Capturing Global Informativeness in Open Domain Keyphrase Extraction

Open-domain KeyPhrase Extraction (KPE) aims to extract keyphrases from documents without domain or quality restrictions, e.g., web pages with variant domains and qualities. Recently, neural methods have shown promising results in many KPE tasks due their powerful capacity for modeling contextual semantics of the given documents. However, we empirically show that most prefer good phraseness, such as short entity-style n-grams, instead globally informative open-domain This paper presents JointKPE, an architecture built on pre-trained language models, which can capture both local phraseness global informativeness when extracting keyphrases. JointKPE learns rank by estimating entire document is jointly trained keyphrase chunking task guarantee candidates. Experiments two large datasets diverse domains, OpenKP KP20k, demonstrate effectiveness different variants scenarios. Further analyses reveal significant advantages predicting long non-entity keyphrases, are challenging previous methods. Our code publicly available at https://github.com/thunlp/BERT-KPE.