Weakly-Supervised Neural Categorization of Wikipedia Articles
Weakly-Supervised Neural Categorization of Wikipedia Articles
复制标题
DOI：
10.1007/978-3-030-34058-2_2
复制
发表时间：
2019-11
期刊：
影响因子：
--
通讯作者：
Xingyu Chen;M. Iwaihara
中科院分区：
文献类型：
--
作者：
Xingyu Chen;M. Iwaihara
深度神经模型由于其强大的表达能力和对特征工程的要求较少，在许多 NLP 任务中越来越受欢迎。神经模型通常需要大量带标签的训练文档。然而，维基百科的某一类别没有包含足够的文章用于训练。即使只给出一个小的标记文档集，弱监督神经文档分类也可以处理这种情况。然而，这些基于 RNN 的方法在处理长文档（例如维基百科文章）时通常会失败，因为很难保留长文档重要部分的记忆。为了克服这些挑战，我们提出了一种称为 WS-Rank 的文本摘要方法，该方法根据文档中与类别相关的关键字和句子位置进行加权，提取文档的关键句子。将我们的 WS-Rank 应用于训练和测试文档并总结为关键句子后，弱监督神经分类在分类结果上显示出显着的改进。
Deep neural models are gaining increasing popularity for many NLP tasks, due to their strong expressive power and less requirement for feature engineering. Neural models often need a large amount of labeled training documents. However, one category of Wikipedia does not contain enough articles for training. Weakly-supervised neural document classification can deal with situations even when only a small labeled document set is given. However, these RNN-based approaches often fail on long documents such as Wikipedia articles, due to hardness to retain memories on important parts of a long document. To overcome these challenges, we propose a text summarization method called WS-Rank, which extracts key sentences of documents with weighting based on class-related keywords and sentence positions in documents. After applying our WS-Rank to training and test documents to summarize then into key sentences, weakly-supervised neural classification shows remarkable improvement on classification results.