Entity at SemEval-2021 Task 5: Weakly Supervised Token Labelling for Toxic Spans Detection
Vaibhav Jain | Mina Naghshnejad
Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021)
Detection of toxic spans - detecting toxicity of contents in the granularity of tokens - is crucial for effective moderation of online discussions. The baseline approach for this problem using the transformer model is to add a token classification head to the language model and fine-tune the layers with the token labeled dataset. One of the limitations of such a baseline approach is the scarcity of labeled data. To improve the results, We studied leveraging existing public datasets for a related but different task of entire comment/sentence classification. We propose two approaches: the first approach fine-tunes transformer models that are pre-trained on sentence classification samples. In the second approach, we perform weak supervision with soft attention to learn token level labels from sentence labels. Our experiments show improvements in the F1 score over the baseline approach. The implementation has been released publicly.
This paper presents a vector initialization approach for the SemEval2020 Task 1: Unsupervised Lexical Semantic Change Detection. Given two corpora belonging to different time periods and a set of target words, this task requires us to classify whether a word gained or lost a sense over time (subtask 1) and to rank them on the basis of the changes in their word senses (subtask 2). The proposed approach is based on using Vector Initialization method to align GloVe embeddings. The idea is to consecutively train GloVe embeddings for both corpora, while using the first model to initialize the second one. This paper is based on the hypothesis that GloVe embeddings are more suited for the Vector Initialization method than SGNS embeddings. It presents an intuitive reasoning behind this hypothesis, and also talks about the impact of various factors and hyperparameters on the performance of the proposed approach. Our model ranks 12th and 10th among 33 teams in the two subtasks. The implementation has been shared publicly.