A Survey on Aspect-Based Sentiment Analysis: Tasks, Methods, and Challenges

In plain words

Sentiment analysis usually asks whether a whole review is positive or negative. Aspect-based sentiment analysis asks a sharper question: what exactly is being praised or criticised, and why. A restaurant review can say the food is great but the service is slow. The method has to pick out each topic and the opinion attached to it.

This survey sorts the many research tasks in this area by which pieces of that judgement they try to find: the topic, the opinion words, the polarity, or combinations of these. It also covers how the field moved to large pre-trained language models, how models carry over to new domains and languages, and which problems remain open. Its goal is a map of the field, not a new system.

TL;DR

• A survey, not a method: it gives a taxonomy of ABSA tasks by the sentiment elements they predict, and it reviews methods per task §1 fig. 2.
• Four elements: aspect category, aspect term, opinion term, sentiment polarity. Single tasks predict one element; compound tasks predict coupled pairs, triplets or quads §2.1 §2.2.
• Covers pre-trained language models (PLMs), cross-domain and cross-lingual transfer, and open challenges up to its writing (v2 dated 2022-11-06) §5 §6 §7.
• No hyperparameters or experiments of its own. Headline F1 figures are approximate summaries of other papers' results, given without dataset names.

Problem

Early ABSA work predicts one sentiment element at a time, which does not capture how the elements relate to each other §1. Compound tasks were introduced to model those relations, but no systematic review covered all ABSA tasks, the impact of PLMs, and cross-domain and cross-lingual transfer together §1. The paper aims to fill that gap §1.

Method

This is a review, so the "method" here is its organising scheme.

1. Four sentiment elements. Target is described by an aspect category c or an aspect term a. Sentiment is described by an opinion term o and a polarity p. Polarity usually takes positive, negative or neutral values. An implicit aspect term is written as a special "null" value §2.1.
2. Single vs compound. A task is single if its output is one element (ATE, ACD, ASC, OTE). It is compound if the output couples elements into pairs (AOPE, ACSA, E2E-ABSA), triplets (ASTE, ACSD) or quads (ASQP) §2.2 fig. 2 fig. 3.
3. Modeling paradigms. Four unified paradigms cover most tasks: sequence classification (SeqClass), token classification (TokenClass), machine reading comprehension (MRC), and sequence-to-sequence (Seq2Seq). A fifth, Pipeline, chains several models §2.3.1–§2.3.5. Each paradigm has one formula: SeqClass maps a whole input through an encoder and a classifier eq. 1; TokenClass labels every token eq. 2; MRC predicts a start and an end position for a span given a query eq. 3; Seq2Seq decodes the output sequence eq. 4. tab. 1 maps each task to representative methods per paradigm.
4. Datasets and metric. The main benchmarks are SemEval-2014/2015/2016 and their derived sets (TOWE, ASTE-Data-V2, ACOS, ABSA-QUAD and others). Evaluation is exact match: a prediction counts only if every predicted element equals the annotation. Accuracy, precision, recall and F1 are then computed §2.4 tab. 2.
5. Task-by-task review. Each task is covered with its input and output format tab. 3 and its method families:
   • ATE (token classification; supervised, semi-supervised, unsupervised) §3.1.
   • ACD (multi-label classification if supervised; mapping extracted terms to categories if unsupervised) §3.2.
   • OTE, split into AOCE and TOWE §3.3.
   • ASC, term-based and category-based, usually built on attention or syntax-aware graph networks §3.4.
   • AOPE: pipeline, MRC, and unified grid or span tagging §4.1.
   • E2E-ABSA: pipeline, joint, and unified tagging §4.2 tab. 4.
   • ACSA: Cartesian product, add-one-dimension, hierarchy, and Seq2Seq §4.3 fig. 4.
   • ASTE: two-stage pipeline, multi-task, unified tagging, MRC, Seq2Seq §4.4 fig. 5.
   • ACSD: decomposed triplet prediction and Seq2Seq §4.5.
   • ASQP: generative quad prediction with templates or opinion trees §4.6.
6. PLMs. Early work used PLM embeddings as drop-in replacements for word vectors. Later work reformulates tasks for the PLM (sentence-pair ASC, MRC, BART or T5 generation) and uses PLMs for augmentation and for parsing §5.
7. Transfer. Cross-domain methods are feature-based, data-based, or combine both. Cross-lingual methods use translation with label projection, cross-lingual embeddings, or multilingual PLMs with zero-shot transfer §6.1 §6.2.
8. Challenges. Larger datasets, multimodal ABSA, one model for many tasks (via shared paradigms), and lifelong ABSA with changing label sets §7.

The one concrete data format the survey gives is the unified tagging scheme for E2E-ABSA. It is shown for one example sentence:

# Unified (collapsed) tagging for E2E-ABSA  [§4.2] [tab. 4]
# One tag per token: boundary (B, I, E, S, O) plus polarity (POS, NEG, NEU)
# Tokens: The AMD Turin Processor seems to always perform better than Intel .
tags = [
    "O", "B-POS", "I-POS", "E-POS",
    "O", "O", "O", "O", "O", "O",
    "S-NEG", "O",
]
# Spans read off the tags: (AMD Turin Processor, POS), (Intel, NEG)  [tab. 4]

Hyperparameters

Not applicable. The paper is a survey; it presents itself as a review of existing work §1 and trains no models of its own. It reports no learning rates, batch sizes or model settings.

Results

Author-reported. The survey does not run experiments. The only numbers are approximate F1 summaries of other works, given without a named dataset or model:

The 25% figure is attributed to another paper's evaluation §5. tab. 2 lists language, domains, annotations and URLs for each dataset; it gives no dataset sizes.

Paper vs code

No official code. The paper links a GitHub repository that collects resources from the survey (paper list, tools, datasets) §1. It holds no implementation of the surveyed methods, so nothing was cloned or checked.

Pitfalls

• Dated coverage. The survey is current to its writing (v2 dated 2022-11-06). Methods and PLMs after that date are absent. Check newer work before relying on the "state of the art" statements.
• Approximate numbers. The F1 figures in the Results table are prose summaries with no table or dataset named. Do not use them as benchmark targets.
• Terminology overlaps. "Aspect", "target" and "opinion target" are used loosely in the literature. The survey uses "aspect term" and "aspect category" for the two formats and "target" or "aspect" as a general word §2.1. Read the source paper's wording before you copy a task name.
• ASC covers two inputs. Aspect sentiment classification is treated as one problem whether the aspect is a term or a category, so a model may be shared across both §3.4.
• Cartesian product cost. For ACSA, the Cartesian-product method makes the training set several times larger than the original §4.3.
• Pipeline error propagation. Pipelines suffer when early errors propagate. The survey argues this is the main reason to prefer unified models for compound tasks §2.3.5.
• Unclear best E2E method. The survey notes that it is still unclear which E2E-ABSA family (pipeline, joint, unified) is best. It cites studies on each side §4.2.
• Dataset size. Most ABSA benchmarks derive from SemEval and are small (the survey says "hundreds of sentences" for some) §7.1. Results on them may vary across random seeds, which the survey notes many papers average over five or ten runs §7.1.
• Dataset terms. tab. 2 lists URLs for each dataset. The survey does not state their terms of use, so check each dataset's own terms before reuse.

Lineage

• Builds on: none. This is a review of the field. It positions itself against earlier ABSA surveys, which covered only the pre-2016 non-neural work or deep-learning single tasks, not compound tasks or PLMs §1.
• Compared against: the earlier ABSA surveys named in §1.
• Superseded by: not known from content in this wiki.