Opinions are key influences on human behaviors and are central to almost all human activities. Our cognition of the world and the decisions we make are considerably conditioned on how others see and evaluate the world. For this reason, sentiment analysis, aiming to automatically characterize human beings’ opinions, stances, and attitudes from textual data, has been actively investigated over the past decades.

Recent advances in deep learning enable breakthroughs in a variety of NLP tasks. However, the huge success highly relies on the availability of massive labeled data, which hinders its potentials to the low-resource scenarios where the labeled data is scarce and costly to obtain. On the contrary, humans possess the ability to recognize new objects or perceive abstract conce...[ Read more ]

Opinions are key influences on human behaviors and are central to almost all human activities. Our cognition of the world and the decisions we make are considerably conditioned on how others see and evaluate the world. For this reason, sentiment analysis, aiming to automatically characterize human beings’ opinions, stances, and attitudes from textual data, has been actively investigated over the past decades.

Recent advances in deep learning enable breakthroughs in a variety of NLP tasks. However, the huge success highly relies on the availability of massive labeled data, which hinders its potentials to the low-resource scenarios where the labeled data is scarce and costly to obtain. On the contrary, humans possess the ability to recognize new objects or perceive abstract concepts with a few examples. The significant gap between human learning ability and deep learning has spurred on a promising direction, namely transfer learning, which aims to leverage knowledge from a source domain, task, or language that is sufficiently labeled to improve the predictive learning in a target one with minimal supervision.

In this thesis, we focus on developing deep transfer learning methodologies for low-resource sentiment analysis (LRSA) at different levels, varying from coarse-grained sentiment analysis (CGSA) to fine-grained opinion mining, i.e., aspect-based sentiment analysis (ABSA) and end-to-end ABSA (E2E-ABSA). To coincide with existing limitations of these different subtasks, we consider different perspectives of knowledge, including cross-domain, cross-task, and cross-lingual settings, to be transferred.

Specifically, we begin with domain adaptation in CGSA and propose to address (1) how to explicitly and automatically identify both domain-invariant and domain-specific information as transferable knowledge to, to a considerable degree, minimize the discrepancy between domains. We further push the boundary to explore the less studied knowledge transfer in fine-grained opinion mining that concerns more with aspect-oriented opinions. In ABSA, we propose a new cross-task and cross-domain setting by considering the effect of aspects with different granularity, where we study (2) how to transfer aspect-specific knowledge cross both different aspect-based tasks and domains. However, the limitation of specifying the input aspects in advance for ABSA hinders its potential applications in practice. This inspires us to continuously explore (3) how to transfer cross-domain knowledge in E2E-ABSA that aims to jointly extract aspects and aspect-oriented sentiments across domains. Due to the diversity of human languages around the world, cross-lingual sentiment analysis (CLSA) remains to be another critical problem, where both the feature space and feature distribution are different across languages. Motivated by human beings’ ability to learn new tasks rapidly with a few examples by extracting accumulated meta-knowledge from previous tasks, we are also curious about (4) whether we can leverage previous cross-lingual transfer experiences to enhance the transfer effectiveness in new cross-lingual tasks. We evaluate and validate the proposed models and algorithms on multiple public and real-world industrial datasets. This thesis will also introduce the research frontier and points out promising research directions for future investigation.