{ "paper_id": "N15-1010", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T14:34:54.043515Z" }, "title": "Not All Character N -grams Are Created Equal: A Study in Authorship Attribution", "authors": [ { "first": "Upendra", "middle": [], "last": "Sapkota", "suffix": "", "affiliation": { "laboratory": "", "institution": "The University of Alabama at Birmingham", "location": { "addrLine": "1300 University Boulevard Birmingham", "postCode": "35294", "region": "AL", "country": "USA" } }, "email": "upendra@cis.uab.edu" }, { "first": "Steven", "middle": [], "last": "Bethard", "suffix": "", "affiliation": { "laboratory": "", "institution": "The University of Alabama at Birmingham", "location": { "addrLine": "1300 University Boulevard Birmingham", "postCode": "35294", "region": "AL", "country": "USA" } }, "email": "bethard@cis.uab.edu" }, { "first": "Manuel", "middle": [], "last": "Montes-Y-G\u00f3mez", "suffix": "", "affiliation": {}, "email": "" }, { "first": "Thamar", "middle": [], "last": "Solorio", "suffix": "", "affiliation": { "laboratory": "", "institution": "University of Houston", "location": { "addrLine": "4800 Calhoun Rd Houston, TX", "postCode": "77004", "country": "USA" } }, "email": "solorio@cs.uh.edu" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "Character n-grams have been identified as the most successful feature in both singledomain and cross-domain Authorship Attribution (AA), but the reasons for their discriminative value were not fully understood. We identify subgroups of character n-grams that correspond to linguistic aspects commonly claimed to be covered by these features: morphosyntax, thematic content and style. We evaluate the predictiveness of each of these groups in two AA settings: a single domain setting and a cross-domain setting where multiple topics are present. We demonstrate that character ngrams that capture information about affixes and punctuation account for almost all of the power of character n-grams as features. Our study contributes new insights into the use of n-grams for future AA work and other classification tasks.", "pdf_parse": { "paper_id": "N15-1010", "_pdf_hash": "", "abstract": [ { "text": "Character n-grams have been identified as the most successful feature in both singledomain and cross-domain Authorship Attribution (AA), but the reasons for their discriminative value were not fully understood. We identify subgroups of character n-grams that correspond to linguistic aspects commonly claimed to be covered by these features: morphosyntax, thematic content and style. We evaluate the predictiveness of each of these groups in two AA settings: a single domain setting and a cross-domain setting where multiple topics are present. We demonstrate that character ngrams that capture information about affixes and punctuation account for almost all of the power of character n-grams as features. Our study contributes new insights into the use of n-grams for future AA work and other classification tasks.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "Authorship Attribution (AA) tackles the problem of determining who, among a set of authors, wrote the document at hand. AA has relevant applications ranging from plagiarism detection (Stamatatos, 2011) to Forensic Linguistics, such as identifying authorship of threatening emails or malicious code. Applied areas such as law and journalism can also benefit from authorship attribution, where identifying the true author of a piece of text (such as a ransom note) may help save lives or catch the offenders.", "cite_spans": [ { "start": 183, "end": 201, "text": "(Stamatatos, 2011)", "ref_id": "BIBREF20" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "We know from state of the art research in AA that the length of the documents and the number of po-tential candidate authors have an important effect on the accuracy of AA approaches (Moore, 2001; Luyckx and Daelemans, 2008; Luyckx and Daelemans, 2010) . We can also point out the most common features that have been used successfully in AA work, including: bag-of-words (Madigan et al., 2005; Stamatatos, 2006) , stylistic features (Zheng et al., 2006; Stamatatos et al., 2000) , and word and character level n-grams (Kjell et al., 1994; Peng et al., 2003; Juola, 2006) .", "cite_spans": [ { "start": 183, "end": 196, "text": "(Moore, 2001;", "ref_id": "BIBREF13" }, { "start": 197, "end": 224, "text": "Luyckx and Daelemans, 2008;", "ref_id": "BIBREF10" }, { "start": 225, "end": 252, "text": "Luyckx and Daelemans, 2010)", "ref_id": "BIBREF11" }, { "start": 371, "end": 393, "text": "(Madigan et al., 2005;", "ref_id": "BIBREF12" }, { "start": 394, "end": 411, "text": "Stamatatos, 2006)", "ref_id": "BIBREF18" }, { "start": 433, "end": 453, "text": "(Zheng et al., 2006;", "ref_id": "BIBREF23" }, { "start": 454, "end": 478, "text": "Stamatatos et al., 2000)", "ref_id": "BIBREF17" }, { "start": 518, "end": 538, "text": "(Kjell et al., 1994;", "ref_id": "BIBREF6" }, { "start": 539, "end": 557, "text": "Peng et al., 2003;", "ref_id": "BIBREF14" }, { "start": 558, "end": 570, "text": "Juola, 2006)", "ref_id": "BIBREF2" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "The utility of bag-of-words features is well understood: they effectively capture correlations between authors and topics (Madigan et al., 2005; Kaster et al., 2005) . The discriminative value of these features is thus directly related to the level of content divergence among authors and among train and test sets.", "cite_spans": [ { "start": 122, "end": 144, "text": "(Madigan et al., 2005;", "ref_id": "BIBREF12" }, { "start": 145, "end": 165, "text": "Kaster et al., 2005)", "ref_id": "BIBREF3" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "The utility of stylistic features is also well understood: they model author preferences for the use of punctuation marks, emoticons, white spaces, and other traces of writing style. Such preferences are less influenced by topic, and directly reflect some of the unique writing patterns of an author.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "Character n-grams are the single most successful feature in authorship attribution (Koppel et al., 2009; Frantzeskou et al., 2007; Koppel et al., 2011) , but the reason for their success is not well understood. One hypothesis is that character n-grams carry a little bit of everything: lexical content, syntactic content, and even style by means of punctuation and white spaces (Koppel et al., 2011) . While this argument seems plausible, it falls short of a rigorous explanation.", "cite_spans": [ { "start": 83, "end": 104, "text": "(Koppel et al., 2009;", "ref_id": "BIBREF7" }, { "start": 105, "end": 130, "text": "Frantzeskou et al., 2007;", "ref_id": "BIBREF1" }, { "start": 131, "end": 151, "text": "Koppel et al., 2011)", "ref_id": "BIBREF8" }, { "start": 378, "end": 399, "text": "(Koppel et al., 2011)", "ref_id": "BIBREF8" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "In this paper, we investigate what in the make-up of these small units of text makes them so powerful. Our goal is two-fold: on the one hand we want to have a principled understanding of character ngrams that will inform their use as features for AA and other tasks; on the other hand we want to make AA approaches more accessible to non-experts so that, for example, they could be acceptable pieces of evidence in criminal cases.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "The research questions we aim to answer are:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "\u2022 Are all character n-grams equally important? For example, are the prefix of 'there', the suffix of 'breathe' and the whole word 'the' all equivalent? More generally, are character n-grams that capture morpho-syntactic information, thematic information and style information equally important?", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "\u2022 Are the character n-grams that are most important for single-domain settings also the most important for cross-domain settings? Which character n-grams are more like bag-of-words features (which tend to track topics), and which are more like stylistic features (which tend to track authors)?", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "\u2022 Do different classifiers agree on the importance of the different types of character n-grams? Are some character n-grams consistently the best regardless of the learning algorithm?", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "\u2022 Are some types of character n-grams irrelevant in AA tasks? Are there categories of character n-grams that we can exclude and get similar (or better) performance than using all n-grams? If there are, are they the same for both singledomain and cross-domain AA settings?", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "Our study shows that using the default bag-ofwords representation of char n-grams results in collapsing sequences of characters that correspond to different linguistic aspects, and that this yields suboptimal prediction performance. We further show that we can boost accuracy by loosing some categories of n-grams. Char n-grams closely related to thematic content can be completely removed without loss of accuracy, even in cases where the train and test sets have the same topics represented, a counter-intuitive argument. Given the wide spread use of char n-grams in text classification tasks, our findings have significant implications for future work in related areas.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "To answer our research questions and explore the value of character n-grams in authorship attribution, we propose to separate character n-grams into ten distinct categories. Unlike previous AA work where all character n-grams were combined into a single bagof-n-grams, we evaluate each category separately to understand its behavior and effectiveness in AA tasks. These categories are related to the three linguistic aspects hypothesized to be represented by character n-grams: morpho-syntax (as represented by affix-like n-grams), thematic content (as represented by word-like n-grams) and style (as represented by punctuation-based n-grams). We refer to these three aspects as super categories (SC).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Categories of Character N -grams", "sec_num": "2" }, { "text": "The following sections describe the different types of n-grams. We use the sentence in Table 1 as a running example for the classes and in Table 2 we show the resulting n-grams in that sentence. For ease of understanding, we replace spaces in n-grams with underscores ( ).", "cite_spans": [], "ref_spans": [ { "start": 87, "end": 94, "text": "Table 1", "ref_id": "TABREF0" }, { "start": 139, "end": 146, "text": "Table 2", "ref_id": "TABREF2" } ], "eq_spans": [], "section": "Categories of Character N -grams", "sec_num": "2" }, { "text": "The actors wanted to see if the pact seemed like an old-fashioned one. ", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Categories of Character N -grams", "sec_num": "2" }, { "text": "Character n-grams are generally too short to represent any deep syntax, but some of them can reflect morphology to some degree. In particular, we consider the following affix-like features by looking at n-grams that begin or end a word:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Affix n-grams", "sec_num": "2.1" }, { "text": "prefix A character n-gram that covers the first n characters of a word that is at least n + 1 characters long.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Affix n-grams", "sec_num": "2.1" }, { "text": "suffix A character n-gram that covers the last n characters of a word that is at least n + 1 characters long.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Affix n-grams", "sec_num": "2.1" }, { "text": "space-prefix A character n-gram that begins with a space. space-suffix A character n-gram that ends with a space.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Affix n-grams", "sec_num": "2.1" }, { "text": "While character n-grams are often too short to capture entire words, some types can capture partial words and other word-relevant tokens. We consider the following such features:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Word n-grams", "sec_num": "2.2" }, { "text": "whole-word A character n-gram that covers all characters of a word that is exactly n characters long.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Word n-grams", "sec_num": "2.2" }, { "text": "mid-word A character n-gram that covers n characters of a word that is at least n + 2 characters long, and that covers neither the first nor the last character of the word.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Word n-grams", "sec_num": "2.2" }, { "text": "multi-word N -grams that span multiple words, identified by the presence of a space in the middle of the n-gram.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Word n-grams", "sec_num": "2.2" }, { "text": "The main stylistic choices that character n-grams can capture are the author's preferences for particular patterns of punctuation. The following features characterize punctuation by its location in the n-gram.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Punctuation n-grams", "sec_num": "2.3" }, { "text": "beg-punct A character n-gram whose first character is punctuation, but middle characters are not.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Punctuation n-grams", "sec_num": "2.3" }, { "text": "mid-punct A character n-gram with at least one punctuation character that is neither the first nor the last character.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Punctuation n-grams", "sec_num": "2.3" }, { "text": "end-punct A character n-gram whose last character is punctuation, but middle characters are not.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Punctuation n-grams", "sec_num": "2.3" }, { "text": "The above ten categories are intended to be disjoint, so that a character n-gram belongs to exactly one of the categories. For n-grams that contain both spaces and punctuation, we first categorize by punctuation and then by spaces. For example, 'e, ' is assigned to the mid-punct category, not the spacesuffix category.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Punctuation n-grams", "sec_num": "2.3" }, { "text": "We have observed that in our data almost 80% of the n-grams in the punct-beg and punct-mid categories contain a space. This tight coupling of punctuation and spaces is due to the rules of English orthography: most punctuation marks require a space following them. The 20% of n-grams that have punctuation but no spaces correspond mostly to the exceptions to this rule: quotation marks, mid-word hyphens, etc. An interesting experiment for future work would be to split out these two types of punctuation into separate feature categories.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Punctuation n-grams", "sec_num": "2.3" }, { "text": "We consider two corpora, a single-domain corpus, where there is only one topic that all authors are writing about, and a multi-domain corpus, where there are multiple different topics. The latter allows us to test the generalization of AA models, by testing them on a different topic from that used for training.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Datasets", "sec_num": "3" }, { "text": "The first collection is the CCAT topic class, a subset of the Reuters Corpus Volume 1 (Lewis et al., 2004) . Although this collection was not gathered for the goal of doing authorship attribution studies, previous work has reported results for AA with 10 and 50 authors (Stamatatos, 2008; Plakias and Stamatatos, 2008; Escalante et al., 2011) . We refer to these as CCAT 10 and CCAT 50, respectively. Both CCAT 10 and CCAT 50 belong to CCAT category (about corporate/industrial news) and are balanced across authors, with 100 documents sampled for each author. Manual inspection of the dataset revealed that some of the authors in this collection consistently used signatures at the end of documents. Also, we noticed some writers use quotations a lot. Con-Corpus #authors #docs #sentences #words /author/topic /doc /doc CCAT 10 10 100 19 425 CCAT 50 50 100 19 415 Guardian1 13 13 53 1034 Guardian2 13 65 10 207 Table 3 : Some statistics about the datasets. sidering these parts of text for measuring the frequencies of character n-grams is not a good idea because signatures provide direct clues about the authorship of document and quotations do not reflect the author's writing style. Therefore, to clean up the CCAT collection, we preprocessed it to remove signatures and quotations from each document. Since the CCAT collection contains documents belonging to only corporate/industrial topic category, this will be our single-domain collection.", "cite_spans": [ { "start": 86, "end": 106, "text": "(Lewis et al., 2004)", "ref_id": "BIBREF9" }, { "start": 270, "end": 288, "text": "(Stamatatos, 2008;", "ref_id": "BIBREF19" }, { "start": 289, "end": 318, "text": "Plakias and Stamatatos, 2008;", "ref_id": "BIBREF15" }, { "start": 319, "end": 342, "text": "Escalante et al., 2011)", "ref_id": "BIBREF0" } ], "ref_spans": [ { "start": 821, "end": 940, "text": "CCAT 10 10 100 19 425 CCAT 50 50 100 19 415 Guardian1 13 13 53 1034 Guardian2 13 65 10 207 Table 3", "ref_id": "TABREF0" } ], "eq_spans": [], "section": "Datasets", "sec_num": "3" }, { "text": "The other collection consists of texts published in The Guardian daily newspaper written by 13 authors in four different topics (Stamatatos, 2013) . This dataset contains opinion articles on the topics: World, U.K., Society, and Politics. Following prior work, to make the collection balanced across authors, we choose at most ten documents per author for each of the four topics. We refer to this corpus as Guardian1. We also consider a variation of this corpus that makes it more challenging but that more closely matches realistic scenarios of forensic investigation that deal with short texts such as tweets, SMS, and emails. We chunk each of the documents by sentence boundaries into five new short documents. We refer to this corpus as Guardian2. Table 3 shows some of the statistics of the CCAT and Guardian corpora and Table 4 presents some of the top character n-grams for each category (taken from an author in the Guardian data, but the top ngrams look qualitatively similar for other authors).", "cite_spans": [ { "start": 128, "end": 146, "text": "(Stamatatos, 2013)", "ref_id": "BIBREF21" } ], "ref_spans": [ { "start": 753, "end": 760, "text": "Table 3", "ref_id": null }, { "start": 827, "end": 834, "text": "Table 4", "ref_id": "TABREF4" } ], "eq_spans": [], "section": "Datasets", "sec_num": "3" }, { "text": "We performed various experiments using different categories of character n-grams. We chose n=3 since our preliminary experiments found character 3-grams to be more effective than other higher level character n-grams. For each category, we considered only those 3-grams that occur at least five times in the training documents. tion models was measured in terms of accuracy. In the single-domain CCAT experiments, accuracy was measured using the train/test partition of prior work.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Experimental Settings", "sec_num": "4" }, { "text": "In the cross-domain Guardian experiments, accuracy was measured by considering all 12 possible pairings of the 4 topics, treating one topic as training data and the other as testing data, and averaging accuracy over these 12 scenarios. This ensured that in the crossdomain experiments, the topics of the training data were always different from that of the test data. We trained support vector machine (SVM) classifiers using the Weka implementation (Witten and Frank, 2005) with default parameters. We also ran some comparative experiments with the Weka implementation of naive Bayes classifiers and the Lib-SVM implementation of SVMs. In the results below, when performance of a single classifier is presented, it is the result of Weka's SVM, which generally gave the best performance. When performance of other classifiers are presented, the classifiers are explicitly indicated.", "cite_spans": [ { "start": 450, "end": 474, "text": "(Witten and Frank, 2005)", "ref_id": "BIBREF22" } ], "ref_spans": [], "eq_spans": [], "section": "Experimental Settings", "sec_num": "4" }, { "text": "In this section, we present various results on authorship attribution tasks using both single as well as cross-domain datasets. We will explore character ngrams in depth and try to understand why they are so effective in discriminating authors. most discriminative. Table 5 shows the accuracy of each type of n-gram for each of the different corpora. Table 5 (a) shows that the top four categories for single-domain AA are: prefix, suffix, space-prefix, and mid-word. These four categories have the best performance on both CCAT 10 and CCAT 50. In contrast, Table 5 (b) shows that the top four categories for cross-domain AA are: prefix, space-prefix, beg-punct, and mid-punct.", "cite_spans": [], "ref_spans": [ { "start": 266, "end": 273, "text": "Table 5", "ref_id": "TABREF6" }, { "start": 351, "end": 358, "text": "Table 5", "ref_id": "TABREF6" }, { "start": 558, "end": 565, "text": "Table 5", "ref_id": "TABREF6" } ], "eq_spans": [], "section": "Experimental Results and Evaluation", "sec_num": "5" }, { "text": "For both single-domain and cross-domain AA, prefix and space-prefix are strong features, and are generally better than the suffix features, perhaps because authors have more control over prefixes in English, while suffixes are often obligatory for grammatical reasons. For cross-domain AA, beg-punct and midpunct are the top features, likely because an author's use of punctuation is consistent even when the topic changes. For single-domain AA, mid-word was also a good feature, probably because it captured lexical information that correlates with authors' preferences towards writing about specific topics. Figure 1 shows an alternate view of these results, graphing the rank of each n-gram type. For computing the rank, the accuracies of the ten different n-gram type classifiers are sorted in decreasing order and ranked from 1 to 10 respectively with ties getting the same rank. For the Guardian corpora, the average rank of each n-gram category was computed by averaging its rank across the 12 possible test/train cross-domain combinations. In both of the single-domain CCAT corpora, the classifier based on prefix n-grams had the top accuracy (rank 1), and the classifier based on mid-punct had the worst accuracy (rank 10). In both of the cross-domain Guardian corpora, on the other hand, mid-punct was among the top-ranked n-gram categories. This suggests that punctuation features generalize the best across topic, but if AA is more of a topic classification task (as in the single-domain CCAT corpora), then punctuation adds little over other features that more directly capture the topic.", "cite_spans": [], "ref_spans": [ { "start": 610, "end": 618, "text": "Figure 1", "ref_id": "FIGREF0" } ], "eq_spans": [], "section": "Which n-gram", "sec_num": "5.1" }, { "text": "Since our cross-domain datasets are small, we performed a small number of planned comparisons using a two-tailed t-test over the accuracies on the Guardian1 and Guardian2 corpora. We found that in both corpora, the best punctuation category (punctmid) is better than the best word category (wholeword) with p < 0.001. In the Guardian2 corpus, the best affix category (space-prefix) is also better than the best word category (whole-word) with p < 0.05, but this does not hold in the Guardian1 corpus (p = 0.14). Also, we observed that in both Guardian1 and Guardian2 datasets, both punct-mid and spaceprefix are better than multi-word (p < 0.01).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Which n-gram", "sec_num": "5.1" }, { "text": "Overall, we see that affix n-grams are generally effective in both single-domain and cross-domain settings, punctuation n-grams are effective in crossdomain settings, and mid-word is the only effective word n-gram, and only in the single-domain setting.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Which n-gram", "sec_num": "5.1" }, { "text": "Importance of Different n-gram Types?", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Do Different Classifiers Agree on the", "sec_num": "5.2" }, { "text": "The previous experiments have shown, for example, that prefix n-grams are universally predictive in AA tasks, that mid-word n-grams are good predictors in single-domain settings, and that beg-punct n-grams are good predictors in cross-domain settings. But are these facts about the n-gram types themselves, or are these results only true for the specific SVM classifiers we trained?", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Do Different Classifiers Agree on the", "sec_num": "5.2" }, { "text": "To see whether certain types of n-grams are fundamentally good or bad, regardless of the classifier, we compare performance of the different n-gram types for three classifiers: Weka SVM classifiers (as used in our other experiments), LibSVM classifiers and Weka's naive Bayes classifiers 1 . Figure 2 shows the n-gram category rankings for all these classifiers 2 for both the single-domain CCAT and the cross-domain Guardian settings.", "cite_spans": [], "ref_spans": [ { "start": 292, "end": 300, "text": "Figure 2", "ref_id": null } ], "eq_spans": [], "section": "Do Different Classifiers Agree on the", "sec_num": "5.2" }, { "text": "Across the different classifiers, the pattern of feature rankings are similar. Table 6 shows the Spearman's rank correlation coefficient (\u03c1) for the per-ngram-type accuracies of each pair of classifiers. We observe fairly high correlations, with \u03c1 above 0.70 for all single-domain pairings, and between 0.44 and 0.81 for cross-domain pairings.", "cite_spans": [], "ref_spans": [ { "start": 79, "end": 86, "text": "Table 6", "ref_id": "TABREF8" } ], "eq_spans": [], "section": "Do Different Classifiers Agree on the", "sec_num": "5.2" }, { "text": "As in Section 5.1, prefix and space-prefix are among the most predictive n-gram types. In the single-domain settings, we again see that suffix and mid-word are also highly predictive, while in the cross-domain settings, we again see that beg-punct and mid-punct are highly predictive. These results all confirm that some types of n-grams are fundamentally more predictive than others, and our results are not specific to the particular type of classifier used. 6.7", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Do Different Classifiers Agree on the", "sec_num": "5.2" }, { "text": "1.6 1.6", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Do Different Classifiers Agree on the", "sec_num": "5.2" }, { "text": "Cross Domain (Guardian) Figure 2 : Average rank of the performance of each n-gram category across different types of classifiers on the single-domain CCAT task (top) and the cross-domain Guardian task (bottom).", "cite_spans": [], "ref_spans": [ { "start": 24, "end": 32, "text": "Figure 2", "ref_id": null } ], "eq_spans": [], "section": "4.4", "sec_num": null }, { "text": "In the previous sections, we have seen that some types of character n-grams are more predictive than others -affix n-grams performed well in both single domain and cross-domain settings and punctuation n-grams performed well in cross-domain settings.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Are Some Character N -grams Irrelevant?", "sec_num": "5.3" }, { "text": "In general, word n-grams were not as predictive as other types of n-grams (with the one exception being mid-word n-grams in the single domain setting). Given this poor performance of word n-grams, a natural question is: could we exclude these features entirely and achieve similar performance? Our goal then is to compare a model trained on affix n-grams and punct n-grams against a model trained on \"all\" n-grams. We consider two definitions of \"all\":", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Are Some Character N -grams Irrelevant?", "sec_num": "5.3" }, { "text": "all-untyped The traditional approach to extracting n-grams where n-gram types are ignored (e.g., 'the' as a whole word is no different from 'the' in the middle of a word)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Are Some Character N -grams Irrelevant?", "sec_num": "5.3" }, { "text": "all-typed The approach discussed in this paper, where n-grams of different types are distinguished (equivalent to the set of all af-fix+punct+word n-grams).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Are Some Character N -grams Irrelevant?", "sec_num": "5.3" }, { "text": "We compare these models trained on all the n-grams to our affix+punct model. Table 7 shows this analysis. For either definition of \"all\", the model that discards all word features achieves performance as high or higher than the model with all of the features, and does so with only about two thirds of the features. This is not too surprising in the cross-domain Guardian tasks, where the word n-grams were among the worst features. On the single-domain CCAT tasks this result is more surprising, since we have discarded the mid-word n-grams, which was one of the best single-domain n-gram types. This indicates that whatever information mid-word is capturing it is also being captured in other ways via affix and punct n-grams. Of all 1024 possible combinations of features, we tried a Table 7 : Results of excluding word n-grams, compared to using all n-grams, either in the traditional approach (untyped n-grams) or in the approach of this paper (typed n-grams). Accuracy (Acc) and the number of features (N in italics) are reported for each classifier. The best accuracy for each dataset is in bold.", "cite_spans": [], "ref_spans": [ { "start": 77, "end": 84, "text": "Table 7", "ref_id": null }, { "start": 787, "end": 794, "text": "Table 7", "ref_id": null } ], "eq_spans": [], "section": "Are Some Character N -grams Irrelevant?", "sec_num": "5.3" }, { "text": "number of different combinations and were unable to identify one that outperformed affix+punct. Overall, this experiment gives compelling evidence that affix and punct n-grams are more important than word n-grams.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Are Some Character N -grams Irrelevant?", "sec_num": "5.3" }, { "text": "We did a manual exploration of our datasets. In our cross-domain dataset, the character 3-gram 'sti' shows up as both prefix and mid-word. All 13 authors use 'sti' frequently as a mid-word n-gram in words such as institution, existing, justice, and distinction. For example:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Analysis", "sec_num": "6" }, { "text": "\u2022 The government's story is that the existing warheads might be deteriorating.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Analysis", "sec_num": "6" }, { "text": "\u2022 For all the justice of many of his accusations, the result is occasionally as dreadful as his title suggests.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Analysis", "sec_num": "6" }, { "text": "But only six authors use 'sti' as a prefix, in examples like:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Analysis", "sec_num": "6" }, { "text": "\u2022 Their mission was to convince tourists that Britain was still open for business.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Analysis", "sec_num": "6" }, { "text": "\u2022 There aren't even any dead people on it, since by the very act of being dead and still famous, they assert their long-term impact.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Analysis", "sec_num": "6" }, { "text": "Thus 'sti' as a prefix is predictive of authorship even though 'sti' as a mid-word n-gram is not. Notably, under the traditional untyped bag-of-n-grams approach, both versions of 'sti' would have been treated the same, and this discriminative power would have been lost.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Analysis", "sec_num": "6" }, { "text": "To use old-fashioned language, she is motherly -a plump, rosy-cheeked woman of Kent, whom nature seemed to have created to raise children.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Analysis", "sec_num": "6" }, { "text": "To use old-fashioned language, she is motherly -a plump, rosy-cheeked woman of Kent, whom nature seemed to have created to raise children. As already demonstrated in Section 5 that af-fix+punct features perform better than using all the features, we would like to use an example from our dataset to visualize the text when features in SC word are discarded. Out of seven categories in affix and punct, we computed in how many of them each character belongs to, three being the maximum possible value. Therefore, we show each character with different opacity level depending on number of categories it belongs to: zero will get white color (word related n-grams), one will get 33% black, two will get 67% black, and three will get 100% black. In Table 8 , we show an example sentence before (first row of Table 8 ) and after (second row of Table 8) showing the opacity level of each character. It is clear that the darkest characters are those around the punctuation characters and those around spaces are second darkest, while the lightest (with 0% darkness) are the ones in the middle of long words. This gives us an idea about the characters in a text that are important for AA tasks.", "cite_spans": [], "ref_spans": [ { "start": 745, "end": 752, "text": "Table 8", "ref_id": "TABREF9" }, { "start": 804, "end": 811, "text": "Table 8", "ref_id": "TABREF9" }, { "start": 839, "end": 847, "text": "Table 8)", "ref_id": "TABREF9" } ], "eq_spans": [], "section": "Analysis", "sec_num": "6" }, { "text": "Various hypotheses have been put forth to explain the \"black magic\" (Kestemont, 2014) behind the success of character n-gram features in authorship attribution. Kestemont (2014) conjectured that their utility was in capturing function words and morphology. Koppel et al. (2009) suggested that they were capturing topic information in single domain settings, and style and syntactic information in cross-domain settings. Our study provides empirical evidence for testing these claims. We did indeed find that the ability of character n-grams to capture morphology is useful, as reflected in the high prediction performance of af-fix n-grams in both single-domain and cross-domain settings. And we found that word n-grams (capturing topic information) were useful in single domain settings, while puct n-grams (capturing style information) were useful in cross-domain settings. We further found that word n-grams are unnecessary, even in single-domain settings. Models based only on affix and punct n-grams performed as well as models with all n-grams regardless of whether it was a single-domain or cross-domain authorship attribution task.", "cite_spans": [ { "start": 68, "end": 85, "text": "(Kestemont, 2014)", "ref_id": "BIBREF5" }, { "start": 161, "end": 177, "text": "Kestemont (2014)", "ref_id": "BIBREF5" }, { "start": 257, "end": 277, "text": "Koppel et al. (2009)", "ref_id": "BIBREF7" } ], "ref_spans": [], "eq_spans": [], "section": "Discussion", "sec_num": "7" }, { "text": "Our findings on the value of selecting n-grams according to the linguistic aspect they represent may also be beneficial in other classification tasks where character n-grams are commonly used. Promising tasks are those related to the stylistic analysis of texts, such as native language identification, document similarity and plagiarism detection.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Discussion", "sec_num": "7" }, { "text": "Morphologically speaking, English is a poor language. The fact that we identified significant differences in performance by selecting n-gram categories that are related to affixation in this poorly inflected language suggests that we may find even larger differences in performance in morphologically richer languages. We leave this research question for future work.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Discussion", "sec_num": "7" }, { "text": "Weka SVM and LibSVM are both support vector machine classifiers, but Weka uses Platt's sequential minimal optimization algorithm while LibSVM uses working set selection with second order information. The result is that they achieve different performance on our AA tasks.2 We also tried a decision tree classifier, C4.5 (J48) from WEKA, and it produced similar patterns (not shown).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null } ], "back_matter": [ { "text": "This research was partially supported by NSF awards 1462141 and 1254108. It was also supported in part by the CONACYT grant 134186 and the WIQ-EI IRSES project (grant no. 269180) within the FP 7 Marie Curie.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Acknowledgements", "sec_num": null } ], "bib_entries": { "BIBREF0": { "ref_id": "b0", "title": "Local histograms of character n-grams for authorship attribution", "authors": [ { "first": "H", "middle": [ "J" ], "last": "Escalante", "suffix": "" }, { "first": "T", "middle": [], "last": "Solorio", "suffix": "" }, { "first": "M", "middle": [], "last": "Montes-Y Gomez", "suffix": "" } ], "year": 2011, "venue": "Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies", "volume": "", "issue": "", "pages": "288--298", "other_ids": {}, "num": null, "urls": [], "raw_text": "H. J. Escalante, T. Solorio, and M. Montes-y Gomez. 2011. Local histograms of character n-grams for au- thorship attribution. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguis- tics: Human Language Technologies, pages 288-298, Portland, Oregon, USA, June. Association for Compu- tational Linguistics.", "links": null }, "BIBREF1": { "ref_id": "b1", "title": "Identifying authorship by byte-level ngrams: The source code author profile (SCAP) method", "authors": [ { "first": "G", "middle": [], "last": "Frantzeskou", "suffix": "" }, { "first": "E", "middle": [], "last": "Stamatatos", "suffix": "" }, { "first": "S", "middle": [], "last": "Gritzalis", "suffix": "" }, { "first": "C", "middle": [ "E" ], "last": "Chaski", "suffix": "" } ], "year": 2007, "venue": "Journal of Digital Evidence", "volume": "6", "issue": "1", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "G. Frantzeskou, E. Stamatatos, S. Gritzalis, and C. E. Chaski. 2007. Identifying authorship by byte-level n- grams: The source code author profile (SCAP) method. Journal of Digital Evidence, 6(1).", "links": null }, "BIBREF2": { "ref_id": "b2", "title": "Authorship attribution", "authors": [ { "first": "P", "middle": [], "last": "Juola", "suffix": "" } ], "year": 2006, "venue": "Foundations and Trends in Information Retrieval", "volume": "1", "issue": "3", "pages": "233--334", "other_ids": {}, "num": null, "urls": [], "raw_text": "P. Juola. 2006. Authorship attribution. Foundations and Trends in Information Retrieval, 1(3):233-334, Decem- ber.", "links": null }, "BIBREF3": { "ref_id": "b3", "title": "Combining text and linguistic document representations for authorship attribution", "authors": [ { "first": "A", "middle": [], "last": "Kaster", "suffix": "" }, { "first": "S", "middle": [], "last": "Siersdorfer", "suffix": "" }, { "first": "G", "middle": [], "last": "Weikum", "suffix": "" } ], "year": 2005, "venue": "SIGIR Workshop: Stylistic Analysis of Text for Information Access (STYLE", "volume": "", "issue": "", "pages": "27--35", "other_ids": {}, "num": null, "urls": [], "raw_text": "A. Kaster, S. Siersdorfer, and G. Weikum. 2005. Com- bining text and linguistic document representations for authorship attribution. In SIGIR Workshop: Stylistic Analysis of Text for Information Access (STYLE, pages 27-35.", "links": null }, "BIBREF4": { "ref_id": "b4", "title": "N-gram based author profiles for authorship attribution", "authors": [ { "first": "V", "middle": [], "last": "Keselj", "suffix": "" }, { "first": "F", "middle": [], "last": "Peng", "suffix": "" }, { "first": "N", "middle": [], "last": "Cercone", "suffix": "" }, { "first": "C", "middle": [], "last": "Thomas", "suffix": "" } ], "year": 2003, "venue": "Proceedings of the Pacific Association for Computational Linguistics", "volume": "", "issue": "", "pages": "255--264", "other_ids": {}, "num": null, "urls": [], "raw_text": "V. Keselj, F. Peng, N. Cercone, and C. Thomas. 2003. N-gram based author profiles for authorship attribution. In Proceedings of the Pacific Association for Computa- tional Linguistics, pages 255-264.", "links": null }, "BIBREF5": { "ref_id": "b5", "title": "Function words in authorship attribution. From black magic to theory", "authors": [ { "first": "M", "middle": [], "last": "Kestemont", "suffix": "" } ], "year": 2014, "venue": "CLFL", "volume": "", "issue": "", "pages": "59--66", "other_ids": {}, "num": null, "urls": [], "raw_text": "M. Kestemont. 2014. Function words in authorship attri- bution. From black magic to theory? In CLFL, pages 59-66, Gothenburg, Sweden, April.", "links": null }, "BIBREF6": { "ref_id": "b6", "title": "Discrimination of authorship using visualization. Information Processing & Management", "authors": [ { "first": "W", "middle": [ "Addison" ], "last": "Bradley Kjell", "suffix": "" }, { "first": "Ophir", "middle": [], "last": "Woods", "suffix": "" }, { "first": "", "middle": [], "last": "Frieder", "suffix": "" } ], "year": 1994, "venue": "", "volume": "30", "issue": "", "pages": "141--150", "other_ids": {}, "num": null, "urls": [], "raw_text": "Bradley Kjell, W.Addison Woods, and Ophir Frieder. 1994. Discrimination of authorship using visualization. Information Processing & Management, 30(1):141 - 150.", "links": null }, "BIBREF7": { "ref_id": "b7", "title": "Computational methods in authorship attribution", "authors": [ { "first": "M", "middle": [], "last": "Koppel", "suffix": "" }, { "first": "J", "middle": [], "last": "Schler", "suffix": "" }, { "first": "S", "middle": [], "last": "Argamon", "suffix": "" } ], "year": 2009, "venue": "Journal of the American Society for Information Science and Technology", "volume": "60", "issue": "1", "pages": "9--26", "other_ids": {}, "num": null, "urls": [], "raw_text": "M. Koppel, J. Schler, and S. Argamon. 2009. Computa- tional methods in authorship attribution. Journal of the American Society for Information Science and Technol- ogy, 60(1):9-26.", "links": null }, "BIBREF8": { "ref_id": "b8", "title": "Authorship attribution in the wild", "authors": [ { "first": "M", "middle": [], "last": "Koppel", "suffix": "" }, { "first": "J", "middle": [], "last": "Schler", "suffix": "" }, { "first": "S", "middle": [], "last": "Argamon", "suffix": "" } ], "year": 2011, "venue": "Language Resources and Evaluation", "volume": "45", "issue": "", "pages": "83--94", "other_ids": {}, "num": null, "urls": [], "raw_text": "M. Koppel, J. Schler, and S. Argamon. 2011. Author- ship attribution in the wild. Language Resources and Evaluation, 45:83-94.", "links": null }, "BIBREF9": { "ref_id": "b9", "title": "RCV1: A new benchmark collection for text categorization research", "authors": [ { "first": "D", "middle": [ "D" ], "last": "Lewis", "suffix": "" }, { "first": "Y", "middle": [], "last": "Yang", "suffix": "" }, { "first": "T", "middle": [ "G" ], "last": "Rose", "suffix": "" }, { "first": "F", "middle": [], "last": "Li", "suffix": "" } ], "year": 2004, "venue": "Journal of Machine Learning Research", "volume": "5", "issue": "", "pages": "361--397", "other_ids": {}, "num": null, "urls": [], "raw_text": "D. D. Lewis, Y. Yang, T. G. Rose, and F. Li. 2004. RCV1: A new benchmark collection for text categorization re- search. Journal of Machine Learning Research, 5:361- 397.", "links": null }, "BIBREF10": { "ref_id": "b10", "title": "Authorship attribution and verification with many authors and limited data", "authors": [ { "first": "K", "middle": [], "last": "Luyckx", "suffix": "" }, { "first": "W", "middle": [], "last": "Daelemans", "suffix": "" } ], "year": 2008, "venue": "Proceedings of the 22nd International Conference on Computational Linguistics (COLING 2008)", "volume": "", "issue": "", "pages": "513--520", "other_ids": {}, "num": null, "urls": [], "raw_text": "K. Luyckx and W. Daelemans. 2008. Authorship attri- bution and verification with many authors and limited data. In Proceedings of the 22nd International Confer- ence on Computational Linguistics (COLING 2008), pages 513-520, Manchester, UK, August.", "links": null }, "BIBREF11": { "ref_id": "b11", "title": "The effect of author set size and data size in authorship attribution", "authors": [ { "first": "K", "middle": [], "last": "Luyckx", "suffix": "" }, { "first": "W", "middle": [], "last": "Daelemans", "suffix": "" } ], "year": 2010, "venue": "Literary and Linguistic Computing", "volume": "", "issue": "", "pages": "1--21", "other_ids": {}, "num": null, "urls": [], "raw_text": "K. Luyckx and W. Daelemans. 2010. The effect of author set size and data size in authorship attribution. Literary and Linguistic Computing, pages 1-21, August.", "links": null }, "BIBREF12": { "ref_id": "b12", "title": "Author identification on the large scale", "authors": [ { "first": "D", "middle": [], "last": "Madigan", "suffix": "" }, { "first": "A", "middle": [], "last": "Genkin", "suffix": "" }, { "first": "S", "middle": [], "last": "Argamon", "suffix": "" }, { "first": "D", "middle": [], "last": "Fradkin", "suffix": "" }, { "first": "L", "middle": [], "last": "Ye", "suffix": "" } ], "year": 2005, "venue": "Proceedings of CSNA/Interface 05", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "D. Madigan, A. Genkin, S. Argamon, D. Fradkin, and L. Ye. 2005. Author identification on the large scale. In Proceedings of CSNA/Interface 05.", "links": null }, "BIBREF13": { "ref_id": "b13", "title": "There's no data like more data (but when will enough be enough?)", "authors": [ { "first": "R", "middle": [], "last": "Moore", "suffix": "" } ], "year": 2001, "venue": "Proceedings of the IEEE International Workshop on Intelligent Signal Processing", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "R. Moore. 2001. There's no data like more data (but when will enough be enough?). In Proceedings of the IEEE International Workshop on Intelligent Signal Processing, Budapest, Hungary.", "links": null }, "BIBREF14": { "ref_id": "b14", "title": "Language independent authorship attribution using character level language models", "authors": [ { "first": "D", "middle": [], "last": "Peng", "suffix": "" }, { "first": "V", "middle": [], "last": "Schuurmans", "suffix": "" }, { "first": "S", "middle": [], "last": "Keselj", "suffix": "" }, { "first": "", "middle": [], "last": "Wang", "suffix": "" } ], "year": 2003, "venue": "10th Conference of the European Chapter of the Association for Computational Linguistics, EACL", "volume": "", "issue": "", "pages": "267--274", "other_ids": {}, "num": null, "urls": [], "raw_text": "Peng, D. Schuurmans, V. Keselj, and S. Wang. 2003. Language independent authorship attribution using character level language models. In 10th Conference of the European Chapter of the Association for Compu- tational Linguistics, EACL, pages 267-274.", "links": null }, "BIBREF15": { "ref_id": "b15", "title": "Tensor space models for authorship attribution", "authors": [ { "first": "S", "middle": [], "last": "Plakias", "suffix": "" }, { "first": "E", "middle": [], "last": "Stamatatos", "suffix": "" } ], "year": 2008, "venue": "Proceedings of the 5th", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "S. Plakias and E. Stamatatos. 2008. Tensor space models for authorship attribution. In Proceedings of the 5th", "links": null }, "BIBREF16": { "ref_id": "b16", "title": "Hellenic Conference on Artificial Intelligence: Theories, Models and Applications", "authors": [], "year": null, "venue": "", "volume": "5138", "issue": "", "pages": "239--249", "other_ids": {}, "num": null, "urls": [], "raw_text": "Hellenic Conference on Artificial Intelligence: Theo- ries, Models and Applications, volume 5138 of LNCS, pages 239-249, Syros, Greece.", "links": null }, "BIBREF17": { "ref_id": "b17", "title": "Automatic text categorization in terms of genre and author", "authors": [ { "first": "E", "middle": [], "last": "Stamatatos", "suffix": "" }, { "first": "G", "middle": [], "last": "Kokkinakis", "suffix": "" }, { "first": "N", "middle": [], "last": "Fakotakis", "suffix": "" } ], "year": 2000, "venue": "Computational Linguistics", "volume": "26", "issue": "4", "pages": "471--495", "other_ids": {}, "num": null, "urls": [], "raw_text": "E. Stamatatos, G. Kokkinakis, and N. Fakotakis. 2000. Automatic text categorization in terms of genre and author. Computational Linguistics, 26(4):471-495, De- cember.", "links": null }, "BIBREF18": { "ref_id": "b18", "title": "Authorship attribution based on feature set subspacing ensembles", "authors": [ { "first": "E", "middle": [], "last": "Stamatatos", "suffix": "" } ], "year": 2006, "venue": "International Journal on Artificial Intelligence tools", "volume": "15", "issue": "5", "pages": "823--838", "other_ids": {}, "num": null, "urls": [], "raw_text": "E. Stamatatos. 2006. Authorship attribution based on fea- ture set subspacing ensembles. International Journal on Artificial Intelligence tools, 15(5):823-838.", "links": null }, "BIBREF19": { "ref_id": "b19", "title": "Author identification: Using text sampling to handle the class imbalance problem. Information Processing and Managemement", "authors": [ { "first": "E", "middle": [], "last": "Stamatatos", "suffix": "" } ], "year": 2008, "venue": "", "volume": "44", "issue": "", "pages": "790--799", "other_ids": {}, "num": null, "urls": [], "raw_text": "E. Stamatatos. 2008. Author identification: Using text sampling to handle the class imbalance problem. Infor- mation Processing and Managemement, 44:790-799.", "links": null }, "BIBREF20": { "ref_id": "b20", "title": "Plagiarism detection using stopword n-grams", "authors": [ { "first": "E", "middle": [], "last": "Stamatatos", "suffix": "" } ], "year": 2011, "venue": "Journal of the American Society for Information Science and Technology", "volume": "62", "issue": "12", "pages": "2512--2527", "other_ids": {}, "num": null, "urls": [], "raw_text": "E. Stamatatos. 2011. Plagiarism detection using stopword n-grams. Journal of the American Society for Informa- tion Science and Technology, 62(12):2512-2527.", "links": null }, "BIBREF21": { "ref_id": "b21", "title": "On the robustness of authorship attribution based on character n-gram features", "authors": [ { "first": "E", "middle": [], "last": "Stamatatos", "suffix": "" } ], "year": 2013, "venue": "Journal of Law & Policy", "volume": "21", "issue": "2", "pages": "421--439", "other_ids": {}, "num": null, "urls": [], "raw_text": "E. Stamatatos. 2013. On the robustness of authorship attribution based on character n-gram features. Journal of Law & Policy, 21(2):421 -439.", "links": null }, "BIBREF22": { "ref_id": "b22", "title": "Data Mining: Practical Machine Learning Tools and Techniques", "authors": [ { "first": "I", "middle": [ "H" ], "last": "Witten", "suffix": "" }, { "first": "E", "middle": [], "last": "Frank", "suffix": "" } ], "year": 2005, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "I. H. Witten and E. Frank. 2005. Data Mining: Practi- cal Machine Learning Tools and Techniques. Morgan Kauffmann, 2nd edition.", "links": null }, "BIBREF23": { "ref_id": "b23", "title": "A framework for authorship identification of online messages: Writing-style features and classification techniques", "authors": [ { "first": "R", "middle": [], "last": "Zheng", "suffix": "" }, { "first": "Jiexun", "middle": [], "last": "Li", "suffix": "" }, { "first": "Hsinchun", "middle": [], "last": "Chen", "suffix": "" }, { "first": "Zan", "middle": [], "last": "Huang", "suffix": "" } ], "year": 2006, "venue": "J. Am. Soc. Inf. Sci. Technol", "volume": "57", "issue": "3", "pages": "378--393", "other_ids": {}, "num": null, "urls": [], "raw_text": "R. Zheng, Jiexun Li, Hsinchun Chen, and Zan Huang. 2006. A framework for authorship identification of on- line messages: Writing-style features and classification techniques. J. Am. Soc. Inf. Sci. Technol., 57(3):378- 393, February.", "links": null } }, "ref_entries": { "FIGREF0": { "num": null, "type_str": "figure", "uris": null, "text": "Average rank of the performance of each n-gram category on the single-domain CCAT tasks (top) and the cross-domain Guardian tasks (bottom)." }, "FIGREF2": { "num": null, "type_str": "figure", "uris": null, "text": ".8 8245 77.2 9715 78.8 5474 CCAT 50 69.2 14461 69.1 17062 69.3 9966 Guardian1 55.6 5689 53.6 6966 57.0 3822 Guardian2 45.9 5687 45.6 6965 48.0 3820" }, "TABREF0": { "html": null, "type_str": "table", "text": "Example sentence to demonstrate the selection of different n-gram categories.", "content": "", "num": null }, "TABREF2": { "html": null, "type_str": "table", "text": "", "content": "
: Example of the n-gram categories (n = 3) for the
sentence in Table 1. The first column represents the super
category (SC). The n-grams that appear in more than one
category are in bold.
", "num": null }, "TABREF4": { "html": null, "type_str": "table", "text": "Top character 3-grams in each category for author 'Catherine Bennet' in the cross-domain training data.", "content": "", "num": null }, "TABREF5": { "html": null, "type_str": "table", "text": "word punct Dataset prefix suffix space-prefix space-suffix multi-word whole-word mid-word beg-punct mid-punct end-punct", "content": "
affix
CCAT 10 74.6 71.071.266.065.848.070.060.235.456.2
CCAT 50 61.9 59.657.051.051.235.461.039.712.436.5
(a) Single Domain
affixwordpunct
Dataset prefix suffix space-prefix space-suffix multi-word whole-word mid-word beg-punct mid-punct end-punct
Guardian1 41.6 36.741.938.132.238.137.843.546.137.3
Guardian2 31.0 26.929.727.023.226.827.233.633.524.5
(b) Cross-Domain
", "num": null }, "TABREF6": { "html": null, "type_str": "table", "text": "Accuracy of AA classifiers trained on each of the character n-gram categories. The top four accuracies for each dataset are in bold.", "content": "
prefixsuffixspace-prefixspace-suffixmulti-wordwhole-wordmid-wordbeg-punctmid-punctend-punct
", "num": null }, "TABREF8": { "html": null, "type_str": "table", "text": "Spearman's rank correlation coefficient (\u03c1) for each pair of classifiers on the single-domain (CCAT) and cross-domain (Guardian) settings.", "content": "", "num": null }, "TABREF9": { "html": null, "type_str": "table", "text": "Example sentence showing the opacity of each character. Darkness of character is determined by the number of categories it belongs to (lowest=lighter, high-est=darkest color). Categories in word are discarded.", "content": "
", "num": null } } } }