emphasized answers in the prompt, ChatGPT tended to enclose them in quotes or brackets; an explicit instruction countered this.Source: C Prompting ChatGPT · p. 6
- The variety instruction targeted repetitive openings such as “This number/digit represents...”.Source: C Prompting ChatGPT · p. 6