PaperlayerView exact revision on arXiv

Source-linked AI summary

Large Language Models for Data Annotation and Synthesis: A Survey

Zhen Tan, Dawei Li, Song Wang, Alimohammad Beigi, Bohan Jiang, Amrita Bhattacharjee, Mansooreh Karami, Jundong Li, Lu Cheng, Huan Liu

arXiv:2402.13446v3cs.CL

TL;DR

Data annotation and synthesis is labor-intensive, costly, and broad, while existing LLM surveys do not focus specifically on this application. This survey organizes LLM-based annotation generation, assessment, utilization, and challenges, concluding that the area offers promising solutions while requiring attention to bias, hallucination, data dependence, and tuning complexity.

  • Problem

    Data annotation and synthesis is labor-intensive and costly, and existing LLM surveys largely do not focus on their specific utility for annotation.

  • Method

    The survey develops a taxonomy of annotation types and reviews methods for generating, assessing, selecting, and utilizing LLM-generated annotations.

  • Results

    The survey reviews LLM annotation methodologies, applications, learning strategies, and technical and ethical challenges across the annotation synthesis pipeline.

  • Takeaways & Limitations

    The work provides a guide for researchers and practitioners exploring LLMs for data annotation and synthesis.

  • Takeaways & Limitations

    LLM-generated annotations can exhibit sampling bias and hallucination, perpetuate social bias, depend on high-quality data, and require sophisticated prompt engineering and fine-tuning.

Abstract

from arXiv · show

Data annotation and synthesis generally refers to the labeling or generating of raw data with relevant information, which could be used for improving the efficacy of machine learning models. The process, however, is labor-intensive and costly. The emergence of advanced Large Language Models (LLMs), exemplified by GPT-4, presents an unprecedented opportunity to automate the complicated process of data annotation and synthesis. While existing surveys have extensively covered LLM architecture, training, and general applications, we uniquely focus on their specific utility for data annotation. This survey contributes to three core aspects: LLM-Based Annotation Generation, LLM-Generated Annotations Assessment, and LLM-Generated Annotations Utilization. Furthermore, this survey includes an in-depth taxonomy of data types that LLMs can annotate, a comprehensive review of learning strategies for models utilizing LLM-generated annotations, and a detailed discussion of the primary challenges and limitations associated with using LLMs for data annotation and synthesis. Serving as a key guide, this survey aims to assist researchers and practitioners in exploring the potential of the latest LLMs for data annotation, thereby fostering future advancements in this critical field.

1 Introduction

Data annotation and synthesis is broad, complex, and resource-intensive, motivating LLMs as tools for automating and improving annotation. The survey focuses specifically on annotation generation, assessment, utilization, and associated societal and technical challenges.

  • Data annotation and synthesis includes labels, contextual information, confidence scores, preferences, entity relationships, semantic roles, temporal sequences, instructions, and responses.
  • Complexity, subjectivity, diversity, domain expertise, and manual resource demands make annotation difficult for current machine learning models.
  • Advanced LLMs such as GPT-4, Gemini, and LLaMA-2 can automate annotation tasks and support consistency across large data volumes.
  • The survey reviews annotation generation across instruction-response, rationale, pairwise feedback, textual feedback, and domain-specific data, including diversity and quality criteria.
  • It also examines annotation assessment, utilization during supervised fine-tuning, alignment tuning, and inference, plus ethical and technical limitations.
  • Unlike broader LLM surveys, this work centers on annotation synthesis as a crucial yet underexplored application domain.

2 Preliminaries

The survey frames annotation synthesis as a pipeline involving an annotator model that produces annotations and a task learner that uses them for downstream tasks.

  • The annotator model A maps input data to annotations, while the task learner L learns from those annotations to accomplish specific tasks.
  • Advanced LLMs such as GPT-4 and LLaMA can serve as annotators, while task learners may be large models or smaller models such as BERT.

3 LLM-Based Annotation Generation

LLMs generate diverse annotation types for language and structured data, including instructions, responses, labels, rationales, feedback, conversations, and structural insights. Research improves synthesis through diversification, refinement, quality control, and varied reasoning structures.

  • Instruction & Response: Instruction and response datasets support LLM fine-tuning and in-context learning, with recent work automating their generation.
  • Instruction & Response: Instruction diversity is increased through sample mixing, seed-based generation, multi-step prompting, iterative explain-then-generate pipelines, paraphrasing, retrieval, and Monte Carlo Tree Search.
  • Instruction & Response: Response quality is improved through detailed prompts, self-consistency, source-grounded intermediate reasoning, and generating fewer mathematical reasoning data points.
  • Label: LLMs are used to automate classification labels, event-extraction annotations, metadata enrichment, and large-scale human-LLM co-annotation.
  • Rationale: Rationales provide reasoning pathways for answer prediction and can be represented as trees, graphs, tables, programs, recursive structures, or concepts.
  • Pairwise Feedback: Pairwise feedback is synthesized by ranking sampled responses, constructing best-versus-worst pairs, updating reward functions, or directly generating responses with different qualities.
  • Textual Feedback: Textual feedback identifies shortcomings or suggests improvements, while debate and peer review support reasoning and evaluation.
  • Other Data Types: Multi-round conversations can construct or enhance dialogue datasets, while LLMs contextualize graph and tabular data to distill structural insights.

LLM-Generated Annotations Assessment

Assessing LLM-generated annotations combines human-led and automated methods, with evaluation criteria varying by task. Selecting useful annotations further relies on rule-based, external-source, and LLM-driven filtering strategies.

  • General Approaches: Annotation quality is assessed through human comparisons with benchmarks, manual checks of originality, accuracy and variety, and comparisons against human-annotated labels.
  • Task-Specific Evaluations: Task-specific evaluation uses token-ranking metrics for knowledge graphs, Self-BLEU for counterfactual diversity, and Pass@k for code generation.
  • LLM-Driven Selection: LLM-as-a-judge can score synthetic responses and apply a threshold to collect high-quality query-response pairs for later training.
  • Selection Methods: Filtering and selection methods are categorized as rule-based filtering, external source utilization, and LLM-driven selection.
  • LLM-Driven Selection: LLMs can rank candidate annotations or judge whether one response genuinely surpasses another in pairwise feedback synthesis.

LLM-Generated Annotations Utilization

The survey reviews how LLM-generated annotations are used to improve models through supervised fine-tuning, alignment tuning, and in-context learning. These strategies support self-improvement, reasoning, preference alignment, prompt refinement, candidate selection, and post hoc response editing.

  • LLM-generated annotations provide labeled data for NLP models at different learning stages.
  • Supervised Fine-Tuning: Supervised fine-tuning uses synthetic annotations for self-evolution, instruction tuning, and improving zero-shot generalization.Some approaches iteratively fine-tune an LLM on self-annotated data.
  • Alignment Tuning: Synthetic pairwise feedback supports alignment through reward modeling, policy training, DPO, filtering, and iterative bootstrapping.Several methods directly apply DPO to pairwise feedback generated by LLMs.
  • In-Context Learning: In-context learning uses LLM-generated task descriptions, demonstrations, and test-case augmentations to refine or expand prompts.
  • Feedback and Refinement: LLM-generated feedback enables candidate elimination, post hoc editing, response refinement, and tool-assisted improvements to reasoning.Some utilization methods incorporate domain tools such as program interpreters.

Societal Impact and Future Work

The survey identifies societal, technical, and efficiency challenges in LLM-based annotation. It discusses ethical risks, model collapse, hallucinations, and computational demands alongside validation, data, and model-efficiency responses.

  • Ethics Consideration: LLM-generated annotations may produce biased and unfair results in high-stakes finance, jurisprudence, and healthcare decisions.The survey also notes potential social disparities from making human annotator roles redundant.
  • Model Collapse: Model collapse describes gradual performance decreases when LLMs train on other LLMs’ outputs.Statistical and functional approximation errors can amplify through successive training cycles.
  • Model Collapse: Diverse, high-quality training data containing substantial human-generated content is presented as a way to prevent model collapse.Accumulating real and machine-generated data is described as maintaining diversity and preventing performance degradation.
  • Hallucinations: Hallucinated annotations undermine reliability and can spread misinformation, motivating automated and manual validation mechanisms.The survey also mentions reverse validation and chain-of-thought explanations as hallucination-related approaches.
  • Efficiency of LLMs: LLM size and complexity create latency, energy, cost, and deployment challenges that motivate pruning, mixture-of-experts, quantization, compression, and distillation.These techniques target resource-constrained environments and can be combined for further efficiency.

7 Conclusion

The survey presents LLM-based annotation and synthesis as a developing NLP research area addressing data scarcity and annotation efficiency. It organizes generation, assessment, and utilization methods, reviews their effects on model training, and catalogs technical and ethical challenges.

  • The survey develops a taxonomy of LLM annotation methodologies spanning annotation generation, assessment, and utilization.It also categorizes techniques and compiles benchmark datasets for ongoing research.
  • LLM-based annotation and synthesis offers approaches to longstanding data-scarcity challenges while improving annotation quality and process efficiency.
  • The survey evaluates how LLM-generated annotations affect machine-learning model training and discusses bias and societal ramifications.
  • The paper concludes by identifying persistent challenges and open questions to guide future investigation.

Limitations

The survey highlights sampling and social biases, dependence on high-quality data, tuning complexity, generalization risks, and computational requirements as limitations of LLM-based annotation.

  • Sampling bias can produce incorrect or hallucinated data, reducing annotation reliability and quality for discriminative tasks.
  • Social Bias and Ethical Dilemmas: Biases in training data may be perpetuated and amplified through annotated data, creating ethical concerns in tasks requiring fairness and impartiality.
  • Dependence on High-Quality Data: LLM annotation depends on large, high-quality datasets whose labor-intensive curation creates a scalability challenge.
  • Complexity in Tuning and Prompt Engineering: Sophisticated prompt engineering and fine-tuning create barriers for practitioners and researchers without extensive NLP and machine-learning expertise.
  • Generalization and Overfitting: Overfitting to training data can limit generalization to unseen data or different contexts, especially in discriminative tasks.
  • Computational and Resource Requirements: The training and deployment requirements of LLMs impose computational and resource demands.

A LLM-assisted Tools and Software for Annotation

LLM-assisted annotation tools support diverse NLP labeling workflows through user-friendly interfaces, modular pipelines, and automated or model-assisted labeling. The surveyed tools aim to simplify annotation, improve label quality, and increase productivity.

  • LLM-assisted annotation tools support diverse NLP tasks, from binary labels to hierarchical structures, while simplifying labeling and improving productivity.They are designed for both engineers and non-technical annotators.
  • LangChain facilitates annotation by structuring complex problems into smaller chained subtasks whose outputs become inputs for later actions.Its agents can interact with external environments and data sources.
  • Stack AI provides a paid platform for visually designing, testing, and deploying LLM-integrated workflows for content creation, data labeling, conversational AI, and document processing.It also uses weakly supervised models to expedite data preparation.
  • UBIAI offers multilingual cloud-based annotation with manual labeling, LLM-assisted zero- and few-shot labeling, model-assisted labeling, and model integration.Its interface targets insight extraction from unstructured documents.
  • Prodigy combines rule-based, statistical, and LLM-assisted annotation for named entities, spans, and classification across text, audio, and vision.

C Collections of Papers on LLM for Data Annotation

The survey’s paper collections organize representative LLM-based annotation research by generation, assessment, and utilization, covering multiple annotation types, methods, venues, and resources. The listed studies span instruction and response synthesis, rationales, information-extraction samples, pairwise feedback, and domain applications.

  • The collections categorize LLM annotation research into generation, assessment, and utilization, with tables listing methods, technologies, venues, and code or data links.
  • Annotation-generation papers cover instruction and response data, rationales, dialogue, information-extraction samples, pairwise feedback, and domain-specific applications.
  • Instruction-generation studies use approaches including model inference, API calling, and model training, with examples such as Self-Instruct and Stanford Alpaca.
  • The collections include resource and cost details, such as $600 for Self-Instruct, $223.55 for SynthIE, and 30 hours on an A100 GPU for 100k InPars-v2 queries.
  • The listed rationale studies use chain-of-thought annotations generated through GPT-3, GPT-4, ChatGPT, or GPT-neox API or inference workflows.
Loading 2402.13446v3…