From Elazar et al.. Evaluated on 2 Winograd corpora (“The X doesn’t fit into Y because it is Z”):
- Winograd Schema Challenge (WSC, 2012) 273 manually curated to minimize influence of priors; e.g. The racecar zoomed by the school bus because it was going so fast (racecars are usually fast).
- WinoGrande (2019), adversarial Winograd
w/ a bunch of train (40k / 9k unbiased), dev (1.2k), test (1.7k) samples.
The best state-of-the-art methods on WinoGrande achieve 59.4-79.1%, which are 15-35% below human performance of 94.0%
Twin sentences in the context of WS:
- The trophy does not fit into the brown suitcase because it is too small;
- The trophy does not fit into the brown suitcase because it is too large
Their explanations for the (falsely?) perceived progress on the WS task:
- lax evaluation criteria; partially alleviate it via group-scoring: model
credited with the worst performing score of a group
For WS, where the task involves a binary classification, we use group scoring over the twin sentences, with accuracy as the per-instance scoring function.
- remaining artifacts in the datasets; alleviate via baseline ctxs (below)
- knowledge and reasoning leakage from large training data; don’t use PLMs (?)
we claim that the vast majority of commonsense knowledge a model obtains should come from sources external to the supervised dataset […] The supervised training set should mainly provide a means for learning the format of the task but not as a source for commonsense knowledge acquisition.
Winograd transformers usually take argmax after replacing it in schemas with
corresponding entities. They test BERT, RoBERTa and ALBERT ([CLS] context [SEP]
entity [SEP]
; entity is e.g. axe/suitcase; output is binary, from [CLS]
):
- Full ctx: The axe would not fit in the suitcase because it’s too large.
- Baseline, no-candidate ctx: would not fit in because it’s too large.
- Baseline, partial sent ctx: because it’s too large.
We note that these two baselines create nonsensical sentences. Therefore, we expect humans to not be able to properly solve them. Thus, a model that achieves higher than random performance on these baselines over a large enough dataset is suspected to rely on spurious correlations.
I mean… Okay, but humans instructed to use their priors could actually properly solve some of these. If I see baseline 2 context, I can just guess.
Final study increases amounts of training data (x-axis) vs. performance (y). ALBERT learning curve is quite steep compared to BERT/RoBERTa (Fig 2). Suggest commonsense knowledge being useful (e.g. steel is hard, relative object sizes); datasets should account for that with careful splits (how? Future work?).
Winogrande debiasing algorithm (AfLite)
- Finetune RoBERTa on a random subset of the data
D1
- Encode rest of the instances
D2
- Train linear classifiers
Li
on random subsetsD2_i
- If more than
k
linearLi
s predict label correctly, discard.
Training details and performance
8 epochs (RTX 2080 for 13 (BERT), 14 (RoBERTa), 62 (ALBERT) min/epoch) on large Winogrande train set & models; Adam (LR of 1e-5); cross-entropy; batch size 8.
As the evaluation is conducted on the dev set, we do not use it to select the best model. Instead, we report the performance with the final model, which is converged based on our observation.