Most of the zero-shot learning (ZSL) algorithms currently use the pre-trained models trained on ImageNet as their feature extractor, which is considered to be an effective method to improve the feature extraction ability of the ZSL models. However, our research found that this practice is difficult to work well if the training data used by the ZSL task differs greatly from ImageNet. Although one can adapt the pre-trained models to the ZSL task with fine-tuning methods, it turns out that the extractors obtained in this way cannot be guaranteed to be friendly to the unseen classes. To solve these problems, we have further studied a biologically inspired feature enhancement framework for ZSL that we proposed earlier and re-fined its biological taxonomy-based selection method for choosing auxiliary datasets. Moreover, we have proposed a word2vec-based selection strategy as a supplement to the biologically inspired selection method for the first time and experimentally proved the inherent unity of these two methods. Extensive experimental results show that our proposed method can effectively improve the generalization ability of the ZSL model and achieve state-of-the-art results on benchmarks. We have also explained the experimental phenomena through the way of feature visualization.