Data-to-text Generation by Splicing Together Nearest Neighbors

Abstract

We propose to tackle data-to-text generation tasks by directly splicing together retrieved segments of text from "neighbor" sourcetarget pairs. Unlike recent work that conditions on retrieved neighbors but generates text token-by-token, left-to-right, we learn a policy that directly manipulates segments of neighbor text, by inserting or replacing them in partially constructed generations. Standard techniques for training such a policy require an oracle derivation for each generation, and we prove that finding the shortest such derivation can be reduced to parsing under a particular weighted context-free grammar. We find that policies learned in this way perform on par with strong baselines in terms of automatic and human evaluation, but allow for more interpretable and controllable generation.

ChartSumm: A Comprehensive Benchmark for Automatic Chart Summarization of Long and Short Summaries

“…We also notice that models sometimes fail to generate informative summaries while also failing to predict anything about data (see the second example in Table 6). In both Bart-Large models, we also find some cases where the models generate information that is fully irrelevant to the chart (i.e., the Hallucination effect [1,8,29]) (see the third example in Table 6).…”

TempLM: Distilling Language Models into Template-Based Generators

“…In contrast, TempLM first finetunes a PLM on the data-to-text task and then exploits the PLM's ability in smoothing the text distribution to tackle the paraphrasing problem. (Wiseman et al, 2021), and inducing latent variables (Wiseman et al, 2018;Li and Rush, 2020;Ye et al, 2020). Much like classic template-based methods, these systems attempt to learn structured representation from diverse human-written text, which is challenging and often requires heuristics for additional supervision.…”

“…While obtaining C * is impossible, we can design approximate clusters C based on the presence of different fields, as is standard in other data-to-text methods (Wiseman et al, 2021).…”

“…Second, TempLM leverages the PLM's generation ability to refine templates, using a novel consensus beam search algorithm. Unlike prior works (Wiseman et al, 2018(Wiseman et al, , 2021, our approach can leverage any PLM to generate templates, allowing us to take advantage of improvements in the data efficiency and fluency of PLMs.…”

Nearest Neighbor Non-autoregressive Text Generation

“…Traditionally, nearest neighbors have been utilized in various tasks, including part-of-speech tagging [6] and example-based machine translation [23]. Recently, nearest neighbors have been successfully applied to AR models, particularly to neural machine translation [13], [15], [37], text summarization [5], [25], and datato-text generation [36]. We adopt the pioneering idea to refine and improve NAR models.…”