Eliciting user needs from mass online reviews is playing a significant role in the product iteration process. Efficient user needs elicitation does achieve considerable benefits for maintaining higher competitiveness and a speedier lifecycle. However, there is inevitably an online review scarcity about new products due to the short time on the market and low buyer recognition compared with commonly used products. This paper proposes a small sample data-driven method for user needs elicitation from online reviews in new product iteration. In the first stage, a scraped initial online review dataset is pre-processed roughly to improve the data quality. And then, reviews are classified into multiple categories according to different topics using ERNIE. In the second stage, each topic-based dataset is reprocessed in detail. Thereafter, the key user needs set is determined and facilitated by extracting key product information phrases from every single dataset using improved SIFRank. Moreover, the case study of a smart cat feeder is carried out to demonstrate the feasibility and potential of the ERNIE-ISIFRank methodology. Finally, comparative experiments are conducted to verify the advantages of the proposed method which is primarily based on the pre-trained language model to enhance the deep understanding of the semantics of online reviews. The experimental results confirm that the proposed method can assist in identifying key user needs with high efficiency.