面向属性抽取的门控动态注意力机制

在现阶段属性抽取研究中,现有注意力建模及训练较刚性(单句一次成型),而单句中不同词汇的上下文存在语境语义的差异,一致的注意力分布缺少动态的适应性.因此,文中提出面向属性抽取的门控动态注意力机制,利用双向长短时记忆网络捕获目标句中每个单词的隐层表示.在注意力模型处理词一级属性预测时,根据目标词及其上下文,计算适应该目标词的注意力分布向量,可以根据上下文的变化自动调整注意力权重的分配.借助门控调整注意力向量流向下一层神经元的信息量,最终使用条件随机场进行属性标记.应用2014-2016语义评估官方数据集验证文中方法的有效性,F1值均有所提高.

Gated Dynamic Attention Mechanism towards Aspect Extraction

In the current aspect extraction researches, the attention modeling and training are fixed, and the sentence is modeled in one time step. However, the semantics of the words vary in contexts, and a fixed attention distribution lacks dynamic adaptability. Therefore, a gated dynamic attention mechanism towards aspect extraction is proposed in this paper. A bidirectional long short term memory network is exploited to obtain hidden representations of words in a target sentence. Then, a specific attention distribution is computed according to the target word and its context while the attention model labelling words. Thus, the attention-weight distribution can be automatically adjusted according to the changes of contexts. Next, a gate is adopted to adjust the quantities of information flowing to the next units. Finally, conditional random field is utilized to label the aspect. The official datasets of 2014-2016 semantic evaluation are employed to verify the effectiveness of the proposed method, and F1 scores are increased.