Task-oriented Dialogue Language Understanding With Fewer Labels

Direct interaction with computers through natural language has long been the vision of AI research,since it is natural to expect computers to understand language and help users achieve specific goals.In this context,task-oriented dialogue systems have been proposed,which aim to accomplish user goals through dialogue.Task-oriented dialogue systems are important AI applications with a wide range of application scenarios,such as personal assistants and intelligent customer service.In task-oriented dialogue systems,dialogue language understanding is the basic and crucial module.Since the operation of the whole system is based on the result of dialog language understanding,the quality of dialog language understanding often directly determines the quality of the whole dialog system.Specifically,dialogue language understanding usually consists of two different subtasks: intent detection and slot tagging.Intent detection classifies user utterances into predefined intent categories and is usually modeled as a short text classification task.For example,given an utterance "What is the weather in Beijing tomorrow",the system needs to identify the user intent as "Query Weather".The slot tagging identifies key slot information from the input sentence,which is usually modeled as a sequence labeling task.In the above example,the system needs to extract the time slot "tomorrow" and the location slot "Beijing".In recent years,impressive progress has been made in dialogue language understanding based on deep learning techniques,which often require large amounts of annotated data for learning.However,in practical application scenarios,large amounts of labeled data are often difficult to obtain,because manual data labeling is often costly and task-oriented dialogues often face new domains that lack data.Consequently,insufficient labeled data has become a serious obstacle for practical application of existing methods.Therefore,how to reduce data dependence and learn quickly using a small number of annotations like human learning has become an important research direction,which is currently receiving much attention for conversational language understanding.In this paper,we explore the key techniques for dialogue language understanding with fewer labels,and target the solutions for the unique challenges of two dialogue understanding subtasks.The main contents are as follows.1.Dynamic thresholding based fewer-labels multi-intent detection.Existing research on learning with fewer labels mainly addresses the simple classification problem in data scarcity scenarios,i.e.,each sample belongs to only one category.However,in dialogue language understanding,users often express multiple intents in a single input table,and such multi-label classification scenarios often lead to the failure of existing single-label methods.To address this problem,this paper proposes a meta-learning and calibration learning based multi-label classification method and significantly improves the effectiveness of multi-intent classification in data-scarcity scenarios.2.Data augmentation based fewer-labels slot tagging.Slot tagging in fewerlabels scenarios poses more challenges compared to sentence classification such as intent detection.This is because word-by-word classification requires much more annotation than whole-sentence classification,significantly driving up the data requirements and learning difficulty.To resolve this conflict between annotation and model requirements,this paper proposes a generative data augmentation method for the slot tagging task,which automatically generates annotated data and expands the training set size.Experimental results show that the proposed data augmentation method can effectively alleviate the model’s dependence on data and significantly improve the performance of slot tagging with fewer labels.3.Label dependency transfer based fewer-labels slot tagging.The slot tagging task,as a sequence labeling task,needs to learn and exploit structural features such as label dependencies.However,in the data scarcity scenario,it is difficult for the model to learn effective structural knowledge from the sparse data,and it is also hard to migrate structural prior-knowledge directly from the rich data domain to the less-labeled domain,because different domains have different labels.To this end,this paper proposes a collapsedtransferring method that provides a solution to this challenge by learning and migrating the dependencies between abstract categories.Experiments demonstrate that the proposed method can effectively exploit the structural prior knowledge and significantly improve the accuracy of fewer-labels slot tagging.4.Joint metric-based learning for fewer-labels intent detection and slot tagging.The intent detection and slot tagging are highly relevant in dialogue language understanding tasks,and joint learning of the two tasks has been widely proven to be effective in improving dialogue language understanding performance.However,learning the connection between the two tasks becomes very difficult with limited sparse data.To address this,this paper proposes a metric learning-based task Con Prom model architecture that learns task relations directly in the metric space.Experiments show that the proposed approach learns a better joint metric space and significantly improves the accuracy of joint dialogue understanding.In summary,by targeting the unique challenges of user intent detection and key slot tagging,this paper explores the solutions for label-efficient learning of dialogue language understanding.While improving the performance of dialogue language understanding,the proposed methods are also widely available in other natural language problems,which can be further extended to more application scenarios and help advance the field of label-efficient natural language processing.