Medical Causation in Defining Emotions

Abstract

Emotions are an essential component of human nature, which can describe a person's health and help determine this condition's causes. Proceeding from this, it becomes obvious that health plays a vital role in forming one of emotion condition, and in the reverse order, any emotion can describe the state of human health. This approach can provide medical personnel with important information about patients: emotions, state of health, and establishing cause and effect relationships. How-ever, the creation of this model is hampered by the lack of large labeled datasets. Thus, the study's main goal is to create a dataset that would have information about the emotional state of a person and causal medical relationships that affect a person's emotional state. We conduct comprehensive data collection and analysis, using state-of-the-art models for assessing emotions, medical extractions of creatures, and determination of cause-and-effect relationships.

Keywords

• eHealth
• Emotion recognition
• Named Entity Recognition
• Cause-effect

Introduction

With online communication progress, emotional information becomes significantly valuable not only for social research but also for medical analysis. Life-threatening, severe symptoms such as coughing, breathing difficulties, heart failure, and fatigue cause a compassionate person’s state, leading to various feelings and emotions, such as surprise to anger or fear to joy, and others. Given emotions, for example, help detect treatment effect and state condition of human.

There are several problems in the research of emotion cause extraction. The most notable is no data for emotion cause analysis. First studies defined as a problem of emotion cause extraction described in [ 1 ]. Studying the experience of given research [ 2 ] improved and released a novel dataset that becomes a benchmark dataset for emotion cause extraction research. The task for emotion cause extraction was also studied in novel researches where the problem was addressed as a clause-level binary classification problem [ 3-5 ]. The next problem stands for the small size of the annotated corpus. Consequently, many deep learning models are not relevant for emotion cause extraction. The last problem in our research is defining the relationship between causes and health. Up-to-date improvement of medical text extraction researches [ 6-11 ] was made possible by applying machine learning techniques for medical named entity recognition (NER) and relation extraction (RE) applying modern models as Conditional Random Field Long Short-Term Memory. However, extracting medical text mining has limitations. To tackle this problem, recent study BioBERT [ 12 ] outperform all previous work and become a state-of-the-art benchmark for NER and RE tasks, which is based on powerful model BERT [ 13 ].

Deep learning opens an extensive range for research in any field and complex field as medicine. However, to work with deep learning, large data is required. Our work aims to build a model for analyzing human emotional behavior according to medical and other causes of this emotion. The main problem we tackled is the lack of quality data. For this purpose, we decide to create a corpus for our future research based on this topic. The paper has the following organization. Section 3 discusses the novel corpus creation, including algorithmic and implementation terms. Section 4 reviews the results, and Section 5 concludes and discus future work. 2

Construction of emotion cause corpus

In this section, we first describe the linguistic phenomenon in emotion expressions. It serves as the inspiration to develop the annotated dataset. We then introduce details of the annotation scheme, followed by the construction of the dataset. Today there is a lack of research and dataset for Emotion Cause, which makes this work relevant. To date, there are two studies by [ 14, 15 ].

2.1 Emotion-cause pair extraction corpus

ECPE corpus was constructed based on ECE corpus, where one utterance belong to one emotion and related to one causes. ECE corpus consist of Chinese news containing 20,000 articles. After removing irrelevant instances, there are 2,105 instances with cause relation. Emotion cause annotated as <cause >, and the emotion as <keywords >. Where, 97.2% of data has one emotion cause, other 2.8% respectively.

Example from data with cause: <keywords >sadness <keywords >, <cause >sadness <cause >, because of sadness excessive without cause: <keywords >fear <keywords >, <cause><cause >, there are lingering palpitations, she still has lingering palpitations.

2.2 Emotion-stimulus data

The Emotion-stimulus data corpus [ 16 ] consists of 820 rows of data including emotion tag and emotion cause. Data annotated in XML format: <cause>and <\ cause>belongs to emotion cause. However, <emotion type>describes emotion. The given study was built with FrameNet tool [ 17 ] into the frame of Ekman’s six emotion classes [ 18 ] and finally annotated by human to verify them.

Example from data: <fear >People are becoming more and more concerned, <cause >about the healthiness of their diet and way of life <\cause >. <\fear >.

2.3 Custom Web Dataset

Despite the fact that data were collected from available corpuses, this is still insufficient for extensive analysis. As a result, additional data was collected from “Psychiatric Treatment Adverse Reactions” (PsyTAR) [ 19 ] dataset and medical forums with total amount – 4000 of data. The difficulty lies in the fact that they are not annotated for causal relationships. This is the main task from which the following steps stand out. We split the emotion cause extraction task into two subtasks with the purpose to get a set of emotion clauses: and set of cause clauses for each document.

For cause relevance, we decide to use a keyword matching pattern. According to [ 12 ], there are six linguistic groups of keywords. They are essentially correlated with causes, as shown in Table 1. By given keywords, the manual corpus will be filtered. ‘for’, ’as’ ‘because’, ‘so’, ‘but’, ’after’ ‘to think about’, ‘to talk about’ V:EpistemicMarkers

‘to hear’, ‘to see’, ‘to know’, ‘to exist’ VI:Others

‘is’, say’, ‘at’, ‘can’

For emotion relevance, we collected data from two datasets: Twitter Emotion Corpus (TEC) [ 20 ] and CrowdFlower (CF) [ 21 ] with total amoubt: 61051 tweets and trained it on four models: Naive Bayes (NB) and Support Vector Machine (SVM), BERT, Multi-label BERT

After prepossessing, the number of examples per emotion decreased significantly, due to significant noises in data. As a result, we manually picked from filtered data 800 examples for each emotion for training and 200 examples for emotions for the test set.

We use the F1 score for evaluation, which calculated:

After training on given models we got we that multi label BERT outperforms different model. For other models results described in Table 2: Example from data: <happy >I feel much better after <cause >taking the headache medicine. <\cause ><\happy >.

As a result of three corpora, we got a single dataset consisting of ECPE, ESD, and CWD, described in Table 3. But since the main goal is to identify the medical reason in a particular emotion, medical relevance will be applied to the assembled dataset.

Sum 615 371 612 882 778 242 3500

For medical relevance we decide to use BioBERT, which significantly outper-forms previous state of the art researches in different types of medical text miningtasks, such as question answering ( MRR by 12.24%), named entity recognition(F1 by 0.62%) and medical relation extraction ( F1 by 2.80%) Medical clauses

After applying medical relevance we got results R as subtraction of given sets: The amount of data that is related to medicine decreased from 3500 to 986 data units, which is about 28% of all data. Final annotated data have XML format annotation. Where, <cause>and <\cause>belongs to the emotion cause. However, <mcause>and <\mcause>to the medical-emotion. For emotion, <emotion type> tag was applied.

Example from data: • For medical cause: <happy>I feel much better after <mcause> taking the headache medicine. <\mcause> <\happy> • For other cause: <sad >I am sad <cause>nobody wants to do it like I have done it for them. <\cause><\sad> 4

Conclusion and discussion

In this paper, we present our work on medical causation in defining emotions. Lack of data for building and training a model was the driving force for creating corpus. We also describe the medical emotion cause extraction method to capture required data consisting of 3 main methods: emotion relevance, cause relevance, medical relevance. For emotion and medical relevance, state-of-the-art BERT models were used. However, cause relevance stands for the key word matching method, which needs improvement in future work. Given corpus helps us create the first model for analyzing and extracting emotional causes related to health and different events. We believe that the proposed work will help better investigate treatment effect and help understand human health’s real state.

References

1. Chen Y. , Lee S. , Li . S, and Huang C. : Emotion cause detection with linguistic constructions . In Proceedings of the 23rd International Conference on Computational Linguistics (COLING) , pages 179 - 187 ( 2010 )

2. Gui L. , Wu D. , Xu R. , Lu Q. , and Zhou Y. : Event-driven emotion cause extraction with corpus construction . In Empirical Methods in Natural Language Processing (EMNLP) , pages 1639 - 1649 ( 2016 )

3. Li X. , Song K. , Feng S. , Wang D. , and Zhang Y.: A co-attention neural network model for emotion cause analysis with emotional context awareness . In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (EMNLP) , 4752 - 4757 ( 2018 )

4. Xu B. , Lin H. , Lin Y. , Diao Y. , Yang L. , and Xu L.: Extracting emotion causes using learning to rank methods from an information retrieval perspective . IEEE Access . ( 2019 )

5. Yu X. , Rong W. , Zhang Z., Ouyang Y , and Xiong Z.: Multiple level hierarchical network-based clause selection for emotion cause extraction . IEEE Access , 9071 - 9079 . ( 2019 )

6. Habibi M. et al.: Deep learning with word embeddings improves biomedical named entity recognition . Bioinformatics , 33 , 37 - 48 ( 2017 )

7. Bhasuran B. and Natarajan J.: Automatic extraction of gene-disease associations from literature using joint ensemble learning . PLoS One , 13 , e0200699 ( 2018 )

8. Giorgi J.M. and Bader G.D. : Transfer learning for biomedical named entity recognition with neural networks . Bioinformatics , 34 , 4087 ( 2018 )

9. Wang X. et al.: Cross-type biomedical named entity recognition with deep multi-task learning . Bioinformatics , 35 , 1745 - 1752 ( 2018 )

10. Lim S. and Kang J. Chemical-gene relation extraction using recursive neural network . Database . ( 2018 )

11. Yoon W. et al.: Collabonet: collaboration of deep neural networks for biomedical named entity recognition . BMC Bioinformatics , 20 , 249 ( 2019 )

12. Lee J. , Yoon W. , Kim S. , Kim D. , Kim S. , So C. , Kang J.: BioBERT: a pre-trained biomedical language representation model for biomedical text mining , Bioinformatics , Volume 36 , 1234 - 1240 ( 2020 )

13. Devlin J . et al.: Bert: pre-training of deep bidirectional transformers for language understanding . In: Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , Volume 1 (Long and Short Papers), Minneapolis, MN, USA. pp. 4171 - 4186 ( 2019 )

14. Xia R. and Ding Z. : 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 , 1003 - 1012 ( 2019 )

15. Ghazi D. , Inkpen D. , and Szpakowicz S.: Detecting emotion stimuli in emotion-bearing sentences . In International Conference on Intelligent Text Processing and Computational Linguistics (CICLing) , pages 152 - 165 ( 2015 )

16. Ghazi W. , Inkpen D. and Szpakowicz S. : Computational Linguistics and Intelligent Text Processing Lecture Notes in Computer Science , Vol. 9042 , 152 - 165 ( 2015 )

17. Fillmore C.J. , Petruck M.R. , Ruppenhofer J. , Wright A .: FrameNet in Action: The Case of Attaching. IJL 16 ( 3 ), 297 - 332 ( 2003 )

18. Ekman P.: An argument for basic emotions . Cognition & Emotion 6 ( 3 ), 169 - 200 ( 1992 )

19. Zolnoori M. et al.: “The PsyTAR dataset: From patients generated narratives to a corpus of adverse drug events and effectiveness of psychiatric medications . ” Data in Brief 24 . ( 2019 )

20. Twitter Emotion Corpus 2012 . http://saifmohammad.com/WebPages/SentimentEmotionLabeledData.html

21. CrowdFlower . 2016 . https://www.figureeight.com /data/sentiment-analysis-emotion-text/.