- Overall Structure
- Detailed Work
- Controlled Experiment
- Schedule
- Related Works
- Instance Analysis
- Discussion
- Analysis
- Story
- Contribution
- Future Work
The project consists of 4 parts. In the first part, we need to do some data mining and data preprocessing on our corpus. In the second part, we try to figure out how to leverage connectives in ECPE. In the third part, we do our experiments in a classic way using explicit relation classifier. In the last part, we check our improvement on the basis of previous SOTA [Zhou, C. et al.].
We use several datasets. For Chinese studies, we use the HIT-CDTB and ECPE dataset [Gui, L. et al.]. For English studies, we use the PDTB2.0, PDTB3.0, Gigaword and ECPE-ENG dataset we translate.
Since Chinese has only a few connectives, We choose to accomplish English study first.
The sample data mining work is available in dev-JunfengRan branch.
We need to do research on the connectives first. While dealing with these datasets, we follow the following style of statistics. We should consider the following four structure types (arg1, arg2), (arg1, conn, arg2), (conn, arg1, arg2), (conn1, arg1, conn2, arg2). We note them simply as type 0, 1, 2 and 3 (caution: we do not consider an type-3 instance as type-1 and type-2 simultaneously). In general, we can rewrite the sections to make emo clause and cau clause as arg1 and arg2 for ECPE dataset and leave the original sequence for other datasets.
For HIT-CDTB dataset and PDTB2.0/3.0 dataset, we want to know their occurence frequency with different discourse relation (top 2 levels with reason, result) and discourse relation distance together with structure types mentioned above for every connectives repectivesly. Previous study [Kishimoto, Y. et al.] use the top 11 relations on the second level are Comparison.Concession, Comparison.Contrast, Contingency.Cause, Contingency.Pragmatic cause, Expansion.Alternative, Expansion.Conjunction, Expansion.Instantiation, Expansion.List, Expansion.Restatement, Temporal.Asynchronous, Temporal.Synchrony. But, we use all top-2-levels relations together with reason and result as our tag, and give our top 10 relations according to our statistics later.
Connective | top level | second level | structure type | distance | frequency |
---|---|---|---|---|---|
because | Contingency | Cause.reason | 1 | 1 | 1000 |
because | Contingency | Cause.reason | 1 | 2 | 10 |
if-then | Contingency | Condition | 3 | 1 | 100 |
while | Temporal | Synchrony | 1 | 1 | 100 |
or | Expansion | Alternative | 1 | 1 | 100 |
For ECPE dataset, we need to translate it into English for English studies. For this purpose, we can translate the corpus clause by clause using translate software or algorithms and then correct the output manually. For dataset itself, we want to know the occurence frequence of connectives in ECP with different discourse relation (Contingency.Cause with reason and result) and distance (+/-, cause minus emotion) between emotion clause and cause clause together with structure types mentioned above for every connectives repectivesly. Note that reason/result is decided by whether the distance is positive (reason) or negative (result).
Old version
Connective | top level | second level | structure type | distance | frequency |
---|---|---|---|---|---|
because | Contingency | Cause.reason | 1 | +1 | 1000 |
because | Contingency | Cause.reason | 1 | -1 | 1000 |
so | Contingency | Cause.result | 1 | -1 | 1000 |
because_of | Contingency | Cause.reason | 1 | +1 | 100 |
due_to-then | Contingency | Cause.result | 3 | -1 | 1 |
New version
Connective | top level | second level | structure type | distance | frequency |
---|---|---|---|---|---|
because | Contingency | Cause.reason | 1 | +1 | 1000 |
because | Contingency | Cause.result | 2 | -1 | 1000 |
so | Contingency | Cause.result | 2 | -1 | 1000 |
because_of | Contingency | Cause.reason | 1 | +1 | 100 |
due_to-then | Contingency | Cause.result | 3 | -1 | 1 |
If we want to further enhance the performance of the Bert encoder or explicit relation classifier, we can use the Gigaword dataset. For this dataset, we need to cleanse the data and divide it into seperate clauses which have connectives. And use this new dataset to do domain pre-training. This is inspired by [Rutherford, A. et al., 2015].
At first, we find the emotion clause using ECPE-MM-R turn 1. And we pair up each emotional clause and each non-emotional clause. Then, we add connectives selected by our data mining work (same amount of connectives from each relation type or based on almost-same sum of percentage) and calculate the probability over the Language Model. We implement this with cosine similarity. Finally, if the sum of probability of causal connectives (probability times percentage) is the greatest, we add connectives with probability into original corpus and process them seperately (maybe choose top 5 or 10 or set probability threshold) later on. However, the loss of this step is unclear. I prefer to say that this step is a automatic step completed by LM so there is no loss we can use.
On the other hand, we could use the BLEU or ROUGE to score the connectives directly. We will try this implementation in the future.
For simple verification without whole analysis of PDTB/HIT-CDTB, we use the following list of connectives [Rutherford, A. et al., 2015]: further, in sum, in the end, overall, similarly, whereas, earlier, in turn, nevertheless, on the other hand, ultimately, accordingly, as a result, because, by comparison, by contrast, consequently, for example, for instance, furthermore, in fact, in other words, in particular, in short, indeed, previously, rather, so, specifically, therefore, also, although, and, as, but, however, in addition, instead, meanwhile, moreover, rather, since, then, thus, while, as long as, if, nor, now that, once, otherwise, unless, until.
We can judge the relation of the new pairs of (emotion arg, connectives, arg) with explicit relation classifier as benchmark. We may use the benchmark of [Ji, Y. et al., 2015] or [Lin, Z. et al., 2009] or SOTA.
Last, we can modify the source code of ECPE-MM-R [Zhou, C. et al.] and try to get the SOTA. We also need to do some complementary experiments about this model, like analyzing the precision and recall rate of every turn.
v1, v2, v3 are based on previous SOTA. v1 uses information in connectives to help identify the cause clauses. v2 uses information in connectives to help check whether emotion is corresponding to cause. v3 uses information in connectives to help both above.
v4, v5 are based on explicit relation classifier. v4 is a one-step progress to figure out relation directly after recognizing the emotion-cause pair. v5 check the bond on the basis of v4.
Note: Assuming six group members work all day long through whole process, so the actual schedule may be much later.
2023.1.4-2023.1.11 Try to run through the ECPE-MM-R source code. For those using the server can try to use the Accelerate library or data parallelism (https://zhuanlan.zhihu.com/p/467103734).
2023.1.11-2023.1.18, 2023.2.1-2023.2.8 (not forced) Analyze the ECPE-MM-R source code and modify it to output the precision and recall rate of every turn. Try to create the framework of adding connectives with probability.
2023.2.8-2023.2.15 Refine the code above. Complete data mining and data preprocessing.
2023.2.15-2023.2.22 Fuse the method above with modified ECPE-MM-R source code to complete our experiment.
2023.2.22-2023.3.1 We can judge the relation of the new pairs of (emotion arg, connectives, arg) with explicit relation classifier as benchmark.
2023.1.4-2023.1.11 Try to run through the ECPE-MM-R source code. For those using the server can try to use the Accelerate library or data parallelism (https://zhuanlan.zhihu.com/p/467103734).
HIT-CDTB
http://ir.hit.edu.cn/hit-cdtb/
PDTB2.0
http://www.ldc.upenn.edu/Catalog/CatalogEntry.jsp?catalogId=LDC2008T05 (origin, not available)
https://github.com/cgpotts/pdtb2
ECPE
http://hlt.hitsz.edu.cn/?page_id=694 (origin, not available)
https://github.com/zhoucz97/ECPE-MM-R
PDTB3.0
https://catalog.ldc.upenn.edu/LDC2019T05
Gigaword
https://catalog.ldc.upenn.edu/LDC2011T07
Main reference:
Kurfalı, M., Östling, R., 2021. Let’s be explicit about that: Distant supervision for implicit discourse relation classification via connective prediction, in: Proceedings of the 1st Workshop on Understanding Implicit and Underspecified Language. Presented at the Proceedings of the 1st Workshop on Understanding Implicit and Underspecified Language, Association for Computational Linguistics, Online, pp. 1–10. https://doi.org/10.18653/v1/2021.unimplicit-1.1
Other reference:
Kishimoto, Y., Murawaki, Y., Kurohashi, S., n.d. Adapting BERT to Implicit Discourse Relation Classification with a Focus on Discourse Connectives 7.
Ji, Y. and Eisenstein, J. (2015). One vector is not enough: Entity-augmented distributed semantics for discourse relations. Transactions of the Association of Computational Linguistics, 3:329–344.
Lin, Z., Kan, M.-Y., and Ng, H. T. (2009). Recognizing implicit discourse relations in the Penn Discourse Treebank. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 343–351.
Rutherford, A., Xue, N., 2015. Improving the Inference of Implicit Discourse Relations via Classifying Explicit Discourse Connectives, in: Proceedings of the 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. Presented at the Proceedings of the 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Association for Computational Linguistics, Denver, Colorado, pp. 799–808. https://doi.org/10.3115/v1/N15-1081
Main reference:
Zhou, C., Song, D., Xu, J., Wu, Z., n.d. A Multi-turn Machine Reading Comprehension Framework with Rethink Mechanism for Emotion-Cause Pair Extraction 10.
Other reference:
Gui, L., Wu, D., Xu, R., Lu, Q., Zhou, Y., n.d. Event-Driven Emotion Cause Extraction with Corpus Construction 11.
Xia, R., Ding, Z., 2019. Emotion-Cause Pair Extraction: A New Task to Emotion Analysis in Texts, in: Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. Presented at the Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, Association for Computational Linguistics, Florence, Italy, pp. 1003–1012. https://doi.org/10.18653/v1/P19-1096
We believe, for those have attributive clause, supplementary relative clause and/or adverbial clause, we can rewrite it into two or more clauses to use our method.
1.Those with attributive clause
Page 1 instance from ECPE SOTA paper.
I thank all the people who have helped me.
p.s. I believe 'I thank all these/those people who have helped me' is better.
We can't directly use our inserting-connectives method to deal with this instance. But we can rewrite this sentence with attributive clause into following new form automaticlly.
I thank all these/those people. These/Those people have helped me.
Now we can use our method on these sentences.
I thank all these/those people. (Because) These/Those people have helped me.
2.Those with supplementary relative clause
Instance 3 from ECPE benchmark dataset.
小男孩 胸口 蔫 瘦 得 让 人 心疼
We can rewrite this and transform it.
小男孩 胸口 蔫 瘦 [SEP] 这 让 人 心疼
这 让 人 心疼 [SEP] (因为) 小男孩 胸口 蔫 瘦
3.Those with adverbial clause
Instance 24 from ECPE benchmark dataset.
再次 因 丈夫 撒手人寰 而 陷入 无限 绝望 中
We can rewrite this and transform it.
再次 因 丈夫 撒手人寰 [SEP] 而 陷入 无限 绝望 中
p.s. I think this is much more likely a wrong segementation. The origin corpus should be noted as above.
1.The original ECPE-MM-R paper didn't give the precision and recall rate of three turns. We will try to accomplish these.
2.I think emotion clause itself cannot be the cause of the emotion. Not decided yet, maybe we can construct a modified dataset.
3.We can translate the ECPE corpus clause by clause using translate software or algorithms. Both are OK, but not decided yet.
4.Why choose lambda = 0.7 in [Zhou, C. et al.]. Maybe 0.7 is nearly equal to the square root of delta = 0.5 and consider the emotion-cause and cause-emotion queries are equally important. More theoretical support and fine-tune experiments are needed.
1.If we use the new gigaword dataset to enhance the proformance of our model. Our model may only study the presentation knowledge under the circumstances of "distance 1".
2.Most explicit relation classifier only trained with the top-2 level losses, so they may not have the ability to judge whether the relation is Cause.reason or Cause.result. We may train another classifier to do this.
1.Many errors in ECPE benchmark dataset. For example, in instance 12-7 '但 我 也 很 感谢 那些 帮助 我 的 网友 义工 宁养院', '宁养院' should be '疗养院' and a wrong segmentation metioned above. Maybe we should build our own English dataset.
Submission format
<type>: <subject>
Submission example
git commit -m 'feat: adding connectives and rewriting the document'
git commit -m 'fix: fixed an issue where an interface was not reachable'
git commit -m 'docs: update readme'
usual type
- feat: add a function
- fix: fix a bug
- pref: optimize performance
- docs: edit a doc
1.We need to do robust analysis and error analysis. And we need to write something about special examples, like action/trigger words and cascade events.
2.Creatively, we can do ablation experiments.
3.We can also do probe experiments on word meaning relevancy and word analogy of connectives.
In some scenarios, in addition to extracting objects and corresponding emotions, deeper point of view mining also requires extracting the causes that cause emotions to assist in subsequent more complex causal analysis.
Emotional cause pair extraction is a task that strongly relies on context, the system needs to model the relationship between emotion and cause, and the cause of emotion in many scenarios may not be in the emotional sentence itself, for example, the reason why multiple rounds of dialogue trigger a certain emotion may be the previous rounds.
Emotional cause pair extraction is also a key task for opinion mining interpretability. Emotional reasoning digs into the specific causes of emotions. The reasoned emotional causes will be used as interpretable information to complement some existing tasks, which will greatly promote both the analytical platform and the dialogue system.
1.We deliver a simple but powerful method to extract the emotion-cause pair, which is explicating the implicit connectives in the original corpus and then can extract the pairs by explicit course relation classifier.
2.Based on the MRC formalization, We develop our model based on a novel and reusable query model. We make detailed analysis of our model.
3.We demonstrate that our proposed method outperforms existing state-of-the-art performance.
4.We make the most detailed analysis in the research field of connectives of PDTB and ECPE to our best knowledge.
1.For natural language generation task, we can add previously unexistable connectives, reason through the logical chain or template, and then delete these connectives.
2.We want to use our method to do more experiments about implicit chapter relation classification, refined connective word selection, and probing connectives representation.
3.We'd like to study the evolutionary patterns of emotional dynamics. Add the implicit connective mining task to explicitly let the model learn the dynamic characteristics of emotions, build models based on context, time series, and event information, and compare their effects on the dynamic evolution of emotions.
4.We believe the temporal relations are event-driven. We want to do experiments about trigger words in ECPE and make use of TempEval.
5.We want to enhance the results of discourse relation analysis on PDTB by automatically labeling the emotional polarity and using our model to identify cause relation in the discourse.
6.We will use cross-domain pre-training to enhance the model on the original basis, such as by collecting sentences with explicit connectives as unsupervised pre-training corpus.
7.Possible subsequent extended experiments include using the probe method to study the semantic relevance and analogy of connectives.