Source-linked AI summary
Lifelong Learning for Sentiment Classification
Zhiyuan Chen, Nianzu Ma, Bing Liu
TL;DR
Sentiment classification lacks a unified lifelong-learning formulation that retains knowledge from prior tasks for future learning. The paper proposes LSC, a Bayesian optimization method using stochastic gradient descent and penalty terms, and reports that it outperforms baselines significantly across diverse product domains.
Problem
The paper addresses how sentiment classifiers can retain knowledge from past tasks and use it to improve learning on future tasks.
Method
LSC uses a Naïve Bayesian optimization framework with stochastic gradient descent and penalty terms to incorporate knowledge from past learning.
Results
LSC achieves the best performance in both natural and balanced class distributions, with statistically significant improvements over all baselines.
Takeaways & Limitations
Experiments on 20 diverse product-review domains support lifelong learning as a promising direction for building better sentiment classifiers.
Abstract
from arXiv · showhide
This paper proposes a novel lifelong learning (LL) approach to sentiment classification. LL mimics the human continuous learning process, i.e., retaining the knowledge learned from past tasks and use it to help future learning. In this paper, we first discuss LL in general and then LL for sentiment classification in particular. The proposed LL approach adopts a Bayesian optimization framework based on stochastic gradient descent. Our experimental results show that the proposed method outperforms baseline methods significantly, which demonstrates that lifelong learning is a promising research direction.
1 Introduction
The paper frames sentiment classification as a lifelong learning problem: retain knowledge from earlier tasks and use it to improve later classifiers. It defines lifelong learning and proposes LSC, an optimization approach that incorporates prior knowledge through penalty terms.
- Lifelong Learning: Lifelong learning uses knowledge gained from past tasks to help learning on a future task.The paper identifies retained information, usable knowledge, acquisition, and knowledge use as central design questions.
- Lifelong Sentiment Classification: Lifelong sentiment classification applies this paradigm to sequential supervised sentiment tasks with positive and negative labels.For each new task, knowledge from earlier tasks is used to learn a better classifier.
- Relation to Transfer Learning: Unlike classic transfer learning, lifelong learning uses results or knowledge mined from past learning rather than only labeled source-domain data.The distinction matters even when the target domain already has labeled training data.
- Proposed Approach: The proposed LSC method uses a Naïve Bayesian framework with stochastic gradient descent and penalty terms to incorporate prior knowledge.The approach focuses on binary sentiment classification but is also applicable to multi-class classification.
- Evaluation: The paper evaluates LSC on a large corpus containing reviews from 20 diverse product domains and reports superiority over baseline methods.This corpus supports extensive evaluation of lifelong sentiment classification.
2 Related Work
The related work situates the paper among lifelong learning, multi-task learning, transfer learning, and domain adaptation approaches. It distinguishes the proposed setting from methods that jointly learn tasks or transfer source-domain information without using knowledge from past learning.
- Lifelong and Multi-Task Learning: The work is related to lifelong learning approaches that exploit invariances or other knowledge across multiple tasks.It is also connected to multi-task learning, which optimizes multiple related tasks simultaneously.
- Transfer Learning and Domain Adaptation: Prior sentiment-classification transfer methods use labeled or unlabeled reviews, shared pivot features, or combined source-target features.These approaches represent transfer learning or domain adaptation in the sentiment-classification context.
3 Proposed LSC Technique
The proposed LSC technique combines Naïve Bayes with stored knowledge from past tasks, optimizing virtual word counts by stochastic gradient descent and knowledge-based penalties.
- Past Information and Knowledge: LSC stores past task results rather than original data, including class-conditional word probabilities and positive/negative word counts.These stored quantities feed a knowledge base containing document-level and domain-level knowledge.
- Knowledge Processing: The knowledge miner aggregates past information, while the knowledge-based learner incorporates it through regularization penalty terms.The method uses these components to embed prior knowledge into optimization.
- Objective Function: The objective optimizes virtual positive and negative word counts so target-domain documents receive high posterior probability for their labeled class and low probability for the other class.The variables X+,w and X−,w distinguish optimized virtual counts from empirical counts N+,w and N−,w.
- Optimization: Stochastic gradient descent updates the virtual counts document by document, using empirical and knowledge-base counts as starting points until convergence.The reported convergence criterion is a difference below 1e−3 between consecutive objective-function values.
- Knowledge-Based Penalties: Penalty terms address domain-dependent sentiment words by emphasizing target-domain counts for discriminative words and reliable domain-level knowledge for words appearing across enough past domains.The method uses target-domain distinguishability and a domain-frequency threshold τ to determine which knowledge is trusted.
4 Experiments
The experiments evaluate lifelong sentiment classification across 20 product domains under natural and balanced class distributions. LSC achieves the best reported performance, and its performance improves as more past domains contribute knowledge.
- Datasets: The corpus contains 20 diverse Amazon product domains, with 1,000 reviews per domain and ratings converted into positive or negative labels.The datasets were created for extensive evaluation and are publicly available from the authors.
- Evaluation settings: The evaluation uses negative-class F1-score for naturally skewed reviews and accuracy for balanced datasets.The balanced setting contains 200 reviews per domain, split evenly between positive and negative examples.
- Experimental design: Each domain serves as the target while the other 19 domains provide past-domain data, with models evaluated using 5-fold cross-validation.The baselines include target-only, source-only, and merged source-target training variants.
- Results: LSC achieves the best performance under both natural and balanced class distributions, significantly outperforming all baselines.The improvements are statistically significant, with p < 0.01 against NB-ST and CLF and p < 0.0001 against the other baselines.
- Effects of past domains: LSC performs better as the number of past domains increases, indicating that it accumulates and uses knowledge from additional domains.Figure 1 reports negative-class F1-score for natural class distributions and accuracy for balanced class distributions.
5 Conclusions
The paper presents lifelong sentiment classification as an optimization approach that uses Bayesian probabilities, stochastic gradient descent, and penalty terms to exploit past knowledge. Experiments across 20 product-review domains demonstrate its effectiveness and support lifelong learning as a promising direction for building better classifiers.
- Method: The proposed lifelong sentiment classification method uses stochastic-gradient-descent optimization within a Bayesian-probability framework.Penalty terms are introduced to exploit knowledge gained from past learning.
- Results: Experiments across 20 diverse product-review domains demonstrate the method’s effectiveness.
- Conclusion: The authors identify lifelong learning as a promising direction for building better classifiers.