使用 Seq2seq 学习生成文本到 Hashtag (Text-to-hashtag Generation using Seq2seq Learning)

In this paper, we studied if models based on BiLSTM and BERT can generate hashtags in Brazilian portuguese that can be used in Ecommerce websites. We processed a corpus of Ecommerce reviews and titles of products as inputs and we generated hashtags as outputs. We evaluate the results using four quantitatives metrics: NIST, BLEU, METEOR and a crowdsourced score. Word Cloud was used as a qualitative metric. Besides all computer metered metrics (NIST, BLEU and METEOR) showed bad results, the crowdsourced showed amazing scores. We concluded that the texts generated by the neural networks are very promising to be used as hashtags of products in Ecommerce websites [1]. The code for this work is available on https://github.com/augustocamargo/text-to-hashtag

翻译：在本文中,我们研究了基于BILSTM和BERT的模型能否在巴西的葡萄牙语中产生可用于电子商务网站的标签。我们处理了一系列电子商务评论和产品标题作为投入,我们制作了产品标题作为产出。我们使用四种定量衡量标准评估了结果:NIST、BLEU、METEOR和多源得分。Word Cloud被用作定性衡量标准。除了所有计算机计量计量标准(NIST、BLEU和METEOR)显示的不良结果外,众源也显示了惊人的得分。我们的结论是,神经网络生成的文本极有可能在电子商务网站上用作产品标签[1]。这项工作的代码可在https://github.com/augustocamargo/text-to-hashtag上查阅。