LIME: Weakly-Supervised Text Classification without Seeds

Proceedings of the 29th International Conference on Computational Linguistics, pages 1083–1088 October 12–17, 2022. 1083 LIME: Weakly-Supervised Text Classification Without Seeds Seongmin Park and Jihwa Lee ActionPower, Seoul, Republic of Korea {seongmin.park, jihwa.lee}@actionpower.kr Abstract In weakly-supervised text classification, only label names act as sources of supervision. Pre- dominant approaches to weakly-supervised text classification utilize a two-phase framework, where test samples are first assigned pseudo- labels and are then used to train a neural text classifier. In most previous work, the pseudo- labeling step is dependent on obtaining seed words that best capture the relevance of each class label. We present LIME1, a framework for weakly-supervised text classification that entirely replaces the brittle seed-word gener- ation process with entailment-based pseudo- classification. We find that combining weakly- supervised classification and textual entailment mitigates shortcomings of both, resulting in a more streamlined and effective classification pipeline. With just an off-the-shelf textual en- tailment model, LIME outperforms recent base- lines in weakly-supervised text classification and achieves state-of-the-art in 4 benchmarks. 1 Introduction Weakly-supervised text classification (Meng et al., 2018) is an important avenue of research in low- resourced text classification. Unlike in traditional text classification, all supervision derives from textual information in category names. Weakly- supervised classification offers a practical approach to classification because it does not necessitate mas- sive amounts of training data. Another distinct aspect of weakly-supervised text classification is that the system has access to the entire test set at evaluation time, instead of encountering test samples sequentially. Exploit- ing this characteristic, recent approaches employ keyword-matching pseudo-labeling schemes to ten- tatively assign class labels to each test sample, be- fore using the information to train a separate classi- fier (Meng et al., 2018; Mekala and Shang, 2020; 1Labels Identified with Maximal Entailment Wang et al., 2021). Pseudo-labels are assigned by counting how many “seed words” of each class are found in the test sample. Keyword matching-based labeling, however, is neither adaptable nor flexible because semantic information embedded in class names cannot be extracted adaptively for distinct classification tasks. Inspired by recent advances in prompt-based text classification (Yin et al., 2019, 2020; Schick and Schütze, 2021), we replace the keyword- based pseudo-labeling step with a more streamlined entailment-based approach. Extensive experiments show that entailment-based classifiers assign more accurate pseudo-labels with greater task adaptabil- ity and much fewer hyperparameters. We find that our method realizes the benefits of both entailment- based classification and self-training. Our contributions are as follows: 1. We present LIME, a novel framework for weakly-supervised text classification that uti- lizes textual entailment. LIME surpasses cur- rent state-of-the-art weakly-supervised meth- ods in all tested benchmarks. 2. We show that self-training with pseudo-labels can mitigate unsolved robustness issues in entailment-based classification (Ma et al., 2021). 3. We experimentally confirm that higher con- fidence in pseudo-labels translates to better classification accuracy in self-training. We also find that a balance between filtering out low-confidence labels and preserving a sizable pseudo-training corpus is important. 2 Background 2.1 Weakly-supervised text classification In weakly-supervised text classification, the …

Content truncated to 500 words