| { | |
| "paper_id": "R15-1049", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T14:56:44.877028Z" | |
| }, | |
| "title": "Predicting the Level of Text Standardness in User-generated Content", | |
| "authors": [ | |
| { | |
| "first": "Nikola", | |
| "middle": [], | |
| "last": "Ljube\u0161i\u0107", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Jo\u017eef Stefan Institute", | |
| "location": {} | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Darja", | |
| "middle": [], | |
| "last": "Fi\u0161er", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "University of Ljubljana", | |
| "location": {} | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Toma\u017e", | |
| "middle": [], | |
| "last": "Erjavec", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Jo\u017eef Stefan Institute", | |
| "location": {} | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Dafne", | |
| "middle": [], | |
| "last": "Marko", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "University of Ljubljana", | |
| "location": {} | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Senja", | |
| "middle": [], | |
| "last": "Pollak", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Jo\u017eef Stefan Institute", | |
| "location": {} | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Iza", | |
| "middle": [], | |
| "last": "\u0160krjanec", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "University of Ljubljana", | |
| "location": {} | |
| }, | |
| "email": "" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "Non-standard language as it appears in user-generated content has recently attracted much attention. This paper proposes that non-standardness comes in two basic varieties, technical and linguistic, and develops a machine-learning method to discriminate between standard and nonstandard texts in these two dimensions. We describe the manual annotation of a dataset of Slovene user-generated content and the features used to build our regression models. We evaluate and discuss the results, where the mean absolute error of the best performing method on a three-point scale is 0.38 for technical and 0.42 for linguistic standardness prediction. Even when using no language-dependent information sources, our predictor still outperforms an OOVratio baseline by a wide margin. In addition, we show that very little manually annotated training data is required to perform good prediction. Predicting standardness can help decide when to attempt to normalise the data to achieve better annotation results with standard tools, and provide linguists who are interested in nonstandard language with a simple way of selecting only such texts for their research.", | |
| "pdf_parse": { | |
| "paper_id": "R15-1049", | |
| "_pdf_hash": "", | |
| "abstract": [ | |
| { | |
| "text": "Non-standard language as it appears in user-generated content has recently attracted much attention. This paper proposes that non-standardness comes in two basic varieties, technical and linguistic, and develops a machine-learning method to discriminate between standard and nonstandard texts in these two dimensions. We describe the manual annotation of a dataset of Slovene user-generated content and the features used to build our regression models. We evaluate and discuss the results, where the mean absolute error of the best performing method on a three-point scale is 0.38 for technical and 0.42 for linguistic standardness prediction. Even when using no language-dependent information sources, our predictor still outperforms an OOVratio baseline by a wide margin. In addition, we show that very little manually annotated training data is required to perform good prediction. Predicting standardness can help decide when to attempt to normalise the data to achieve better annotation results with standard tools, and provide linguists who are interested in nonstandard language with a simple way of selecting only such texts for their research.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Abstract", | |
| "sec_num": null | |
| } | |
| ], | |
| "body_text": [ | |
| { | |
| "text": "User-generated content (UGC) is becoming an increasingly frequent and important source of hu-man knowledge and people's opinions (Crystal, 2011) . Language use in social media is characterised by special technical and social circumstances, and as such deviates from the norm of traditional text production. Researching the language of social media is not only of great value to (socio)linguists, but also beneficial for improving automatic processing of UGC, which has proven to be quite difficult (Sproat, 2001) . Consistent decreases in performance on noisy texts have been recorded in the entire text processing chain, from PoS-tagging, where the state-of-the-art Stanford tagger achieves 97% accuracy on Wall Street Journal texts, but only 85% accuracy on Twitter data (Gimpel et al., 2011) , to parsing, where doubledigit decreases in accuracy have been recorded for 4 state-of-the-art parsers on social media texts (Petrov and McDonald, 2012) .", | |
| "cite_spans": [ | |
| { | |
| "start": 129, | |
| "end": 144, | |
| "text": "(Crystal, 2011)", | |
| "ref_id": "BIBREF1" | |
| }, | |
| { | |
| "start": 498, | |
| "end": 512, | |
| "text": "(Sproat, 2001)", | |
| "ref_id": "BIBREF12" | |
| }, | |
| { | |
| "start": 773, | |
| "end": 794, | |
| "text": "(Gimpel et al., 2011)", | |
| "ref_id": "BIBREF4" | |
| }, | |
| { | |
| "start": 921, | |
| "end": 948, | |
| "text": "(Petrov and McDonald, 2012)", | |
| "ref_id": "BIBREF11" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "Non-standard linguistic features have been analysed both qualitatively and quantitatively (Eisenstein, 2013; Hu et al., 2013; Baldwin et al., 2013) and they have been taken into account in automatic text processing applications, which either strive to normalise non-standard features before submitting them to standard text processing tools (Han et al., 2012) , adapt standard processing tools to work on non-standard data (Gimpel et al., 2011) or, in task-oriented applications, use a series of simple pre-processing steps to tackle the most frequent UGC-specific phenomena (Foster et al., 2011) .", | |
| "cite_spans": [ | |
| { | |
| "start": 90, | |
| "end": 108, | |
| "text": "(Eisenstein, 2013;", | |
| "ref_id": "BIBREF2" | |
| }, | |
| { | |
| "start": 109, | |
| "end": 125, | |
| "text": "Hu et al., 2013;", | |
| "ref_id": "BIBREF6" | |
| }, | |
| { | |
| "start": 126, | |
| "end": 147, | |
| "text": "Baldwin et al., 2013)", | |
| "ref_id": "BIBREF0" | |
| }, | |
| { | |
| "start": 341, | |
| "end": 359, | |
| "text": "(Han et al., 2012)", | |
| "ref_id": "BIBREF5" | |
| }, | |
| { | |
| "start": 423, | |
| "end": 444, | |
| "text": "(Gimpel et al., 2011)", | |
| "ref_id": "BIBREF4" | |
| }, | |
| { | |
| "start": 575, | |
| "end": 596, | |
| "text": "(Foster et al., 2011)", | |
| "ref_id": "BIBREF3" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "However, to the best of our knowledge, the level of (non-)standardness of UGC has not yet been measured to improve the corpus pre-processing pipeline or added to the corpus as an annotation layer in comprehensive corpus-linguistic analyses. In this paper, we present an experiment in which we manually annotated and analysed the (non-)standardness level of Slovene tweets, forum messages and news comments. The findings were then used to train a regression model that automatically predicts the level of text standardness in the entire corpus. We believe this information will be highly useful in linguistic analyses as well as in all stages of text processing, from more accurate sampling to build representative corpora to choosing the best tools for processing the collected documents, either with tools trained on standard language or with tools specially adapted for nonstandard language varieties.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "The paper is organised as follows. Section 2 presents the dataset. Section 3 introduces the features used in subsequent experiments, while Section 4 describes the actual experiments and their results, with an emphasis on feature evaluation, the gain when using external resources, and an analysis of the performance on specific subcorpora. The paper concludes with a discussion of the results and plans for future work.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "This section presents the dataset used in subsequent experiments, starting with our corpus of user-generated Slovene and the sampling used to extract the dataset for manual annotation. We then explain the motivation behind having two dimensions of standardness, and describe the process of manual dataset annotation.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The Dataset", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "The dataset for the reported experiments is taken from our corpus of user-generated Slovene, which currently contains three types of text: tweets, forum posts, and news site comments. The complete corpus contains just over 120 million tokens.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The Corpus of User-generated Slovene", | |
| "sec_num": "2.1" | |
| }, | |
| { | |
| "text": "Tweets were collected with the TweetCaT tool (Ljube\u0161i\u0107 et al., 2014b) , which was constructed specifically for compiling Twitter corpora of smaller languages. The tool uses the Twitter API and a small lexicon of language specific Slovene words to first identify the users that predominantly tweet in Slovene, as well as their friends and followers. TweetCaT continuously collected the users' tweets for a period of almost two years, also updating the list of users. This resulted in the Slovene tweet subcorpus, which contains 61 million tokens. Currently, most of the collected tweets were written between 2013 and 2014. It should be noted that the majority of these tweets did not turn out to be user-generated content, but rather news feeds, advertisements, and similar material produced by professional authors.", | |
| "cite_spans": [ | |
| { | |
| "start": 45, | |
| "end": 69, | |
| "text": "(Ljube\u0161i\u0107 et al., 2014b)", | |
| "ref_id": "BIBREF9" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The Corpus of User-generated Slovene", | |
| "sec_num": "2.1" | |
| }, | |
| { | |
| "text": "For forum posts and news site comments, six popular Slovene sources were chosen as they were the most widely used and contained the most texts. The selected forums focus on the topics of motoring, health, and science, respectively. The selected news sites pertain to the national Slovene broadcaster RTV Slovenija, and the most popular left-wing and right-wing weekly magazines. Because the crawled pages differ in terms of structure, separate text extractors were developed using the Beautiful Soup 1 module, which enables writing targeted structure extractors from HTML documents. This allowed us to avoid compromising corpus content with large amounts of noise typically present in these types of sources, e.g. adverts and irrelevant links. It also enabled us to structure the texts and extract relevant metadata from them.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The Corpus of User-generated Slovene", | |
| "sec_num": "2.1" | |
| }, | |
| { | |
| "text": "The forum posts contribute 47 million tokens to the corpus, while the news site comments amount to 15 million tokens. As with tweets, the majority of the collected comments were posted between 2013 and 2014. The forum posts cover a wider time span, with similar portions of text coming from each of the years between 2006 and 2014.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The Corpus of User-generated Slovene", | |
| "sec_num": "2.1" | |
| }, | |
| { | |
| "text": "The corpus is also automatically annotated. The texts were first tokenised and the word tokens normalised (standardised) using the method of Ljube\u0161i\u0107 et al. (2014a) , which employs characterbased statistical machine translation. The CSMT translation model was trained on 1000 keywords taken from the Slovene tweet corpus (compared to a corpus of standard Slovene) and their manually determined standard equivalents. Then, using the models for standard Slovene the standardised word tokens were PoS-tagged and lemmatised.", | |
| "cite_spans": [ | |
| { | |
| "start": 141, | |
| "end": 164, | |
| "text": "Ljube\u0161i\u0107 et al. (2014a)", | |
| "ref_id": "BIBREF8" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The Corpus of User-generated Slovene", | |
| "sec_num": "2.1" | |
| }, | |
| { | |
| "text": "For the experiments reported in this paper, we constructed a dataset containing individual texts that were semi-randomly sampled from the corpus of tweets, forum posts and comments. The dataset was then manually annotated.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Samples for Manual Annotation", | |
| "sec_num": "2.2" | |
| }, | |
| { | |
| "text": "To guarantee a balanced dataset, we selected equal proportions (one third) of texts for each text type. For forum posts and comments, we included equal proportions of each of their six sources. In order to obtain a balanced dataset in terms of language (non-)standardness from a corpus heavily skewed towards standard language, we used a heuristic to roughly estimate the degree of text (non-)standardness, which makes use of the corpus normalisation procedure. For each text, we computed the ratio between the number of word tokens that have been changed by the automatic normalisation, and its overall length in words. If this ratio was 0.1 or less, the text was considered as standard, otherwise it was considered as non-standard. The dataset was then constructed so that it contained an equal number of \"standard\" and \"non-standard\" texts. It should be noted that this is only an estimate, and that the presented method does not depend on an exact balance. Different rough measures of standardness could also be taken, e.g. a simple ratio of out-of-vocabulary words to all words, given a lexicon of standard word forms.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Samples for Manual Annotation", | |
| "sec_num": "2.2" | |
| }, | |
| { | |
| "text": "It is far from easy to tell how \"standard\" a certain text is. While it could be regarded as a single dimension of a text (as is usually the case with e.g. sentiment annotation), standardness turns out to comprise a very disparate set of features. For example, some authors use standard spelling, but no capital letters. Others make many typos, while some will typeset their text in a standard fashion, but use colloquial or dialectal lexis and morphology.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Dimensions of Standardness", | |
| "sec_num": "2.3" | |
| }, | |
| { | |
| "text": "To strike a balance between the adequacy and the complexity of the annotation, we decided to use two dimensions of standardness: technical and linguistic. The score for technical text standardness focuses on word capitalisation, the use of punctuation, and the presence of typos or repeated characters in the words. The score for linguistic standardness, on the other hand, takes into account the knowledge of the language by the authors and their more or less conscious decisions to use nonstandard language, involving spelling, lexis, morphology, and word order.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Dimensions of Standardness", | |
| "sec_num": "2.3" | |
| }, | |
| { | |
| "text": "These two dimensions are meant to be straightforward enough to be applied by the annotators, informative enough for NLP tools to appropriately apply possibly different normalisation meth-ods, and relevant enough to linguists for filtering relevant texts when researching non-standard language.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Dimensions of Standardness", | |
| "sec_num": "2.3" | |
| }, | |
| { | |
| "text": "The annotators, who were postgraduate students of linguistics, were presented with annotation guidelines and criteria for annotating the two dimensions of non-standardness. Each given text was to be annotated in terms of both dimensions using a score between 1 (standard) and 3 (very non-standard), with 2 marking slightly nonstandard texts. We used a three-score system as the task is not (and can hardly be) very precisely defined. Using this scale would also allow us to better observe inter-annotator agreement. In addition, for learning standardness models, we used a regression model, which returns a degree of standardness, rather than a classification one. In this particular case, a slightly more fine-grained scoring is beneficial.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Manual Annotation and Resulting Dataset", | |
| "sec_num": "2.4" | |
| }, | |
| { | |
| "text": "To give an idea of the types of features taken into account for each dimension, two examples of short texts are presented below:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Manual Annotation and Resulting Dataset", | |
| "sec_num": "2.4" | |
| }, | |
| { | |
| "text": "\u2022 T=1 / L=3", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Manual Annotation and Resulting Dataset", | |
| "sec_num": "2.4" | |
| }, | |
| { | |
| "text": "Original: Ma men se zdi tole s poimenovanji oz s poslovenjenjem imen mest\u010dist mem. Standardised: Meni se zdi to s poimenovanji oz. s poslovenjenjem imen mest\u010disto mimo. English: To-me Refl. it-seems this with naming i.e. with making-into-Slovene names ofcities completely wrong. Differences: Colloquial particle (\"ma\"), colloquial form of pronoun (\"tole\" vs. \"to\"), phonetic transcription of dialectal word forms (\"men\" vs. \"meni\", \"\u010dist\" vs. \"\u010disto\", \"mem\" vs. \"mimo\")", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Manual Annotation and Resulting Dataset", | |
| "sec_num": "2.4" | |
| }, | |
| { | |
| "text": "\u2022 T=3 / L=1 Original: se pravi,da predvideva\u0161 razveljavitev Standardised: Se pravi, da predvideva\u0161 razveljavitev? English: Refl. this-means, that you-foresee annuling? Differences: No capital letter at the start of sentence, no space after the comma, no sentence-final punctuation.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Manual Annotation and Resulting Dataset", | |
| "sec_num": "2.4" | |
| }, | |
| { | |
| "text": "The annotators were told to mark with 0 those texts that were out of scope for the experiment, e.g. if they were written in a foreign language, automatically generated (such as news or advert lead-ins in tweets) or if they contained no linguistic material (e.g. only URLs, hashtags, and emoticons). These texts were then not included in the manually annotated dataset.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Manual Annotation and Resulting Dataset", | |
| "sec_num": "2.4" | |
| }, | |
| { | |
| "text": "After a training session in which a small set of texts was annotated and discussed by all annotators, the experimental data was annotated in two campaigns. A first batch of 904 text instances was annotated, each by a single annotator, and was subsequently used as the development data in our experiments. For the second batch, each of 402 text instances was annotated by two annotators. In 8 of these instances, the difference between the annotations made by separate annotators in at least one dimension was two. This means that the first annotator marked a text as standard in at least one dimension, while the other marked it as very nonstandard. This is why these data points were removed from the dataset, leaving 394 instances that constituted the testing set for the experiments. The response variables for the experiments were computed as the average of the values given by two annotators.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Manual Annotation and Resulting Dataset", | |
| "sec_num": "2.4" | |
| }, | |
| { | |
| "text": "We defined 29 features to describe the technical and linguistic text properties. The features can be grouped in two main categories. Characterbased features (listed in Table 1 and described in 3.1) concern the incorrect use of punctuation and spaces, character repetition, the ratio of alphabetic vs. non-alphabetic characters, vowels vs. consonants, etc. Token-based features (listed in Table 2 and described in 3.2) describe word properties. Some are very general, e.g. proportions of very short words, capitalised words, etc., while others depend on the use of language-specific lexicons and mostly compute the proportion of words not included in these lexicons.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 168, | |
| "end": 175, | |
| "text": "Table 1", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 388, | |
| "end": 396, | |
| "text": "Table 2", | |
| "ref_id": "TABREF1" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "The Feature Space", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "This category contains features dealing either with the use of punctuation and brackets or the use of alphanumeric characters.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Character-based Features", | |
| "sec_num": "3.1" | |
| }, | |
| { | |
| "text": "In terms of punctuation and brackets, we calculate the ratio of punctuation compared to all characters, ratio of paragraphs ending with an end-ofsentence punctuation sign, and the ratio of spaces preceding or not following a punctuation sign. Similarly, we calculate the ratio of opening or closing brackets that are preceded and followed by spaces.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Character-based Features", | |
| "sec_num": "3.1" | |
| }, | |
| { | |
| "text": "For the alphanumeric characters, we calculate the ratios of alphabetic and alphanumeric characters in the text, the ratio of uppercase letters, and the ratios of sentences and paragraphs starting with an uppercase letter. One feature is based on the ratio between vowels and consonants in the text, while another encodes the ratio of characters from the Slovene alphabet. Two other features are based on repeating characters, one covering any character, the other focusing on alphabetic characters only. ", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Character-based Features", | |
| "sec_num": "3.1" | |
| }, | |
| { | |
| "text": "In this category, we discriminate between stringbased and lexicon-based features, the latter being dependent on external data sources.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Token-based Features", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "In terms of string-based features, we compute the ratio of title-case and upper-case words, as well as word repetitions. Another feature is the ratio of words composed only of consonants. We also consider the ratio of very short words.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Token-based Features", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "A large part of lexicon-based features uses the Sloleks lexicon 2 (Krek and Erjavec, 2009) , consisting of Slovene words with all their word forms. The lexicon consists of 961,040 forms since Slovene is a morphologically rich language and each lexeme has many possible word forms.", | |
| "cite_spans": [ | |
| { | |
| "start": 66, | |
| "end": 90, | |
| "text": "(Krek and Erjavec, 2009)", | |
| "ref_id": "BIBREF7" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Token-based Features", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "The features based on this resource are the ratio of out-of-vocabulary (OOV) words (sloleks), the ratio of words that are OOVs, but are missing a vowel character (sloleks_vowel), the ratio of short words that are OOVs (sloleks_short), and the number of lower-case forms covered by a title-or upper-case entry in the lexicon only (sloleks_names).", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Token-based Features", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "We experimented with another source of lexical information -the KRES balanced corpus of standard Slovene (Logar Berginc et al., 2012) . We produced two lexicons from the corpus, one consisting of all letter-only tokens occurring at least ten times (70,249 entries, kresleks_10), and the other with the frequency threshold of 100 (4,339 entries, kresleks_100). We used both resources to calculate OOV ratios.", | |
| "cite_spans": [ | |
| { | |
| "start": 105, | |
| "end": 133, | |
| "text": "(Logar Berginc et al., 2012)", | |
| "ref_id": "BIBREF10" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Token-based Features", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "Finally, we used a very small lexical resource of 195 most frequent non-standard forms of Slovene (nonstdlex). We produced this resource by calculating the log-likelihood statistic of each token from our corpus with respect to its frequency in the KRES corpus. We manually inspected the 250 highest-ranked tokens, cleaning out 55 irrelevant entries.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Token-based Features", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "In this section, we describe our regressor optimisation and evaluation, the analysis of feature coefficients, the dependence on external information sources, the learning curve of the problem, and the independence from the text genre.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Experiments and Results", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "In the first set of experiments, we used the development set to perform grid search hyperparameter optimisation via 10-fold cross-validation on the SVR regressor using an RBF kernel. As our scoring function throughout the paper, we use the mean absolute error as it is more resistant to outliers than the mean squared error, and is also easier to interpret.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Regressor Optimisation", | |
| "sec_num": "4.1" | |
| }, | |
| { | |
| "text": "The results obtained from the optimised regressor, presented in Table 3 , showed that the task of predicting technical standardness is simpler than that of predicting linguistic standardness, which was expected.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 64, | |
| "end": 71, | |
| "text": "Table 3", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Regressor Optimisation", | |
| "sec_num": "4.1" | |
| }, | |
| { | |
| "text": "Dimension Mean absolute error technical 0.451 \u00b1 0.033 linguistic 0.544 \u00b1 0.033 Table 3 : Results obtained from the dev set", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 79, | |
| "end": 86, | |
| "text": "Table 3", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Regressor Optimisation", | |
| "sec_num": "4.1" | |
| }, | |
| { | |
| "text": "Once we had optimised our hyperparameters on the development set, we performed an evaluation of the system on our test set. Given that the test set was double-annotated, with opposite-label instances removed and neighbouring labels averaged, we expected a lower level of error compared to the development data.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Test Set Evaluation", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "In Table 4 we compare our system with multiple baselines. The first two baselines (baseline_linear and baseline_SVR) are supervised equivalents to what researchers mostly use in practice -the OOV ratio heuristic. Those baselines use only one feature -the OOV ratio on the Sloleks lexicon (sloleks). The first baseline (baseline_linear) is algorithmically simpler as it uses linear regression, thereby linearly mapping the [0-1] range of the OOV ratio heuristic to the expected [1, 3] range of our response variables. The second baseline (base-line_SVR) uses SVR with an RBF kernel.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 3, | |
| "end": 10, | |
| "text": "Table 4", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Test Set Evaluation", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "The last two baselines are random baselines that produce random numbers in the [1,3] range. The baseline_test was evaluated on our test set and the baseline_theoretical was evaluated on another randomly generated sequence of values in the [1, 3] range. Both baselines were evaluated on drastically longer test sets (either by repeating our test set or by generating longer sequences of random numbers) to produce accurate estimates.", | |
| "cite_spans": [ | |
| { | |
| "start": 239, | |
| "end": 242, | |
| "text": "[1,", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 243, | |
| "end": 245, | |
| "text": "3]", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Test Set Evaluation", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "We can observe that the mean absolute error of our final system did, as expected, go down in comparison to the 10-folding result obtained on the development data from Table 3 . There are two reasons for this: 1) most incorrect annotations in the testing data were removed, and 2) the 3-level scale was transformed to a 5-level scale. The error level on the technical dimension is still lower compared to the linguistic dimension, although the distance between those two dimensions has shrunk from 0.09 points to 0.05 points.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 167, | |
| "end": 174, | |
| "text": "Table 3", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Test Set Evaluation", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "Comparing our system to baseline_linear on the linguistic dimension shows that using more variables than just the OOV ratio and training a nonlinear regressor does produce a much better system, with an error reduction of more than 0.17 points. When using a non-linear regressor as a baseline (baseline_SVR), the error difference falls to 0.12 points, which argues for using non-linear regressors on this problem.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Test Set Evaluation", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "For the technical dimension, as expected, the OOV ratio heuristic is not optimal, producing, differently than when using all the features, similar or worse results compared to the linguistic dimension.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Test Set Evaluation", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "The two random baselines show that both our system and the OOV baselines are a safe distance away from these weak baselines.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Test Set Evaluation", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "Beside the fact that using multiple features enhances our results, we want to stress that using a supervised system should not be questioned at all, since the output of heuristics such as the OOV ra-tio is very hard to interpret by the final corpus user, in contrast to the [1, 3] Table 4 : Final evaluation and comparison with the baselines via mean absolute error.", | |
| "cite_spans": [ | |
| { | |
| "start": 274, | |
| "end": 277, | |
| "text": "[1,", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 278, | |
| "end": 280, | |
| "text": "3]", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 281, | |
| "end": 288, | |
| "text": "Table 4", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Test Set Evaluation", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "Our next experiment focused on the usefulness of specific features by training a linear kernel SVR on standardised data and analysing its coefficients. We thereby inspected which variables demonstrate the highest prediction strength for each of our two dimensions. For the technical dimension, the most prominent features are the ratio of alphabetic characters, the number of character repetitions, the ratio of upper-case characters and the ratio of spaces after punctuation.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Feature Coefficients", | |
| "sec_num": "4.3" | |
| }, | |
| { | |
| "text": "On the other hand, for the linguistic dimension, the most prominent features are the OOV rate given a standard lexicon (sloleks), the OOV rate given a lexicon of non-standard forms (nonstdlex), and the ratio of short tokens.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Feature Coefficients", | |
| "sec_num": "4.3" | |
| }, | |
| { | |
| "text": "As expected, for the technical dimension, character-based features are of greater importance. As for the linguistic dimension, token-based features, especially the lexicon-based ones, carry more weight.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Feature Coefficients", | |
| "sec_num": "4.3" | |
| }, | |
| { | |
| "text": "Our next experiment looked into how much information is obtained from external information sources used in lexicon-based features. With this experiment, we wanted to measure our dependence on these resources and the level of prediction quality one can expect if some of those resources are not available.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "External Information Sources", | |
| "sec_num": "4.4" | |
| }, | |
| { | |
| "text": "The results obtained with information sources that influence prediction quality the most are given in Table 5 . While the technical prediction in general does not suffer a lot when removing external information sources, the linguistic one does suffer an 0.11-point increase in error when all the resource-dependent features are removed (none). The most informative resource is the lexicon of standard language (sloleks), yielding a 0.07-out-of-0.11-points error reduction. Producing a lexicon from a corpus of standard language (kresleks_100) does not come close to that, producing just a 0.02-point error reduction. While the small lexicon of non-standard forms (nonstdlex) does reduce the error by 0.06 points, using all standard-lexicon-related features (sloleks_all) comes 0.03 points closer to the final result obtained with all features (all). Table 5 : Dependence of prediction quality on external information sources", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 102, | |
| "end": 109, | |
| "text": "Table 5", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 850, | |
| "end": 857, | |
| "text": "Table 5", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "External Information Sources", | |
| "sec_num": "4.4" | |
| }, | |
| { | |
| "text": "This set of experiments focused on the impact of the amount of data available for training on our prediction quality. We compared three predictors: the technical one, the linguistic one and the linguistic one without using any external information sources. The three learning curves are depicted in Figure 1 . They show that useful results can be obtained with just a few hundred annotated instances. The learning of the technical and linguistic dimensions seem to be equally hard when there are up to 100 instances available for learning, the technical dimension taking off after that. The linguistic dimension is obviously harder to learn when external information sources are not used. Both learning curves seem to be parallel, showing that a larger amount of training data, at least for the features defined, cannot compensate for the lack of external knowledge.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 299, | |
| "end": 307, | |
| "text": "Figure 1", | |
| "ref_id": "FIGREF2" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Learning Curve", | |
| "sec_num": "4.5" | |
| }, | |
| { | |
| "text": "Our final experiment focused on the dependence of the results on the genre of the training and testing data. We took into consideration the three genres present in the corpus: tweets, news comments and forum posts. On each side, training and testing, we experimented with using either data from Table 6 : Impact of the genre of training (rows) and testing data (columns)", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 295, | |
| "end": 302, | |
| "text": "Table 6", | |
| "ref_id": "TABREF4" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Genre Dependence", | |
| "sec_num": "4.6" | |
| }, | |
| { | |
| "text": "In the technical dimension, we observe best results when training and testing data comes from the same genre. There are no significant differences between the genres.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Genre Dependence", | |
| "sec_num": "4.6" | |
| }, | |
| { | |
| "text": "In the linguistic dimension, Twitter data proves to be easiest to perform prediction on, and forum data the most complicated. Interestingly, the best predictions are not made if training data comes from the same genre, but if all genres are combined.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Genre Dependence", | |
| "sec_num": "4.6" | |
| }, | |
| { | |
| "text": "In this paper, we presented a supervised-learning approach to predicting the text standardness level. While we differentiated between two dimensions of standardness, the technical and the linguistic one, we explained both with the same 29 features, most of which were independent from external information sources.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusion", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "We showed that we outperform the supervised baselines that rely on the traditionally used OOV ratio only. We outperformed those baselines even when not using any external information sources, which makes our predictor highly language-independent.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusion", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "Both predictors outperformed the supervised baselines when only a few tens of instances are available for training. Most of the learning is performed on the first 500 instances. This makes building the training set for a language an easy task that can be completed in day or two.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusion", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "While the single most informative external information source was the lexicon of standard language, adding information from very small lexicons of frequent non-standard forms or from automatically transformed lexicons of standard language significantly improved the results.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusion", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "Finally, we showed that the predictors are, in general, genre-independent. The technical dimension is slightly more genre-dependent than the linguistic one. While predicting linguistic standardness on tweets is the simplest task, predicting the same on forums proves to be much more difficult.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusion", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "Future work includes applying more transformations to the lexicon of standard language than just vowel dropping, inspecting the language independence of features without relying on manually annotated data in the target language, and using lexical information from the training data to improve prediction.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusion", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "http://www.crummy.com/software/ BeautifulSoup/", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Sloleks is available under the CC BY-NC-SA license at http://www.slovenscina.eu/.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| } | |
| ], | |
| "back_matter": [ | |
| { | |
| "text": "The work described in this paper was funded by the Slovenian Research Agency, project J6-6842 and by the European Fund for Regional Development 2007 -2013 .", | |
| "cite_spans": [ | |
| { | |
| "start": 149, | |
| "end": 154, | |
| "text": "-2013", | |
| "ref_id": "BIBREF2" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Acknowledgments", | |
| "sec_num": null | |
| } | |
| ], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "title": "How Noisy Social Media Text, How Diffrnt Social Media Sources", | |
| "authors": [ | |
| { | |
| "first": "Timothy", | |
| "middle": [], | |
| "last": "Baldwin", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Paul", | |
| "middle": [], | |
| "last": "Cook", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Marco", | |
| "middle": [], | |
| "last": "Lui", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Andrew", | |
| "middle": [], | |
| "last": "Mackinlay", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Li", | |
| "middle": [], | |
| "last": "Wang", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2013, | |
| "venue": "Sixth Intl. Joint Conference on NLP", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "356--364", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Timothy Baldwin, Paul Cook, Marco Lui, Andrew MacKinlay, and Li Wang. 2013. How Noisy Social Media Text, How Diffrnt Social Media Sources. In Sixth Intl. Joint Conference on NLP, pages 356-364.", | |
| "links": null | |
| }, | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "title": "Internet Linguistics: A Student Guide. Routledge", | |
| "authors": [ | |
| { | |
| "first": "David", | |
| "middle": [], | |
| "last": "Crystal", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2011, | |
| "venue": "", | |
| "volume": "10001", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "David Crystal. 2011. Internet Linguistics: A Student Guide. Routledge, New York, NY, 10001, 1st edi- tion.", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "What to Do About Bad Language on the Internet", | |
| "authors": [ | |
| { | |
| "first": "Jacob", | |
| "middle": [], | |
| "last": "Eisenstein", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2013, | |
| "venue": "NAACL-HLT", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "359--369", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Jacob Eisenstein. 2013. What to Do About Bad Lan- guage on the Internet. In NAACL-HLT, pages 359- 369. ACL, June.", | |
| "links": null | |
| }, | |
| "BIBREF3": { | |
| "ref_id": "b3", | |
| "title": "From News to Comment: Resources and Benchmarks for Parsing the Language of Web 2.0", | |
| "authors": [ | |
| { | |
| "first": "Jennifer", | |
| "middle": [], | |
| "last": "Foster", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Ozlem", | |
| "middle": [], | |
| "last": "Cetinoglu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Joachim", | |
| "middle": [], | |
| "last": "Wagner", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Joseph", | |
| "middle": [ | |
| "Le" | |
| ], | |
| "last": "Roux", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Joakim", | |
| "middle": [], | |
| "last": "Nivre", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Deirdre", | |
| "middle": [], | |
| "last": "Hogan", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Josef", | |
| "middle": [], | |
| "last": "Van Genabith", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2011, | |
| "venue": "IJCNLP", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "893--901", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Jennifer Foster, Ozlem Cetinoglu, Joachim Wagner, Joseph Le Roux, Joakim Nivre, Deirdre Hogan, and Josef van Genabith. 2011. From News to Com- ment: Resources and Benchmarks for Parsing the Language of Web 2.0. In IJCNLP, pages 893-901, Chiang Mai, Thailand. Asian Federation of NLP.", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "title": "Part-of-Speech Tagging for Twitter: Annotation, Features, and Experiments", | |
| "authors": [ | |
| { | |
| "first": "Kevin", | |
| "middle": [], | |
| "last": "Gimpel", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Nathan", | |
| "middle": [], | |
| "last": "Schneider", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "O'", | |
| "middle": [], | |
| "last": "Brendan", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Dipanjan", | |
| "middle": [], | |
| "last": "Connor", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Daniel", | |
| "middle": [], | |
| "last": "Das", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Jacob", | |
| "middle": [], | |
| "last": "Mills", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Michael", | |
| "middle": [], | |
| "last": "Eisenstein", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Dani", | |
| "middle": [], | |
| "last": "Heilman", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Jeffrey", | |
| "middle": [], | |
| "last": "Yogatama", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Noah", | |
| "middle": [ | |
| "A" | |
| ], | |
| "last": "Flanigan", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Smith", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2011, | |
| "venue": "ACL (Short Papers)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "42--47", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "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. In ACL (Short Papers), pages 42-47. ACL.", | |
| "links": null | |
| }, | |
| "BIBREF5": { | |
| "ref_id": "b5", | |
| "title": "Automatically Constructing a Normalisation Dictionary for Microblogs", | |
| "authors": [ | |
| { | |
| "first": "Bo", | |
| "middle": [], | |
| "last": "Han", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Paul", | |
| "middle": [], | |
| "last": "Cook", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Timothy", | |
| "middle": [], | |
| "last": "Baldwin", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2012, | |
| "venue": "EMNLP-CoNLL", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "421--432", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Bo Han, Paul Cook, and Timothy Baldwin. 2012. Automatically Constructing a Normalisation Dic- tionary for Microblogs. In EMNLP-CoNLL, pages 421-432, Jeju Island, Korea.", | |
| "links": null | |
| }, | |
| "BIBREF6": { | |
| "ref_id": "b6", | |
| "title": "Dude, srsly?: The surprisingly formal nature of twitter's language", | |
| "authors": [ | |
| { | |
| "first": "Yuheng", | |
| "middle": [], | |
| "last": "Hu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Kartik", | |
| "middle": [], | |
| "last": "Talamadupula", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Subbarao", | |
| "middle": [], | |
| "last": "Kambhampati", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2013, | |
| "venue": "ICWSM", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Yuheng Hu, Kartik Talamadupula, and Subbarao Kambhampati. 2013. Dude, srsly?: The sur- prisingly formal nature of twitter's language. In ICWSM.", | |
| "links": null | |
| }, | |
| "BIBREF7": { | |
| "ref_id": "b7", | |
| "title": "Standardised Encoding of Morphological lexica for Slavic languages", | |
| "authors": [ | |
| { | |
| "first": "Simon", | |
| "middle": [], | |
| "last": "Krek", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Toma\u017e", | |
| "middle": [], | |
| "last": "Erjavec", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2009, | |
| "venue": "MONDILEX Second Open Workshop", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "24--33", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Simon Krek and Toma\u017e Erjavec. 2009. Standardised Encoding of Morphological lexica for Slavic lan- guages. In MONDILEX Second Open Workshop, pages 24-9, Kyiv, Ukraine. National Academy of Sciences of Ukraine.", | |
| "links": null | |
| }, | |
| "BIBREF8": { | |
| "ref_id": "b8", | |
| "title": "Standardizing Tweets with Character-Level Machine Translation", | |
| "authors": [ | |
| { | |
| "first": "Nikola", | |
| "middle": [], | |
| "last": "Ljube\u0161i\u0107", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Toma\u017e", | |
| "middle": [], | |
| "last": "Erjavec", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Darja", | |
| "middle": [], | |
| "last": "Fi\u0161er", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2014, | |
| "venue": "CICLing", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "164--75", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Nikola Ljube\u0161i\u0107, Toma\u017e Erjavec, and Darja Fi\u0161er. 2014a. Standardizing Tweets with Character-Level Machine Translation. In CICLing, Lecture notes in computer science, pages 164-75. Springer.", | |
| "links": null | |
| }, | |
| "BIBREF9": { | |
| "ref_id": "b9", | |
| "title": "TweetCaT: a Tool for Building Twitter Corpora of Smaller Languages", | |
| "authors": [ | |
| { | |
| "first": "Nikola", | |
| "middle": [], | |
| "last": "Ljube\u0161i\u0107", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Darja", | |
| "middle": [], | |
| "last": "Fi\u0161er", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Toma\u017e", | |
| "middle": [], | |
| "last": "Erjavec", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2014, | |
| "venue": "Ninth LREC", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Nikola Ljube\u0161i\u0107, Darja Fi\u0161er, and Toma\u017e Erjavec. 2014b. TweetCaT: a Tool for Building Twitter Cor- pora of Smaller Languages. In Ninth LREC, Reyk- javik. ELRA.", | |
| "links": null | |
| }, | |
| "BIBREF10": { | |
| "ref_id": "b10", | |
| "title": "Korpusi slovenskega jezika Gigafida, KRES, ccGigafida in ccKRES: gradnja, vsebina, uporaba. Zbirka Sporazumevanje. Trojina, zavod za uporabno slovenistiko: Fakulteta za dru\u017ebene vede", | |
| "authors": [ | |
| { | |
| "first": "Miha", | |
| "middle": [], | |
| "last": "Nata\u0161a Logar Berginc", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Marko", | |
| "middle": [], | |
| "last": "Gr\u010dar", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Toma\u017e", | |
| "middle": [], | |
| "last": "Brakus", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Erjavec", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2012, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Nata\u0161a Logar Berginc, Miha Gr\u010dar, Marko Brakus, Toma\u017e Erjavec, \u0160pela Arhar Holdt, and Simon Krek. 2012. Korpusi slovenskega jezika Gigafida, KRES, ccGigafida in ccKRES: gradnja, vsebina, uporaba. Zbirka Sporazumevanje. Trojina, zavod za uporabno slovenistiko: Fakulteta za dru\u017ebene vede, Ljubljana.", | |
| "links": null | |
| }, | |
| "BIBREF11": { | |
| "ref_id": "b11", | |
| "title": "Overview of the 2012 Shared Task on Parsing the Web. First Workshop on Syntactic Analysis of Non-Canonical Language (SANCL)", | |
| "authors": [ | |
| { | |
| "first": "Slav", | |
| "middle": [], | |
| "last": "Petrov", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Ryan", | |
| "middle": [], | |
| "last": "Mcdonald", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2012, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Slav Petrov and Ryan McDonald. 2012. Overview of the 2012 Shared Task on Parsing the Web. First Workshop on Syntactic Analysis of Non-Canonical Language (SANCL), 59.", | |
| "links": null | |
| }, | |
| "BIBREF12": { | |
| "ref_id": "b12", | |
| "title": "Normalization of Non-Standard Words", | |
| "authors": [ | |
| { | |
| "first": "Richard", | |
| "middle": [], | |
| "last": "Sproat", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2001, | |
| "venue": "Computer Speech & Language", | |
| "volume": "15", | |
| "issue": "3", | |
| "pages": "287--333", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Richard Sproat. 2001. Normalization of Non-Standard Words. Computer Speech & Language, 15(3):287- 333, July.", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "FIGREF2": { | |
| "uris": null, | |
| "num": null, | |
| "type_str": "figure", | |
| "text": "Mean absolute error as a function of training data size just one genre or from all genres together. On the training side, we made sure to use the same number of instances in each experiment. The results of the genre dependence experiment are presented in" | |
| }, | |
| "TABREF1": { | |
| "num": null, | |
| "content": "<table/>", | |
| "type_str": "table", | |
| "html": null, | |
| "text": "Overview of token-based features" | |
| }, | |
| "TABREF2": { | |
| "num": null, | |
| "content": "<table><tr><td/><td colspan=\"2\">Technical Linguistic</td></tr><tr><td>final system</td><td>0.377</td><td>0.424</td></tr><tr><td>baseline_linear</td><td>0.594</td><td>0.597</td></tr><tr><td>baseline_SVR</td><td>0.584</td><td>0.548</td></tr><tr><td>baseline_test</td><td>0.713</td><td>0.749</td></tr><tr><td>baseline_theoretical</td><td>0.889</td><td>0.889</td></tr></table>", | |
| "type_str": "table", | |
| "html": null, | |
| "text": "range defined in this paper." | |
| }, | |
| "TABREF4": { | |
| "num": null, | |
| "content": "<table><tr><td/><td/><td>Technical</td></tr><tr><td/><td colspan=\"2\">Tweet Comment Forum</td><td>All</td></tr><tr><td>Tweet</td><td>0.384</td><td colspan=\"2\">0.451 0.485 0.440</td></tr><tr><td colspan=\"2\">Comment 0.519</td><td colspan=\"2\">0.389 0.400 0.437</td></tr><tr><td>Forum</td><td>0.514</td><td colspan=\"2\">0.408 0.370 0.431</td></tr><tr><td>All</td><td>0.426</td><td colspan=\"2\">0.382 0.417 0.409</td></tr><tr><td/><td/><td>Linguistic</td></tr><tr><td/><td colspan=\"2\">Tweet Comment Forum</td><td>All</td></tr><tr><td>Tweet</td><td>0.410</td><td colspan=\"2\">0.452 0.503 0.455</td></tr><tr><td colspan=\"2\">Comment 0.453</td><td colspan=\"2\">0.429 0.510 0.465</td></tr><tr><td>Forum</td><td>0.444</td><td colspan=\"2\">0.458 0.500 0.467</td></tr><tr><td>All</td><td>0.395</td><td colspan=\"2\">0.439 0.507 0.448</td></tr></table>", | |
| "type_str": "table", | |
| "html": null, | |
| "text": "" | |
| } | |
| } | |
| } | |
| } |