We propose a novel approach for Estonian lemmatization that enriches the seq2seq neural lemmatization model with lemma candidates generated by the rule-based VABAMORF morphological analyser. In this way, the neural decoder can benefit from the additional input considering that it has a high likelihood of including the correct lemma. We develop our model by stacking two interconnected layers of attention in the decoder—one attending to the input word and another to the candidates obtained from the morphological analyser. We show that the lexicon-enhanced model achieves statistically significant improvements in lemmatization compared to baseline models not utilizing additional lemma information and achieves a new best result on lemmatization on the Estonian UD test set.