STACKEX

不同的文本本质上应对应不同数量的关键短语。现有的神经关键词生成模型很大程度上缺少这种需求。在这项研究中，我们从建模和评估的角度来解决这个问题。我们首先提出了一个循环生成模型，该模型生成多个关键短语作为分隔符分隔的序列。通过操纵解码器隐藏状态，两种新技术进一步增强了生成多样性。与以前的方法相比，我们的模型能够生成不同的关键词并控制输出数量。我们进一步提出了两个针对可变数字生成的评估指标。我们还引入了一个新的数据集 StackEx，该数据集扩展了关键短语生成任务中唯一现有的类型（即学术写作）。凭借先前和新的评估指标，我们的模型在所有数据集上都优于强大的基线。

Different texts inherently demand different quantities of key phrases. Existing neural keyphrase generation models largely fail to meet this critical requirement. In this study, we address this issue from the perspectives of both modeling and evaluation. We first propose a recurrent generative model that produces multiple keyphrases as a delimiter-separated sequence. Two novel techniques further enhance generation diversity by manipulating the decoder's hidden states. Compared with prior approaches, our model can generate diverse keyphrases and control the number of outputs. We further propose two evaluation metrics specifically tailored for variable-number generation. We also introduce a new dataset, StackEx, which expands the only existing genre for keyphrase generation tasks—namely academic writing. Using both existing and newly proposed evaluation metrics, our model outperforms strong baseline models across all datasets.

搜集汇总

背景与挑战

背景概述

STACKEX是一个用于关键词生成研究的文本数据集，旨在解决现有模型在生成可变数量关键短语方面的不足。它扩展了学术写作领域的任务，并引入了新的生成模型与评估指标，以提升多样性和控制能力。