【作者】 于宽；

【导师】 刘桂霞；

【作者基本信息】 吉林大学 ， 软件工程， 2021， 硕士

【摘要】 意图识别是一个句子分类任务,用来识别句子意图。而槽位填充则是一个序列标注问题,所谓的槽位就是这个词的语义信息。现在有许多成熟的理论去研究这个任务,在应用上已经有许多比较成熟的成果,比如聊天机器人阿里小蜜等。它们能根据和你的对话判断你说话的意图并给出答案。在类似淘宝客服这种机构扮演着重要的角色,节省了大量人力物力。当然意图识别领域发展也存在着许多不足甚至是障碍,比如数据集匮乏等问题。本文主要对各种意图识别槽位填充的模型进行研究,多角度对比模型优劣并对新的模型进行了研究。现在意图识别槽位填充的主流研究方法都是基于联合模型的研究,把本来的两个任务通过参数共享的方式联合学习,取得了不错的进展。本文的研究也聚焦于联合模型。以下是本文的主要研究内容:(1)本文将TOKEN算法应用于中文数据集SMP2019。TOKEN算法是在英文数据集ATIS和SNIPS上的意图识别和槽位填充算法。ATIS,SNIPS数据集与SMP2019数据集存在着语言和任务上的差异,无法将TOKEN算法直接应用在SMP2019数据集。因此本文参考TOKEN算法的原代码和论文,改写并扩充了SMP2019的官方baseline,将TOKEN算法成功应用于SMP2019数据集;(2)目前意图识别和槽位填充的研究大多集中于英文数据集ATIS和SNIPS,本文通过7种模型在中文数据集SMP2019上的对比实验,对其结果进行分析,探讨了每一个模型的优点和缺点,以及模型间融合的可能性;(3)本文将SF-ID算法和TOKEN算法结合起来。TOKEN算法的特点在于把句子级分类任务转化为字级别分类任务,所以本文引入字意图向量和字领域向量的概念来表示一个字所包含的句子意图信息和句子领域信息。在SF-ID算法中,还是把意图分类和领域分类看做句子级分类任务,并没有考虑字的意图信息和领域信息与字的槽位信息的依赖关系,只使用句子级向量进行信息融合。这种融合方式会产生很多的冗余信息。本文放弃句子级向量,使用字级别向量进行信息融合,很大程度上避免冗余信息带来的影响。实验中本文方法在准确率和f1值上都取得了不错的提升;(4)基于TOKEN算法使用意图信息增强领域信息。通过实验发现意图分类的分类效果最好。所以本文结合TOKEN算法,将意图分类和领域分类两个句子级分类任务看做字级分类任务。而槽位填充任务是一个序列标注问题,也可以理解为字级分类任务。这样,就可以类比意图分类任务和槽位分类任务的关系,用相似的方法得到基于意图信息增强的领域信息。实验效果相比较前面的方法在领域分类准确率和槽位标注f1值上都有一定程度的提升;(5)考虑到中文数据集的特点,使用百度的ERNIE代替BERT作为预训练模型,然后加上本文的模型,发现实验结果仍然有一定提升。更多还原

【Abstract】 Intention recognition is a sentence classification task that recognizes sentence intent.Slot filling is a sequence labeling problem.The so-called slot is the semantic information of the word.There are many mature theories to study this task,and there have been many mature results in application,such as the chat robot Ali Xiaomi.They can judge your intentions and give answers based on the conversation with you.It plays an important role in organizations like Taobao customer service,saving a lot of manpower and material resources.Of course,there are many shortcomings and even obstacles in the development of intent recognition,such as the problem of data sets.This paper mainly studies various models of intent identification slot filling,compares the pros and cons of the models from multiple angles,and studies the new models.At present,the mainstream research methods for slot filling in intent recognition are based on the research of the joint model.The original two tasks are jointly learned through parameter sharing,and good progress has been made.The research in this article also focuses on the joint model.The following is the main research content of this article:(1)This article applies the TOKEN algorithm to the Chinese data set SMP2019.The TOKEN algorithm is an intent recognition and slot filling algorithm on the English data sets ATIS and SNIPS.ATIS,there are language and task differences between the SNIPS data set and the SMP2019 data set,and the TOKEN algorithm cannot be directly applied to the SMP2019 data set.Therefore,this article refers to the original code and paper of the TOKEN algorithm,rewrites and expands the official baseline of SMP2019,and succes SFully applies the TOKEN algorithm to the SMP2019 data set;(2)At present,the research on intention recognition and slot filling is mostly concentrated on the English data set ATIS and SNIPS.This paper analyzes the results of 7 models on the Chinese data set SMP2019 and discusses the advantages of each model.And shortcomings,and the possibility of fusion between models;(3)This article combines the SF-ID algorithm with the TOKEN algorithm.The characteristic of the TOKEN algorithm is to transform sentence-level classification tasks into word-level classification tasks,so this paper introduces the concepts of word intention vector and word domain vector to represent the sentence intention information and sentence domain information contained in a word.In the SF-ID algorithm,intent classification and domain classification are still regarded as sentence-level classification tasks,and the intent information of the word and the dependence of the domain information and the slot information of the word are not considered,but the sentence-level vector is used for information fusion.This fusion method will generate a lot of redundant information.This paper abandons sentence-level vectors and uses word-level vectors for information fusion,which largely avo IDs the impact of redundant information.In the experiment,the method in this paper has achieved a good improvement in accuracy and f1;(4)Use intent information to enhance domain information based on the TOKEN algorithm.It is found through experiments that the classification effect of intention classification is the best.Therefore,this paper combines the TOKEN algorithm to treat the two sentence-level classification tasks of intent classification and domain classification as multiple word-level classification tasks.The slot filling task is a sequence labeling problem,which can also be understood as a word-level classification task.In this way,the relationship between the intent classification task and the slot classification task can be analogized,and the domain information enhanced based on the intent information can be obtained in a similar way.Compared with the experimental results,the previous methods have a certain degree of improvement in the field classification accuracy and the slot label f1 value;(5)Taking into account the characteristics of the chinese data set,use baidu’s ERNIE instead of BERT as the pre-training model.Then add the model of this article,it is found that the experimental results still have a certain improvement.更多还原