Early results for Named Entity Recognition with Conditional Random Fields, Feature Induction and Web-Enhanced Lexicons

A. McCallumWei Li

article2003CoNLL1,373 citations

Proposes an efficient feature induction technique and automated web-based lexicon expansion to improve the performance and scalability of Conditional Random Fields on named entity recognition tasks.

Listen

Extracting structured information such as person names, locations, and organizations from unstructured text is an essential task for intelligence analysis, search engines, and business automation. While statistical sequence models benefit heavily from diverse contextual cues and word lists, standard approaches often struggle with computational bottlenecks when combining thousands of overlapping word patterns. In addition, manually compiling the extensive, domain-specific vocabularies required to accurately recognize varied entities is expensive and time-consuming.

The article demonstrates an automated method to efficiently construct compact, high-performing sequence models using Conditional Random Fields—a statistical framework for labeling sequential data—combined with automated feature induction and web-based lexicon generation. The objective is to maximize entity extraction accuracy while dramatically cutting the number of model parameters and reducing the human effort needed to build specialized word lists.

To achieve this, the authors developed an automated feature selection process tailored for sequence models, using approximations that focus computations only on misclassified words to select the most impactful word combinations. In parallel, they introduced a web-mining approach called WebListing (leveraging Google Sets) that automatically expands small seed lists of known entities into comprehensive lexicons by identifying regular formatting structures across the internet. The combined framework was evaluated on the benchmark CoNLL-2003 shared task using complex news corpora in English and German across four entity categories: persons, locations, organizations, and miscellaneous items.

The evaluation produced several notable results. First, automated feature selection achieved an overall balanced accuracy score (F1) of 84.04% on the English test set while requiring only 6,423 features. Second, this compact model drastically outperformed a baseline relying on fixed, predefined feature combinations, which attained only 73.34% F1 despite generating approximately one million features. Third, performance was strongest on English person names (90.51% F1) and location names (87.44% F1), while organization and miscellaneous categories presented greater difficulty. Finally, German test accuracy reached 68.11% F1, though it was constrained by limited non-English web lexicon support.

These findings indicate that automated feature selection can simultaneously cut computational storage requirements by over 99% and significantly boost extraction accuracy compared to static feature sets. The automated expansion of vocabularies via the web substantially reduces manual engineering time, enabling rapid deployment of entity recognition models across new domains with minimal upfront labeling effort.

For future implementation, technical teams should adopt automated feature selection over static feature expansion to reduce model training complexity and footprint. Organizations should also develop customized, robust web-scraping tools for lexicon generation rather than relying on third-party utilities. However, decision-makers should note that the reported results represent an early-stage implementation with minimal tuning and hand-filtered lexicons, and performance in non-English languages will remain lower until multilingual web-mining pipelines are expanded.

University of Massachusetts Amherst automation classifiers compact model named entity recognition conditional random fields feature induction

McCallum et al (2003).pdf

Paper: Introduction to the CoNLL-2003 Shared Task: Language-Independent Named Entity Recognition, Erik F. Tjong Kim Sang et al. (2003). This paper establishes the CoNLL-2003 shared task dataset and benchmark protocol for English and German named entity recognition on which the source evaluates its model.
Paper: Shallow Parsing with Conditional Random Fields, Fei Sha et al. (2003). This foundational study demonstrates scalable convex optimization and feature engineering for conditional random fields in sequence labeling tasks.
Paper: Maximum Entropy Markov Models for Information Extraction and Segmentation, A. McCallum et al. (2000). This work introduces Maximum Entropy Markov Models for information extraction, providing key conceptual foundations and motivating the transition to conditional random fields.
Paper: Discriminative Training Methods for Hidden Markov Models: Theory and Experiments with Perceptron Algorithms, Michael Collins (2002). This paper develops discriminative sequence tagging using perceptron-based algorithms, introducing foundational concepts for discrete feature-based sequential prediction.
Paper: Text Chunking using Transformation-Based Learning, Lance A. Ramshaw et al. (1995). This paper formulates phrase chunking and entity boundary identification as sequential token tagging problems, establishing the representation framework used in named entity extraction.
Paper: Message Understanding Conference- 6: A Brief History, Ralph Grishman et al. (1996). This paper outlines the origins and standardized evaluation metrics of the Message Understanding Conferences that defined modern named entity recognition benchmarks.

Paper: Incorporating Non-local Information into Information Extraction Systems by Gibbs Sampling, Jenny Rose Finkel et al. (2005). This paper extends conditional random field sequence extraction by incorporating non-local document-level consistency constraints using Gibbs sampling.
Paper: Design Challenges and Misconceptions in Named Entity Recognition, Lev-Arie Ratinov et al. (2009). This paper systematically evaluates the architectural decisions, chunk representations, and gazetteer integrations that build upon early statistical NER models.
Paper: Word Representations: A Simple and General Method for Semi-Supervised Learning., Joseph Turian et al. (2010). This work advances semi-supervised feature integration for NER by replacing discrete lexicons with induced continuous word representations and clusters.
Paper: Bidirectional LSTM-CRF Models for Sequence Tagging, Zhiheng Huang et al. (2015). This work modernizes sequence tagging by coupling bidirectional LSTM representations with a CRF decoding layer, replacing manual and induced sparse features.
Paper: Neural Architectures for Named Entity Recognition, Guillaume Lample et al. (2016). This work dispenses with handcrafted gazetteers and engineered lexicons by introducing neural BiLSTM-CRF and transition-based architectures evaluated on the CoNLL-2003 benchmark.
Paper: Named Entity Recognition with Bidirectional LSTM-CNNs, Jason P. C. Chiu et al. (2015). This paper demonstrates combining bidirectional LSTMs with CNN character representations and gazetteer matching to advance state-of-the-art performance on CoNLL-2003 NER.
Paper: End-to-end Sequence Labeling via Bi-directional LSTM-CNNs-CRF, Xuezhe Ma et al. (2016). This paper presents an end-to-end BiLSTM-CNN-CRF architecture that eliminates the need for manual feature induction and lexicon engineering in sequence labeling.
Paper: Contextual String Embeddings for Sequence Labeling, A. Akbik et al. (2018). This paper develops contextual character-level string embeddings that resolve polysemy and cross-lingual limitations in sequence labeling, achieving breakthrough results on CoNLL-2003.
Paper: A Survey on Deep Learning for Named Entity Recognition, Jing Li et al. (2018). This comprehensive survey contextualizes the evolution from classical CRF and lexicon-based NER methods to deep neural and pretrained contextual architectures.

Cover for Early results for Named Entity Recognition with Conditional Random Fields, Feature Induction and Web-Enhanced Lexicons

Topics

automation classifiers compact model conditional random field feature selection feature space genetic selection intelligence analysts named entity recognition test samples

Keywords

named entity recognition conditional random fields feature induction web-enhanced lexicons sequence model graphical models

1 Introduction
2 Conditional Random Fields
4 Web-augmented Lexicons
5 Results
Acknowledgments
References

Knowls

Knowl 1 — Feature Induction for Linear-Chain Conditional Random Fields
model/method
Feature induction for linear-chain Conditional Random Fields (CRFs) builds feature conjunctions iteratively to identify only those combinations that significantly improve conditional log-likelihood, rather than enumerating an intractable combinatorial space of all time-shifted feature combinations.

The procedure starts with a model containing no features and conducts multiple induction rounds. In each round:
1. Candidate features are formed as singleton observational tests and binary conjunctions of atomic tests with each other and with existing model features, enabling arbitrary-length conjunctions.
2. The candidate features that maximize the improvement in regularized log-likelihood (the gain) are selected. During gain estimation for candidate $g$ with weight $\mu$ , the parameters $\Lambda$ of previously included features are held fixed.
3. Tractability in sequence models is achieved via two approximations:
  - Mean-Field Token Decomposition: Gain evaluation replaces full dynamic programming across the sequence with a token-level mean-field calculation, decoupling state transitions while using exact state marginal posteriors $P_\Lambda(s_t \mid \mathbf{o})$ obtained from a standard forward-backward pass of the base model.
  - Error-Focused Gain Calculation: The candidate gain is evaluated exclusively over training tokens that are mislabeled by the current model, dramatically reducing computation since the majority of tokens are correctly labeled even early in training.
4. Instead of adding a single feature per round, a batch of top-scoring features exceeding a gain threshold is added, and the combined model is partially retrained with a small number of L-BFGS iterations (rather than to convergence) to prevent overfitting before the next induction round.
Knowl 2 — Mean-Field and Error-Focused Likelihood Gain for Candidate Features
equation
When evaluating a candidate feature $g$ with prospective weight $\mu$ to be added to a Conditional Random Field with existing parameter vector $\Lambda$ , the gain $G_\Lambda(g, \mu)$ represents the change in penalized log-likelihood $L_{\Lambda + g, \mu} - L_\Lambda$ . Using a mean-field token approximation evaluated over only the $M$ tokens currently mislabeled by the model, the gain is computed as:

$G_\Lambda(g, \mu) = \sum_{i=1}^M \log\left( \frac{\exp(\mu g(s_{t(i)}, \mathbf{o}(i), t(i)))}{Z_{o(i)}(\Lambda, g, \mu)} \right) - \frac{\mu^2}{2\sigma^2} = M \mu \tilde{E}[g] - \sum_{i=1}^M \log\left(E_\Lambda[\exp(\mu g) \mid \mathbf{o}(i)]\right) - \frac{\mu^2}{2\sigma^2}$

where:
- $\{o(i) : i = 1, \dots, M\}$ is the set of tokens mislabeled by the current model across the training set, where $o(i)$ occurs at sequence position $t(i)$ within input sequence $\mathbf{o}(i)$ ,
- $s_{t(i)}$ is the true target state at position $t(i)$ ,
- $Z_{o(i)}(\Lambda, g, \mu) = \sum_s P_\Lambda(s \mid \mathbf{o}(i)) \exp(\mu g(s, \mathbf{o}(i), t(i)))$ is the token-level normalization factor,
- $P_\Lambda(s \mid \mathbf{o}(i)) = \frac{\alpha_{t(i)}(s \mid \mathbf{o}(i)) \beta_{t(i)+1}(s \mid \mathbf{o}(i))}{Z_{\mathbf{o}(i)}}$ is the exact marginal posterior distribution over state $s$ at token position $t(i)$ , computed via forward-backward dynamic programming under the current CRF parameters $\Lambda$ ,
- $\sigma^2$ is the variance parameter of the zero-mean Gaussian prior on feature weights,
- $\tilde{E}[g]$ is the empirical expectation of $g$ over the mislabeled tokens, and
- $E_\Lambda[\cdot \mid \mathbf{o}(i)]$ is the expectation under the model's marginal state posterior distribution at token $o(i)$ .
The candidate feature gain is $G_\Lambda(g) = \max_\mu G_\Lambda(g, \mu)$ . While $\mu$ cannot be solved in closed form, its optimal value is computed using Newton's method in approximately 12 iterations.
Knowl 3 — Iterative Feature Induction Algorithm for Conditional Random Fields
algorithm
The algorithm constructs feature conjunctions for linear-chain CRFs by alternately generating candidates, computing approximate gains on mislabeled tokens, and partially training feature weights.
Input: Labeled training dataset $\mathcal{D} = \{(\mathbf{o}^{(j)}, \mathbf{s}^{(j)})\}_{j=1}^N$ , atomic test set $\mathcal{A}$ , Gaussian prior variance $\sigma^2$ , maximum features per round $K = 1000$ , gain threshold $\theta = 5.0$ , L-BFGS iterations per round $I = 10$
Output: Induced feature set $\mathcal{F}$ and trained CRF parameter vector $\Lambda$
Initialize $\mathcal{F} \leftarrow \emptyset$ , $\Lambda \leftarrow \mathbf{0}$
repeat
    Compute token state posteriors $P_\Lambda(s_t \mid \mathbf{o}^{(j)})$ for all sequences in $\mathcal{D}$ using forward-backward dynamic programming
    Identify the set of currently mislabeled tokens $\mathcal{E} = \{o(i) : i = 1, \dots, M\}$ across all sequences in $\mathcal{D}$
    Generate candidate conjunction set $\mathcal{C} \leftarrow \mathcal{A} \cup \{ a \wedge f : a \in \mathcal{A}, f \in \mathcal{F} \}$ , restricted to conjunctions from the 1000 highest-gain atomic and existing features
    for each candidate feature $g \in \mathcal{C}$ do
        Find $\mu^* = \arg\max_\mu G_\Lambda(g, \mu)$ using Newton's method over $\mathcal{E}$
        Set candidate gain $G_\Lambda(g) \leftarrow G_\Lambda(g, \mu^*)$
    end for
    Select top candidates $\mathcal{S} \leftarrow \{ g \in \mathcal{C} : G_\Lambda(g) \ge \theta \text{ and } g \text{ is among top } K \text{ candidates by gain} \}$
    if $\mathcal{S}$ is empty then
        break
    end if
     $\mathcal{F} \leftarrow \mathcal{F} \cup \mathcal{S}$
    Initialize weights for new features in $\Lambda$ to their corresponding $\mu^*$
    Optimize all parameters in $\Lambda$ for $I$ iterations of L-BFGS on $\mathcal{D}$
until stopping criterion is met
return $\mathcal{F}, \Lambda$
Knowl 4 — Linear-Chain Conditional Random Field Model and Log-Likelihood Gradient
equation
A first-order linear-chain Conditional Random Field (CRF) models the conditional probability of an output state sequence $\mathbf{s} = (s_1, s_2, \dots, s_T)$ given an observed input sequence $\mathbf{o} = (o_1, o_2, \dots, o_T)$ as:

$P_\Lambda(\mathbf{s} \mid \mathbf{o}) = \frac{1}{Z_\mathbf{o}} \exp \left( \sum_{t=1}^T \sum_k \lambda_k f_k(s_{t-1}, s_t, \mathbf{o}, t) \right)$

where $f_k(s_{t-1}, s_t, \mathbf{o}, t)$ is an arbitrary feature function over adjacent states and the observation sequence, $\lambda_k \in \Lambda$ is the learned weight for feature $f_k$ , and $Z_\mathbf{o}$ is the sequence normalization factor summing over all state paths:

$Z_\mathbf{o} = \sum_{\mathbf{s}'} \exp \left( \sum_{t=1}^T \sum_k \lambda_k f_k(s'_{t-1}, s'_t, \mathbf{o}, t) \right)$

Given a training set $\mathcal{D} = \{(\mathbf{o}^{(1)}, \mathbf{s}^{(1)}), \dots, (\mathbf{o}^{(N)}, \mathbf{s}^{(N)})\}$ where training labels specify unambiguous state paths, the weights $\Lambda = \{\lambda_k\}$ are set to maximize the regularized log-likelihood with a zero-mean Gaussian prior of variance $\sigma^2$ :

$L_\Lambda = \sum_{j=1}^N \log\left( P_\Lambda(\mathbf{s}^{(j)} \mid \mathbf{o}^{(j)}) \right) - \sum_k \frac{\lambda_k^2}{2\sigma^2}$

The first-derivative with respect to weight $\lambda_k$ optimized by quasi-Newton methods (such as L-BFGS) is:

$\frac{\partial L_\Lambda}{\partial \lambda_k} = \sum_{j=1}^N C_k(\mathbf{s}^{(j)}, \mathbf{o}^{(j)}) - \sum_{j=1}^N \sum_{\mathbf{s}} P_\Lambda(\mathbf{s} \mid \mathbf{o}^{(j)}) C_k(\mathbf{s}, \mathbf{o}^{(j)}) - \frac{\lambda_k}{\sigma^2}$

where $C_k(\mathbf{s}, \mathbf{o}) = \sum_{t=1}^T f_k(s_{t-1}, s_t, \mathbf{o}, t)$ represents the total empirical or model count of feature $k$ across sequence path $\mathbf{s}$ .
Knowl 5 — WebListing Lexicon Induction Method
model/method
WebListing is a semi-supervised lexicon expansion method that generates task-specific named entity word lists by exploiting structural HTML regularities on the World Wide Web rather than extracting text from narrow local linguistic contexts.

The method operates as follows:
1. Seed entities belonging to target semantic categories (such as person names, organizations, political parties, and geographic locations) are extracted from labeled training data.
2. A web search service (such as GoogleSets) is queried using the seed entities to identify web pages where the seed terms appear in identical HTML formatting structures (such as list items <li>, table rows, or structured delimiter patterns).
3. Additional terms appearing within the same formatting patterns on those pages are harvested and aggregated into expanded candidate lexicons.
4. Lexicons undergo optional minimal hand-filtering and are used as binary membership test features in sequential taggers.
Knowl 6 — Feature Space Representation for Named Entity Recognition
experimental setup
In the CoNLL-2003 named entity recognition task, atomic features are evaluated at observation offsets $t + \Delta$ with $\Delta \in \{-2, -1, 0, +1, +2\}$ relative to current token position $t$ :
- Token identity: Exact word at position $t + \Delta$ .
- Character regular expressions: 16 orthographic pattern tests indicating capitalization and digit configurations:
  - A: exactly one uppercase letter [A-Z]
  - A+: one or more uppercase letters
  - Aa+: an uppercase letter followed by one or more lowercase letters [a-z]
  - Aa+Aa*: camel-case patterns with multiple capital letter segments
  - A.: an uppercase letter followed by a period
  - D+: one or more digits [0-9]
- Hand-crafted lexicons: 8 lists covering honorifics, days of the week, and months.
- Web-sourced lexicons: 15 lists downloaded from websites covering country names, publicly traded companies, surnames, stopwords, and universities.
- WebListing lexicons: 25 lists generated by WebListing from seed entities, covering people names, organizations, non-governmental organizations (NGOs), and nationalities.
- First-mention features: For capitalized words, a firstmention prefix prepended to all the above tests if the current token appeared earlier in the document.
- German-specific features: Character bi-grams and tri-grams, combined with 5 small lexicons (due to limited multilingual support in GoogleSets).

Knowl 7 — CoNLL-2003 Named Entity Recognition Benchmark Performance

data/table

The linear-chain CRF model with induced features and web-augmented lexicons was evaluated on the CoNLL-2003 English and German shared task datasets for entity classes: Location (LOC), Miscellaneous (MISC), Organization (ORG), and Person (PER).

Language / Entity	Development Set			Test Set
	Precision	Recall	F1	Precision	Recall	F1
English
LOC	93.82%	91.78%	92.79%	87.23%	87.65%	87.44%
MISC	83.99%	78.52%	81.17%	74.44%	71.37%	72.87%
ORG	84.23%	82.03%	83.11%	79.52%	78.33%	78.92%
PER	92.64%	93.65%	93.14%	91.05%	89.98%	90.51%
English Overall	89.84%	88.10%	88.96%	84.52%	83.55%	84.04%
German
LOC	68.55%	68.84%	68.69%	71.92%	69.28%	70.57%
MISC	72.66%	45.25%	55.77%	69.59%	42.69%	52.91%
ORG	70.64%	54.88%	61.77%	63.85%	48.90%	55.38%
PER	82.21%	64.31%	72.17%	90.04%	74.14%	81.32%
German Overall	73.60%	59.01%	65.50%	75.97%	61.72%	68.11%

The model was trained using a Gaussian prior variance $\sigma^2 = 0.5$ , inducing up to 1000 features per round (down to a minimum gain threshold of 5.0) every 10 iterations of L-BFGS. English performance reached 84.04% F1 overall on the test set using 6,423 induced features. German performance reached 68.11% F1 on the test set using character n-grams and 5 small lexicons.

Knowl 8 — Feature Induction versus Fixed Conjunction Patterns
empirical result
On the CoNLL-2003 English named entity recognition test set, automated feature induction demonstrates substantial improvements in both tagging accuracy and model compactness compared to using predefined fixed conjunction patterns:
- Feature Induction: Achieves an overall F1 score of 84.04% using a compact model containing 6,423 induced features.
- Fixed Conjunction Patterns: Achieves an overall F1 score of 73.34% while requiring approximately 1,000,000 features.
Automated feature induction improves test F1 score by 10.7 percentage points while reducing feature set size by over two orders of magnitude.

Coverage note — No substantial contributed material was omitted.

References

1.A. Borthwick, J. Sterling, E. Agichtein, and R. Grishman. 1998. Exploiting diverse knowledge sources via maximum entropy in named entity recognition. In Proceedings of the Sixth Workshop on Very Large Corpora, Association for Computational Linguistics.
2.M. Collins and Y. Singer. 1999. Unsupervised models for named entity classification. In Proceedings of the Joint SIGDAT Conference on Empirical Methods in Natural Language Processing and Very Large Corpora.
3.Stephen Della Pietra, Vincent J. Della Pietra, and John D. Lafferty. 1997. Inducing features of random fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 19(4):380–393.
4.Rosie Jones, Andrew McCallum, Kamal Nigam, and Ellen Riloff. 1999. Bootstrapping for text learning tasks. In IJCAI-99 Workshop on Text Mining: Foundations, Techniques and Applications.
5.John Lafferty, Andrew McCallum, and Fernando Pereira. 2001. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. In Proc. ICML.
6.Robert Malouf. 2002. A comparison of algorithms for maximum entropy parameter estimation. In Sixth Workshop on Computational Language Learning (CoNLL-2002).
7.Andrew McCallum and Fang-Fang Feng. 2003. Chinese word segmentation with conditional random fields and integrated domain knowledge. In Unpublished Manuscript.
8.Andrew McCallum. 2003. Efficiently inducing features of conditional random fields. In Nineteenth Conference on Uncertainty in Artificial Intelligence (UAI03). (Submitted).
9.Adwait Ratnaparkhi. 1996. A maximum entropy model for part-of-speech tagging. In Eric Brill and Kenneth Church, editors, Proceedings of the Conference on Empirical Methods in Natural Language Processing, pages 133–142. Association for Computational Linguistics.
10.Fei Sha and Fernando Pereira. 2003. Shallow parsing with conditional random fields. In Proceedings of Human Language Technology, NAACL.

Citation

MLA

McCallum, A., and W. Li. “Early Results for Named Entity Recognition with Conditional Random Fields, Feature Induction and Web-enhanced Lexicons”. Proceedings of the Seventh Conference on Natural Language Learning at HLT-NAACL 2003 -, vol. 4, 2003, pp. 188–91, https://doi.org/10.3115/1119176.1119206.

APA

McCallum, A., & Li, W. (2003). Early results for named entity recognition with conditional random fields, feature induction and web-enhanced lexicons. Proceedings of the Seventh Conference on Natural Language Learning at HLT-NAACL 2003 -, 4, 188–191. https://doi.org/10.3115/1119176.1119206

Chicago

McCallum, A., and W. Li. 2003. “Early Results for Named Entity Recognition with Conditional Random Fields, Feature Induction and Web-enhanced Lexicons”. Proceedings of the Seventh Conference on Natural Language Learning at HLT-NAACL 2003 - 4: 188–91. https://doi.org/10.3115/1119176.1119206.

Harvard

McCallum, A. and Li, W. (2003) “Early results for named entity recognition with conditional random fields, feature induction and web-enhanced lexicons”, Proceedings of the seventh conference on Natural language learning at HLT-NAACL 2003 -. Association for Computational Linguistics, pp. 188–191. Available at: https://doi.org/10.3115/1119176.1119206.

Vancouver

1. McCallum A, Li W (2003) Early results for named entity recognition with conditional random fields, feature induction and web-enhanced lexicons. In: Proceedings of the seventh conference on Natural language learning at HLT-NAACL 2003 -. Association for Computational Linguistics, pp 188–191

BibTeX

@inproceedings{McCallum_2003, title={Early results for named entity recognition with conditional random fields, feature induction and web-enhanced lexicons}, volume={4}, url={http://dx.doi.org/10.3115/1119176.1119206}, DOI={10.3115/1119176.1119206}, booktitle={Proceedings of the seventh conference on Natural language learning at HLT-NAACL 2003  -}, publisher={Association for Computational Linguistics}, author={McCallum, Andrew and Li, Wei}, year={2003}, pages={188–191} }

Metadata:Crossref

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

License: https://creativecommons.org/licenses/by-nc-sa/4.0/

Early results for Named Entity Recognition with Conditional Random Fields, Feature Induction and Web-Enhanced Lexicons

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — Feature Induction for Linear-Chain Conditional Random Fields

Knowl 2 — Mean-Field and Error-Focused Likelihood Gain for Candidate Features

Knowl 3 — Iterative Feature Induction Algorithm for Conditional Random Fields

Knowl 4 — Linear-Chain Conditional Random Field Model and Log-Likelihood Gradient

Knowl 5 — WebListing Lexicon Induction Method

Knowl 6 — Feature Space Representation for Named Entity Recognition

Knowl 7 — CoNLL-2003 Named Entity Recognition Benchmark Performance

Knowl 8 — Feature Induction versus Fixed Conjunction Patterns

References

Citation

BibTeX

Access the Paper