基于方面的情感分析综述:任务、方法与挑战 · Literature survey in natural language processing (NLP). We read the arXiv v2 version (22 PDF pages, dated 6 Nov 2022); page numbers below are PDF pages 1–22, not the journal pages 11019–11038. 自然语言处理(natural language processing, NLP)领域的文献综述。我们读的是 arXiv v2 版本(共 22 页 PDF,日期 6 Nov 2022);下文页码均为 PDF 页码 1–22,而非期刊页码 11019–11038。 · Ch 2, 8, 13 · 40 min
Zhang, W., Li, X., Deng, Y., Bing, L., & Lam, W. (2023). A survey on aspect-based sentiment analysis: Tasks, methods, and challenges. IEEE Transactions on Knowledge and Data Engineering, 35(11), 11019–11038. https://doi.org/10.1109/TKDE.2022.3230975
| Pages 页码 | What to read | 读什么 |
|---|---|---|
| 1–2 | Introduction and the four sentiment elements with Figure 1. | 引言,以及配合图 1 的四个情感要素。 |
| 7 | Table 3: every task shown on one sentence. | 表 3:用同一句话展示所有任务。 |
| 9–10 | Section 4.3 on ACSA and Figure 4: the task our method matches. | 第 4.3 节 ACSA 与图 4:与我们方法对应的任务。 |
| 12–13 | Section 5 on pre-trained models, including the robustness warning. | 第 5 节预训练模型,包括稳健性警告。 |
| 14 | Section 7.1 on small datasets and the cross-lingual gap in 6.2. | 第 7.1 节关于小数据集,以及第 6.2 节的跨语言缺口。 |
Pages · 页码 1–2
Pages · 页码 2–5
Pages · 页码 6–8
Pages · 页码 8–12
Pages · 页码 12–13
Pages · 页码 13–14
Pages · 页码 14–15
| Exhibit 图表 | How to read it | 怎么看 |
|---|---|---|
| Figure 1 (p. 2) | One sentence, "The pizza is delicious", broken into its four sentiment elements. Read it as the template for any coded review. | 把一句“The pizza is delicious”拆成四个情感要素。可把它当作任何评论编码的模板。 |
| Figure 2 (p. 3) | Tree of all ABSA tasks with representative models. The top branch is single tasks; the bottom branch is compound tasks. | 全部 ABSA 任务的分类树,附代表性模型。上半部分是单一任务,下半部分是复合任务。 |
| Table 1 (p. 4) | Which modelling paradigm (SeqClass, TokenClass, MRC, Seq2Seq, pipeline) has been used for each task. A dash means no representative work. | 每个任务用过哪种建模范式(SeqClass、TokenClass、MRC、Seq2Seq、流水线)。短横线表示没有代表性研究。 |
| Table 2 (p. 5) | Common benchmark datasets with language, domain and which elements are labelled (a, c, p, o). Only ASC-QA and ASAP are Chinese. | 常用基准数据集,列出语言、领域和标注的要素(a、c、p、o)。其中只有 ASC-QA 和 ASAP 是中文。 |
| Table 3 (p. 7) | Input and output for every task on one sentence: "The pizza is delicious, but the service is terrible". The clearest page for telling tasks apart. | 用同一句“The pizza is delicious, but the service is terrible”展示每个任务的输入和输出。这是区分各任务最清楚的一页。 |
| Figure 3 (p. 8) | Map linking the four elements to the tasks that predict them. ACSA sits between aspect category and sentiment polarity. | 把四个要素与预测它们的任务连起来的关系图。ACSA 位于方面类别与情感极性之间。 |
| Table 4 (p. 9) | Joint versus unified tagging for E2E-ABSA on one laptop-processor sentence. Each word carries a boundary tag and a polarity tag. | 以一句关于笔记本处理器的句子,对比 E2E-ABSA 的联合标注与统一标注。每个词都带边界标签和极性标签。 |
| Figure 4 (p. 10) | Four unified ways to do ACSA, including the extra "N/A" column for categories not mentioned. Panel (4) generates the answer as a sentence. | ACSA 的四种统一做法,包括为未提及类别加一列“N/A”。第 (4) 格直接把答案生成为一句话。 |
| Figure 5 (p. 11) | The same triplet task solved four ways: pipeline, tagging, MRC and Seq2Seq. Shows how framing changes the model. | 同一个三元组任务的四种解法:流水线、标注、MRC 和 Seq2Seq。展示问题框架如何改变模型。 |
| Term 术语 | Meaning | 解释 |
|---|---|---|
| aspect-based sentiment analysis (ABSA) 基于方面的情感分析 | Finding sentiment towards each part or feature of a product, not the review as a whole. | 分析用户对产品每个部分或特性的情感,而不是整条评论的总体情感。 |
| aspect term 方面词 | The target word that actually appears in the text, such as "pizza". | 文本中实际出现的评价对象词,例如“pizza”。 |
| aspect category 方面类别 | A label from a fixed list chosen for the domain, such as food or service. | 从该领域预先定好的类别清单中选出的标签,例如食物或服务。 |
| opinion term 观点词 | The words that carry the feeling, such as "delicious" or "terrible". | 承载感受的词,例如“delicious”或“terrible”。 |
| sentiment polarity 情感极性 | The direction of feeling: positive, negative or neutral. | 情感的方向:正面、负面或中性。 |
| implicit aspect 隐含方面 | A target that is meant but not named; the term is recorded as "null". | 评论所指但没有明说的对象;方面词记为“null”。 |
| single vs compound ABSA task 单一与复合 ABSA 任务 | Single tasks predict one element; compound tasks predict linked pairs, triplets or quads. | 单一任务只预测一个要素;复合任务预测相互关联的二元组、三元组或四元组。 |
| aspect category sentiment analysis (ACSA) 方面类别情感分析 | Detect which predefined categories a text discusses and the polarity of each. | 识别文本谈到哪些预定义类别,以及每个类别的情感极性。 |
| aspect sentiment quad prediction (ASQP) 方面情感四元组预测 | Predict category, term, polarity and opinion together; the hardest task. | 同时预测类别、方面词、极性和观点词;是最难的任务。 |
| pre-trained language model (PLM) 预训练语言模型 | A model such as BERT trained on huge text first, then fine-tuned for a task. | 像 BERT 这样先在海量文本上训练、再针对具体任务微调的模型。 |
| Seq2Seq (sequence-to-sequence) 序列到序列 | A model that reads text and writes an answer as new text. | 读入文本并以新文本形式写出答案的模型。 |
| exact-match F1 完全匹配 F1 | A prediction counts as correct only if every element equals the human label; F1 balances precision and recall. | 只有所有要素都与人工标注一致才算正确;F1 综合了精确率和召回率。 |
| zero-shot / few-shot 零样本/少样本 | Working with no, or only a few, labelled training examples. | 在没有或只有很少标注训练样本的情况下完成任务。 |
| cross-lingual ABSA (XABSA) 跨语言 ABSA | Applying a model trained in one language to another language. | 把用一种语言训练的模型用到另一种语言上。 |
| Value 数值 | What it measures | 含义 | Page 页 |
|---|---|---|---|
| 2203.01054v2 | arXiv identifier of the version we read | 我们所读版本的 arXiv 编号 | p. 1 |
| 6 Nov 2022 | Date stamped on the arXiv v2 version | arXiv v2 版本标注的日期 | p. 1 |
| 2016 | The earliest ABSA survey covered studies before 2016 only | 最早的 ABSA 综述只覆盖 2016 年前的研究 | p. 2 |
| SemEval-2014 | First of the three most-used benchmark datasets (laptops, restaurants) | 三个最常用基准数据集中的第一个(笔记本电脑、餐厅) | p. 5 |
| SemEval-2016 | The only multilingual SemEval benchmark in Table 2 | 表 2 中唯一的多语言 SemEval 基准 | p. 5 |
| 18 | Predefined aspect categories in the Chinese ASAP dataset | 中文 ASAP 数据集的预定义方面类别数 | p. 5 |
| four | Number of single ABSA tasks, one per sentiment element | 单一 ABSA 任务的数量(four),每个要素一个 | p. 6 |
| 80% | Approximate F1 of supervised aspect term extraction on benchmarks | 有监督方面词抽取在基准上的 F1 约为 80% | p. 6 |
| 70% | Approximate F1 of current E2E-ABSA methods on benchmarks | 目前 E2E-ABSA 方法在基准上的 F1 约为 70% | p. 9 |
| 40% | Approximate F1 of the best quad-prediction (ASQP) models | 最佳四元组预测(ASQP)模型的 F1 约为 40% | p. 12 |
| 25% | Performance drop (more than) of PLM-based models on adversarial ASC tests | PLM 模型在对抗性 ASC 测试中的性能下降(超过 25%) | p. 13 |
| hundreds of sentences | Typical size of existing ABSA benchmark datasets | 现有 ABSA 基准数据集的典型规模(几百句) | p. 14 |
| five | Random seeds usually averaged when reporting results (often five or ten) | 报告结果时通常取平均的随机种子数(often five or ten,常为五或十个) | p. 14 |
Written by the group from the full text. Numbers and page references were checked against the source. Licensed cases and reports are summarised, not reproduced.
由小组根据全文整理,数字与页码已对照原文核对。授权案例与报告只作摘要,不转载原文。