1{2 "document_id": "lb_qasper_93004cdb0e6d24f5de0568ed952fd50655e42900dc9dbfdb",3 "family": "public_qasper",4 "split": "test",5 "text": "Introduction\nAccurate language identification (LID) is the first step in many natural language processing and machine comprehension pipelines. If the language of a piece of text is known then the appropriate downstream models like parts of speech taggers and language models can be applied as required.\nLID is further also an important step in harvesting scarce language resources. Harvested data can be used to bootstrap more accurate LID models and in doing so continually improve the quality of the harvested data. Availability of data is still one of the big roadblocks for applying data driven approaches like supervised machine learning in developing countries.\nHaving 11 official languages of South Africa has lead to initiatives (discussed in the next section) that have had positive effect on the availability of language resources for research. However, many of the South African languages are still under resourced from the point of view of building data driven models for machine comprehension and process automation.\nTable TABREF2 shows the percentages of first language speakers for each of the official languages of South Africa. These are four conjunctively written Nguni languages (zul, xho, nbl, ssw), Afrikaans (afr) and English (eng), three disjunctively written Sotho languages (nso, sot, tsn), as well as tshiVenda (ven) and Xitsonga (tso). The Nguni languages are similar to each other and harder to distinguish. The same is true of the Sotho languages.\nThis paper presents a hierarchical naive Bayesian and lexicon based classifier for LID of short pieces of text of 15-20 characters long. The algorithm is evaluated against recent approaches using existing test sets from previous works on South African languages as well as the Discriminating between Similar Languages (DSL) 2015 and 2017 shared tasks.\nSection SECREF2 reviews existing works on the topic and summarises the remaining research problems. Section SECREF3 of the paper discusses the proposed algorithm and Section SECREF4 presents comparative results.\nRelated Works\nThe focus of this section is on recently published datasets and LID research applicable to the South African context. An in depth survey of algorithms, features, datasets, shared tasks and evaluation methods may be found in BIBREF0.\nThe datasets for the DSL 2015 & DSL 2017 shared tasks BIBREF1 are often used in LID benchmarks and also available on Kaggle . The DSL datasets, like other LID datasets, consists of text sentences labelled by language. The 2017 dataset, for example, contains 14 languages over 6 language groups with 18000 training samples and 1000 testing samples per language.\nThe recently published JW300 parallel corpus BIBREF2 covers over 300 languages with around 100 thousand parallel sentences per language pair on average. In South Africa, a multilingual corpus of academic texts produced by university students with different mother tongues is being developed BIBREF3. The WiLI-2018 benchmark dataset BIBREF4 for monolingual written natural language identification includes around 1000 paragraphs of 235 languages. A possibly useful link can also be made BIBREF5 between Native Language Identification (NLI) (determining the native language of the author of a text) and Language Variety Identification (LVI) (classification of different varieties of a single language) which opens up more datasets. The Leipzig Corpora Collection BIBREF6, the Universal Declaration of Human Rights and Tatoeba are also often used sources of data.\nThe NCHLT text corpora BIBREF7 is likely a good starting point for a shared LID task dataset for the South African languages BIBREF8. The NCHLT text corpora contains enough data to have 3500 training samples and 600 testing samples of 300+ character sentences per language. Researchers have recently started applying existing algorithms for tasks like neural machine translation in earnest to such South African language datasets BIBREF9.\nExisting NLP datasets, models and services BIBREF10 are available for South African languages. These include an LID algorithm BIBREF11 that uses a character level n-gram language model. Multiple papers have shown that 'shallow' naive Bayes classifiers BIBREF12, BIBREF8, BIBREF13, BIBREF14, SVMs BIBREF15 and similar models work very well for doing LID. The DSL 2017 paper BIBREF1, for example, gives an overview of the solutions of all of the teams that competed on the shared task and the winning approach BIBREF16 used an SVM with character n-gram, parts of speech tag features and some other engineered features. The winning approach for DSL 2015 used an ensemble naive Bayes classifier. The fasttext classifier BIBREF17 is perhaps one of the best known efficient 'shallow' text classifiers that have been used for LID .\nMultiple papers have proposed hierarchical stacked classifiers (including lexicons) that would for example first classify a piece of text by language group and then by exact language BIBREF18, BIBREF19, BIBREF8, BIBREF0. Some work has also been done on classifying surnames between Tshivenda, Xitsonga and Sepedi BIBREF20. Additionally, data augmentation BIBREF21 and adversarial training BIBREF22 approaches are potentially very useful to reduce the requirement for data.\nResearchers have investigated deeper LID models like bidirectional recurrent neural networks BIBREF23 or ensembles of recurrent neural networks BIBREF24. The latter is reported to achieve 95.12% in the DSL 2015 shared task. In these models text features can include character and word n-grams as well as informative character and word-level features learnt BIBREF25 from the training data. The neural methods seem to work well in tasks where more training data is available.\nIn summary, LID of short texts, informal styles and similar languages remains a difficult problem which is actively being researched. Increased confusion can in general be expected between shorter pieces of text and languages that are more closely related. Shallow methods still seem to work well compared to deeper models for LID. Other remaining research opportunities seem to be data harvesting, building standardised datasets and creating shared tasks for South Africa and Africa. Support for language codes that include more languages seems to be growing and discoverability of research is improving with more survey papers coming out. Paywalls also seem to no longer be a problem; the references used in this paper was either openly published or available as preprint papers.\nMethodology\nThe proposed LID algorithm builds on the work in BIBREF8 and BIBREF26. We apply a naive Bayesian classifier with character (2, 4 & 6)-grams, word unigram and word bigram features with a hierarchical lexicon based classifier.\nThe naive Bayesian classifier is trained to predict the specific language label of a piece of text, but used to first classify text as belonging to either the Nguni family, the Sotho family, English, Afrikaans, Xitsonga or Tshivenda. The scikit-learn multinomial naive Bayes classifier is used for the implementation with an alpha smoothing value of 0.01 and hashed text features.\nThe lexicon based classifier is then used to predict the specific language within a language group. For the South African languages this is done for the Nguni and Sotho groups. If the lexicon prediction of the specific language has high confidence then its result is used as the final label else the naive Bayesian classifier's specific language prediction is used as the final result. The lexicon is built over all the data and therefore includes the vocabulary from both the training and testing sets.\nThe lexicon based classifier is designed to trade higher precision for lower recall. The proposed implementation is considered confident if the number of words from the winning language is at least one more than the number of words considered to be from the language scored in second place.\nThe stacked classifier is tested against three public LID implementations BIBREF17, BIBREF23, BIBREF8. The LID implementation described in BIBREF17 is available on GitHub and is trained and tested according to a post on the fasttext blog. Character (5-6)-gram features with 16 dimensional vectors worked the best. The implementation discussed in BIBREF23 is available from https://github.com/tomkocmi/LanideNN. Following the instructions for an OSX pip install of an old r0.8 release of TensorFlow, the LanideNN code could be executed in Python 3.7.4. Settings were left at their defaults and a learning rate of 0.001 was used followed by a refinement with learning rate of 0.0001. Only one code modification was applied to return the results from a method that previously just printed to screen. The LID algorithm described in BIBREF8 is also available on GitHub.\nThe stacked classifier is also tested against the results reported for four other algorithms BIBREF16, BIBREF26, BIBREF24, BIBREF15. All the comparisons are done using the NCHLT BIBREF7, DSL 2015 BIBREF19 and DSL 2017 BIBREF1 datasets discussed in Section SECREF2.\nResults and Analysis\nThe average classification accuracy results are summarised in Table TABREF9. The accuracies reported are for classifying a piece of text by its specific language label. Classifying text only by language group or family is a much easier task as reported in BIBREF8.\nDifferent variations of the proposed classifier were evaluated. A single NB classifier (NB), a stack of two NB classifiers (NB+NB), a stack of a NB classifier and lexicon (NB+Lex) and a lexicon (Lex) by itself. A lexicon with a 50% training token dropout is also listed to show the impact of the lexicon support on the accuracy.\nFrom the results it seems that the DSL 2017 task might be harder than the DSL 2015 and NCHLT tasks. Also, the results for the implementation discussed in BIBREF23 might seem low, but the results reported in that paper is generated on longer pieces of text so lower scores on the shorter pieces of text derived from the NCHLT corpora is expected.\nThe accuracy of the proposed algorithm seems to be dependent on the support of the lexicon. Without a good lexicon a non-stacked naive Bayesian classifier might even perform better.\nThe execution performance of some of the LID implementations are shown in Table TABREF10. Results were generated on an early 2015 13-inch Retina MacBook Pro with a 2.9 GHz CPU (Turbo Boosted to 3.4 GHz) and 8GB RAM. The C++ implementation in BIBREF17 is the fastest. The implementation in BIBREF8 makes use of un-hashed feature representations which causes it to be slower than the proposed sklearn implementation. The execution performance of BIBREF23 might improve by a factor of five to ten when executed on a GPU.\nConclusion\nLID of short texts, informal styles and similar languages remains a difficult problem which is actively being researched. The proposed algorithm was evaluated on three existing datasets and compared to the implementations of three public LID implementations as well as to reported results of four other algorithms. It performed well relative to the other methods beating their results. However, the performance is dependent on the support of the lexicon.\nWe would like to investigate the value of a lexicon in a production system and how to possibly maintain it using self-supervised learning. We are investigating the application of deeper language models some of which have been used in more recent DSL shared tasks. We would also like to investigate data augmentation strategies to reduce the amount of training data that is required.\nFurther research opportunities include data harvesting, building standardised datasets and shared tasks for South Africa as well as the rest of Africa. In general, the support for language codes that include more languages seems to be growing, discoverability of research is improving and paywalls seem to no longer be a big problem in getting access to published research.",6 "char_len": 11930,7 "document_hash": "1200dfbaf212cc5b9f2feabc2e0f4d857b6cca56cbc761e5f0d2467fa3c67e5f",8 "questions": [9 {10 "question_id": "q00",11 "qtype": "qasper",12 "question": "Which languages are similar to each other?"13 }14 ]15}