Feature-Less End-to-End Nested Term Extraction

Yu Yuan; Yuze Gao

REVIEW 3 major objections 7 minor 24 references

Feature-Less End-to-End Nested Term Extraction

T0 review · 3 major / 7 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A span classifier plus ranker extracts nested domain terms from segmented raw text, reaching F1 0.7789 on GENIA without syntactic features.

desk verdict Genuinely new span-ranking approach to nested term extraction, but the reported F1 leans on a test-set-informed output threshold; fix the evaluation and it's worth publishing. read the letter →

arxiv 1908.05426 v1 pith:2G42C66C submitted 2019-08-15 cs.CL cs.LGstat.ML

Yuze Gao , Yu Yuan This is my paper · ORCID

classification cs.CLcs.LGstat.ML

keywords automatictermextractionnestedtermsspanclassificationrankingdeeplearningGENIAcorpusfeature-free

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that automatic term extraction can be done end-to-end by enumerating candidate spans, classifying each as term-like or not, and ranking the survivors, without needing part-of-speech tags, parse trees, or hand-built features. On the GENIA biology corpus, the ranker achieves F1 0.7789, above the two cited baselines (0.707 and 0.7143), and it can return nested and overlapping terms that sequence-labeling methods miss. The authors also show that added features like POS tags and ELMo contribute little, supporting the claim that raw segmented text suffices. The main caveat they state is that output size is controlled by a corpus-level term ratio, which limits portability to new domains.

What carries the argument

The central object is the span representation, a vector built for every candidate span up to length k by concatenating five components: a span node MLP over the span's hidden states, a term-attention head, the start and end word hidden states, a sentence-level attention vector, and a length embedding. A binary classifier first prunes the candidate set to 'true positive spans'; a regression module then scores the survivors, and the top K scored spans are returned, with K = α·|TotalWords| (α=0.23 in the experiments). The sentence attention and term attention are learned mechanisms that let the model weigh context words and head words without external features.

What would settle it

Run the trained ranker on a held-out domain whose true term ratio is far from 0.23 (for instance, a legal corpus with sparse terminology) without retuning α, and check whether F1 falls sharply; a second test would compare this threshold rule against a threshold-free ranker that predicts per-sentence term counts, to see if the gap stems from α rather than from the span representations.

Watch

Extended reading notes

Core claim

The paper claims that automatic term extraction can be treated as span classification followed by span ranking, and that this works without extra syntactic features. On the GENIA corpus, the ranker reaches F1 0.7789 with all features and F1 0.7751 with random embeddings, both above the compared baselines of 0.707 and 0.7143. Because all spans up to a fixed length are candidates, nested and overlapping terms are scored independently, which sequence-labelling ATE systems cannot do.

Load-bearing premise

The load-bearing assumption is that the number of terms in any text is roughly α times the number of words (α=0.23); when a new domain has a different term density, the fixed output threshold K=α·|TotalWords| distorts the results because the threshold is not learned from the text.

Editorial extensions

If this is right

Nested terms like [0,4] and [0,5] in one sentence can both be returned because the model scores each span separately rather than forcing a non-overlapping label sequence.
The gap between the classifier (recall near 0.96) and the ranker (recall 0.83) shows that the ranking step, not span detection, is where true positives are lost in the current design.
Extra feature sources add little: the full feature combination raises F1 by only 0.0038 over random embeddings in the ranker, so a feature-free deployment is a small trade.
The output-size threshold K=0.23·|TotalWords| embeds a prior about term density; on a corpus with a different term ratio, the model's precision or recall will shift even if the learned span scores are unchanged.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A natural next step, not tested in the paper, is to replace the corpus-level threshold with a learned, per-sentence decision rule; the ranker's scores could be calibrated as probabilities and cut adaptively.
The same span-enumeration plus ranker design would apply directly to nested named-entity recognition, where overlapping mentions are common, since the loss function does not encode ATE-specific assumptions.
Because ELMo barely moved the numbers, the authors suspect hard features may be more useful than soft embeddings; that claim is testable by swapping the word LSTM for a transformer encoder in the same architecture.
The O(n·k) span enumeration grows with sentence length; on longer documents a learned span proposer would be needed, but the classifier-plus-ranker pipeline itself would remain unchanged.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

I gave it a read. The new thing is real: span classification plus ranking, borrowed from coreference, applied to nested term extraction. That's a legitimate gap in ATE, and the paper makes a decent case that you don't need hand-built features. Code and data are on GitHub, which earns credit.

The soft spot is the output threshold, and it's load-bearing. K = α × total words with α = 0.23. The α is set from the corpus-wide term ratio computed before the split, so the test sentences are baked into that number. Fig. 5 then sweeps α on the test set and reports results near the true density. The final F1 is therefore partly an artifact of knowing how many terms are in the test data. This doesn't sink the approach — the ranker's ordering is still doing something, and the classifier alone gets high recall — but it does compromise the comparison to baselines, who don't get the same test-informed count. The fix is straightforward: estimate α from train or dev, or report threshold-free ranking metrics like average precision. The authors themselves admit in the conclusion that the threshold doesn't transfer to unknown domains, so this is a known limitation rather than a hidden one.

Other issues are minor: Eq. 7 says sum when it means mean, Eq. 4 has an undefined denominator, evaluation is a single random seed with no variance, and the gains over baselines are modest. None of these are fatal, but they need cleanup.

Who is this for? People working on terminology extraction or nested entity/term recognition. It's a simple, reproducible baseline that opens up nested ATE and invites better evaluation. It deserves a serious referee, not a desk reject, but it needs a revision before publication. I'd engage with it — the idea is cheap to test and the weakness is fixable.

Referee Report

3 major / 7 minor

Summary. The paper proposes an end-to-end neural architecture for automatic term extraction (ATE) that treats term extraction as span classification followed by ranking. Given a sentence, all spans up to a maximum length k are scored by a classifier, then a ranker scores the surviving positive spans, and the top K = α·|TotalWords| spans are emitted. The model is evaluated on the GENIA corpus, where it supports nested term extraction and, with raw text plus optional features, achieves a ranker F1 of 0.7789, above two cited baselines (Wang et al. F1=0.707, Yuan et al. F1=0.7143). The central claim is that nested terms can be extracted without syntactic features or hand-crafted feature engineering.

Significance. If the reported result holds, the contribution is a simple and reproducible span-classification-plus-ranking pipeline for nested ATE, a task for which few systems exist. The paper shares code and data, and it explicitly reports the threshold α and the corpus term-ratio statistics, which is useful for replicability. The architecture is straightforward, building on span representations from coreference resolution and targeted-attention mechanisms, and it makes the falsifiable prediction that a feature-free ranker can match or exceed feature-based baselines. However, the significance is currently limited by the evaluation protocol: the final output count K is informed by the gold term density of the test data, and all results come from a single random seed with no error bars. These issues bear directly on whether the claimed empirical advantage is attributable to the model rather than to the evaluation setup.

major comments (3)

[§4.1, Eq. (13), Fig. 5] The output threshold K is derived from the gold term density of the full corpus before the data split, which includes the test sentences. Section 4.1 reports the corpus-wide term ratio 99,111/490,766 ≈ 0.202 and states that α is set to 0.23, 'a little bigger than true distribution.' Because the 0.9/0.05/0.05 split is performed after this statistic is computed, the test gold terms are included in α. Eq. (13) then fixes the number of emitted spans as K = α·|TotalWords|, and Fig. 5 sweeps α on the test set, finding the best F1 near the point where K-num crosses True-Term-num. The reported F1 = 0.7789 is therefore obtained with a test-informed output count, and the comparison to Wang et al. and Yuan et al. in Table 4 is unfair unless those baselines are also given test-informed thresholds. Because the ranker's output is top-K by score, the magnitude of the reported F1 gain over the baselines may be an artifact of knowing how many terms to emit rather than an intrinsic ranking advantage. The authors already note in the Conclusion that threshold-based output is domain-specific; the more serious issue is that the specific α used in the reported experiments is computed from the evaluation corpus. This can be fixed by estimating α from the training or development partition only, by reporting threshold-free ranking metrics (e.g., average precision, R-precision), or by presenting full precision-recall curves over α.
[§4.2, Tables 3 and 4] All experiments are run with a single random seed (Table 3 lists Random Seed 626) and no error bars or significance tests. The differences among the ranker configurations are small: Random Embedding reaches F1=0.7751, while +ALL reaches F1=0.7789. Without multiple seeds or a paired comparison, the central claim that the model 'does not crucially need extra features' is not statistically supported; the near-equality of the random-embedding and feature-augmented rankers is consistent with noise. The paper should report means and standard deviations over at least five seeds and perform a paired test (e.g., bootstrap or Wilcoxon) for the key comparisons, especially for the claimed improvement over the two baselines.
[§3.2, Eqs. (4) and (7)] The model equations contain undefined notation and mathematical typos that impair reproducibility. Equation (7) defines ĥ_m as a 'mean vector' but computes a sum over the span tokens without dividing by the span length. Equation (4) has an undefined denominator: the sum runs over x = i..j while the summed term is h_k·v_t^T, with h_k not bound to the summation index. These issues should be corrected so that the span-head attention and the sentence-targeted attention are fully specified and implementable from the text.

minor comments (7)

[§3.2, Eq. (8)] The notation is inconsistent: hs[x] is used in the numerator and hs[k] in the denominator, but earlier the sentence hidden states are indexed as hs_i. Please use a single indexing scheme throughout.
[§4.2, Table 4] The ranker with Random Embedding already achieves F1=0.7751, only 0.0038 below the +ALL configuration. This suggests that the output threshold K, rather than the learned span representations, dominates the final F1; the paper should discuss this explicitly and report ranking-oriented metrics that separate threshold effects from representation quality.
[§4.3] The term-ratio analysis in Fig. 5 is informative but is performed entirely on the test set. The text should clarify that this is an analysis of test-set behavior, not a model-selection procedure, and should provide development-set results for selecting α.
[§4.1, Table 3] The hyperparameter table lists 'Early Stop 26' but does not explain the criterion (e.g., number of epochs without improvement on the development loss). Please specify the early-stopping rule and how the 'best model' is selected.
[§3.2, 'Additional Features'] The paragraph on optional features uses an asterisk but no footnote, and it does not specify the dimension of the ELMo embeddings or how multiple feature sources are concatenated. This makes the '+ALL' configuration in Table 4 difficult to reproduce.
[Global] There are numerous typos and formatting issues, including 'degisned' in the Conclusion, 'emdedding' in Section 4.2, and 'T able' in several table captions. A careful proofreading pass is needed.
[Fig. 2] The caption 'TermLength&Numbers&Percentage' is not self-explanatory; the figure should be labeled with clear axis names and a legend distinguishing the count, percentage, and cumulative lines.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the model is trained on gold spans and evaluated on a held-out split; the alpha threshold is a transparent hyperparameter, not a prediction derived from the data.

full rationale

The paper presents a standard supervised span-classification and ranking pipeline, not a formal derivation whose conclusion equals its inputs. Span candidates are enumerated from the input sentence (Sec. 3.1), represented from learned word/character features (Sec. 3.2), classified (Eq. 11), scored (Eq. 12), and thresholded by a ranker (Eq. 13). The ranker is trained with a gold-term loss (Eq. 15) on the training split and evaluated on a held-out test split (Sec. 4.1). The reported F1 values are therefore empirical measurements of the trained model, not quantities forced by construction. The self-citations ([15], [18]) are background references to the authors' earlier work on pruning and targeted attention; neither is load-bearing for the main ATE claim, and the paper contains no imported uniqueness theorem or ansatz-by-citation. The most plausible concern is the threshold K = alpha * |TotalWords|: Sec. 4.1 sets alpha = 0.23 from the corpus-wide term ratio (99,111/490,766 approximately 0.202), and Sec. 4.3 sweeps alpha on the test set, observing that F1 peaks near the true term ratio. The paper itself concedes in Sec. 5 that threshold-based output is not applicable to unknown domains because the term distribution and ratio are unknown. This is a legitimate evaluation-protocol weakness: the output count is informed by test-set term density, and the comparison to baselines may be unfair if those baselines do not receive similarly test-informed thresholds. However, knowing roughly how many spans to emit does not determine which spans are correct; the model must still learn to rank true terms above distractor spans. The feature-free claim is additionally supported by the within-model comparison of classifier versus ranker and the feature-ablation rows in Table 4. Thus the issue affects generalizability and benchmark fairness, not the logical derivation, and does not constitute self-definitional or fit-as-prediction circularity.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The model introduces no new physical or conceptual entities. Its main external dependencies are the assumption that terms are contiguous in-sentence spans of bounded length, and the need for a domain-specific term ratio α to set the output threshold.

free parameters (2)

term ratio α = 0.23
Controls the number of output spans via K = α·|TotalWords| (Eq. 13); set close to but above the observed ratio 0.202 in GENIA to favor recall, explicitly tuned on the corpus.
maximum span length k = 15 (used in main experiments)
All candidate spans are truncated to length ≤ k; the paper reports experiments for k=1..15 and says 97.2% of terms are length 1-5, but the main results use a larger cap that is not listed in Table 3.

assumptions (3)

domain assumption Every term is a contiguous span of tokens within a single sentence of length ≤ k.
Section 3.1 defines candidates as all spans [i,j] with j-i ≤ k; discontinuous terms or terms crossing sentence boundaries are not representable.
domain assumption The term ratio α is known or can be estimated for the deployment domain.
Eq. 13 sets K = α·|TotalWords|; the Conclusion states this threshold 'is not so applicable on unknown or unfamiliar domain or data'.
domain assumption GENIA annotations provide correct, complete term boundaries.
The model is trained and evaluated on GENIA 3.02, treating its term spans as gold; annotation noise directly transfers to the results.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Feature-Less End-to-End Nested Term Extraction." pith.science (2026). https://pith.science/paper/2G42C66C

@misc{pith2026190805426,
  author       = {Pith},
  title        = {Pith review of: Feature-Less End-to-End Nested Term Extraction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2G42C66C}},
  note         = {Machine review of arXiv:1908.05426}
}

read the original abstract

In this paper, we proposed a deep learning-based end-to-end method on the domain specified automatic term extraction (ATE), it considers possible term spans within a fixed length in the sentence and predicts them whether they can be conceptual terms. In comparison with current ATE methods, the model supports nested term extraction and does not crucially need extra (extracted) features. Results show that it can achieve high recall and a comparable precision on term extraction task with inputting segmented raw text.

Figures

Figures reproduced from arXiv: 1908.05426 by the authors.

**Figure 1.** Model Architecture We will elaborate in next two parts (Sentence Features and Span Representation) for how to build the span representation in details. Sentence Features This part describes how the sentence sequence features are built from the raw segmented sentence (Refer to the red rectangle part in [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗

**Figure 2.** Term lengths & Numbers & Percentage distribution in the corpus There are total 99,111 terms (distribution indicated in the red line) in 18,539 sentences (total 490,766 words). In these terms, 22675 terms are nested in or overlapped with other terms. 76436 terms are independent. The max length of the terms is 22, with most terms (97.2%) have a length range from 1 to 5. The term ratio is 99111/490766 ≈ 0.202. However,… view at source ↗

**Figure 3.** Classifier on lengths(Testset) 0 2 4 6 8 10 12 14 16 0.4 0.6 0.8 Precision Recall F1 [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 6.** Figure 6: True Positive Terms Distribution The recall [Red Line] increases gradually with the increasing term ratio, and also the precision [Blue Line] decreases due to the increasing of candidates space [Purple Line, K-num]. When the ratio approach 0.2 (the actual term distribu…

**Figure 7.** Figure 7: Samples on Test-Set 5 Conclusion and Future Work We proposed a deep learning-based end-to-end term extraction method in this paper. It employs classification and ranking on the span (n-grams) candidates in the sentences. Compared with current methods, it supports the n…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 20 canonical work pages

[1]

Stankovi Ranka, Krstev Cvetana, Obradovi Ivan, Lazi Biljana and Trtovac Aleksan- dra: ”Rule-based automatic multi-word term extraction and lemmatization.” Pro- ceedings of the Tenth International Conference on Language Resources and Evalu- ation (LREC 2016). 2016

work page 2016
[2]

Katerina Frantzi, Sophia Ananiadou and Hideki Mima: ”Automatic recognition of multi-word terms: the c-value/nc-value method.” International Journal on Digital Libraries, 3(2):115130. 2000

work page 2000
[3]

Ziqi Zhang, Jie Gao, and Fabio Ciravegna: ”Jate 2.0: Java automatic term extraction with apache solr.” Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC 2016), 2016

work page 2016
[4]

Lishuang Li, Yanzhong Dang, Jing Zhang and Dan Li: ”Domain term extraction based on conditional random ﬁelds combined with active learning strategy.” Journal of Information & Computational Science, vol. 9, no. 7, pp. 19311940, 2012

work page 2012
[5]

IEEE, 2017

Yu Yuan, Jie Gao, and Yue Zhang: ”Supervised learning for robust term extraction.” 2017 International Conference on Asian Language Processing (IALP). IEEE, 2017

work page 2017
[6]

Association for Compu- tational Linguistics, 2004

GuoDong Zhou and Jian Su: ”Exploring deep knowledge resources in biomedical name recognition.” Proceedings of the International Joint Workshop on Natural Language Processing in Biomedicine and its Applications. Association for Compu- tational Linguistics, 2004

work page 2004
[7]

Rogelio Nazar and Maria Teresa Cabre: ”Supervised learning algorithms applied to terminology extraction.” In Proceedings of the 10th Terminology and Knowledge Engineering Conference, pages 209217, 2012

work page 2012
[8]

Merley da Silva Conrado, Thiago A. Salgueiro Pardo and Solange Oliveira Rezende: ”A machine learning approach to automatic term extraction us- ing a rich feature set.” Proceedings of the NAACL HLT 2013 Student Re- search Workshop, Atlanta, Georgia, June 2013, pp. 1623. [Online]. Available: http://www.aclweb.org/anthology/N13-2003

work page 2013

Show all 24 references

[9]

Maren Kucza, Jan Niehues and Sebastian Stker: ”Term Extraction via Neural Se- quence Labeling a Comparative Evaluation of Strategies Using Recurrent Neural Networks.” Proceedings pf Interspeech. 2018

2018
[10]

Rui Wang, Wei Liu, and Chris McDonald: ”Featureless domain-speciﬁc term ex- traction with minimal labelled data.” Proceedings of the Australasian Language Technology Association Workshop 2016. 2016

2016
[11]

”Analyzing the eﬀectiveness and applicability of co-training.” Cikm

Nigam, Kamal, and Rayid Ghani. ”Analyzing the eﬀectiveness and applicability of co-training.” Cikm. Vol. 5. 2000

2000
[12]

14 Gao and Yu

Kenton Lee, Luheng He, Mike Lewis and Luke Zettlemoyer: ”End-to-end neural coreference resolution.” arXiv preprint arXiv:1707.07045 (2017). 14 Gao and Yu

2017 arXiv
[13]

Jie Yang, Yue Zhang: ”NCRF++: An Open-source Neural Sequence Labeling Toolkit.” Proceedings of Association for Computational Linguistics ACL (2018), 2018

2018
[14]

Hinton: ”Imagenet classiﬁcation with deep convolutional neural networks.” Advances in neural information process- ing systems

Krizhevsky, Alex, Ilya Sutskever and Geoﬀrey E. Hinton: ”Imagenet classiﬁcation with deep convolutional neural networks.” Advances in neural information process- ing systems. 2012

2012
[15]

Yuze Gao, and Tong Xiao: ”A Comparison of Pruning Methods for CYK-based Decoding in Machine Translation.” Proceedings of China Workshop on Machine Translation (CWMT). 2015

2015
[16]

Hochreiter Sepp and Jrgen Schmidhuber: ”Long short-term memory.” Neural com- putation 9.8 (1997): 1735-1780

1997
[17]

Jiangming Liu and Yue Zhang: ”Attention modeling for targeted sentiment.” Pro- ceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers. 2017

2017
[18]

Yuze Gao, Yue Zhang and Tong Xiao: ”Implicit Syntactic Features for Targeted Sentiment Analysis.” Proceedings of IJCNLP (2017), 2017

2017
[19]

Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee and Luke Zettlemoyer: ”Deep contextualized word representa- tions.” arXiv preprint arXiv:1802.05365 (2018)

2018 arXiv
[20]

and Jimmy Ba: ”Adam: A method for stochastic optimiza- tion.” arXiv preprint arXiv:1412.6980 (2014)

Kingma Diederik P. and Jimmy Ba: ”Adam: A method for stochastic optimiza- tion.” arXiv preprint arXiv:1412.6980 (2014)

2014 arXiv
[21]

19 Suppl 1, pp

Jin-Dong Kim, Tomoko Ohta, Yuka Teteisi and Junichi Tsujii: ”Genia corpusa se- mantically annotated corpus for bio-textmining.” Bioinformatics (Oxford, England), vol. 19 Suppl 1, pp. i1802, 02 2003

2003
[22]

Nitish Srivastava, Geoﬀrey Hinton, Alex Krizhevsky, Ilya Sutskever and Ruslan Salakhutdinov: ”Dropout: a simple way to prevent neural networks from overﬁtting.” The journal of machine learning research, 15(1), 1929-1958. 2014

1929
[23]

”Bert: Pre-training of deep bidirectional transformers for language understanding.” arXiv preprint arXiv:1810.04805 (2018)

Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. ”Bert: Pre-training of deep bidirectional transformers for language understanding.” arXiv preprint arXiv:1810.04805 (2018)

2018 arXiv
[24]

”Language models are unsupervised multitask learners.” OpenAI Blog 1, no

Radford Alec, Jeﬀrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. ”Language models are unsupervised multitask learners.” OpenAI Blog 1, no. 8 (2019)

2019

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

Stankovi Ranka, Krstev Cvetana, Obradovi Ivan, Lazi Biljana and Trtovac Aleksan- dra: ”Rule-based automatic multi-word term extraction and lemmatization.” Pro- ceedings of the Tenth International Conference on Language Resources and Evalu- ation (LREC 2016). 2016

work page 2016

[2] [2]

Katerina Frantzi, Sophia Ananiadou and Hideki Mima: ”Automatic recognition of multi-word terms: the c-value/nc-value method.” International Journal on Digital Libraries, 3(2):115130. 2000

work page 2000

[3] [3]

Ziqi Zhang, Jie Gao, and Fabio Ciravegna: ”Jate 2.0: Java automatic term extraction with apache solr.” Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC 2016), 2016

work page 2016

[4] [4]

Lishuang Li, Yanzhong Dang, Jing Zhang and Dan Li: ”Domain term extraction based on conditional random ﬁelds combined with active learning strategy.” Journal of Information & Computational Science, vol. 9, no. 7, pp. 19311940, 2012

work page 2012

[5] [5]

IEEE, 2017

Yu Yuan, Jie Gao, and Yue Zhang: ”Supervised learning for robust term extraction.” 2017 International Conference on Asian Language Processing (IALP). IEEE, 2017

work page 2017

[6] [6]

Association for Compu- tational Linguistics, 2004

GuoDong Zhou and Jian Su: ”Exploring deep knowledge resources in biomedical name recognition.” Proceedings of the International Joint Workshop on Natural Language Processing in Biomedicine and its Applications. Association for Compu- tational Linguistics, 2004

work page 2004

[7] [7]

Rogelio Nazar and Maria Teresa Cabre: ”Supervised learning algorithms applied to terminology extraction.” In Proceedings of the 10th Terminology and Knowledge Engineering Conference, pages 209217, 2012

work page 2012

[8] [8]

Merley da Silva Conrado, Thiago A. Salgueiro Pardo and Solange Oliveira Rezende: ”A machine learning approach to automatic term extraction us- ing a rich feature set.” Proceedings of the NAACL HLT 2013 Student Re- search Workshop, Atlanta, Georgia, June 2013, pp. 1623. [Online]. Available: http://www.aclweb.org/anthology/N13-2003

work page 2013