Probing Factually Grounded Content Transfer with Factual Ablation

Findings of the Association for Computational Linguistics: ACL 2022, pages 3732 - 3746 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Probing Factually Grounded Content Transfer with Factual Ablation Peter West† Chris Quirk‡ Michel Galley‡ Yejin Choi †⋆ †Paul G. Allen School of Computer Science & Engineering, University of Washington ‡ Microsoft Research, Redmond, WA, USA ⋆Allen Institute for Artificial Intelligence {pawest;yejin}cs.washington.edu {chrisq;mgalley}@microsoft.com Abstract Despite recent success, large neural models often generate factually incorrect text. Com- pounding this is the lack of a standard auto- matic evaluation for factuality–it cannot be meaningfully improved if it cannot be mea- sured. Grounded generation promises a path to solving both of these problems: models draw on a reliable external document (grounding) for factual information, simplifying the chal- lenge of factuality. Measuring factuality is also simplified–to factual consistency, testing whether the generation agrees with the ground- ing, rather than all facts. Yet, without a stan- dard automatic metric for factual consistency, factually grounded generation remains an open problem. We study this problem for content transfer, in which generations extend a prompt, using in- formation from factual grounding. Particularly, this domain allows us to introduce the notion of factual ablation for automatically measuring factual consistency: this captures the intuition that the model should be less likely to produce an output given a less relevant grounding docu- ment. In practice, we measure this by present- ing a model with two grounding documents, and the model should prefer to use the more factually relevant one. We contribute two eval- uation sets to measure this. Applying our new evaluation, we propose multiple novel methods improving over strong baselines. 1 Introduction Large pretrained models have shown impressive ef- fectiveness at longstanding tasks and benchmarks. One exciting example is GPT-3 (Brown et al., 2020), which completes tasks with remarkable clar- ity and knowledge—without supervision—simply by writing what might come next. Yet significant challenges prevent these models from helping hu- mans write real documents. For example, in Fig- ure 1 GPT-3 attempts to auto-complete the next sentence of a prompt regarding auto racer Ralph Context Speedway management would maintain their pol-icy to not race on Sundays until 1974 Indianapolis 500—1974. After a heartbreaking loss in 1912 Indianapolis 500—1912, Ralph DePalma succeeds in victory for 1915. DePalma was accompanied by riding mechanic Louis Fontaine. GPT-3 → The 1915 race was the first to have a post-race distance of more than 500 miles GPT-2tuned → Depalma died on march 31, 1915, at his home in south Pasadena, California, of cancer. Grounding SOUTH PASADENA, Calif., March 31 (AP)–Ralph DePalma, pioneer auto racer who probably won more races than any other driver in history, died today of cancer. His age was 72. GPT-2LT→ He was the first driver to win the World War I-era American championship. GPT-2PMI-add → Depalma died of cancer at his home in south Pasadena, California, at the age of 72. Figure 1: Generation with different models continuing a Wikipedia article. GPT-3 has no grounding, while the other 3 models use one …

Content truncated to 500 words