In this section we provide notes on reproducing components that were implemented from scientific publications.
Small-Text is specifically intended to provide a robust set of reusable pre-implemented components which support the reproduction of scientific experiments.
However, in the end the correctness of your experiments is a serious matter and must still be assured:
Never assume the implementation is perfect : We might miss a special case as well (or break something accidentally). In other cases underlying functions from other libraries might have changed / might have a bug.
Never assume default parameters are what you want : It might be possible that this is the case, in most cases we will try to achieve this, but is it your responsibility to verify this. (In cases where a paper describes multiple configurations, the perfect default parameters might just not be possible.)
Nevertheless, using a shared code base (and posing questions / providing feedback on github) will reduce the risk of errors compared to re-implementing these strategies yourself, especially when more and more people will have reviewed the code.
Expected Gradient Length
technically works with all common neural networks. In the context of active learning for
text classification it has been shown to work in combination with the
KimCNN model [ZLW17]
and also with transformer models [EHG+20]. For transformer models, however, this strategy is computationally expensive.
While the distance metric to be used is interchangeable, the original publication [SS17] used euclidean distance.
OverallUncertainty: The original implementation used the full unlabeled set as stopping set. Moreover, it is not stated whether the specified threshold refer to normalized or unnormalized entropy values.