| { | |
| "paper_id": "L18-1015", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T11:36:33.931907Z" | |
| }, | |
| "title": "Multi-Dialect Arabic POS Tagging: A CRF Approach", | |
| "authors": [ | |
| { | |
| "first": "Kareem", | |
| "middle": [], | |
| "last": "Darwish", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Google Inc", | |
| "location": {} | |
| }, | |
| "email": "[email protected]" | |
| }, | |
| { | |
| "first": "Hamdy", | |
| "middle": [], | |
| "last": "Mubarak", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Google Inc", | |
| "location": {} | |
| }, | |
| "email": "[email protected]" | |
| }, | |
| { | |
| "first": "Mohamed", | |
| "middle": [], | |
| "last": "Eldesouki", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Google Inc", | |
| "location": {} | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Ahmed", | |
| "middle": [], | |
| "last": "Abdelali", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Google Inc", | |
| "location": {} | |
| }, | |
| "email": "[email protected]" | |
| }, | |
| { | |
| "first": "Younes", | |
| "middle": [], | |
| "last": "Samih", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Google Inc", | |
| "location": {} | |
| }, | |
| "email": "[email protected]" | |
| }, | |
| { | |
| "first": "Randah", | |
| "middle": [], | |
| "last": "Alharbi", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Google Inc", | |
| "location": {} | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Mohammed", | |
| "middle": [], | |
| "last": "Attia", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Google Inc", | |
| "location": {} | |
| }, | |
| "email": "[email protected]" | |
| }, | |
| { | |
| "first": "Walid", | |
| "middle": [], | |
| "last": "Magdy", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Google Inc", | |
| "location": {} | |
| }, | |
| "email": "[email protected]" | |
| }, | |
| { | |
| "first": "Laura", | |
| "middle": [], | |
| "last": "Kallmeyer", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Google Inc", | |
| "location": {} | |
| }, | |
| "email": "[email protected]" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Qcri", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Google Inc", | |
| "location": {} | |
| }, | |
| "email": "" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "This paper introduces a new dataset of POS-tagged Arabic tweets in four major dialects along with tagging guidelines. The data, which we are releasing publicly, includes tweets in Egyptian, Levantine, Gulf, and Maghrebi, with 350 tweets for each dialect with appropriate train/test/development splits for 5-fold cross validation. We use a Conditional Random Fields (CRF) sequence labeler to train POS taggers for each dialect and examine the effect of cross and joint dialect training, and give benchmark results for the datasets. Using clitic n-grams, clitic metatypes, and stem templates as features, we were able to train a joint model that can correctly tag four different dialects with an average accuracy of 89.3%.", | |
| "pdf_parse": { | |
| "paper_id": "L18-1015", | |
| "_pdf_hash": "", | |
| "abstract": [ | |
| { | |
| "text": "This paper introduces a new dataset of POS-tagged Arabic tweets in four major dialects along with tagging guidelines. The data, which we are releasing publicly, includes tweets in Egyptian, Levantine, Gulf, and Maghrebi, with 350 tweets for each dialect with appropriate train/test/development splits for 5-fold cross validation. We use a Conditional Random Fields (CRF) sequence labeler to train POS taggers for each dialect and examine the effect of cross and joint dialect training, and give benchmark results for the datasets. Using clitic n-grams, clitic metatypes, and stem templates as features, we were able to train a joint model that can correctly tag four different dialects with an average accuracy of 89.3%.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Abstract", | |
| "sec_num": null | |
| } | |
| ], | |
| "body_text": [ | |
| { | |
| "text": "Part-of-speech (POS) tagging is important for a variety of applications such as parsing, information extraction, and machine translation. Though much work has focused on POS tagging of Modern Standard Arabic (MSA), work on Dialectal Arabic (DA) POS tagging is rather scant with POS tagged corpora for most dialects being nonexistent or of limited availability. Dialectal POS tagging is becoming increasingly important due to the ubiquity of social media, where users typically write in their dialects to match how they speak in their daily interactions. Dialectal text poses interesting challenges such as lack of spelling standards, pervasiveness of transformative morphological operations, such as word merging and letter substitution or deletion, in addition to lexical borrowing from foreign languages. Existing work on dialectal POS tagging focuses on building resources and tools for each dialect separately (Duh and Kirchhoff, 2005; Habash et al., 2013) . The rationale for the separation is that different dialects have different affixes, make different lexical and word ordering choices, and are influenced by different foreign languages. However, performing reliable dialect identification to properly route text to the appropriate POS tagger may be problematic, because conventional dialectal identification may lead to results lower than 90% (Darwish et al., 2014) . Thus, building a POS tagger that performs reliably across multiple dialects without the need for dialect identification is desirable. In this paper, we present new POS-tagging annotations on a dialectal dataset that is composed of social media text from Twitter for four major Arabic dialects, namely Egyptian (EGY), Levantine (LEV), Gulf (GLF), and Maghrebi (MGR) (Eldesouki et al., 2017; Samih et al., 2017a) . For each dialect, we tagged 350 tweets using an extended version of the Farasa tagset (Darwish et al., 2017) . We extended the tagset to account for tweet-specific tokens, namely hashtags, user mentions, emoticons and emojis, and URL's. We created 5-fold partitions for cross-validation with 70/10/20 train/dev/test splits for each dialect. We used the new dataset to train dialectal POS taggers for each dialect separately to test the effectiveness of the taggers on test data from the same dialect or from different dialects. We also experimented with cross-dialect and joint training to see if POS tagging of one dialect can benefit from data from other dialects. We show that joint models can perform on average at par with dialect specific models. For all our experiments we used a Conditional Random Fields (CRF) sequence labeler. The contributions of this paper are as follows:", | |
| "cite_spans": [ | |
| { | |
| "start": 914, | |
| "end": 939, | |
| "text": "(Duh and Kirchhoff, 2005;", | |
| "ref_id": "BIBREF7" | |
| }, | |
| { | |
| "start": 940, | |
| "end": 960, | |
| "text": "Habash et al., 2013)", | |
| "ref_id": "BIBREF12" | |
| }, | |
| { | |
| "start": 1354, | |
| "end": 1376, | |
| "text": "(Darwish et al., 2014)", | |
| "ref_id": "BIBREF5" | |
| }, | |
| { | |
| "start": 1744, | |
| "end": 1768, | |
| "text": "(Eldesouki et al., 2017;", | |
| "ref_id": "BIBREF8" | |
| }, | |
| { | |
| "start": 1769, | |
| "end": 1789, | |
| "text": "Samih et al., 2017a)", | |
| "ref_id": "BIBREF17" | |
| }, | |
| { | |
| "start": 1878, | |
| "end": 1900, | |
| "text": "(Darwish et al., 2017)", | |
| "ref_id": "BIBREF6" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1." | |
| }, | |
| { | |
| "text": "1. We present new dialectal POS tagging annotations on a multi-dialectal tweet dataset.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1." | |
| }, | |
| { | |
| "text": "2. We report benchmark results on DA POS tagging.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1." | |
| }, | |
| { | |
| "text": "3. We show that we can develop an effective joint model for POS tagging of different dialects without the need for dialect identification or dialect specific models.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1." | |
| }, | |
| { | |
| "text": "The scarcity of dialectal resources have hampered research in the area of DA, despite efforts from large institutions (ex. LDC) and programs (ex. TIDES, GALE and BOLT). Limited resources were made available for researchers with limited size and coverage. CallHome Egyptian Colloquial Arabic (ECA) 1 was the first attempt for a corpus to address this shortage released in 1997. The corpus is a collection of transcripts that cover five to ten minute segments taken from 120 unscripted telephone conversations between native speakers of Egyptian Arabic. Levantine Colloquial Arabic 2 as well as the Iraqi Arabic Conversational Telephone Speech 3 are two additional resources to cover dialectal Arabic that were built between 2004 and 2006 (Maamouri et al., 2004) . The most recent dataset in this series was BOLT Egyptian Arabic SMS/Chat and Transliteration dataset 4 , which is considerably the largest resource for dialectal Arabic (over a million words) even though it covers only Egyptian. The availability of these resources is stinted given the license requirement. On the other hand, researchers used ad-hoc resources or small datasets that were curated locally and not widely available. Graja et al. (2010) (Roth et al., 2008) , extending it to cover Egyptian as well.", | |
| "cite_spans": [ | |
| { | |
| "start": 737, | |
| "end": 760, | |
| "text": "(Maamouri et al., 2004)", | |
| "ref_id": "BIBREF13" | |
| }, | |
| { | |
| "start": 1193, | |
| "end": 1212, | |
| "text": "Graja et al. (2010)", | |
| "ref_id": "BIBREF9" | |
| }, | |
| { | |
| "start": 1213, | |
| "end": 1232, | |
| "text": "(Roth et al., 2008)", | |
| "ref_id": "BIBREF16" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Background", | |
| "sec_num": "2." | |
| }, | |
| { | |
| "text": "The emergence of social media platforms and their support to languages other than Latin -bidirectional/left-to-right or right-to-left helped dialectal presence to be more apparent than ever. This created a new need for newer resource with wider coverage. Our work attempts to fill some of this gap by providing a collection of tweet data that covers 4 major Arabic dialects as well as POS annotation of the data, which is unique and a first to be open for researchers.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Background", | |
| "sec_num": "2." | |
| }, | |
| { | |
| "text": "We used the dialectal Arabic dataset described by Eldesouki et al. 2017and Samih et al. (2017b) , which includes a set of 350 tweets for four major Arabic dialects that were manually segmented. The size of the dataset is as follows:", | |
| "cite_spans": [ | |
| { | |
| "start": 75, | |
| "end": 95, | |
| "text": "Samih et al. (2017b)", | |
| "ref_id": "BIBREF18" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data Description", | |
| "sec_num": "3." | |
| }, | |
| { | |
| "text": "Dialect No of Tweets No of Words Egyptian (EGY) 350 7,481 Levantine (LEV) 350 7,221 Gulf (GLF) 350 6,767 Maghrebi (MGR) 350 6,400", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data Description", | |
| "sec_num": "3." | |
| }, | |
| { | |
| "text": "The words in the dataset were segmented in place without any modification or standardization attempts (ex. CODA (Habash et al., 2012) ), and the segmentation guidelines aimed to generate a number of segments that match the correct number of POS tags for a word.", | |
| "cite_spans": [ | |
| { | |
| "start": 112, | |
| "end": 133, | |
| "text": "(Habash et al., 2012)", | |
| "ref_id": "BIBREF11" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data Description", | |
| "sec_num": "3." | |
| }, | |
| { | |
| "text": "We used the POS tagset described by Darwish et al. (2017) which has 18 tags for MSA POS tagging, and we added 2 dialect-specific tags (namely PROG PART, and NEG PART), and 4 tweet-specific tags (namely HASH, EMOT, MENTION, and URL). Segmentation and POS tagging were applied on the original raw text without any correction as suggested by Eldesouki et al. (2017) to overcome the need for standardization of different dialectal writings proposed in CODA by Habash et al. (2012) . For example the word (wmbyqwlw$) \"and they are not saying\" is segmented as (w+m+b+yqwl+w+$) and tagged as: CONJ+PART+PROG PART+V+PRON+NEG PART. Words are white-space and punctuation separated while hashtags, emotions, mentions and URL's are considered as single words without internal segmentation. Data is formatted in CoNLL format: Words are split into tokens (clitics), and POS is provided for each token. In our annotation scheme, tokens, words, and sentences are separated by token boundary tag (TB), word boundary tag (WB), and end of sentence tag (EOS) respectively as shown in Table 2 . Tagging was performed by a native speaker for each dialect. Then, multiple rounds of quality control and revision were performed to obtain high accuracy and consistency across dialects.", | |
| "cite_spans": [ | |
| { | |
| "start": 36, | |
| "end": 57, | |
| "text": "Darwish et al. (2017)", | |
| "ref_id": "BIBREF6" | |
| }, | |
| { | |
| "start": 339, | |
| "end": 362, | |
| "text": "Eldesouki et al. (2017)", | |
| "ref_id": "BIBREF8" | |
| }, | |
| { | |
| "start": 456, | |
| "end": 476, | |
| "text": "Habash et al. (2012)", | |
| "ref_id": "BIBREF11" | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 1064, | |
| "end": 1071, | |
| "text": "Table 2", | |
| "ref_id": "TABREF3" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Data Description", | |
| "sec_num": "3." | |
| }, | |
| { | |
| "text": "Index Token POS 0 (b) (present cont. particle) PROG PART 0 TB TB 0 (Hb) \"I love\" Figure 1 compares the distribution of POS tags in the four dialects against a sample of 350 MSA sentences from ATB with similar number of words (7,385 words). From this figure, some interesting observations can be made. For example, the four dialects are generally similar to each other in their POS distribution, while MSA shows substantial divergence. For example, nouns, adjectives, prepositions, numbers, and definite articles appear more frequently in MSA than in dialects, while on the other hand dialects show higher frequency of verbs, pronouns and particles. Our justification for this noticeable disparity is that the POS distribution is affected by the genre. The MSA text is from the formal news domain with a special focus on facts and entities, while the dialects are informal expressions with a focus on events, attitudes, and conversations. Another observa-tion is that MSA has more noun suffixes and grammatical case endings, while dialects have more progressive particles and negation suffixes. This variance is related more to the linguistic nature of the language rather than the genre.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 81, | |
| "end": 89, | |
| "text": "Figure 1", | |
| "ref_id": "FIGREF1" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Data Description", | |
| "sec_num": "3." | |
| }, | |
| { | |
| "text": "V 0 WB WB 1 (AsmE) \"I listen\" V 1 WB WB .. .. .. n EOS EOS", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data Description", | |
| "sec_num": "3." | |
| }, | |
| { | |
| "text": "For the experiments that we conducted, we used the CRF++ implementation of a CRF sequence labeler with L2 regularization and default value of 10 for the generalization parameter \"C\". We conducted three sets of experiments.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Experimental Setup", | |
| "sec_num": "4.1." | |
| }, | |
| { | |
| "text": "In the first, we tested the effectiveness of different features including different size contexts, metatypes, and stem templates, which we describe later.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Experimental Setup", | |
| "sec_num": "4.1." | |
| }, | |
| { | |
| "text": "In the second, we used the training and dev parts for each split for every dialect for training and then we used the test parts for all dialects testing. In these experiments, we were interested in knowing the POS tagging effectiveness when training and testing data are from the same dialect or from different dialects. High results for cross dialect training and testing may indicate closeness between dialects.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Experimental Setup", | |
| "sec_num": "4.1." | |
| }, | |
| { | |
| "text": "In the third set of experiments, we trained on all the train and dev parts of all the dialects jointly and tested on test sets of all the dialects. We were interested in determining if POS tagging from one dialect can benefit from added training data from other dialects. For our experiments, given a sequence of clitics", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Experimental Setup", | |
| "sec_num": "4.1." | |
| }, | |
| { | |
| "text": "c n ...c \u22122 , c \u22121 , c 0 , c 1 , c 2 ...c m ,", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Experimental Setup", | |
| "sec_num": "4.1." | |
| }, | |
| { | |
| "text": "where we assumed perfect segmentation, we used the following features for each clitic (c 0 ): \u2022 Clitic metatypes. We defined a set of 10 \"metatypes\" that we heuristically determined. They include: Hashtag (if clitic starts with \"#\"); Mention (if clitic starts with \"@\"); URL (if clitic starts with \"http\"); Emoticon/emoji (if it appears in a list of 2,730 emoticons/emojis that we constructed); Retweet (if the clitic is \"RT\"); Foreign (if it contains non-Arabic letters); Number (if it matches Arabic or Hindi numerals or a gazetteer of written out numbers that we obtained from Farasa (Abdelali et al., 2016) ); Punctuation (if it matches punctuations in the UTF8 codepage); Arabic (if it contains Arabic letters only); and Other for all other clitics. Using metatypes was shown to be effective for MSA POS tagging (Darwish et al., 2017) .", | |
| "cite_spans": [ | |
| { | |
| "start": 587, | |
| "end": 610, | |
| "text": "(Abdelali et al., 2016)", | |
| "ref_id": "BIBREF0" | |
| }, | |
| { | |
| "start": 817, | |
| "end": 839, | |
| "text": "(Darwish et al., 2017)", | |
| "ref_id": "BIBREF6" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Experimental Setup", | |
| "sec_num": "4.1." | |
| }, | |
| { | |
| "text": "\u2022", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Experimental Setup", | |
| "sec_num": "4.1." | |
| }, | |
| { | |
| "text": "\u2022 Clitic stem templates. Arabic words are typically derived from a closed set of roots that are placed in socalled stem templates to generate stems. For example, the root ktb can be fit in the template CCAC to generate the stem ktAb (book). Stem templates may overwhelmingly have one POS tag (e.g., yCCC is overwhelmingly a V) or favor one tag over another (e.g., CCAC is more likely a NOUN than an ADJ). This was shown to be effective for MSA POS tagging (Darwish et al., 2017), and we were curious to see if this would be effective for dialects also, particularly given the overlap between MSA and dialectal Arabic. We used Farasa to determine stem templates (Abdelali et al., 2016) .", | |
| "cite_spans": [ | |
| { | |
| "start": 661, | |
| "end": 684, | |
| "text": "(Abdelali et al., 2016)", | |
| "ref_id": "BIBREF0" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Experimental Setup", | |
| "sec_num": "4.1." | |
| }, | |
| { | |
| "text": "For all the experiments, we trained on the training and dev parts and tested on the test part. As mentioned earlier, we also randomly selected 350 MSA sentences from Arabic Penn Treebank (ATB) and treated MSA as a language variety. Doing so would allow us to observe the divergence of dialects from MSA and the relative effectiveness of using a small dataset compared to much more data.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Experimental Setup", | |
| "sec_num": "4.1." | |
| }, | |
| { | |
| "text": "We conducted the the following sets of experiments: Set 1: In this set, we examined the effectiveness of the different features by when training and testing on the same dialect. We evaluated the following feature sets:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Evaluation", | |
| "sec_num": "4.2." | |
| }, | |
| { | |
| "text": "\u2022 Baseline (BL): clitic n-grams only, where we used the aforementioned combination of clitic unigrams, bigrams, and trigrams.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Evaluation", | |
| "sec_num": "4.2." | |
| }, | |
| { | |
| "text": "\u2022 Baseline + stem template (+ST)", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Evaluation", | |
| "sec_num": "4.2." | |
| }, | |
| { | |
| "text": "\u2022 Baseline + metatype (+MT)", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Evaluation", | |
| "sec_num": "4.2." | |
| }, | |
| { | |
| "text": "\u2022 A combination of clitic unigrams and bigrams + stem template + metatype (+ST+MT (2g))", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Evaluation", | |
| "sec_num": "4.2." | |
| }, | |
| { | |
| "text": "\u2022 A combination of clitic unigrams, bigrams, and trigrams + stem template + metatype (+ST+MT (3g))", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Evaluation", | |
| "sec_num": "4.2." | |
| }, | |
| { | |
| "text": "\u2022 A combination of clitic unigrams, bigrams, trigrams, and 4-grams + stem template + metatype (+ST+MT (4g)) Table 3 reports on the word-level accuracy for the different experiments.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 108, | |
| "end": 115, | |
| "text": "Table 3", | |
| "ref_id": "TABREF5" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Evaluation", | |
| "sec_num": "4.2." | |
| }, | |
| { | |
| "text": "To demonstrate the difference between word-level accuracy, which we report here, and clitic-level accuracy, consider the phrase (h+ylEb b+Al+kwr+m -\"he will play with the ball\") with correct POS tags FU-TURE PART+V PREP+DET+NOUN+NSUFF. If tagger erroneously tagged the phrase as FUTURE PART+V PREP+DET+ADJ+NSUFF, then word-level accuracy would be 1/2 while clitic level accuracy would be 5/6. Since we used 5-fold cross validation, we report on the average across all folds. As the results show, using clitic n-grams only yielded the lowest results. Using stem template and metatype features improved results over using clitic n-grams alone with the combination of both features leading to even greater gain. When combined with stem template and metatype features, a combination of clitic unigrams and bigrams (2g) yielded the best results edging the use of higher order n-grams. Set 2: Next, we were interested in determining crossdialect training results to see if dialects can learn from each other and whether models from one dialect can generalize to other dialects. For all experiments, we used a combination of clitic unigrams and bigrams with stem template and metatype features (+ST+MT (2g)). Table 4 reports on cross-dialect results. Not surprisingly, the best results for each dialect were obtained when the training and test sets were from the same dialect. Among the different dialects (excluding MSA), Maghrebi suffered the most when the training set was from another dialect, and conversely training on Maghrebi yielded the worst results for all the other dialects. This may indicate that Egyptian, Levantine, and Gulf were closer together and Maghrebi was most dissimilar to all of them. Further, training on MSA and testing on dialects yielded significantly lower results compared to training on dialects and testing on MSA. This may imply that the affixes that we observed in dialects are a superset of those observed in MSA. Set 3: Lastly, we were curious to see if dialects can benefit from the addition of data from other dialects during training. Thus we combined the training and dev parts for all dialects and tested on the test part of each dialect. Table 5 reports on the results of joint learning with and without the inclusion of MSA. As the results show, Egyptian and Levantine benefited from the additional training data, while Maghrebi, Gulf, and MSA did not. The difference in accuracy (either positive or negative) ranged between 0.2% and 0.7%. On average across dialects only, the addition of MSA data marginally affected POS tagging effectiveness for different dialects. We suspect that if we use MSA tweets, instead of news sentences that we obtained from ATB, to match the genre of the dialect data would lead to greater improvement. Having a joint model that performs at par or better than dialect specific models across dialects is highly advantageous as it would avoid the need for dialect identification. The results show that a joint model may outperform dialect-specific models.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 1202, | |
| "end": 1209, | |
| "text": "Table 4", | |
| "ref_id": "TABREF6" | |
| }, | |
| { | |
| "start": 2175, | |
| "end": 2183, | |
| "text": "Table 5", | |
| "ref_id": "TABREF8" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Evaluation", | |
| "sec_num": "4.2." | |
| }, | |
| { | |
| "text": "0 10 20 N O U N D E T N S U F F P R E P A D J P U N C V P R O N C O N J P A R T C A S E N U M F O R E I G N A D V F U T _ P A R T P R O G _ P A R T N E G _ P A R T H A S H E M O T M E N T I O N U R L Ratio %", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Evaluation", | |
| "sec_num": "4.2." | |
| }, | |
| { | |
| "text": "To evaluate the strengths and weaknesses of our system, we analyzed the predicted results and the various types of errors made by the system. For such, we assessed the top 10 error types for each dialect and MSA for which correct POS of whole words is different than the guessed POS by the system. These errors represent 74%, 70%, 78%, 72% and 85% of all errors in EGY, LEV, GLF, MGR and MSA respectively. Next, we compiled together these errors and sorted them according to their average. Results are shown in Figure 1 . to incorrect classification of nouns as verbs or adjectives and vice versa. This ratio increases to 65% in MSA. Table 6 shows examples of the top error types across all dialects and MSA which represent 71% of all errors. Some of these errors are due to the fact that the words were not seen in training data. However, many of these words, such as (mnTqyp -\"logical\") and ($wArE -\"streets\"), are words that overlap between MSA and dialects and would exist in a large MSA corpus, such as ATB. Habash and Rambow (2006) and Mubarak (2017) reported an overlap of 60% for LEV verbs and 66% for EGY respectively. This suggests that domain adaptation with MSA data or using word embeddings that are trained on a large Arabic corpus would help overcome such errors.", | |
| "cite_spans": [ | |
| { | |
| "start": 1013, | |
| "end": 1037, | |
| "text": "Habash and Rambow (2006)", | |
| "ref_id": "BIBREF10" | |
| }, | |
| { | |
| "start": 1042, | |
| "end": 1056, | |
| "text": "Mubarak (2017)", | |
| "ref_id": "BIBREF14" | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 511, | |
| "end": 519, | |
| "text": "Figure 1", | |
| "ref_id": "FIGREF1" | |
| }, | |
| { | |
| "start": 634, | |
| "end": 641, | |
| "text": "Table 6", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Error Analysis", | |
| "sec_num": "5." | |
| }, | |
| { | |
| "text": "Another source of errors is due to the lack of writing standards and handling words without correcting their spelling mistakes as in words (w<sm -\"and name\") and (m>sAh -\"tragedy\"). ", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Error Analysis", | |
| "sec_num": "5." | |
| }, | |
| { | |
| "text": "In this paper, we introduced a new dataset for POS tagging of four major Arabic dialects that we constructed from tweets. We plan to provide the data freely to the community including our training, dev, and test splits. We also built POS taggers for the four dialects using a CRF sequence labeler using clitic n-gram features, stem templates, and clitic metatypes. Further, we show that we can train a joint model using data from all the dialects to train a POS tagger with comparable results to mono-dialectal training and testing, alleviating the need for dialect identification prior to POS tagging. For future work, we plan to explore two distinct directions, namely:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusion and Future Work", | |
| "sec_num": "6." | |
| }, | |
| { | |
| "text": "\u2022 the use Brown clusters (Brown et al., 1992) . Brown clustering is a hierarchical clustering of words based on their context and produces a kind of word embeddings that can be learned from large unlabeled texts. The rationale for using it here is that similar words, particularly those that share the same POS tag, tend to appear in similar contexts (Owoputi et al., 2013; Stratos and Collins, 2015 ).", | |
| "cite_spans": [ | |
| { | |
| "start": 25, | |
| "end": 45, | |
| "text": "(Brown et al., 1992)", | |
| "ref_id": "BIBREF3" | |
| }, | |
| { | |
| "start": 351, | |
| "end": 373, | |
| "text": "(Owoputi et al., 2013;", | |
| "ref_id": "BIBREF15" | |
| }, | |
| { | |
| "start": 374, | |
| "end": 399, | |
| "text": "Stratos and Collins, 2015", | |
| "ref_id": "BIBREF19" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusion and Future Work", | |
| "sec_num": "6." | |
| }, | |
| { | |
| "text": "\u2022 the use of Deep Neural Networks (DNN). DNNs have the advantage of alleviating the need for specific feature engineering including long distance relationships. Further, character-level models may be able to learning morphological patterns automatically.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusion and Future Work", | |
| "sec_num": "6." | |
| }, | |
| { | |
| "text": "https://catalog.ldc.upenn.edu/LDC97T19 2 https://catalog.ldc.upenn.edu/LDC2005S14 3 https://catalog.ldc.upenn.edu/LDC2006T16", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| } | |
| ], | |
| "back_matter": [], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "title": "Farasa: A fast and furious segmenter for arabic", | |
| "authors": [ | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Abdelali", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Darwish", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "N", | |
| "middle": [], | |
| "last": "Durrani", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Mubarak", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2016, | |
| "venue": "HLT-NAACL Demos", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "11--16", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Abdelali, A., Darwish, K., Durrani, N., and Mubarak, H. (2016). Farasa: A fast and furious segmenter for arabic. In HLT-NAACL Demos, pages 11-16.", | |
| "links": null | |
| }, | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "title": "Mining the web for the induction of a dialectical arabic lexicon", | |
| "authors": [ | |
| { | |
| "first": "R", | |
| "middle": [], | |
| "last": "Al-Sabbagh", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "R", | |
| "middle": [], | |
| "last": "Girju", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2010, | |
| "venue": "Proceedings of the Seventh International Conference on Language Resources and Evaluation (LREC 2010)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "288--293", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Al-Sabbagh, R. and Girju, R. (2010). Mining the web for the induction of a dialectical arabic lexicon. In Proceed- ings of the Seventh International Conference on Lan- guage Resources and Evaluation (LREC 2010), pages 288-293.", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "A multidialectal parallel corpus of arabic", | |
| "authors": [ | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Bouamor", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "N", | |
| "middle": [], | |
| "last": "Habash", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Oflazer", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2014, | |
| "venue": "Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC 2014)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "1240--1245", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Bouamor, H., Habash, N., and Oflazer, K. (2014). A multi- dialectal parallel corpus of arabic. In Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC 2014), pages 1240-1245.", | |
| "links": null | |
| }, | |
| "BIBREF3": { | |
| "ref_id": "b3", | |
| "title": "Class-based n-gram models of natural language", | |
| "authors": [ | |
| { | |
| "first": "P", | |
| "middle": [ | |
| "F" | |
| ], | |
| "last": "Brown", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "P", | |
| "middle": [ | |
| "V" | |
| ], | |
| "last": "Desouza", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "R", | |
| "middle": [ | |
| "L" | |
| ], | |
| "last": "Mercer", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "V", | |
| "middle": [ | |
| "J D" | |
| ], | |
| "last": "Pietra", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [ | |
| "C" | |
| ], | |
| "last": "Lai", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1992, | |
| "venue": "Computational linguistics", | |
| "volume": "18", | |
| "issue": "4", | |
| "pages": "467--479", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Brown, P. F., Desouza, P. V., Mercer, R. L., Pietra, V. J. D., and Lai, J. C. (1992). Class-based n-gram models of natural language. Computational linguistics, 18(4):467- 479.", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "title": "A multidialect, multi-genre corpus of informal written arabic", | |
| "authors": [ | |
| { | |
| "first": "R", | |
| "middle": [], | |
| "last": "Cotterell", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "C", | |
| "middle": [], | |
| "last": "Callison-Burch", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2014, | |
| "venue": "Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC 2014)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "241--245", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Cotterell, R. and Callison-Burch, C. (2014). A multi- dialect, multi-genre corpus of informal written arabic. In Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC 2014), pages 241-245.", | |
| "links": null | |
| }, | |
| "BIBREF5": { | |
| "ref_id": "b5", | |
| "title": "Verifiably effective arabic dialect identification", | |
| "authors": [ | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Darwish", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Sajjad", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Mubarak", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2014, | |
| "venue": "EMNLP", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "1465--1468", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Darwish, K., Sajjad, H., and Mubarak, H. (2014). Verifi- ably effective arabic dialect identification. In EMNLP, pages 1465-1468.", | |
| "links": null | |
| }, | |
| "BIBREF6": { | |
| "ref_id": "b6", | |
| "title": "Arabic pos tagging: Dong't abandon feature engineering just yet", | |
| "authors": [ | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Darwish", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Mubarak", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Abdelali", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Eldesouki", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2017, | |
| "venue": "WANLP", | |
| "volume": "2017", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Darwish, K., Mubarak, H., Abdelali, A., and Eldesouki, M. (2017). Arabic pos tagging: Dong't abandon feature en- gineering just yet. WANLP 2017 (co-located with EACL 2017), page 130.", | |
| "links": null | |
| }, | |
| "BIBREF7": { | |
| "ref_id": "b7", | |
| "title": "Pos tagging of dialectal arabic: a minimally supervised approach", | |
| "authors": [ | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Duh", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Kirchhoff", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2005, | |
| "venue": "Proceedings of the acl workshop on computational approaches to semitic languages", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "55--62", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Duh, K. and Kirchhoff, K. (2005). Pos tagging of dialectal arabic: a minimally supervised approach. In Proceed- ings of the acl workshop on computational approaches to semitic languages, pages 55-62. Association for Com- putational Linguistics.", | |
| "links": null | |
| }, | |
| "BIBREF8": { | |
| "ref_id": "b8", | |
| "title": "Arabic multi-dialect segmentation: bi-lstm-crf vs", | |
| "authors": [ | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Eldesouki", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Samih", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Abdelali", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Attia", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Mubarak", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Darwish", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Kallmeyer", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2017, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": { | |
| "arXiv": [ | |
| "arXiv:1708.05891v1" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Eldesouki, M., Samih, Y., Abdelali, A., Attia, M., Mubarak, H., Darwish, K., and Kallmeyer, L. (2017). Ara- bic multi-dialect segmentation: bi-lstm-crf vs. svm. arXiv:1708.05891v1 [cs.CL].", | |
| "links": null | |
| }, | |
| "BIBREF9": { | |
| "ref_id": "b9", | |
| "title": "Lexical study of a spoken dialogue corpus in tunisian dialect", | |
| "authors": [ | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Graja", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Jaoua", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Hadrich", | |
| "middle": [], | |
| "last": "Belguith", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2010, | |
| "venue": "The international arab conference on information technology (ACIT)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "benghazi--libya", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Graja, M., Jaoua, M., and Hadrich Belguith, L. (2010). Lexical study of a spoken dialogue corpus in tunisian dialect. In The international arab conference on infor- mation technology (ACIT), benghazi-libya.", | |
| "links": null | |
| }, | |
| "BIBREF10": { | |
| "ref_id": "b10", | |
| "title": "Magead: a morphological analyzer and generator for the arabic dialects", | |
| "authors": [ | |
| { | |
| "first": "N", | |
| "middle": [], | |
| "last": "Habash", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "O", | |
| "middle": [], | |
| "last": "Rambow", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2006, | |
| "venue": "Proceedings of the 21st International Conference on Computational Linguistics and the 44th annual meeting of the Association for Computational Linguistics", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "681--688", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Habash, N. and Rambow, O. (2006). Magead: a mor- phological analyzer and generator for the arabic dialects. In Proceedings of the 21st International Conference on Computational Linguistics and the 44th annual meeting of the Association for Computational Linguistics, pages 681-688. Association for Computational Linguistics.", | |
| "links": null | |
| }, | |
| "BIBREF11": { | |
| "ref_id": "b11", | |
| "title": "Conventional orthography for dialectal arabic", | |
| "authors": [ | |
| { | |
| "first": "N", | |
| "middle": [], | |
| "last": "Habash", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [ | |
| "T" | |
| ], | |
| "last": "Diab", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "O", | |
| "middle": [], | |
| "last": "Rambow", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2012, | |
| "venue": "LREC", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "711--718", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Habash, N., Diab, M. T., and Rambow, O. (2012). Conven- tional orthography for dialectal arabic. In LREC, pages 711-718.", | |
| "links": null | |
| }, | |
| "BIBREF12": { | |
| "ref_id": "b12", | |
| "title": "Morphological analysis and disambiguation for dialectal arabic", | |
| "authors": [ | |
| { | |
| "first": "N", | |
| "middle": [], | |
| "last": "Habash", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "R", | |
| "middle": [], | |
| "last": "Roth", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "O", | |
| "middle": [], | |
| "last": "Rambow", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "R", | |
| "middle": [], | |
| "last": "Eskander", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "N", | |
| "middle": [], | |
| "last": "Tomeh", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2013, | |
| "venue": "Hlt-Naacl", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "426--432", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Habash, N., Roth, R., Rambow, O., Eskander, R., and Tomeh, N. (2013). Morphological analysis and disam- biguation for dialectal arabic. In Hlt-Naacl, pages 426- 432.", | |
| "links": null | |
| }, | |
| "BIBREF13": { | |
| "ref_id": "b13", | |
| "title": "Dialectal arabic telephone speech corpus: Principles, tool design, and transcription conventions", | |
| "authors": [ | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Maamouri", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Buckwalter", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "C", | |
| "middle": [], | |
| "last": "Cieri", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2004, | |
| "venue": "NEMLAR International Conference on Arabic Language Resources and Tools", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "22--23", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Maamouri, M., Buckwalter, T., and Cieri, C. (2004). Di- alectal arabic telephone speech corpus: Principles, tool design, and transcription conventions. In NEMLAR In- ternational Conference on Arabic Language Resources and Tools, Cairo, pages 22-23.", | |
| "links": null | |
| }, | |
| "BIBREF14": { | |
| "ref_id": "b14", | |
| "title": "Analysis and quantitative study of egyptian dialect on twitter", | |
| "authors": [ | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Mubarak", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2017, | |
| "venue": "17TH INTERNATIONAL CONFERENCE ON WEB ENGINEERING", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Mubarak, H. (2017). Analysis and quantitative study of egyptian dialect on twitter. In 17TH INTERNATIONAL CONFERENCE ON WEB ENGINEERING. ICWE 2017.", | |
| "links": null | |
| }, | |
| "BIBREF15": { | |
| "ref_id": "b15", | |
| "title": "Improved part-ofspeech tagging for online conversational text with word clusters", | |
| "authors": [ | |
| { | |
| "first": "O", | |
| "middle": [], | |
| "last": "Owoputi", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "B", | |
| "middle": [], | |
| "last": "O'connor", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "C", | |
| "middle": [], | |
| "last": "Dyer", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Gimpel", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "N", | |
| "middle": [], | |
| "last": "Schneider", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "N", | |
| "middle": [ | |
| "A" | |
| ], | |
| "last": "Smith", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2013, | |
| "venue": "Proceedings of NAACL-HLT 2013", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "380--390", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Owoputi, O., O'Connor, B., Dyer, C., Gimpel, K., Schnei- der, N., and Smith, N. A. (2013). Improved part-of- speech tagging for online conversational text with word clusters. In Proceedings of NAACL-HLT 2013, pages 380-390. Association for Computational Linguistics.", | |
| "links": null | |
| }, | |
| "BIBREF16": { | |
| "ref_id": "b16", | |
| "title": "Arabic morphological tagging, diacritization, and lemmatization using lexeme models and feature ranking", | |
| "authors": [ | |
| { | |
| "first": "R", | |
| "middle": [], | |
| "last": "Roth", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "O", | |
| "middle": [], | |
| "last": "Rambow", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "N", | |
| "middle": [], | |
| "last": "Habash", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Diab", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "C", | |
| "middle": [], | |
| "last": "Rudin", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2008, | |
| "venue": "Proceedings of the Conference of American Association for Computational Linguistics (ACL08)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Roth, R., Rambow, O., Habash, N., Diab, M., and Rudin, C. (2008). Arabic morphological tagging, diacritiza- tion, and lemmatization using lexeme models and feature ranking. Proceedings of the Conference of American As- sociation for Computational Linguistics (ACL08).", | |
| "links": null | |
| }, | |
| "BIBREF17": { | |
| "ref_id": "b17", | |
| "title": "A neural architecture for dialectal arabic segmentation", | |
| "authors": [ | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Samih", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Attia", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Eldesouki", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Mubarak", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Abdelali", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Kallmeyer", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Darwish", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2017, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Samih, Y., Attia, M., Eldesouki, M., Mubarak, H., Abdelali, A., Kallmeyer, L., and Darwish, K. (2017a). A neural architecture for dialectal arabic segmentation. WANLP 2017 (co-located with EACL 2017), page 46.", | |
| "links": null | |
| }, | |
| "BIBREF18": { | |
| "ref_id": "b18", | |
| "title": "Learning from relatives: Unified dialectal arabic segmentation", | |
| "authors": [ | |
| { | |
| "first": "Y", | |
| "middle": [], | |
| "last": "Samih", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Eldesouki", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Attia", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Darwish", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Abdelali", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Mubarak", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Kallmeyer", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2017, | |
| "venue": "Proceedings of the 21st Conference on Computational Natural Language Learning", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "432--441", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Samih, Y., Eldesouki, M., Attia, M., Darwish, K., Abdelali, A., Mubarak, H., and Kallmeyer, L. (2017b). Learning from relatives: Unified dialectal arabic segmentation. In Proceedings of the 21st Conference on Computational Natural Language Learning (CoNLL 2017), pages 432- 441.", | |
| "links": null | |
| }, | |
| "BIBREF19": { | |
| "ref_id": "b19", | |
| "title": "Simple semisupervised pos tagging", | |
| "authors": [ | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Stratos", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Collins", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2015, | |
| "venue": "VS@ HLT-NAACL", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "79--87", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Stratos, K. and Collins, M. (2015). Simple semi- supervised pos tagging. In VS@ HLT-NAACL, pages 79- 87.", | |
| "links": null | |
| }, | |
| "BIBREF20": { | |
| "ref_id": "b20", | |
| "title": "The arabic online commentary dataset: an annotated dataset of informal arabic with high dialectal content", | |
| "authors": [ | |
| { | |
| "first": "O", | |
| "middle": [ | |
| "F" | |
| ], | |
| "last": "Zaidan", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "C", | |
| "middle": [], | |
| "last": "Callison-Burch", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2011, | |
| "venue": "Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies: short papers", | |
| "volume": "2", | |
| "issue": "", | |
| "pages": "37--41", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Zaidan, O. F. and Callison-Burch, C. (2011). The arabic online commentary dataset: an annotated dataset of in- formal arabic with high dialectal content. In Proceed- ings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technolo- gies: short papers-Volume 2, pages 37-41. Association for Computational Linguistics.", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "FIGREF0": { | |
| "type_str": "figure", | |
| "uris": null, | |
| "text": "Clitic n-grams. a combination of clitic unigram features {c \u22121 ;c 0 ;c 1 }, bigram features {c \u2212", | |
| "num": null | |
| }, | |
| "FIGREF1": { | |
| "type_str": "figure", | |
| "uris": null, | |
| "text": "Distribution of POS tags per dialect and MSA. Tags are ordered according to average usage in MSA and DA", | |
| "num": null | |
| }, | |
| "FIGREF2": { | |
| "type_str": "figure", | |
| "uris": null, | |
| "text": "Distribution of Error types", | |
| "num": null | |
| }, | |
| "TABREF2": { | |
| "text": "Dialect-specific and tweet-specific POS tags", | |
| "content": "<table/>", | |
| "html": null, | |
| "num": null, | |
| "type_str": "table" | |
| }, | |
| "TABREF3": { | |
| "text": "Data format for segmentation and POS tagging", | |
| "content": "<table/>", | |
| "html": null, | |
| "num": null, | |
| "type_str": "table" | |
| }, | |
| "TABREF4": { | |
| "text": "On the average, 50% of all errors in DA are due", | |
| "content": "<table><tr><td colspan=\"3\">MSA EGY LEV GLF MGR</td></tr><tr><td>BL 90.0</td><td>88.8</td><td>80.8 81.9 81.7</td></tr><tr><td>+ST 92.8</td><td>91.4</td><td>84.3 85.9 84.2</td></tr><tr><td>+MT 90.7</td><td>90.3</td><td>84.5 83.5 86.1</td></tr><tr><td>+ST+MT (2g) 93.6</td><td>92.9</td><td>87.9 87.8 88.3</td></tr><tr><td>+ST+MT (3g) 93.1</td><td>92.4</td><td>87.5 87.3 88.0</td></tr><tr><td>+ST+MT (4g) 92.5</td><td>91.6</td><td>86.8 86.6 87.2</td></tr></table>", | |
| "html": null, | |
| "num": null, | |
| "type_str": "table" | |
| }, | |
| "TABREF5": { | |
| "text": "", | |
| "content": "<table><tr><td colspan=\"3\">: Per dialect training: baseline (BL), stem templates (+ST),</td></tr><tr><td colspan=\"3\">and metatypes (+MT), and combined (+ST+MT) varying clitic n-</td></tr><tr><td>grams (2g, 3g, and 4g).</td><td/><td/></tr><tr><td/><td/><td>Training Set</td></tr><tr><td colspan=\"3\">Test Set MSA EGY LEV GLF MGR</td></tr><tr><td>MSA 93.6</td><td>76.1</td><td>76.1 76.9 72.7</td></tr><tr><td>EGY 54.5</td><td>92.9</td><td>74.3 78.1 72.7</td></tr><tr><td>LEV 52.0</td><td>74.7</td><td>87.9 73.5 69.8</td></tr><tr><td>GLF 58.7</td><td>78.8</td><td>76.7 87.8 74.4</td></tr><tr><td>MGR 50.8</td><td>71.1</td><td>73.1 70.1 88.3</td></tr><tr><td>Avg 61.9</td><td>78.7</td><td>77.6 77.3 75.6</td></tr></table>", | |
| "html": null, | |
| "num": null, | |
| "type_str": "table" | |
| }, | |
| "TABREF6": { | |
| "text": "Cross dialect training using clitic bigrams (2g), stem templates, and metatypes as features.", | |
| "content": "<table/>", | |
| "html": null, | |
| "num": null, | |
| "type_str": "table" | |
| }, | |
| "TABREF8": { | |
| "text": "Results of joint learning", | |
| "content": "<table/>", | |
| "html": null, | |
| "num": null, | |
| "type_str": "table" | |
| } | |
| } | |
| } | |
| } |