Supervised Gradual Machine Learning for Aspect Category Detection

TL;DR

This paper tackles Aspect Category Detection (ACD) under limited labeled data by introducing a supervised Gradual Machine Learning (GML) framework that couples deep neural networks (DNNs) for semantic relation modeling with a factor-graph based gradual inference mechanism. Semantic relations between sentence pairs are captured via a category-specific DNN and a Bert-based binary model, and are encoded as binary factors in a GML graph to progressively propagate labels from easy to hard instances. The method yields state-of-the-art performance on SemEval 2014/2016, MAMS, and SentiHood, with demonstrated data efficiency (e.g., competitive results at 30% of labeled data) and robustness to the number of relations $k_b$ used in inference. The approach offers a practical path to high-accuracy ACD in settings with scarce labeled data and could be extended to other fine-grained text classification tasks.

Abstract

Aspect Category Detection (ACD) aims to identify implicit and explicit aspects in a given review sentence. The state-of-the-art approaches for ACD use Deep Neural Networks (DNNs) to address the problem as a multi-label classification task. However, learning category-specific representations heavily rely on the amount of labeled examples, which may not readily available in real-world scenarios. In this paper, we propose a novel approach to tackle the ACD task by combining DNNs with Gradual Machine Learning (GML) in a supervised setting. we aim to leverage the strength of DNN in semantic relation modeling, which can facilitate effective knowledge transfer between labeled and unlabeled instances during the gradual inference of GML. To achieve this, we first analyze the learned latent space of the DNN to model the relations, i.e., similar or opposite, between instances. We then represent these relations as binary features in a factor graph to efficiently convey knowledge. Finally, we conduct a comparative study of our proposed solution on real benchmark datasets and demonstrate that the GML approach, in collaboration with DNNs for feature extraction, consistently outperforms pure DNN solutions.

Figures (6)

• Figure 1: An illustrative example of the GML paradigm. In brief, GML consists three steps. Firstly, easy instance labeling relies on simple rules or unsupervised learning to automatically label some instances in the task. Secondly, feature extraction and influence modeling extract common features from labeled and unlabeled instances and model their influence on labels. Lastly, gradual inference labels instances with increasing difficulty using iterative inference, choosing the instance with the highest degree of evidential certainty at each iteration until all instances are labeled.
• Figure 2: An example of the proposed solution, which consists of two main steps. Firstly, it identifies common semantic features among instances by modeling two types of relations: BERT-based and KNN-based relations. These semantic connections between instances provide a high-level description of whether two instances discuss the same aspect. Secondly, the instances are gradually labeled based on their level of difficulty. The gradual labeling process begins with the easiest instances and progresses to the more challenging ones. The information gained from labeling the easier instances is then used to guide the labeling of the harder ones. This way, the labels assigned to the easier instances serve as a reference for labeling the more difficult ones, resulting in a more efficient and accurate labeling process.
• Figure 3: The distribution of category Menu's instances based on final layer of QACG-BERT.
• Figure 4: An example of gradual inference of the running example.
• Figure 5: The performance of GML on each category of SemEval and MAMS compared with QACG-BERT in terms of Macro-F1.

Theorems & Definitions (1)

• Definition 1: Aspect Category Detection