基于重建误差的任务型对话未知意图检测

现有未知意图检测模型通常将语句映射到向量空间,并使用局部异常因子算法定义密度较小的特征点为未知意图,但经交叉熵损失训练的已知意图特征簇更加狭长,簇内的整体间距、密度和分散情况不均匀,进而增加了检测难度。针对上述问题,提出一种基于自动编码器重建误差的未知意图检测模型。在训练阶段,使用融入标签知识的联合损失函数训练已知意图分类器,使已知意图特征类间距离大且类内距离小,并利用这些特征训练一个仅能获取已知意图信息的自动编码器。在测试阶段,利用自动编码器将重建误差较大的样本视为未知意图,其余样本视为已知意图正常分类。在SNIPS数据集上的实验结果表明,在已知意图占比为25%、50%、75%时,该模型的Macro F1得分相比于表现最优的增强语义的高斯混合损失基线模型分别提升了16.93%、1.14%和2.37%,能够检测到更多的未知意图样本,同时在类别分布极不平衡的ATIS数据集上也有较好的性能表现。

Unknown Intent Detection for Task-Oriented Dialogs Based on Reconstruction Error

Existing unknown intent detection models map utterances to the vector space and use the Local Outlier Factor(LOF) algorithm to define the feature points with low density as the unknown intent.However, known intent feature clusters trained by the cross-entropy loss are narrower and longer, which causes the overall spacing in density, and the dispersion in the clusters is not sufficiently uniform for detection.This study proposes an unknown intent detection model based on the autoencoder reconstruction error to solve the above problems.During the training stage, the model uses a joint loss function with label knowledge to train a known intent classifier, which forces the distribution of known intent features to minimize the intraclass distance and maximize the interclass distance.It then uses these features to train an autoencoder that only contains information regarding the known intent.During the testing stage, samples with significant reconstruction errors are regarded as unknown intentions using the automatic encoder, and the other samples are regarded as normal classifications of known intentions.Experiments on the SNIPS dataset show that the Macro F1 score of the proposed model increases by 16.93%, 1.14%, and 2.37% compared with the Semantic-Enhanced large-margin Gaussian mixture loss(SEG) model of the best performance in the baseline models when proportions of the known intents are 25%, 50%, and 75%, respectively.Moreover, the proposed model can detect more unknown samples.Furthermore, the proposed model exhibits improved performance on the ATIS dataset, in which the intent distribution is highly unbalanced.