We examine the algebraic and geometric properties of a uni-directional GRU and word embeddings trained end-to-end on a text classification task. A hyperparameter search over word embedding dimension, GRU hidden dimension, and a linear combination of the GRU outputs is performed. We conclude that words naturally embed themselves in a Lie group and that RNNs form a nonlinear representation of the group. Appealing to these results, we propose a novel class of recurrent-like neural networks and a word embedding scheme.
Explainable Natural Language Processing with Matrix Product States
Quantum Technology Foundation (Thailand)(泰国量子技术基金会);Thailand Center of Excellence in Physics, Ministry of Higher Education, Science, Research and Innovation(高等教育、科学、研究和创新部泰国卓越物理中心);Chulalongkorn University(朱拉隆功大学);Mahidol University(玛希隆大学);Blueqat Inc.(Blueqat公司);Technical University of Crete(克里特理工大学);National University of Singapore(新加坡国立大学)
RNN Architecture Learning with Sparse Regularization
Carnegie Mellon University(卡内基梅隆大学);Allen Institute for Artificial Intelligence(艾伦人工智能研究所);University of Washington(华盛顿大学)
NLP的神经模型通常使用大量参数来达到最先进的性能,这可能导致过多的内存使用和运行时间增加。我们提出了一种结构学习方法来学习稀疏的,参数有效的NLP模型。我们的方法将组套索应用于有理RNN(Peng et al.,2018),一个与加权有限状态自动机(WFSA)密切相关的模型家族。我们利用了RNN对权重的自然分组,因此组套索惩罚直接删除了WFSA状态,大大减少了模型中的参数数量。我们在一些情感分析数据集上进行的实验,使用GloVe和BERT嵌入,表明我们的方法可以学习参数较少的神经结构,而不会牺牲相对于参数丰富的基线的性能。我们的方法还突出了理性RNN的可解释性。我们表明,稀疏化这样的模型,使他们更容易可视化,我们提出的模型,完全依赖于少至三个WFSA修剪后超过90%的权重。我们公开发布代码。
Investigating how well contextual features are captured by bi-directional recurrent neural network models
Adobe Research, Big Data Experience Lab(Adobe 研究院大数据体验实验室);National Center for Text Mining, The University of Manchester(曼彻斯特大学国家文本挖掘中心);IIT Guwahati(印度理工学院古瓦哈提分校)