Representation Learning for Resource-Constrained Keyphrase Generation
share
State-of-the-art keyphrase generation methods generally depend on large annotated datasets, limiting their performance in domains with constrained resources. To overcome this challenge, we investigate strategies to learn an intermediate representation suitable for the keyphrase generation task. We introduce salient span recovery and salient span prediction as guided denoising language modeling objectives that condense the domain-specific knowledge essential for keyphrase generation. Through experiments on multiple scientific keyphrase generation benchmarks, we show the effectiveness of the proposed approach for facilitating low-resource and zero-shot keyphrase generation. Furthermore, we observe that our method especially benefits the generation of absent keyphrases, approaching the performance of SOTA methods trained with large training sets.
READ FULL TEXT