{
  "name": "Part-of-Speech Tagger Accuracy Benchmarks on Penn Treebank WSJ",
  "description": "Comparative evaluation of statistical, graphical, and neural part-of-speech tagging architectures on the Penn Treebank Wall Street Journal benchmark corpus (Sections 23-24), compiled from primary computational linguistics literature and standard NLP references.",
  "page": "https://casrai.org/guides/part-of-speech-tagging",
  "publisher": "CASRAI",
  "license": "https://creativecommons.org/licenses/by/4.0/",
  "dateModified": "2026-11-16T18:45:33Z",
  "temporalCoverage": "1993/2026",
  "spatialCoverage": "Global",
  "columns": [
    {
      "name": "Model Architecture",
      "unit": "",
      "description": "Sequence labeling architecture and representative implementation"
    },
    {
      "name": "Paradigm",
      "unit": "",
      "description": "Underlying statistical or machine learning paradigm"
    },
    {
      "name": "Overall Token Accuracy",
      "unit": "%",
      "description": "Percentage of test word tokens correctly assigned gold-standard POS tags"
    },
    {
      "name": "Unknown Word Accuracy",
      "unit": "%",
      "description": "Tagging accuracy on out-of-vocabulary tokens absent from the training set"
    },
    {
      "name": "Sentence Accuracy",
      "unit": "%",
      "description": "Percentage of complete sentences tagged without a single error"
    },
    {
      "name": "Decoding Method",
      "unit": "",
      "description": "Inference and decoding algorithm utilized at test time"
    },
    {
      "name": "Primary Reference",
      "unit": "",
      "description": "Authoritative publication reporting benchmark performance"
    }
  ],
  "rows": [
    {
      "Model Architecture": "Most Frequent Class (Unigram Baseline)",
      "Paradigm": "Lexical Lookup Baseline",
      "Overall Token Accuracy": 90.1,
      "Unknown Word Accuracy": 15,
      "Sentence Accuracy": 12.4,
      "Decoding Method": "Greedy Lexical Argmax",
      "Primary Reference": "Jurafsky & Martin (2024)"
    },
    {
      "Model Architecture": "Trigrams'n'Tags (TnT) HMM",
      "Paradigm": "Generative Second-Order Markov Model",
      "Overall Token Accuracy": 96.7,
      "Unknown Word Accuracy": 85.5,
      "Sentence Accuracy": 54,
      "Decoding Method": "Viterbi Dynamic Programming",
      "Primary Reference": "Brants (2000)"
    },
    {
      "Model Architecture": "Maximum Entropy Markov Model (MEMM)",
      "Paradigm": "Locally Normalized Discriminative Model",
      "Overall Token Accuracy": 96.6,
      "Unknown Word Accuracy": 86.1,
      "Sentence Accuracy": 53.8,
      "Decoding Method": "Viterbi Beam Search",
      "Primary Reference": "Ratnaparkhi (1996)"
    },
    {
      "Model Architecture": "Linear-Chain Conditional Random Field (CRF)",
      "Paradigm": "Globally Normalized Undirected Graphical Model",
      "Overall Token Accuracy": 97.1,
      "Unknown Word Accuracy": 88.3,
      "Sentence Accuracy": 56.2,
      "Decoding Method": "Exact Viterbi Decoding",
      "Primary Reference": "Sha & Pereira (2003)"
    },
    {
      "Model Architecture": "Cyclic Dependency Network (Stanford Tagger)",
      "Paradigm": "Bidirectional Feature-Rich MaxEnt Network",
      "Overall Token Accuracy": 97.24,
      "Unknown Word Accuracy": 89.01,
      "Sentence Accuracy": 57.1,
      "Decoding Method": "Loopy Belief Propagation / Viterbi",
      "Primary Reference": "Toutanova et al. (2003)"
    },
    {
      "Model Architecture": "BiLSTM-CRF with Character Embeddings",
      "Paradigm": "Deep Bidirectional Recurrent Neural Network",
      "Overall Token Accuracy": 97.55,
      "Unknown Word Accuracy": 90.2,
      "Sentence Accuracy": 59.3,
      "Decoding Method": "Viterbi on CRF Transition Matrix",
      "Primary Reference": "Ma & Hovy (2016)"
    },
    {
      "Model Architecture": "Fine-Tuned RoBERTa / BERT-Base",
      "Paradigm": "Pretrained Bidirectional Transformer",
      "Overall Token Accuracy": 97.96,
      "Unknown Word Accuracy": 93.4,
      "Sentence Accuracy": 63.1,
      "Decoding Method": "Subword Classification Softmax Argmax",
      "Primary Reference": "Devlin et al. (2019)"
    }
  ],
  "sources": [
    {
      "title": "Speech and Language Processing (3rd ed. draft)",
      "url": "https://web.stanford.edu/~jurafsky/slp3/8.pdf",
      "accessed": "2026-11-16"
    },
    {
      "title": "Building a Large Annotated Corpus of English: The Penn Treebank",
      "url": "https://aclanthology.org/J93-2004/",
      "accessed": "2026-11-16"
    },
    {
      "title": "TnT: A Statistical Part-of-Speech Tagger",
      "url": "https://aclanthology.org/A00-1031/",
      "accessed": "2026-11-16"
    },
    {
      "title": "Feature-Rich Part-of-Speech Tagging with a Cyclic Dependency Network",
      "url": "https://aclanthology.org/N03-1033/",
      "accessed": "2026-11-16"
    }
  ]
}