Domain adaptation, adapting models from domains rich in labeled training datato domains poor in such data, is a fundamental NLP challenge. We introduce aneural network model that marries together ideas from two prominent strands ofresearch on domain adaptation through representation learning: structuralcorrespondence learning (SCL, (Blitzer et al., 2006)) and autoencoder neuralnetworks. Particularly, our model is a three-layer neural network that learnsto encode the nonpivot features of an input example into a low-dimensionalrepresentation, so that the existence of pivot features (features that areprominent in both domains and convey useful information for the NLP task) inthe example can be decoded from that representation. The low-dimensionalrepresentation is then employed in a learning algorithm for the task. Moreover,we show how to inject pre-trained word embeddings into our model in order toimprove generalization across examples with similar pivot features. On the taskof cross-domain product sentiment classification (Blitzer et al., 2007),consisting of 12 domain pairs, our model outperforms both the SCL and themarginalized stacked denoising autoencoder (MSDA, (Chen et al., 2012)) methodsby 3.77% and 2.17% respectively, on average across domain pairs.
展开▼
机译:NLP面临的基本挑战是,域自适应(将模型从富含标签的训练数据的域适应到缺乏此类数据的域)进行适应。我们介绍了一种神经网络模型,该模型将通过表示学习的两个主要领域研究思路融合在一起:结构对应学习(SCL,(Blitzer et al。,2006))和自动编码器神经网络。特别地,我们的模型是一个三层神经网络,该网络学习以将输入示例的非枢轴特征编码为低维表示,以便存在枢轴特征(在两个域中都突出并且可以为NLP任务传达有用信息的特征)在该示例中,可以根据该表示进行解码。然后将低维表示形式用于任务的学习算法。此外,我们展示了如何将预训练的词嵌入注入我们的模型中,以提高具有相似枢轴特征的示例之间的通用性。在包含12个域对的跨域产品情感分类(Blitzer等,2007)的任务上,我们的模型优于SCL和归一化堆叠式去噪自动编码器(MSDA,(Chen等,2012))方法,降低了3.77%和跨域对的平均值分别为2.17%。
展开▼