The (Un)expected Effects of Applying Standard Cleansing Models to Human Ratings on Compositionality

Human ratings are an important source for evaluating computational models that predict compositionality, but like many data sets of human semantic judgements, are often fraught with uncertainty and noise. However, despite their importance, to our knowledge there has been no extensive look at the effects of cleansing methods on human rating data. This paper assesses two standard cleansing approaches on two sets of compositionality ratings for German noun-noun compounds, in their ability to produce compositionality ratings of higher consistency, while reducing data quantity. We find (i) that our ratings are highly robust against aggressive filtering; (ii) Z-score filtering fails to detect unreliable item ratings; and (iii) Minimum Subject Agreement is highly effective at detecting unreliable subjects.