[Objective] This paper aims to automatically recognize key sentences describing the research topics of scientific papers. [Methods] First, we used paper sections as the unit to organize sentence sets. Then, we calculated the WMD distance between sentences by trained domain word embeddings. Third, we optimized the iterative process of TextRank algorithm, and used external features to adjust sentence's weights. Finally, we identified the core topic sentences according to the sentence's weights descendingly. [Results] We examined the proposed method with scientific papers on climate changes and compared it with the traditional TextRank algorithm. The recognition efficiency (F-value) was about 5% higher than that of the TextRank algorithm. [Limitations] The extraction of sentence features needs to be improved, and word embedding training and related parameters of the proposed method need to be further optimized. [Conclusions] The improved TextRank algorithm, could effectively recognize inner core sentences of scientific paper sections. It could recognize core topic sentences of a paper with the adjusted weights of external features.