联合多任务学习的对话情感分类和行为识别

Dialogue Sentiment Classification and Act Recognition Based on Multi-Task Learning

对话情感分类和对话行为识别是对话系统中的两个子任务,旨在预测对话中每个语句的情感标签和行为标签.这两个任务受多种因素的影响而密切相关,而现有的模型没有合理利用对话中包含的显式和隐式信息,如说话者信息,时间信息,标签信息等,并且两个相关任务之间缺乏有效的交互.为了解决上述问题,本文提出了一个新的多任务学习模型,即说话者感知跨任务协同交互图网络(Speaker-aware Cross-task Co-interactive Graph Network,SA-CCGN).该模型首先捕捉了说话者随时间变化的情感和行为线索,以生成说话者感知的语句表示,然后通过跨任务协同交互图网络来充分建模对话内的信息传播和任务间的信息交互,其中,通过构建一个有向无环图来模拟一个对话的信息传播,每一次图传播后,使用协同交互层对两个任务进行适当的交互.最后,在解码时引入标签信息,即标签之间的区分度和关联性,对模型训练进行约束.在两个公开数据集上的实验结果表明,该模型相较于目前最先进的联合模型,在对话情感分类任务上的F1值分别提高了4.57%和3.33%,在对话行为识别任务上的F1值分别提高了2.15%和0.63%,而参数量和内存使用降低了约1/2.

Currently,social media platforms allow users to universally express their opinions and sentiments due to their convenience and openness.As one of the most common ways of communication,dialogue contains rich information and sentiment expression of participants.Dialogue sentiment classification and dialogue act recognition are two sub-tasks in dialogue systems that aim to predict the sentiment and act label of each utterance in a dialogue.In the past few years,these two tasks gained attention from the NLP community due to the increase of public availability of dialogue data.They can be used to analyze dialogues that take place on social media or other scenes and provide support for downstream tasks,such as dialogue response generation.They can also aid in analyzing dialogues in real times,which can be public opinion monitoring,interviews,psychological consulting and more.These two tasks are influenced by multiple factors and closely related.However,existing models do not make reasonable use of the explicit and implicit information contained in a dialogue,such as speaker information,temporal information,and label information,and simply or coarse-grained modeling the interaction of two tasks.To solve the above problems,this paper proposes a new multi-task learning model,namely Speaker-aware Cross-task Co-interactive Graph Network(SA-CCGN).The model first captures speaker-aware sentiment and act cues along with the time to generate speaker-aware utterance representations,and then adequately models information propagation within a conversation and information interaction between tasks through a cross-task co-interactive graph network,where information propagation of a conversation is modeled by constructing a directed acyclic graph,and after each graph propagation,appropriate interaction between two tasks is performed using the co-interactive layer.Finally,the label information is introduced,i.e.,differentiation and correlation between labels,which can constrain the model training when decoding.Specifically,in the multi-loss decoder,the supervised contrastive learning loss is used to make the learned representation of different labels more differentiated and the conditional random field loss is used to constrain the generation of adjacent label sequences,then the final sentiment and act label of each utterance are obtained.In order to prove the effectiveness of the model in this paper,experiments were conducted on the two public two-way dialogue datasets:DailyDialog dataset and Mastodon dataset,and we compare our proposed method with a variety of state-of-the-art methods,including dialogue sentiment classification methods,dialogue act recognition methods and joint-train methods.Experimental results on two public datasets show that our model outperforms the current state-of-theart joint model Co-GAT,with an improvement of 4.57%and 3.33%in F1 scores for the dialogue sentiment classification task and 2.15%and 0.63%in F1 scores for the dialogue act recognition task on the two datasets,respectively,while reducing the number of parameters and memory usage by about 1/2.The performance of SA-CCGN on two public datasets exceeds the best results in the known literature.Experiments show that this method can effectively utilize dialogue information,and has obvious advantages in dialogue sentiment classification task and dialogue act recognition task compared to previous methods.