Extractive Summarization with Text Generator

Standard extractive systems suffer from the lack of gold training signals since existing corpora solely provide document and human-written summary pairs while disregarding extractive labels. As a result, existing methods resort to imperfect pseudo-labels that are both biased and error-prone, thereby hindering the learning process of extractive models. In contrast, text generators which are commonly employed in abstractive summarization can effortlessly overcome this predicament on account of flexible sequence-to-sequence architectures. Motivated to bypass this inherent limitation, we investigate the possibility of conducting extractive summarization with text generators. Through extensive experiments covering six summarization benchmarks, we show that high-quality extractive summaries can be assembled via approximating the outputs (abstractive summaries) of these generators. Moreover, we find that the approximate summaries correlate positively with the auxiliary summaries (i.e. better generation model enables the production of better extractive summaries). Our results signify a new paradigm for training extractive summarizers i.e. learning with generation (abstractive) objectives rather than extractive schemes.