Machines have a tougher time summarizing text than you’d think, at least where the summarization is abstractive rather than extractive. While the extraction requires merely concatenating sentences, abstraction involves the task of paraphrasing using novel sentences. Progress has been made in the news domain recently, perhaps owing to the abundance of corpora on which algorithmic systems can be trained. But robust summarization of most other writing forms remains an unsolved problem.

Motivated by this, a team at Google Brain investigated an abstractive summarization system dubbed SummAE that’s largely unsupervised, meaning it’s able to generalize from a small amount of training data to unseen textual examples. While it couldn’t summarize beyond single five-sentence paragraphs, the researchers claim it “significantly” improves upon the baseline and represents a “major” step in the direction of human-level performance.

The data set and code are freely available on GitHub, along with the configuration settings for the best model.

“As one of the very first works approaching single-document [abstract summarization], we propose a novel neural model — SummAE,” wrote the coauthors. “[We believe it] is therefore desirable to have models capable of automatically summarizing documents abstractively with little to no supervision.”

SummAE contains a denoising autoencoder that encodes (that is, generates numerical representations of) sentences and paragraphs of the target text in a shared space. Guided by a decoder whose input is prepended with a token signaling whether to decode a sentence or a paragraph, the system generates summaries by decoding each sentence from the encoded paragraphs.

The researchers discovered that most traditional approaches to training the auto-encoder resulted in long, multi-sentence summaries. To encourage it to learn higher-level concepts disentangled from their original expression, the team employed two denoising approaches — randomly masking tokens and permuting the order of sentences within paragraphs — that increased the number of training examples substantially. They also experimented with an adversarial critic component that could distinguish between sentences and paragraphs, in addition to two pretraining tasks that encouraged the encoder to learn how sentences narratively followed within a paragraph.

The researchers trained three different variations of SummAE¬†on the ROCStories, a corpus of self-contained, diverse, non-technical, and concise prose. They split the original 98,159 training stories into three separate collections — a training set, a validation set, and a test set — and collected three human summaries each for 500 validation examples and 500 test examples.

After 100,000 training steps with pretraining, the team reports that the best model significantly outperformed a baseline extractive sentence generator on the Recall-Oriented Understudy for Gisting Evaluation (ROUGE), a set of metrics devised to evaluate automatic summarization. Moreover, they say that in a qualitative study involving evaluators recruited through Amazon’s Mechanical Turk, volunteers rated one of the three SummAE models’ summaries “fluent” and “information-relevant” 80% of the time.

“The paragraph reconstructions show some coherence, although with some disfluencies and factual inaccuracies that are common with neural generative models,” wrote the coauthors. “Since the summaries are decoded from the same latent vector as the reconstructions, improving them could lead to more accurate summaries.”