Your suspicion is correct. I worked on such a dataset paper and worked to "beat" other methods on well-accepted benchmarks with dubious accuracy scores. One fun

Your suspicion is correct. I worked on such a dataset paper and worked to "beat" other methods on well-accepted benchmarks with dubious accuracy scores. One fundamental issue is that outside of POS tagging, there isn't a notion of empirical truth to measure against, only a small sample of what a "normal person would think." This stands in contrast to computer vision, whereby in a task such as monocular depth perception from a single frame, you can always measure against a Lidar-scanned depth map. The system can still overfit on the benchmark and they do, but at least the baseline truth itself is not in dispute. But questions such as: is this the "appropriate" response to a query is too open to interpretation.