基于标签感知增强的社交媒体心理亚健康归因方法

doi:10.3969/j.issn.1000-5641.2025.01.010

摘要/Abstract

摘要：

在大量的网络社交媒体中, 存在一些表达了潜在的心理健康障碍和精神疾病的帖子, 根据帖子文本识别用户产生心理健康障碍的原因是一项重要任务. 观察这些帖子发现, 其上下文中存在标签共现现象, 即上下文中同时出现了多个候选标签的语义, 干扰了标签表征的建模与预测. 为缓解该现象带来的影响, 提出了一种标签感知增强分类的方法, 该方法利用大规模预训练语言模型识别潜在的候选标签, 并通过估计样本独立的标签语义强度作为增强数据以消减共现标签带来的噪声, 基于增强数据构建了性能良好的预训练语言模型分类器. 在数据集Intent_SDCNL和SAD上进行的实验验证了该方法的有效性.

关键词: 心理亚健康归因, 自然语言处理, 提示学习

Abstract:

Online social media are frequently used by people as a way of expressing their thoughts and feelings. Among the vast amounts of online posts, there may be more concerning ones expressing potential grievances and mental illnesses. Identifying these along with potential causes of mental health problems is an important task. Observing these posts, it is found that there is a label co-occurrence phenomenon in contexts, i.e., the semantics of multiple candidate labels appear in the context of one sample, which interferes with the modeling and prediction of label patterns. To mitigate the impact of this phenomenon, we propose a label-aware data augmentation method, which leverages large-scale pre-trained language models with excellent text comprehension capability to identify potential candidate labels, abates the noise from irrelevant co-occurring labels by estimating sample-independent label semantic strengths, and constructs well-performing classifiers with pre-trained language models. Extensive experiments validate the effectiveness of our model on the recent datasets Intent_SDCNL and SAD.

Key words: causal analysis of mental health, natural language processing, prompt learning

中图分类号:

TP391.1

梁怡萍, 肖路巍, 王琳琳. 基于标签感知增强的社交媒体心理亚健康归因方法[J]. 华东师范大学学报（自然科学版）, 2025, 2025(1): 124-137.

Yiping LIANG, Luwei XIAO, Linlin WANG. Label-perception augmented causal analysis of mental health over social media[J]. J* E* C* N* U* N* S*, 2025, 2025(1): 124-137.

图/表 10

图1

表1

表2

Intent_SDCNL与SAD数据集上使用的任务指令$ {p} $和元提示$ {\delta },{\eta } $"

内容	Intent_SDCNL	SAD
任务指令$ p $的内容	Post: [Post]. Based on the post, Is [Label_desc] the cause of this individual’s mental health issues? You must only return “yes” or “no”.	Post: [Post]. Based on the post, Is [Label_desc] the source of stress of this individual in everyday life? You must only return “yes” or “no”.
元提示$ \delta $的内容	I’m trying to recognize what causes mental health issues from social media posts. My current cause is: “[Label_desc]” But this cause gets the following examples wrong: [error_string] give only [num_feedbacks] reasons why the cause could have gotten these examples wrong. Wrap each reason with $ \langle$START$\rangle $ and $\langle $END$\rangle $	I’m trying to recognize what causes everyday life stress from social media posts. My current cause is: ”[Label_desc]” But this cause gets the following examples wrong: [error_string] give only [num_feedbacks] reasons why the cause could have gotten these examples wrong. Wrap each reason with $\langle $START$\rangle $ and $\langle $END$\rangle $
元提示$ \eta $的内容	I’m trying to recognize what causes mental health issues from social media posts. My current cause is: “[Label_desc]” But this cause gets the following examples wrong: [error_string] Based on these examples the problem with this cause is that [gradient] Based on the above information, wrote [steps_per_gradient] different improved cause phrases without interpretations. Each cause is wrapped with $\langle $START$\rangle $ and $\langle $END$\rangle $	I’m trying to recognize what causes everyday life stress from social media posts. My current cause is: “[Label_desc]” But this cause gets the following examples wrong: [error_string] Based on these examples the problem with this cause is that [gradient] Based on the above information, wrote [steps_per_gradient] different improved cause phrases without interpretations. Each cause is wrapped with $\langle $START$\rangle $ and $\langle $END$\rangle $

表2

表3

表4

表5

图2

图3

图4

图5

参考文献 28

1	MAURIELLO M L, LINCOLN T, HON G, et al. SAD: A stress annotated dataset for recognizing everyday stressors in SMS-like conversational systems [C]// Extended Abstracts of the 2021 CHI Conference on Human Factors in Computing Systems. 2021: 399.
2	GARG M, SAXENA C, SAHA S, et al. CAMS: An annotated corpus for causal analysis of mental health issues in social media posts [C]// Proceedings of the Thirteenth Language Resources and Evaluation Conference. 2022: 6387-6396.
3	SAXENA C, GARG M, ANSARI G. Explainable causal analysis of mental health on social media data [C]// International Conference on Neural Information Processing. 2022: 172-183.
4	YANG K, JI S, ZHANG T, et al. Towards interpretable mental health analysis with large language models [C]// Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023: 6056-6077.
5	YANG K, ZHANG T, KUANG Z, et al. MentaLLaMA: Interpretable mental health analysis on social media with large language models [EB/OL]. (2023-09-24)[2024-01-06]. https://arxiv.org/abs/2309.13567.
6	ZHOU Y, MURESANU A I, HAN Z, et al. Large language models are human-level prompt engineers [EB/OL]. (2023-03-10)[2024-01-06]. https://arxiv.org/abs/2211.01910.
7	XU N, QIAO C, GENG X, et al.. Instance-dependent partial label learning. Advances in Neural Information Processing Systems, 2021, 34, 27119- 27130.
8	RIBEIRO M T, SINGH S, GUESTRIN C. “Why should I trust you?”: Explaining the predictions of any classifier [C]// Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. 2016: 1135-1144.
9	ZHAO W X, ZHOU K, LI J, et al. A survey of large language models [EB/OL]. (2023-11-24)[2024-01-06]. https://arxiv.org/abs/2303.18223.
10	TOUVRON H, LAVRIL T, IZACARD G, et al. LLaMA: Open and efficient foundation language models [EB/OL]. (2023-02-27)[2024-01-06]. https://arxiv.org/abs/2302.13971.
11	VASWANI A, SHAZEER N, PARMAR N, et al. Attention is all you need[J]. Proceedings of the 31st International Conference on Neural Information Processing Systems. 2017: 6000-6010.
12	DEVLIN J, CHANG M W, LEE K, et al. BERt: Pre-training of deep bidirectional transformers for language understanding [EB/OL]. (2019-05-24)[2024-01-06]. https://arxiv.org/abs/1810.04805.
13	LIU P, YUAN W, FU J, et al.. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys, 2023, 55 (9): 195- 35.
14	LU Y, BARTOLO M, MOORE A, et al. Fantastically ordered prompts and where to find them: Overcoming few-shot prompt order sensitivity [C]// Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics. 2022: 8086-8098.
15	HOFFMANN J, BORGEAUD S, MENSCH A, et al.. An empirical analysis of compute-optimal large language model training. Advances in Neural Information Processing Systems, 2022, 35, 30016- 30030.
16	BROWN T, MANN B, RYDER N, et al.. Language models are few-shot learners. Advances in Neural Information Processing Systems, 2020, 33, 1877- 1901.
17	CHUNG H W, HOU L, LONGPRE S, et al. Scaling instruction-finetuned language models [EB/OL]. (2022-12-06)[2024-01-06]. https://arxiv.org/abs/2210.11416.
18	WEI J, WANG X, SCHUURMANS D, et al.. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 2022, 35, 24824- 24837.
19	BLEI D M, KUCUKELBIR A, MCAULIFFE J D.. Variational inference: A review for statisticians. Journal of the American Statistical Association, 2017, 112 (518): 859- 877.
20	HOFFMAN M, BLEI D M, WANG C, et al.. Stochastic variational inference. Journal of Machine Learning Research, 2013, 14, 1303- 1347.
21	RANGANATH R, GERRISH S, BLEI D. Black box variational inference [C]// Proceedings of the 17th International Conference on Articial Intelligence and Statistics. 2014: 814-822.
22	OPPER M, SAAD D. Advanced Mean Field Methods: Theory and Practice [M]. Cambridge: Massachusetts Institute of Technology Press, 2001.
23	PRYZANT R, ITER D, LI J, et al. Automatic prompt optimization with “gradient descent” and beam search [C]// Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023: 7957-7968.
24	ZHU X. Semi-supervised Learning with Graphs [M]. Pennsylvania, USA: Carnegie Mellon University Press, 2005.
25	KIPF T N, WELLING M. Variational graph auto-encoders [EB/OL]. (2016-11-21)[2024-01-06]. https://arxiv.org/abs/1611.07308.
26	KIPF T N, WELLING M. Semi-supervised classification with graph convolutional networks [EB/OL]. (2017-02-22)[2024-01-06]. https://arxiv.org/abs/1609.02907.
27	FIGURNOV M, MOHAMED S, MNIH A. Implicit reparameterization gradients [C]// Proceedings of the 32nd International Conference on Neural Information Processing Systems. 2018: 439-450.
28	POPESCU M C, BALAS V E, PERESCU-POPESCU L, et al.. Multilayer perceptron and neural networks. WSEAS Transactions on Circuits and Systems, 2009, 8 (7): 579- 588.

数据集	训练集的样本数量/条	测试集的样本数量/条	标签	标签数/个
Intent_SDCNL	1461	370	无理由(NR, 19%) 虐待与偏见(BA, 13%) 药物(M, 12%) 人际关系(R, 20%) 职业(JC, 12%) 异化(A, 23%)	6
SAD	5870	980	财务困难(FP, 9%) 情感波动(ET, 10%) 家庭事务(FI, 11%) 学校(S, 11%) 健康, 疲劳或身体疼痛(H, 11%) 社交关系(SR, 9%) 其他(O, 14%) 日常决策(ED, 5%) 工作(W, 19%)	9

模型	微观F₁/%		模型	微观F₁/%
模型	Intent_SDCNL	SAD	模型	Intent_SDCNL	SAD
CNN^[4]	34.63	38.45	LLaMA2-7B^[5]	30.47	49.60
GRU^[4]	29.33	34.79	MentaLLaMA-chat-13B^[5]	45.52	63.62
BiLSTM^[4]	29.49	38.50	ChatGPT_CoT _emo^[4]	45.99	63.56
MentalRoBERTa^[4]	47.62	68.44	GPT-4_Few-shot^[5]	42.37	55.68
BART-Large^[5]	43.80	59.60	RoBERTa_base^[5]	36.54	67.53
T5-Large^[5]	40.20	58.10	MentaLLaMA-7B^[5]	32.52	49.93

模型	分类器	准确度/%		宏观精确率/%		宏观召回率/%		宏观F₁/%
模型	分类器	Intent_SDCNL	SAD	Intent_SDCNL	SAD	Intent_SDCNL	SAD	Intent_SDCNL	SAD
chatglm3-6B	RoBERTa	62.21±0.27	81.24±1.49	68.76±1.98	76.30±1.12	59.20±0.43	73.72±0.53	59.90±0.36	74.03±0.50
	Longformer	65.13±0.77	82.24±1.26	68.63±0.87	77.49±1.94	60.87±0.28	74.80±1.74	61.69±0.39	75.00±1.91
	FLAN-T5	50.34±1.09	80.47±0.82	62.78±5.94	75.08±1.77	45.95±0.44	72.76±1.33	45.60±0.45	72.52±0.72
	GPT-2	56.50±0.30	81.32±2.02	57.48±0.64	76.02±3.04	52.88±1.01	73.78±1.29	52.72±1.31	73.92±1.96
Qwen-7B	RoBERTa	57.30±0.42	78.14±0.99	66.75±5.09	75.99±2.30	50.71±1.41	72.75±0.77	49.90±1.49	73.20±0.81
	Longformer	53.30 ±0.52	79.07±1.35	58.33±2.02	73.61±2.17	46.44±0.48	71.33±0.65	45.89±0.36	71.14±1.33
	FLAN-T5	42.36±0.88	77.08±1.47	53.07±2.24	73.56±2.93	33.89±0.57	68.50±2.21	32.18±1.01	68.19±2.86
	GPT-2	54.69±1.58	76.22±1.76	54.18±4.49	69.92±3.27	48.10±1.56	68.54±1.48	47.36±1.50	68.57±1.64
Baichuan-13B-Chat	RoBERTa	64.99±0.69	85.03±1.24	64.30±0.11	83.56±0.35	62.63±0.62	79.11±0.23	61.76±0.57	79.97±0.25
	Longformer	65.62±0.75	83.40±0.89	64.75±1.38	79.42±1.71	63.34±1.59	76.41±1.60	62.73±2.04	76.62±1.57
	FLAN-T5	55.52±0.82	82.60±1.96	56.84±0.55	78.46±0.65	51.83±0.80	74.24±2.10	53.03±0.84	74.78±2.13
	GPT-2	61.11±1.83	79.67±2.11	62.18±1.76	75.80±2.87	58.16±2.61	72.46±1.43	58.55±2.00	72.09±1.97

变体	分类器	准确度/%		宏观F₁/%
变体	分类器	Intent_SDCNL	SAD	Intent_SDCNL	SAD
w/o 增强	RoBERTa	58.67 ± 1.86	79.81 ± 1.73	56.55 ± 1.43	71.98 ± 2.13
	Longformer	59.86 ± 1.44	78.08 ± 1.69	56.52 ± 1.97	72.12 ± 1.39
	FLAN-T5	41.26 ± 2.43	71.86 ± 0.24	36.60 ± 1.20	62.41 ± 1.95
	GPT-2	52.35 ± 0.75	72.52 ± 0.85	48.65 ± 0.50	65.57 ± 0.69
w/o 标签语义搜索 (Baichuan-13B-Chat)	RoBERTa	62.67 ± 1.28	83.24 ± 1.49	59.78 ± 2.04	76.43 ± 0.97
	Longformer	63.91 ± 0.76	80.70 ± 1.04	61.37 ± 1.03	73.52 ± 1.02
	FLAN-T5	53.17 ± 1.69	79.35 ± 1.31	47.73 ± 3.30	70.65 ± 1.92
	GPT-2	58.39 ± 1.27	78.52 ± 2.01	55.93 ± 2.12	69.68 ± 1.84

[1]	肖开研, 廉洁. 基于多核支持向量机的句子分类算法[J]. 华东师范大学学报（自然科学版）, 2023, 2023(6): 85-94.
[2]	徐秋荣, 朱鹏, 罗轶凤, 董启文. 金融领域中文命名实体识别研究进展[J]. 华东师范大学学报（自然科学版）, 2021, 2021(5): 1-13.
[3]	穆肇南, 刘梦珠, 孙界平, 王成. 基于演化算法的唐诗自动生成系统研究[J]. 华东师范大学学报（自然科学版）, 2020, 2020(6): 129-139.
[4]	郭晓哲, 彭敦陆, 张亚彤, 彭学桂. GRS: 一种面向电商领域智能客服的生成-检索式对话模型[J]. 华东师范大学学报（自然科学版）, 2020, 2020(5): 156-166.
[5]	王嘉宁, 何怡, 朱仁煜, 刘婷婷, 高明. 基于远程监督的关系抽取技术[J]. 华东师范大学学报（自然科学版）, 2020, 2020(5): 113-130.
[6]	韩程程, 李磊, 刘婷婷, 高明. 语义文本相似度计算方法[J]. 华东师范大学学报（自然科学版）, 2020, 2020(5): 95-112.
[7]	陈远哲, 匡俊, 刘婷婷, 高明, 周傲英. 共指消解技术综述[J]. 华东师范大学学报(自然科学版), 2019, 2019(5): 16-35.