情绪分析是细粒度的情感分析任务，其目的是通过训练机器学习模型来判别文本中蕴含了何种情绪，是当前自然语言处理领域中的研究热点.情绪分析可细分为情绪分类与情绪回归两个任务.针对情绪回归任务，提出一种基于对抗式神经网络的多维度情绪回归方法.所提出的对抗式神经网络由3部分组成：特征抽取器、回归器、判别器.该方法旨在训练多个特征抽取器和回归器，以对输入文本的不同情绪维度进行打分.特征抽取器接受文本为输入，从文本中抽取针对不同情绪维度的特征；回归器接受由特征抽取器输出的特征为输入，对文本的不同情绪维度打分；判别器接受由特征抽取器输出的特征为输入，以判别输入的特征是针对何情绪维度.该方法借助判别器对不同的特征抽取器进行对抗式训练，从而获得能够抽取出泛化性更强的针对不同情绪维度的特征抽取器.在EMOBANK多维度情绪回归语料上的实验结果表明，该方法在EMOBANK新闻领域和小说领域的情绪回归上均取得了较为显著的性能提升，并在r值上超过了所有的基准系统，其中包括文本回归领域的先进系统.

Emotion analysis, which aims to determine the emotion contained in a piece of text via training a machine learning model, is a fine-grained sentiment analysis task. Emotion analysis can be divided into two tasks:Emotion classification and emotion regression. In this paper, an adversarial neural network is proposed for multi-dimensional emotion regression task. The proposed network consists of three modules:Feature extractors, regressors, and a discriminator. The network aims to train multiple feature extractors and regressors to score for multiple emotion dimensions for a textual input. Feature extractors take a text as inputs, and extract different feature vectors for different emotion dimensions. Regressors take extracted feature vectors as inputs to score for multiple emotion dimensions. The discriminator take an extracted feature vector as its input, and discriminate for which emotion dimension the feature vector is extracted. The proposed approach conducts adversarial training between different feature extractors via the discriminator in order to training feature extractors which can extract more generalized features for multiple emotion dimensions. Empirical studies on EMOBANK corpus demonstrate the notable improvements in r-value achieved by the proposed approach on EMOBANK readers' and writers' emotion regression in news domain and fictions domain compared to all baseline systems, including several state-of-the-art text regression systems.

TP391

国家自然科学基金（61672366，61751206）