IHS_RD: Lexical Normalization for English Tweets

Dmitry Supranovich IHS Inc. / IHS Global Belarus 131 Starovilenskaya St 220123, Minsk, Belarus Dmitry.Supranovich@ihs.com Abstract This paper describes the Twitter lexical nor- malization system submitted by IHS R&D Belarus team for the ACL 2015 workshop on noisy user-generated text. The proposed sys- tem consists of two components: a CRF- based approach to identify possible normali- zation candidates, and a post-processing step in an attempt to normalize words that do not have normalization variants in the lexicon. Evaluation on the test data set showed that our unconstrained system achieved the F- measure of 0.8272 (rank 1 out of 5 submis- sions for the unconstrained mode, rank 2 out of all 11 submissions). 1 Introduction Social media texts found in such services as Twitter or Facebook have a great data-mining potential, as they offer real-time data that can be useful to monitor public opinion on brands, products, events, etc. However, current Natural Language Processing systems are usually opti- mized for clean data, which is not the type of data found in social media texts, as they are often noisy, containing a lot of slang, typos, and ab- breviations. Normalizing such text is challenging. We want to achieve high recall, making as many correc- tions as possible, but not at the expense of preci- sion – words should not be incorrectly normal- ized. Previous approaches to this task incorporated different tools and methods: dictionaries, lan- guage models, finite state transducers, and ma- chine translation models. Some of the methods are unsupervised, though often requiring adjust- ment of parameters based on annotated data (Han and Baldwin (2011), Liu et al. (2011), and Gouws et al. (2011)). Some are supervised, like that in Chrupała (2014), making use of a Condi- tional Random Field (Lafferty et al., 2001) to Viachaslau Patsepnia IHS Inc. 55 Cambridge pkwy, Suite 601 Cambridge, MA 02142, USA Slava.Patsepnia@ihs.com learn the sequences of edit operations from la- belled data. In this paper, we present an approach based on the usage of normalization lexicons and a CRF model for identifying potential candidates. 2 Task Description 2.1 Dataset The corpus provided by the organizers consists of 2950 annotated tweets. The annotations follow these guidelines (Baldwin et al., 2015): • Non-standard words are normalized to one or more canonical English words based on a pre-defined lexicon. For instance, l o v e should be normalized to love (many-to-one normalization), tmrw to tomorrow (one-to- one normalization), and cu to see you (one-to-many normalization). Additional- ly, IBM should be left untouched as it is in the lexicon and it is in its canonical form, and the informal lol should be expanded to laughing out loud . • Non-standard words may be either out-of- vocabulary (OOV) tokens (e.g., tmrw for tomorrow ) or in-vocabulary (IV) tokens (e.g., wit for with in “I will come wit you”). • Only alphanumeric tokens (e.g., 2 , 4eva and tmrw ) and apostrophes used in con- tractions (e.g., yoou've ) are considered for normalization. Tokens including hyphens, single quotes and other types of contrac- tions should be ignored. • Domain specific entities are ignored even if they are in non-standard forms, e.g., #ttyl , @nyc • It is possible for a tweet to have no non- standard tokens but still require normaliza- tion (e.g., the example of wit above), and it is also possible for the tweet to require no normalization whatsoever.

78 Proceedings of the ACL 2015 Workshop on Noisy User-generated Text , pages 78–81, Beijing, China, July 31, 2015. c � 2015 Association for Computational Linguistics

• Proper nouns should be left untouched, even if they are not in the given lexicon (e.g., Twitter ). • All normalizations should use the Ameri- can spelling (e.g., tokenize rather than to- kenise ). 2.2 Evaluation Evaluation was to be carried out according to Precision, Recall, and F1 metrics. 3 Experimental Setup First, a normalization lexicon was generated from the given training data, enriched with the data from several sources: • Word pairs extracted from the datasets used for lexical normalization (Han, 2011; Liu, 2011) • The online social media abbreviation list of Beal (2015) 1 . Compared to the previous workshops with one-to-one normaliza- tions, the current task also considers one- to-many normalizations, and obviously not all abbreviations are present in the training data, so the use of a list of social media abbreviations can be vital to the system. At the current stage of development the sys- tem is unable to differentiate between several normalization variants; thus, entries with multi- ple possible variants were reviewed to make the most suitable variant first in the list (entries that are most frequent in datasets are placed first, any ties were manually reviewed). Second, a CRF model was trained. The labels chosen were CAND and NOT_CAND, reflecting potential normalization candidates and words that should not be normalized, respectively. The following features were used: Token: This feature represents the string of the current token. Context Feature: The token to the left and the token to the right are used as two context fea- tures. The surrounding words usually convey useful information about a token which helps in predicting the correct tag for each token. Alphanumeric feature: This feature checks whether the token adheres to the annotation guidelines and makes sure that non-adhering to- kens are not marked as potential candidates. 1 http://www.webopedia.com/quick_ref/textmessageabbr eviations.asp Normalization dictionary feature: This fea- ture checks whether the token is present in the generated normalization lexicon. Canonical lexicon feature : This feature indi- cates whether or not the token is present in the canonical lexicon provided by the workshop or- ganizers. Word length and number of vowels : Two separate features as well as their correlation, al- lowing to tag words with uncommon length- vowel correlation, like bcz, pls, etc. Edit distance feature: marks a token that is within an edit distance of 2 or less from any word in the canonical lexicon. Third, the text is normalized: • All tokens tagged as potential candidates by the CRF model are normalized to their lexicon variants. • All alphanumeric words are normalized to the American spelling with the VarCon tool (Atkinson, 2015) 2 . This includes the tokens which are already normalized using the lexicon. • We have also tried to improve the normal- ization results by using a did-you-mean (DYM) module that is currently being de- veloped at IHS R&D team. The DYM module corrects user queries/sentences with misspellings by providing corrected variant(s) with a confidence measure (in- cluding no correction variant with the cor- responding confidence measure). The DYM module is an SVM model trained on a set of features for each of the multiple candidates generated for an input que- ry/sentence. We used the following fea- tures: error model score, Levenshtein dis- tance, language model score, the ratio of common noun vocabulary words, the ratio of proper noun vocabulary words, and the number of changes in non-lowercase words. An error model score was obtained from an autocompletion and autocorrec- tion module (AAM) for which an index was built from 12.4M documents (scien- tific papers - 42.1%, Wikipedia articles - 23.5%, patents - 19.4%, social texts - 8%, and news - 7%). The 2-gram language model was built from 177K patents (1.36G words and 2.6M vocabulary). Since we did not have enough time to tailor both DYM and AAM modules for social text pro- cessing, DYM and AAM modules were 2 http://wordlist.aspell.net/varcon/

used for this Twitter lexical normalization system as is, being actually tailored for technical and scientific texts. 3.1 Results and error analysis Testing was performed on the provided corpus of 1967 tweets. Table 1 shows the performance of our CRF candidate model with different features: • A baseline model with only token, context and alphanumeric features. • A baseline model with the normalization dictionary and the canonical lexicon fea- tures added. • A model with all features enabled. Table 2 reflects our submitted normalization result and a result without the DYM module de- scribed above. 1651 1113 917 945 787 773 518 518 Precision (CRF | Final) Recall (CRF | Final) F1 (CRF | Final) Tokens + Context + Alphanumeric 0.991 | 0.8782 0.57 | 0.6013 0.7237 | 0.7139 Added diction- ary features 0.907 | 0.8376 0.824 | 0.8133 0.8635 | 0.8253 All features 0.915 | 0.8469 0.817 | 0.8083 0.8632 | 0.8272

Table 1. Result metrics of candidate CRF model with different features (and its impact on the re- sult after normalization using a submitted sys- tem). 1555 1061 1065 945 283 1023 773 Precision Recall F1 Lexicon Normalization + DYM (submitted) 0.8469 0.8083 0.8272 Lexicon Normalization without DYM 0.8765 0.7949 0.8337

Table 2. Result metrics of two normalization sys- tem configurations. The DYM feature does a good job correcting typos and removing excessive duplicate letters ( beutiful 4 beautiful , tosee 4 to see , and smileeeeee 4 smile ). However, even with a high confidence threshold, quite a number of words are normalized excessively, mainly those in non- English (or partially English) tweets, e.g. jeil 4 jail , hoje 4 hope , and wasan 4 was an , in addi- tion to some incorrect normalizations like parkd 4 park (instead of parked ) or hundread 4 hun- dreds (instead of hundred ). These mistakes are frequent, and an increase in recall does not out- weigh a loss in precision; thus, the F-measure without the DYM feature in its current state is even a little bit higher than our submitted system with it. Lowering the confidence threshold brings more correct normalizations, but due to the na- ture of tweets even more incorrect ones, leading to an overall drop in F1 score. Nevertheless, we decided to use and submit the system with DYM, since we believe the text normalized this way is more suitable for further use. Attempts were made to improve the perfor- mance of the DYM module as well as to select the correct candidate from a normalization lexi- con if there is more than one variant present (ur 4 you’re, your, you). For example, language detection works well on regular search queries and could potentially forbid the normalization of words in non-English tweets. However, it proved to be not helpful for tweets – the messages are short, some of them are a mixture of English and some other language (thus, if there is a normali- zation restriction on such tweets, potential Eng- lish normalizations are lost), and slang- and ab- breviation-rich tweets are hard to analyse. A lan- guage model was used in an attempt to select a correct normalization from multiple variants, but this did not prove to be effective, likely because the model used was not focused on social media texts. We see room for potential improvement in tuning the DYM tool to social media texts, as well as in filtering non-English words from nor- malization candidates, experimenting with lan- guage models tailored to social media texts and further enriching the lexicon with new normali- zation data. 4 Conclusion In this paper, we presented a system designed for participation in shared task #2 of the ACL 2015 workshop on noisy user-generated text. Our sys- tem makes use of CRF for identifying potential candidates, lexicons to normalize them and a DYM module as a post-processing step to further correct some of the misspelled words. Our sys- tem ranked second among all 11 submissions with 0.8272 F-measure and ranked first among 5 submissions for the unconstrained mode.

References Kevin Atkinson. VarCon. Vers. 2015.02.15. Web . 01 Apr. 2015. http://wordlist.aspell.net/varcon/ 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. In Proceedings of the Workshop on Noisy User-generated Text (WNUT 2015) , Beijing, China. Vangie Beal. Text messaging and online chat abbre- viations. Web . 01 Apr. 2015. http://www.webopedia.com/quick_ref/textmessage abbreviations.asp Grzegorz Chrupała. 2014. Normalizing tweets with edit scripts and recurrent neural embeddings. In Proceedings of the 52nd Annual Meeting of the As- sociation for Computational Linguistics (ACL 2014) , pages 680–686, Baltimore, USA. Stephan Gouws, Dirk Hovy, and Donald Metzler. 2011. Unsupervised mining of lexical variants from noisy text. In Proceedings of the First workshop on Unsupervised Learning in NLP , pages 82–90, Ed- inburgh, UK. Bo Han and Timothy Baldwin. 2011. Lexical nor- malization of short text messages: Makn sens a #twitter. In Proceedings of the 49th Annual Meet- ing of the Association for Computational Linguis- tics (ACL 2011) , pages 368-378, Portland, USA. 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 In- ternational Conference on Machine Learning , ICML'01, pages 282–289. Morgan Kaufmann Pub- lishers Inc., San Francisco, CA, USA 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 49th Annual Meeting of the Association for Computational Linguistics (ACL 2011) , pages 71- 76, Portland, USA. Fei Liu, Fuliang Weng, and Xiao Jiang. 2012. A broad-coverage normalization system for social media language. In Proceedings of the 50th Annual Meeting of the Association for Computational Lin- guistics (ACL 2012) , pages 1035-1044, Jeju Island, Korea. Yi Yang and Jacob Eisenstein. 2013. A log-linear model for unsupervised text normalization. In Pro- ceedings of Conference on Empirical Methods in Natura Language Processing (EMNLP) , pages 61- 72, Seattle, USA.