Named entity recognition (NER) is a common task in the field of natural language processing, but it remains more challenging in Chinese due to the lack of natural delimiters. Recently, lots of works incorporate external lexicon into character-level Chinese NER, which focus on how to integrate the matched words in the lexicon into a specific model like LSTM or Transformer. However, in this case, the performance strongly depends on the quality of lexicon and the matching between lexicon and corpora. In reality, there are definitely some noises in the words provided by lexicon, being unhelpful for Chinese NER. To address this issue, in this paper, we propose a simple but effective multi-task learning method with helpful word selection for lexicon-enhanced Chinese NER. One task is to score the matched words and select top-K more helpful ones of them. The other task is to integrate the selected words by multi-head attention network and further implement Chinese NER by character-level sequence labeling. The two tasks are jointly learned with the same encoder. A series of experiments are conducted on three public datasets, demonstrating that the proposed method outperforms the recent advanced baselines.