因为这篇论文是在组会的时候的精讲论文,做了ppt进行讲述的,所以直接把ppt粘过来进行总结(偷下懒(;´д`)ゞ)
文章目录
1. Keypoint
2. Corpus Analysis
This paper is based on the work proposed in another paper(《Emotion Cause Events: Corpus Construction and Analysis》), and the latter offer a more detailed description to the corpus, so if you want to look deep into the detail of the constructed corpus, you can click the link to jump. But there are also some other points that are different from the latter paper and require attention.
As shown in Table 2, There may be more than one clause that are the cause of one certain emotion. So if we just identify the task of emotion cause detection as a single-label problem, the model will not achieve high performance as we wish. In terms of the multi-clause distribution table, this paper propose a multi-label model to try to detect multi-clause causes.
3. Linguistic Patterns
The linguistic features have been described in the paper 《Emotion Cause Events: Corpus Construction and Analysis》, and this paper offers the method how to use these linguistic features to locate the position of an emotion cause.
4. Generalized Patterns
Rule-based patterns usually achieve a rather high accuracy, but suffer from low coverage. This lead to the problem of
low generalizetion ability. To avoid this shortcoming, this paper extracted generalized features automatically according to the rules shown in the above Table 4.
5. Multi-label Model
To be honest, I am not very clear how the multi-label model works. This paper do not give a detailed enough explanation about the model. In my views, it may just have one classifier that the output is a vector and every element in the vector corresponds to the probability of a combination shwn in Table 2 such as “Left_2_1_0”. Or in another conditon, there can be many classifiers and every classifier just has one scalar output that means the probability of a combination to be cause of a certain emotion
6. Summary
Though the multi-label model looks a bit unclear, we can not deny the work of this paper on the linguistic features and patterns. At the same time, the method that automatically extracts generalized patterns to solve the problem of low coverage is very innovative