Internet-augmented language models through few-shot prompting for open-domain question answering

Lazaridou, Angeliki; Gribovskaya, Elena; Stokowiec, Wojciech; Grigorev, Nikolai

doi:10.48550/arxiv.2203.05115

Cited by 34 publications

(59 citation statements)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Retrieving sources using a search engine (Lazaridou et al, 2022;Thoppilan et al, 2022) that is kept up-to-date -and supplying them to the language model in a nonparametric fashion -can enable improved temporal generalization over a purely parametric model (Borgeaud et al, 2021;Lewis et al, 2021a;Liska et al, 2022). It also enables the system to attempt questions implying the present date, like "which country got the most medals in the last winter olympics?…”

Section: Gophercitementioning

confidence: 99%

“…Many Question-Answering papers have developed "deep learning"-based retrieval systems with KNN lookups (Borgeaud et al, 2021;Guu et al, 2020;Lewis et al, 2021a). Instead, we follow Komeili et al (2021); Lazaridou et al (2022); ; Thoppilan et al (2022) in calling out to production search engines to find relevant sources, leveraging their access to the entire web, convenience of use, and frequent updates. In particular, we simply forward the input question to Google Search, and show as much context as possible from the resulting documents to the language model.…”

Section: Conditioning and Retrievalmentioning

confidence: 99%

“…Appendix H has our prompt templates and further details. For a more thorough study on prompting language models with search engine results to increase factuality, see Lazaridou et al (2022).…”

Section: Bootstrapping Via Promptingmentioning

confidence: 99%

“…Three closely related pieces of work have recently been released (Lazaridou et al, 2022;Thoppilan et al, 2022). We outline similarities and differences below.…”

Section: Similarities and Differences Compared To Recent Workmentioning

confidence: 99%

“…Similarly to WebGPT, we apply Reinforcement Learning from Human Preferences to improve the quality of our supervised system. Lazaridou et al (2022) do not do any finetuning and rely only on prompting. • Learning to query LaMDA and WebGPT train agents to learn to query a search engine, and can query multiple times for a given input.…”

Section: Similarities and Differences Compared To Recent Workmentioning

confidence: 99%

See 4 more Smart Citations

Teaching language models to support answers with verified quotes

Menick¹,

Trebacz²,

Mikulik³

et al. 2022

Preprint

View full text Add to dashboard Cite

Recent large language models often answer factual questions correctly. But users can't trust any given claim a model makes without fact-checking, because language models can hallucinate convincing nonsense. In this work we use reinforcement learning from human preferences (RLHP) to train "open-book" QA models that generate answers whilst also citing specific evidence for their claims, which aids in the appraisal of correctness. Supporting evidence is drawn from multiple documents found via a search engine, or from a single user-provided document. Our 280 billion parameter model, GopherCite, is able to produce answers with high quality supporting evidence and abstain from answering when unsure. We measure the performance of GopherCite by conducting human evaluation of answers to questions in a subset of the NaturalQuestions and ELI5 datasets. The model's response is found to be high-quality 80% of the time on this Natural Questions subset, and 67% of the time on the ELI5 subset. Abstaining from the third of questions for which it is most unsure improves performance to 90% and 80% respectively, approaching human baselines. However, analysis on the adversarial TruthfulQA dataset shows why citation is only one part of an overall strategy for safety and trustworthiness: not all claims supported by evidence are true.

show abstract

Section: Gophercitementioning

confidence: 99%

Section: Conditioning and Retrievalmentioning

confidence: 99%

Section: Bootstrapping Via Promptingmentioning

confidence: 99%

“…Three closely related pieces of work have recently been released (Lazaridou et al, 2022;Thoppilan et al, 2022). We outline similarities and differences below.…”

Section: Similarities and Differences Compared To Recent Workmentioning

confidence: 99%

Section: Similarities and Differences Compared To Recent Workmentioning

confidence: 99%

See 3 more Smart Citations

Teaching language models to support answers with verified quotes

Menick¹,

Trebacz²,

Mikulik³

et al. 2022

Preprint

View full text Add to dashboard Cite

show abstract

Application of 3D Printing and WebGL-Based 3D Visualisation Technology in Imaging Teaching of Ankle Joints

Dai

Guo

et al. 2021

J. Shanghai Jiaotong Univ. (Sci.)

View full text Add to dashboard Cite

With the rapid development of medical technology, 3D printing technology with realistic representation can perfectly display static human anatomy, while 3D visualisation technology based on Web Graphics Library (WebGL) can promote the rigid replication characteristics of traditional teaching models and express the dynamic spatial relationship between different anatomical structures. Medical students traditionally have less cognition of ankle ligament sprains. In this study, computed tomography (CT) and magnetic resonance imaging (MRI) data of the ankle joints of volunteers were used to print models of the ankle bone, tendon, and ligament using 3D printing technology, and a real-time interactive 3D digital model of the functional ankle joint was designed using 3D visualisation based on WebGL and 2D image real-time rendering technology for interactive teaching. The utility of the 3D printing model combined with the WebGL-based 3D digital teaching model was evaluated in comparison with traditional teaching methods in 24 medical students. The results showed that the total score of students in the experimental group (mean ± SD, 79.48 ± 12.93) was significantly better than that of the control group (61.00 ± 14.94) with P < 0.05. The practical test scores of the experimental group (18.00 ± 2.70) were significantly higher than those of the control group (13.67 ± 4.96) with P < 0.05. In the satisfaction survey, the feedback questionnaire showed that the interactive teaching model of 3D printing technology combined with WebGL-based 3D visualisation technology was recognised by students in terms of quality and overall satisfaction. In addition, female students who used 3D printing combined with WebGL-based 3D visualisation technology as learning aids had a greater difference in practical test scores from the control group than male students. This study has demonstrated that the interactive teaching mode of 3D printing combined with WebGL-based 3D visualisation technology is beneficial to the teaching of medical imaging, enriching the learning experience of students, and increasing the interaction between teachers and students.

show abstract