Matches in SemOpenAlex for { <https://semopenalex.org/work/W3212363653> ?p ?o ?g. }
- W3212363653 endingPage "5102" @default.
- W3212363653 startingPage "5087" @default.
- W3212363653 abstract "Exposure bias has been regarded as a central problem for auto-regressive language models (LM). It claims that teacher forcing would cause the test-time generation to be incrementally distorted due to the training-generation discrepancy. Although a lot of algorithms have been proposed to avoid teacher forcing and therefore alleviate exposure bias, there is little work showing how serious the exposure bias problem actually is. In this work, we focus on the task of open-ended language generation, propose metrics to quantify the impact of exposure bias in the aspects of quality, diversity, and consistency. Our key intuition is that if we feed ground-truth data prefixes (instead of prefixes generated by the model itself) into the model and ask it to continue the generation, the performance should become much better because the training-generation discrepancy in the prefix is removed. Both automatic and human evaluations are conducted in our experiments. On the contrary to the popular belief in exposure bias, we find that the the distortion induced by the prefix discrepancy is limited, and does not seem to be incremental during the generation. Moreover, our analysis reveals an interesting self-recovery ability of the LM, which we hypothesize to be countering the harmful effects from exposure bias." @default.
- W3212363653 created "2021-11-22" @default.
- W3212363653 creator A5000092497 @default.
- W3212363653 creator A5039081803 @default.
- W3212363653 creator A5051747323 @default.
- W3212363653 creator A5064766573 @default.
- W3212363653 date "2021-11-01" @default.
- W3212363653 modified "2023-09-23" @default.
- W3212363653 title "Exposure Bias versus Self-Recovery: Are Distortions Really Incremental for Autoregressive Text Generation?" @default.
- W3212363653 cites W1515851193 @default.
- W3212363653 cites W1861492603 @default.
- W3212363653 cites W2064675550 @default.
- W3212363653 cites W2078861931 @default.
- W3212363653 cites W2099471712 @default.
- W3212363653 cites W2101105183 @default.
- W3212363653 cites W2133459682 @default.
- W3212363653 cites W2402268235 @default.
- W3212363653 cites W2542835211 @default.
- W3212363653 cites W2581637843 @default.
- W3212363653 cites W2616969219 @default.
- W3212363653 cites W2620623908 @default.
- W3212363653 cites W2742947407 @default.
- W3212363653 cites W2757836268 @default.
- W3212363653 cites W2798025142 @default.
- W3212363653 cites W2807747378 @default.
- W3212363653 cites W2890969459 @default.
- W3212363653 cites W2908747729 @default.
- W3212363653 cites W2963084599 @default.
- W3212363653 cites W2963096510 @default.
- W3212363653 cites W2963141266 @default.
- W3212363653 cites W2963163972 @default.
- W3212363653 cites W2963248296 @default.
- W3212363653 cites W2963403868 @default.
- W3212363653 cites W2963456134 @default.
- W3212363653 cites W2963595537 @default.
- W3212363653 cites W2963620441 @default.
- W3212363653 cites W2963730239 @default.
- W3212363653 cites W2963951265 @default.
- W3212363653 cites W2964110616 @default.
- W3212363653 cites W2964268978 @default.
- W3212363653 cites W2964345285 @default.
- W3212363653 cites W2973049837 @default.
- W3212363653 cites W2981323647 @default.
- W3212363653 cites W2985008383 @default.
- W3212363653 cites W2985308740 @default.
- W3212363653 cites W2995404354 @default.
- W3212363653 cites W2996068536 @default.
- W3212363653 cites W2996287690 @default.
- W3212363653 cites W2998277219 @default.
- W3212363653 cites W3020264409 @default.
- W3212363653 cites W3022463379 @default.
- W3212363653 cites W3035239386 @default.
- W3212363653 cites W3035252911 @default.
- W3212363653 cites W3099180151 @default.
- W3212363653 cites W3114346701 @default.
- W3212363653 cites W3156476125 @default.
- W3212363653 cites W3175092301 @default.
- W3212363653 cites W648786980 @default.
- W3212363653 hasPublicationYear "2021" @default.
- W3212363653 type Work @default.
- W3212363653 sameAs 3212363653 @default.
- W3212363653 citedByCount "0" @default.
- W3212363653 crossrefType "proceedings-article" @default.
- W3212363653 hasAuthorship W3212363653A5000092497 @default.
- W3212363653 hasAuthorship W3212363653A5039081803 @default.
- W3212363653 hasAuthorship W3212363653A5051747323 @default.
- W3212363653 hasAuthorship W3212363653A5064766573 @default.
- W3212363653 hasConcept C119857082 @default.
- W3212363653 hasConcept C132010649 @default.
- W3212363653 hasConcept C134306372 @default.
- W3212363653 hasConcept C138885662 @default.
- W3212363653 hasConcept C141603448 @default.
- W3212363653 hasConcept C146849305 @default.
- W3212363653 hasConcept C149782125 @default.
- W3212363653 hasConcept C154945302 @default.
- W3212363653 hasConcept C15744967 @default.
- W3212363653 hasConcept C159877910 @default.
- W3212363653 hasConcept C162324750 @default.
- W3212363653 hasConcept C187736073 @default.
- W3212363653 hasConcept C188147891 @default.
- W3212363653 hasConcept C197115733 @default.
- W3212363653 hasConcept C2776436953 @default.
- W3212363653 hasConcept C2780451532 @default.
- W3212363653 hasConcept C2985684807 @default.
- W3212363653 hasConcept C33923547 @default.
- W3212363653 hasConcept C41008148 @default.
- W3212363653 hasConcept C41895202 @default.
- W3212363653 hasConceptScore W3212363653C119857082 @default.
- W3212363653 hasConceptScore W3212363653C132010649 @default.
- W3212363653 hasConceptScore W3212363653C134306372 @default.
- W3212363653 hasConceptScore W3212363653C138885662 @default.
- W3212363653 hasConceptScore W3212363653C141603448 @default.
- W3212363653 hasConceptScore W3212363653C146849305 @default.
- W3212363653 hasConceptScore W3212363653C149782125 @default.
- W3212363653 hasConceptScore W3212363653C154945302 @default.
- W3212363653 hasConceptScore W3212363653C15744967 @default.
- W3212363653 hasConceptScore W3212363653C159877910 @default.
- W3212363653 hasConceptScore W3212363653C162324750 @default.