New Datasets, New Evaluation Measures and an Improved Unsupervised Method
ATE:aspect term extraction
ABSA:aspect based sentiment analysis
文章假设搜索引擎获取到用户对某件实体(a particular target entity)的评价
ABAS系统主要包括三个子任务:
1) Aspect term extraction 2)Aspect term sentiment estimation分类 3) Aspect aggregation
文章主要关注点在:aspect term extraction(ATE)
文章的contribution:1) 过去的数据集存在的问题:来自某特定领域或者是很少target entities的评价或者不包含aspect term的注释,所以文章提供三个新的数据集(restaurants, laptops, hotels),并且有gold annotations of all aspect term occurrences,measured inter-annotator agreement注释间的一致性 2) 普遍使用的evaluation measures不是所有都是satisfactory的,例如,经常使用的precision, recall, 和 F-measure 通过计算aspect terms的距离,频繁出现的aspect term和不频繁出现的是equal weight的,然而经常讨论的aspect terms应该是更重要的。文章提出了权重不同的precision和recall 3) 方法
查了一下inter-annotator agreement,链接是:https://corpuslinguisticmethods.wordpress.com/2014/01/15/what-is-inter-annotator-agreement/
Inter-annotator agreement is a measure of how well two (or more) annotators can make the same annotation decision for a certain category.
Aspect term extraction methods:
1) baseline: dubbed FREQ,返回最频繁的不同的名词和名词短语 2) Hu and Liu的方法:给baseline增加pruning mechanisms(剪枝机制),发现更多的aspect terms (dubbed H&L)3)对H&L方法的扩展,增加了pruning step(dubbed H&L+w2v)4)类似的(dubbed FREQ+w2v)
所有方法都是unsupervised
FREQ baseline:返回频率最高的名词和名词短语,并排序
H&L的方法:首先提取不同的名词和名词短语,作为aspect term的备选。然后通过连接成对或三个同时出现在一个句子中的aspect terms生成更长的candidate aspect terms。所有aspect term按照decreasing p-support排序,p-support是包含apect term句子的个数,除去某个含有子term的,例如aspect term有“battery life”和"battery",那么在句子"The battery life was good"计算在"battery life"的p-support,而不计算在"batter"的p-support中。通过剪枝进行自纠正,首先抛弃"non-conpact"的multi-word distinct aspect terms,例如"battery life screen" appears in non-compact form in "battery life is way better than screen";然后,如果某个candidate distinct aspect term t的p-support比3小,t is subsumed(包括) by another candidate distinct aspect term t撇,那么t删掉。然后,一组"opinion of adjectives"被组成,对每个句子和每个candidate distinct aspect term t 出现在句子中的,句子中距离t最近的adjective增加到一组opinion adjectives中,然后句子被重新扫描,如果句子中不包含任何candidate aspect term但是包括一个opinion adjective,然后最接近opinion adjective的名词添加到candidate distinct aspect terms。
H&L+W2V:输入变成continuous vector space representations of words,使用神经网络,剪枝步骤使用最频繁的十个candidate distinct aspect terms,然后计算每个向量的centroid,称为domain centroid;相似的,计算the Brown Corpus(news category)中的最频繁的20个词,除去停用词和短于3个字符的词,称为common language centroid。任意candidate distinct aspect term的vector的距离更接近common language centroid而不是the domain centroid,会被删除。接近common language centroid的是common words,而接近domain centroids 的是domain-specific concepts,更有可能是aspect terms。
FREQ+W2V:增加了pruning step,同H&L+W2V,距离。
实验结果:
网友评论