通过数据增强和知识蒸来增强攻击性语言检测
Jiawen Deng1,1, Zhuang Chen1, Hao Sun1
1The CoAI group, DCST; Institute for Artificial Intelligence; State Key Lab of Intelligent Technology and Systems; Beijing National Research Center for Information Science and Technology; Tsinghua University, Beijing 100084, China.
Research (Washington, D.C.)
|September 20, 2023
概括
这项研究介绍了AugCOLD,这是一百万个样本数据集,用于改进检测中文攻击性语言. 一个新的多层蒸框架提高了模型性能和稳定性,以实现更安全的在线通信.
科学领域:
- 自然语言处理自然语言处理.
- 计算语言学 计算语言学
- 人工智能的人工智能
背景情况:
- 攻击性语言的检测对于社交媒体和安全的AI部署至关重要.
- 与英语资源相比,中国现有的攻击性语言数据集在规模和范围上是有限的.
- 这种数据稀缺性阻碍了中国冒犯性语言检测器的准确性,特别是在复杂或新的案例中.
研究的目的:
- 为了解决现有的中国攻击性语言数据集的局限性.
- 开发一个大规模的,无监督的数据集,用于训练更强大的探测器.
- 提高中国攻击性语言检测模型的性能和概括能力.
主要方法:
- 介绍了AugCOLD (增强的中文攻击性语言数据集),这是通过数据抓取和模型生成创建的100万样本无监督数据集.
- 采用多层次知识蒸框架来利用无监督数据.
- 利用公开可用的数据集来训练多个教师模型,然后将软标签分配给AugCOLD,以便将知识传输到学生网络 (最终检测器).
主要成果:
- 在攻击性语言检测性能方面显著改进.
- 在各种测试集中展示了攻击性语言检测器的增强概括性和稳定性,包括具有挑战性的硬案例.
- 用AugCOLD数据集验证了拟议的多层蒸方法的有效性.
结论:
- AugCOLD数据集和多层蒸框架有效地解决了中国攻击性语言数据的稀缺问题.
- 拟议的方法显著提高了中国攻击性语言检测器的准确性,概括性和稳定性.
- 这项工作有助于更安全的在线通信和在中国语境中负责任地部署大型语言模型.
相关概念视频
02:07Improving Translational Accuracy
11.5K
Base complementarity between the three base pairs of mRNA codon and the tRNA anticodon is not a failsafe mechanism. Inaccuracies can range from a single mismatch to no correct base pairing at all. The free energy difference between the correct and nearly correct base pairs can be as small as 3 kcal/ mol. With complementarity being the only proofreading step, the estimated error frequency would be one wrong amino acid in every 100 amino acids incorporated. However, error frequencies observed in...
11.5K
01:24Generalization, Discrimination, and Extinction
604
Generalization, discrimination, and extinction are key concepts in operant conditioning that influence how behaviors are learned and maintained.
Generalization occurs when a behavior reinforced in one context is performed in similar situations. For instance, a student who studies diligently for calculus and receives excellent grades might apply the same study habits to psychology and history, expecting similar results. Generalization shows how learning in one setting can influence behavior in...
Generalization occurs when a behavior reinforced in one context is performed in similar situations. For instance, a student who studies diligently for calculus and receives excellent grades might apply the same study habits to psychology and history, expecting similar results. Generalization shows how learning in one setting can influence behavior in...
604
01:22Language Development
394
Children master language quickly and with relative ease, supported by both biological predisposition and reinforcement. B. F. Skinner (1957) proposed that language is learned through reinforcement, while Noam Chomsky (1965) argued that language acquisition mechanisms are biologically determined.
The critical period for language acquisition suggests that the ability to acquire language is at its peak early in life. As people age, this proficiency decreases. Language development begins very...
The critical period for language acquisition suggests that the ability to acquire language is at its peak early in life. As people age, this proficiency decreases. Language development begins very...
394
01:09Censoring Survival Data
125
Survival analysis is a statistical method used to analyze time-to-event data, often employed in fields such as medicine, engineering, and social sciences. One of the key challenges in survival analysis is dealing with incomplete data, a phenomenon known as "censoring." Censoring occurs when the event of interest (such as death, relapse, or system failure) has not occurred for some individuals by the end of the study period or is otherwise unobservable, and it might have many different...
125
01:00Detection of Gross Error: The Q Test
6.2K
When one or more data points appear far from the rest of the data, there is a need to determine whether they are outliers and whether they should be eliminated from the data set to ensure an accurate representation of the measured value. In many cases, outliers arise from gross errors (or human errors) and do not accurately reflect the underlying phenomenon. In some cases, however, these apparent outliers reflect true phenomenological differences. In these cases, we can use statistical methods...
6.2K
01:05Difference from Background: Limit of Detection
6.4K
The limit of detection (LOD) is the smallest amount of analyte that can be distinguished from the background noise. The LOD value corresponds to the concentration at which the analyte signal is three times larger than the standard deviation of the blank signal. Below this value, the analyte signal cannot be differentiated from the background noise. It is calculated by dividing the calibration slope by 3 times the standard deviation of the blank signals.
The LOD indicates the presence or absence...
The LOD indicates the presence or absence...
6.4K