A Question Answering Approach to Emotion Cause Extraction

Presentation on theme: "A Question Answering Approach to Emotion Cause Extraction"— Presentation transcript:

1 A Question Answering Approach to Emotion Cause Extraction
Lin Gui, Jiannan Hu, Yulan He, Ruifeng Xu*, Qin Lu, Jiachen Du 大家早上好，今天介绍一个我们使用一种新的记忆网络对文本中的情感原因进行抽取的工作

2 Outlines Related Works Our Approach Experiments and Evaluation
Introduction Related Works Our Approach Experiments and Evaluation Conclusions and Future Works

3 Introduction Previous researches on emotion recognition in text usually focus on classification and elements (feeler, expression et al.) extraction. But there are few works on extracting the reason behind emotional expressions. Document: 我的手机昨天丢了，我现在很难过。 (I lost my phone yesterday, and I feel sad now. ) Emotion：Sad Emotional Expression: 很难过 Emotion Cause: 我的手机昨天丢了 传统文本情感分析任务主要关注文本情感分类任务与文本情感要素抽取，但是现有工作很少关注如何抽取文本情感表达背后所的直接原因及其根本原因。 例如当给定文本 时 情感分类可以将上述文本识别为为悲伤情感 情感表达抽取可以将文本中的表达“很难过”提取出来 而我们所研究的情感原因抽取任务的目标是根据文本信息及其中包含的情感表达抽取出情感原因。

4 Related Works There is a few researches on emotion cause detection:
Sophia M. Y. Lee – Rule-based Emotion Cause(COLING 2010) Gao – OCC-model based Emotion Cause Detection(Expert system with applications 2015) Ghazi – CRF-based Emotion Cause Detection (CICLING 2015) But these works didn’t release public datasets for researching 下面介绍情感原因抽取的相关研究进展 情感原因发现与抽取最早由Sohia等研究者提出，他们所提出的方法相是由语言学规则驱动的 Gao等人在此基础上引入OCC情感认知模型，扩充了原因识别规则 除了规则驱动方法外，Ghazi等人利用机器学习及CRF序列标注模型对情感原因抽取进行建模 上述工作对情感原因抽取从不同的角度进行了研究， 但是这些工作都只关注语言学特征及浅层语义特征，同时上述研究所使用的数据集都未公开为研究者使用

5 Related Works Gui et al firstly released a manually annotated dataset for emotion cause extraction on EMNLP 2016. We collected more than 20,000 news from Sina social news and selected documents with one emotional expression and at least one emotion causes . All the emotional keywords and causes are annotated by W3C emotional markup language <instance id= “29”> …… <clause id= “5” cause= “N” keywords= “N”> <text>这是当时丹阳最大的理发店。</text> </clause> <clause id= “6” cause= “N” keywords= “N”> <text>我在那儿获得了好多证书和荣誉。”(</text> <clause id= “7” cause= “Y” keywords= “N”> <text>说起自己的荣誉(</text> <cause id= “1” type= “v” begin= “0” length= “7”>说起自己的荣誉</cause> <clause id= “8” cause= “N” keywords= “Y”> <text>朱某很是自豪(</text> <keywords emotion= “happiness” keywords-begin= “5” keywords-length= “2”>自豪</keywords> </instance> 为了更好地研究情感原因抽取问题，Gui于2016年EMNLP发布了首个开放的情感原因抽取数据库 数据从20000条新浪社会新闻获取，我们从中选取2105个带有明显情感表达与与至少一个情感原因的文档。所有数据均采用W3C 情感标注语言标注 标注实例如右图所示

6 Related Works [Gui et al. 2016] used dependency parser to parse the text to tree structures. The emotion cause extraction task was converted to tree classification task. Then two variants of Tree-Kernel SVMs were used to classify the tress. This model heavily depends on accurate dependency trees. And also this model cannot extract phrase-level emotion causes. 针对此数据集，我们使用基于依存分析器将文本解析为树状结构，情感原因抽取任务此时被转化为了树结构分类任务，即从文本中所有字句对应的依存分析树进行分类 这里的树核定义在一个递归结构上，但是没有考虑终端节点，即词语信息相关节点 通过两种方式引入终端节点： 引入基于终端节点词语近义词的改进树核 引入基于词表示的多核函数 我们所使用的方法严重依赖于依存分析的准确性，而且同时我们的方法只能适用于子句级别，无法在细粒度的短语级别的

7 Our Approach Emotion Cause Detection is analogue to Question Answering： Emotional Text as Reading Text Emotional Words as Question Emotion Cause as Answer 针对这些问题，我们提出一种全新的情感原因建模框架，我们将情感原因抽取任务看做问答任务 将抽取任务中的情感文本看做问答任务中的阅读文本 将已经提取出的情感词看做问答任务中的问题语句即query语句 将情感原因分类结果看做问答任务中的答案，此处答案只存在Yes与No两种答案

8 Memory Network (Memnet)
To model the process of question answering, we use Memory Networks [Sukhbaatar et al. 2015] as our base model. And the multiple-hop was used to better represent the text. One-Hop Architecture Multiple-Hop Architecture 为了对上述类问答任务进行建模，我们使用memory network作为我们的基础模型 通过词向量表示学习将待阅读文本子句中每个词语表达为e_1,e_2,...,e_k，将其存储在memory network的记忆单元中。同时，待判断的情感词的词向量记做$E$作为注意力单元。 此时，可以通过内积操作，计算情感词与子句中每一个词语的关系,然后，通过softmax将这一数量关系映射到0到1之间进行归一化,作为每个词的权重，用于计算文本的语义表示。 最后，利用语义表示的加权和计算最终的输出。

9 Convolutional Memory Network (Conv-Memnet)
In order to capture context information for clauses, we propose a new architecture which cotain more memory slot to model the context with a convolutional operation. the weight of word wi in the i-th position considers both the previous word wi 1 and the fol-lowing word wi+1 by a convolutional operation: 在此基础上，我们发现每个词的注意力权重不仅仅是由当前词语所决定的，词语的上下文信息也是一个计算过程中的重要依据 所以我们提出了一种类似卷积的注意力加权方法 在这种注意力计算方法中，每个词语的注意力是由当前词、前文词及后文词共同决定的，并且在加权过程中根绝上下文注意力对不同位置的词语进行加权，最终获得一个以短语窗口为单位的加权结果进行最终输出。

10 Experiments and Evaluation
We conduct experiments on a simplified Chinese emotion cause corpus (Gui et al., 2016)∗, the only publicly available dataset on this task to the best of our knowledge. It is commonly accepted so that we can compare our results with others. If a proposed emotion cause clause covers the annotated answer, the word sequence is considered correct. The precision, recall, and F-measure are defined by Item Number Documents 2,105 Clauses 11,799 Emotion Causes 2,167 Documents with 1 emotions 2,046 Documents with 2 emotions 56 Documents with 3 emotions 3 最终的评估方法是对测试文本进行子句切分，并依次判断每一个子句是否含有情感原因的分类任务。 选取了语料库中90\%的数据作为训练数据，10\%的数据作为测试数据，进行交叉验证。为了获得相对可信的实验结果，在本章的相关实验中，独立随机切分训练集与测试集25次，对比其平均性能指标。这里采用的评估指标是基于准确率（Precision），召回率（Recall）和F值（F-measure）的评价标准。 ∗Available at: id=694

11 Experiments and Evaluation
Performance compared with baselines RB (Rule based method): The rule based method proposed in (Lee et al., 2010) CB (Common-sense based method): We use the Chinese Emotion Cognition Lexicon (Xu et al., 2013) as the common-sense knowledge base. RB+CB+ML (Machine learning method trained from rule-based features and facts from a common-sense knowledge base): This methods was previously proposed for emotion cause classification in (Chen et al., 2010). SVM: This is a SVM classifier using the unigram, bigram and trigram features Word2vec: This is a SVM classifier using word representations learned by Word2vec (Mikolov et al., 2013) as features. Multi-kernel: This is the state-of-the-art method using the multi-kernel method (Gui et al., 2016) to identify the emotion cause. We use the best performance reported in their paper. CNN: The convolutional neural network for sentence classification (Kim, 2014).

12 Results Performance compared with baselines
从实验结果而言，本文提出方法相较于传统Memory Network有较大提升 同时本文方法相比较于我们之前提出的基于树核的方法也有2个百分点的巨大提升

13 Results Key phrases extracted from attention mechanism：

14 Results In order to evaluate the quality of keywords extracted by memory networks, we define a new metric on the keyword level of emotion cause extraction. The keyword is defined as the word which obtains the highest attention weight in the identified clause. If the keywords extracted by our algorithm is located within the boundary of annotation, it is treated as correct.

15 Conclusion We proposed a novel QA based framework for emotion cause extraction A new convolutional operation was introduced to memory networks Thoughtful experiments showed efficiency of our method

16 Appendix To better work on emotion cause extraction, we organized “NTCIR emotion cause extraction shared task” . We annotated more documents in both English and Chinese for this task. The data will be released as soon as possible.

17