评价的阶段使用了sentiment classifcation task。Two popular non-domain-specifc sentiment lexicons as well as state-of-the-art machine-learning and deep-learning
models are chosen as benchmarks, and the experimental results show that our sentiment lexicons outperform the benchmarks with
statistically signifcant diferences, thus proving the efectiveness of the proposed approach.
Experiments
We mainly evaluate diferent lexicons and approaches using document-level classifcation tasks in the domain of online product reviews.
For hybrid sentiment classifcation methods, we consider the features of the document vector representation as the lexicon. We use the F1-measure as our main evaluation index and choose NB and SVM as the classifers.
除了上面用sentiment classification进行验证之外,还引入coverage (V), usage (U), and average(T)来评价lexicon。
test set里一共包含有N个不同的 words,其中有T个sentiment words,用于训练分类模型的lexicon的大小是S。因此coverage=T/N (感觉这个公式有问题,T和N都是定值啊),usage of lexicon=T/S (这个T是test set里包含的,和实际用到的lexicon里的词完全没有关系啊)
we discuss the classifcation performance in terms of the coverage (V), usage (U), and average(T)。Te results for the test set are listed in Tables 9 and 10
Summary:
现在对于review sentiment的分析大部分基于sentiment lexicons。因此,如何生成高质量的domain sentiment lexicon是一个问题。这篇文章考虑sentiment words和product feature(from reviews)的realtion,构建lexicion。评价的阶段使用了sentiment classifcation task。
Resource:
Paper information:
Notes:
评价的阶段使用了sentiment classifcation task。Two popular non-domain-specifc sentiment lexicons as well as state-of-the-art machine-learning and deep-learning models are chosen as benchmarks, and the experimental results show that our sentiment lexicons outperform the benchmarks with statistically signifcant diferences, thus proving the efectiveness of the proposed approach.
We mainly evaluate diferent lexicons and approaches using document-level classifcation tasks in the domain of online product reviews.
For hybrid sentiment classifcation methods, we consider the features of the document vector representation as the lexicon. We use the F1-measure as our main evaluation index and choose NB and SVM as the classifers.
其中中文的产品评论有3个domain,查看每个类型的f1有多少
这部分使用了downstream的task来测试lexicon的效果。只不过我要用的话并不是分类问题,而是NER,f1也是可以用的。
4.2. Experimental Design
在实验部分,对sentiment classification进行测试。首先不使用lexicon的方法有两个:bag-of-words和word2vec(BOW和W2V)。然后使用lexicon的方法用两个有名的lexicon:DUTIR,SentiWordNet。最后是自己的Domain-Specifc Lexicon (DS)
除了上面用sentiment classification进行验证之外,还引入coverage (V), usage (U), and average(T)来评价lexicon。
test set里一共包含有N个不同的 words,其中有T个sentiment words,用于训练分类模型的lexicon的大小是S。因此coverage=T/N (感觉这个公式有问题,T和N都是定值啊),usage of lexicon=T/S (这个T是test set里包含的,和实际用到的lexicon里的词完全没有关系啊)
XU:这里的converage,usage,average分别模仿了precision, recall, f1。但是公式的定义有问题。
F1 = 2 * (precision * recall) / (precision + recall)
4.3. Results and Discussion
Lexicon Coverage
we discuss the classifcation performance in terms of the coverage (V), usage (U), and average(T)。Te results for the test set are listed in Tables 9 and 10
Model Graph:
Result::
Thoughts:
Next Reading: