---
title: "USZEGED — paper by Berend and Tasnádi (2015)"
description: "The paper describes team USZEGED's system for the W-NUT 2015 shared task on lexical… 0.7564 F1. 18 references, cited by 4 papers here."
authors: ["Gábor Berend", "Ervin Tasnádi"]
year: 2015
venue: "Proceedings of the Workshop on Noisy User-generated Text"
source: "https://aclanthology.org/W15-4318/"
doi: "10.18653/v1/w15-4318"
licence: "CC BY-NC-SA 3.0 (ACL Anthology, before 2016)"
licence_url: "https://creativecommons.org/licenses/by-nc-sa/3.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/berend-2015-uszeged-correction-type-sensitive-normalization/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/berend-2015-uszeged-correction-type-sensitive-normalization/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2015](https://md.research.elamur.ai/free/en/papers/2015/page/3/) / USZEGED: Correction Type-sensitive Normalization of English Tweets Using Efficiently Indexed n-gram Statistics

# USZEGED: Correction Type-sensitive Normalization of English Tweets Using Efficiently Indexed n-gram Statistics

**Authors:** [Gábor Berend](https://md.research.elamur.ai/free/en/authors/gabor-berend/) and Ervin Tasnádi · **Year:** 2015 · **Venue:** [Proceedings of the Workshop on Noisy User-generated Text](https://md.research.elamur.ai/free/en/venues/wnut/) ([Workshop on Noisy User-generated Text 2015](https://md.research.elamur.ai/free/en/venues/wnut/2015/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY-NC-SA 3.0 (ACL Anthology, before 2016)](https://creativecommons.org/licenses/by-nc-sa/3.0/)

**Links:** [Publisher's page](https://aclanthology.org/W15-4318/) · [PDF](https://aclanthology.org/W15-4318.pdf) · [DOI 10.18653/v1/w15-4318](https://doi.org/10.18653/v1/w15-4318) · [OpenAlex](https://openalex.org/W2250475966)

**Citations:** 12 citations (OpenAlex, 2026-10-14) · [4](https://md.research.elamur.ai/free/en/papers/berend-2015-uszeged-correction-type-sensitive-normalization/cited-by/) in our library, of the 149,439 papers whose bibliographies we read

## Summary
The paper describes team USZEGED's system for the W-NUT 2015 shared task on lexical normalisation of English tweets. Standard NLP tools break down on noisy social-media text, so the goal is to map each non-standard token to its in-vocabulary (IV) form. The system has two stages.

1. **Correction type.** A linear-chain CRF assigns each token one of six labels: no change, missing apostrophe, missing whitespace, close one-to-one rewrite, distant one-to-one rewrite, or abbreviation expansion. Its features come from surface cues, several normalisation lexicons, Twitter POS tags and Google 1T frequencies.
2. **Candidate choice.** The predicted type selects which candidate set to build. The winning candidate is the one that best fits the left and right neighbours, judged by bigram statistics from a large Twitter n-gram corpus indexed for fast wildcard lookups.

The underlying intuition is that other users write the standard form in the same contexts. The system reached F1 0.8052 on the test set, second among unconstrained entries. The gap between training and test scores was small, suggesting the system generalises well.

## Key points
- **Data:** the training set has 2,950 tweets (44,385 tokens) and the test set has 1,967 tweets (29,421 tokens). About 91% of tokens need no correction.
- **Correction-type classifier:** a CRF trained with CRFsuite (L-BFGS, L1 regularisation).
  - Token accuracy was 0.9830 on train and 0.9746 on test.
  - Fully correct tweets were 0.7902 on train and 0.7143 on test.
  - Macro F was 0.8171 on test. Missing-whitespace was the weakest class, with test F 0.5546.
- **Lexical resources:**
  - SCOWL from aspell
  - normalisation lexicons compiled by Han et al. (2012) and, separately, by Liu et al. (2012)
  - a 5,307-entry noslang.com list
  - the Twitter POS tagger of Gimpel et al. (2011), whose G tag marks multi-word abbreviations
- **Candidate ranking:** candidates are scored by P(t′|left)·P(right|t′) over the Twitter n-gram corpus of Herdağdelen (2013), indexed with the LIT indexer. When the dictionaries offer exactly one suggestion, that suggestion overrides the n-gram choice.
- **Overall test score:** P 0.8606, R 0.7564, F1 0.8052. On train, F1 was 0.8156.

## Interesting ideas
- The system first classifies what kind of fix a token needs and only then generates candidates of that kind, which keeps each candidate generator small and specialised.
- Their edit distance is asymmetric: inserting characters into the noisy form costs nothing, and a few phonetic endings (z→s, a→er) are free, because abbreviations mostly drop letters.
- Missing spaces are detected by enumerating splits whose parts are all dictionary words and accepting a split only if it is very frequent in a large n-gram resource.
- The correct form is picked by querying "left neighbour * right neighbour" against an indexed corpus and keeping the candidate that best fits both sides.
- Evaluation is reported per correction type and at whole-tweet level, not only as overall token accuracy, which exposes rare but hard classes.

## For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

- **I1** (tokens): crate en could tag noisy tokens with a repair type (apostrophe, split, abbreviation, spelling) before lemmatisation, so each repair stays explainable and is applied only where it belongs.
- **I2** (tokens): a direction-aware distance like this would let the lemmatiser link clipped or informal spellings to dictionary lemmas with high recall and a clear reason for each match.
- **I3** (tokens): the tokenizer could split run-together words or hashtags the same way, keeping the winning split and its frequency as a traceable justification.
- **I4** (learn): such wildcard context queries could act as one cheap extra witness when building silver data, flagging tokens whose surface form is unusual for its context.
- **I5** (learn): evaluation reports could add per-phenomenon breakdowns and whole-sentence correctness, so a weak rare category is not hidden by a strong majority class.

*Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.*

## Institutions

- [University of Szeged](https://md.research.elamur.ai/free/en/universities/university-of-szeged/) — University, Szeged, Hungary · 12 works · 1,507 citations

📚 Shelves (5): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 16,667 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 16,194 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 16,039 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 15,891 works · [🖱️ Human-Computer Interaction](https://md.research.elamur.ai/free/en/shelves/computer-science/human-computer-interaction/) · 141 works

🧭 Topics (1): [English](https://md.research.elamur.ai/free/en/topics/english/) · 2,263 works

💡 Concepts (1): [💡 N-gram](https://md.research.elamur.ai/free/en/concepts/n-gram/) · 26 works

## Cite

Cite as: Gábor Berend and Ervin Tasnádi. 2015. USZEGED: Correction Type-sensitive Normalization of English Tweets Using Efficiently Indexed n-gram Statistics. Proceedings of the Workshop on Noisy User-generated Text.

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@inproceedings{berend2015uszeged,
  title = {{USZEGED: Correction Type-sensitive Normalization of English Tweets Using Efficiently Indexed n-gram Statistics}},
  author = {Gábor Berend and Ervin Tasnádi},
  booktitle = {Proceedings of the Workshop on Noisy User-generated Text},
  year = {2015},
  doi = {10.18653/v1/w15-4318},
  url = {https://doi.org/10.18653/v1/w15-4318}
}
```

- [← previous UMMU@QALB-2015 Shared Task: Character and Word level SMT pipeline for Automatic Error Correction of Arabic Text](https://md.research.elamur.ai/free/en/papers/bougares-2015-ummu-qalb-2015-shared-task/)
- [next → Answer Extraction for Why Arabic Questions Answering Systems: EWAQ](https://md.research.elamur.ai/free/en/papers/al-khawaldeh-2019-answer-extraction-why-arabic-questions/)
- [↑ back to papers of 2015](https://md.research.elamur.ai/free/en/papers/2015/page/3/)

## References

18 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Timothy Baldwin, Marie Catherine de Marneffe, Bo Han, Young-Bum Kim, Alan Ritter, and Wei Xu. 2015. [Shared tasks of the 2015 workshop on noisy user-generated text: Twitter lexical normalization and named entity recognition](https://md.research.elamur.ai/free/en/papers/baldwin-2015-shared-tasks-2015-workshop-noisy/). In Proceedings of the Workshop on Noisy User-generated Text (WNUT 2015), Beijing, China.
2. Hakan Ceylan and Rada Mihalcea. 2011. An efficient indexer for large n-gram corpora. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies: Systems Demonstrations, HLT ’11, pages 103– 108, Stroudsburg, PA, USA. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Hakan+Ceylan+and+Rada+Mihalcea%2E+2011%2E+An+efficient+indexer+for+large+n%2Dgram+corpora%2E+In+Proceedings+of+the+49th+Annual+Meeting+of+the+Association+for+Computational+Linguistics%3A+Human+Language+Technolo)
3. Leon Derczynski, Alan Ritter, Sam Clark, and Kalina Bontcheva. 2013. [Twitter part-of-speech tagging for all: Overcoming sparse and noisy data](https://md.research.elamur.ai/free/en/papers/derczynski-2013-twitter-part-speech-tagging-all/). In Proceedings of the International Conference on Recent Advances in Natural Language Processing. Association for Computational Linguistics.
4. Kevin Gimpel, Nathan Schneider, Brendan O’Connor, Dipanjan Das, Daniel Mills, Jacob Eisenstein, Michael Heilman, Dani Yogatama, Jeffrey Flanigan, and Noah A. Smith. 2011. [Part-of-speech tagging for twitter: Annotation, features, and experiments](https://md.research.elamur.ai/free/en/papers/gimpel-2011-part-speech-tagging-twitter-annotation/). In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies: Short Papers - Volume 2, HLT ’11, pages 42–47, Stroudsburg, PA, USA. Association for Computational Linguistics. [↗](https://doi.org/10.21236/ada547371)
5. Bo Han and Timothy Baldwin. 2011. [Lexical normalisation of short text messages: Makn sens a #twitter](https://md.research.elamur.ai/free/en/papers/han-2011-lexical-normalisation-short-text-messages/). In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies - Volume 1, HLT ’11, pages 368–378, Stroudsburg, PA, USA. Association for Computational Linguistics.
6. Bo Han, Paul Cook, and Timothy Baldwin. 2012. Automatically constructing a normalisation dictionary for microblogs. In Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLP-CoNLL 2012), pages 421– 432, Jeju Island, Korea. [find](https://scholar.google.com/scholar?q=Bo+Han%2C+Paul+Cook%2C+and+Timothy+Baldwin%2E+2012%2E+Automatically+constructing+a+normalisation+dictionary+for+microblogs%2E+In+Proceedings+of+the+2012+Joint+Conference+on+Empirical+Methods+in+Natural+Language)
7. Bo Han, Paul Cook, and Timothy Baldwin. 2013. Lexical normalization for social media text. ACM Trans. Intell. Syst. Technol., 4(1):5:1–5:27, February. [↗](https://doi.org/10.1145/2414425.2414430)
8. Ama Herdadelen. 2013. Twitter n-gram corpus with demographic metadata. Language Resources and Evaluation, 47(4):1127–1147. [find](https://scholar.google.com/scholar?q=Ama+Herdadelen%2E+2013%2E+Twitter+n%2Dgram+corpus+with+demographic+metadata%2E+Language+Resources+and+Evaluation%2C+47%284%29%3A1127%E2%80%931147%2E)
9. John D. Lafferty, Andrew McCallum, and Fernando C. N. Pereira. 2001. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. In Proceedings of the Eighteenth International Conference on Machine Learning, ICML ’01, pages 282–289, San Francisco, CA, USA. Morgan Kaufmann Publishers Inc. [find](https://scholar.google.com/scholar?q=John+D%2E+Lafferty%2C+Andrew+McCallum%2C+and+Fernando+C%2E+N%2E+Pereira%2E+2001%2E+Conditional+random+fields%3A+Probabilistic+models+for+segmenting+and+labeling+sequence+data%2E+In+Proceedings+of+the+Eighteenth+Interna)
10. Alex Lamb, Michael J. Paul, and Mark Dredze. 2013. Separating fact from fear: Tracking flu infections on twitter. In In NAACL. [find](https://scholar.google.com/scholar?q=Alex+Lamb%2C+Michael+J%2E+Paul%2C+and+Mark+Dredze%2E+2013%2E+Separating+fact+from+fear%3A+Tracking+flu+infections+on+twitter%2E+In+In+NAACL%2E)
11. Fei Liu, Fuliang Weng, Bingqing Wang, and Yang Liu. 2011. Insertion, Deletion, or Substitution? Normalizing Text Messages without Pre-categorization nor Supervision. In Proceedings of the 49 th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 71– 76, Stroudsburg, PA, USA. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Fei+Liu%2C+Fuliang+Weng%2C+Bingqing+Wang%2C+and+Yang+Liu%2E+2011%2E+Insertion%2C+Deletion%2C+or+Substitution%3F+Normalizing+Text+Messages+without+Pre%2Dcategorization+nor+Supervision%2E+In+Proceedings+of+the+49+th+Annual)
12. Fei Liu, Fuliang Weng, and Xiao Jiang. 2012. A Broad-Coverage Normalization System for Social Media Language. Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics: Long Papers-Volume 1, pages 1035–1044. [find](https://scholar.google.com/scholar?q=Fei+Liu%2C+Fuliang+Weng%2C+and+Xiao+Jiang%2E+2012%2E+A+Broad%2DCoverage+Normalization+System+for+Social+Media+Language%2E+Proceedings+of+the+50th+Annual+Meeting+of+the+Association+for+Computational+Linguistics%3A+L)
13. Naoaki Okazaki. 2007. Crfsuite: a fast implementation of conditional random fields (crfs). [find](https://scholar.google.com/scholar?q=Naoaki+Okazaki%2E+2007%2E+Crfsuite%3A+a+fast+implementation+of+conditional+random+fields+%28crfs%29%2E)
14. Alan Ritter, Sam Clark, Mausam, and Oren Etzioni. 2011. [Named entity recognition in tweets: An experimental study](https://md.research.elamur.ai/free/en/papers/ritter-2011-named-entity-recognition-tweets-experimental/). In Proceedings of the Conference on Empirical Methods in Natural Language Processing, EMNLP ’11, pages 1524–1534, Stroudsburg, PA, USA. Association for Computational Linguistics.
15. Alan Ritter, Mausam, Oren Etzioni, and Sam Clark. 2012. Open domain event extraction from twitter. In Proceedings of the 18th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’12, pages 1104–1112, New York, NY, USA. ACM. [find](https://scholar.google.com/scholar?q=Alan+Ritter%2C+Mausam%2C+Oren+Etzioni%2C+and+Sam+Clark%2E+2012%2E+Open+domain+event+extraction+from+twitter%2E+In+Proceedings+of+the+18th+ACM+SIGKDD+International+Conference+on+Knowledge+Discovery+and+Data+Mining)
16. Alan Ritter, Evan Wright, William Casey, and Tom Mitchell. 2015. Weakly supervised extraction of computer security events from twitter. In Proceedings of the 24th International Conference on World Wide Web, WWW ’15, pages 896–905, Republic and Canton of Geneva, Switzerland. International World Wide Web Conferences Steering Committee. [find](https://scholar.google.com/scholar?q=Alan+Ritter%2C+Evan+Wright%2C+William+Casey%2C+and+Tom+Mitchell%2E+2015%2E+Weakly+supervised+extraction+of+computer+security+events+from+twitter%2E+In+Proceedings+of+the+24th+International+Conference+on+World+Wid)
17. Takeshi Sakaki, Makoto Okazaki, and Yutaka Matsuo. 2010. Earthquake shakes twitter users: Real-time event detection by social sensors. In Proceedings of the 19th International Conference on World Wide Web, WWW ’10, pages 851–860, New York, NY, USA. ACM. [find](https://scholar.google.com/scholar?q=Takeshi+Sakaki%2C+Makoto+Okazaki%2C+and+Yutaka+Matsuo%2E+2010%2E+Earthquake+shakes+twitter+users%3A+Real%2Dtime+event+detection+by+social+sensors%2E+In+Proceedings+of+the+19th+International+Conference+on+World+Wide)
18. Yi Yang and Jacob Eisenstein. 2013. A Log-Linear Model for Unsupervised Text Normalization. Proceedings of the Empirical Methods on Natural Language Processing (EMNLP), pages 61–72. [find](https://scholar.google.com/scholar?q=Yi+Yang+and+Jacob+Eisenstein%2E+2013%2E+A+Log%2DLinear+Model+for+Unsupervised+Text+Normalization%2E+Proceedings+of+the+Empirical+Methods+on+Natural+Language+Processing+%28EMNLP%29%2C+pages+61%E2%80%9372%2E)

*Updated 1 November 2026*