Query-Based Keyphrase Extraction from Long Documents

Transformer-based architectures in natural language processing force input size limits that can be problematic when long documents need to processed. This paper overcomes this issue for keyphrase extraction by chunking the while keeping a global context as query defining topic which relevant keyphrases should extracted. The developed system employs pre-trained BERT model and adapts it estimate probability given text span forms keyphrase. We experimented using various sizes on two popular datasets, Inspec SemEval, large novel dataset. presented results show shorter with longer one without documents.