Research on Toxic Speech Detection Based on Large Language Models

Toxic Speech (TS) is widely spread in social media and is increasingly harmful to the online networking environment. At the same time, words and their variants containing implicit toxic semantics are influenced by the humanistic and social environment of the language and are extremely rich in form and shape, and the lack of background knowledge related to the "discourse body" of existing research has led to the confusion of Implicit Toxic Speech (ITS) with Non-Toxic Speech (NTS). Therefore, this paper proposes the Prompt-enhanced Attention Head Fusion (PAHF) toxic text classification model based on the generative language model. We optimize the hidden layer vectors by intervening in the hidden layer information through prompt, then filter the vectors by combining linear probes to locate those vectors in the hidden layer information that are strongly correlated with the encoding of the background knowledge, and finally construct an attention fusion mechanism to fuse the filtered features, which provides a higher detection accuracy compared with the existing algorithms.