| { | |
| "paper_id": "R13-1001", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T14:56:37.225262Z" | |
| }, | |
| "title": "ASMA: A System for Automatic Segmentation and Morpho-Syntactic Disambiguation of Modern Standard Arabic", | |
| "authors": [ | |
| { | |
| "first": "Muhammad", | |
| "middle": [], | |
| "last": "Abdul-Mageed", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Indiana University Bloomington", | |
| "location": { | |
| "region": "IN", | |
| "country": "USA" | |
| } | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Mona", | |
| "middle": [], | |
| "last": "Diab", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "George", | |
| "location": { | |
| "addrLine": "Washington University Washington DC", | |
| "country": "USA" | |
| } | |
| }, | |
| "email": "[email protected]" | |
| }, | |
| { | |
| "first": "Sandra", | |
| "middle": [], | |
| "last": "K\u00fcbler", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Indiana University Bloomington", | |
| "location": { | |
| "region": "IN", | |
| "country": "USA" | |
| } | |
| }, | |
| "email": "[email protected]" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "In this paper, we present ASMA, a fast and efficient system for automatic segmentation and fine grained part of speech (POS) tagging of Modern Standard Arabic (MSA). ASMA performs segmentation both of agglutinative and of inflectional morphological boundaries within a word. In this work, we compare ASMA to two state of the art suites of MSA tools: AMIRA 2.1 (Diab et al., 2007; Diab, 2009) and MADA+TOKAN 3.2. (Habash et al., 2009). ASMA achieves comparable results to these two systems' state-of-theart performance. ASMA yields an accuracy of 98.34% for segmentation, and an accuracy of 96.26% for POS tagging with a rich tagset and 97.59% accuracy with an extremely reduced tagset.", | |
| "pdf_parse": { | |
| "paper_id": "R13-1001", | |
| "_pdf_hash": "", | |
| "abstract": [ | |
| { | |
| "text": "In this paper, we present ASMA, a fast and efficient system for automatic segmentation and fine grained part of speech (POS) tagging of Modern Standard Arabic (MSA). ASMA performs segmentation both of agglutinative and of inflectional morphological boundaries within a word. In this work, we compare ASMA to two state of the art suites of MSA tools: AMIRA 2.1 (Diab et al., 2007; Diab, 2009) and MADA+TOKAN 3.2. (Habash et al., 2009). ASMA achieves comparable results to these two systems' state-of-theart performance. ASMA yields an accuracy of 98.34% for segmentation, and an accuracy of 96.26% for POS tagging with a rich tagset and 97.59% accuracy with an extremely reduced tagset.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Abstract", | |
| "sec_num": null | |
| } | |
| ], | |
| "body_text": [ | |
| { | |
| "text": "Arabic raises various challenges to natural language processing (NLP): Arabic is a morphologically rich language (Tsarfaty et al., 2010) , where significant information concerning syntactic units is expressed at the word level, which makes part of speech (POS) tagging a challenge since it involves morpho-syntactic disambiguation, including features like voice, number, gender (Diab, 2007; .", | |
| "cite_spans": [ | |
| { | |
| "start": 113, | |
| "end": 136, | |
| "text": "(Tsarfaty et al., 2010)", | |
| "ref_id": "BIBREF19" | |
| }, | |
| { | |
| "start": 378, | |
| "end": 390, | |
| "text": "(Diab, 2007;", | |
| "ref_id": "BIBREF7" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "We address the problem of full morphosyntactic disambiguation of words in context. We devise a system, ASMA, that performs both inflectional morpheme segmentation and agglutinative clitic segmentation. For example, given a surface word in context such as (wabiHasanaAtihim, Eng. 'and by their virtues') 1 , 1 For Arabic examples, we use both the Arabic script and the Buckwalter Arabic transliteration scheme (Buckwalter, 2004) .", | |
| "cite_spans": [ | |
| { | |
| "start": 307, | |
| "end": 308, | |
| "text": "1", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 409, | |
| "end": 427, | |
| "text": "(Buckwalter, 2004)", | |
| "ref_id": "BIBREF1" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "ASMA provides the following segmentation:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "wa bi Hasan aAti him, with the prefixal clitics (wa bi, Eng. 'and' 'by'), the stem (Hasan), the inflection morpheme (aAti), and the suffixal pronominal morpheme (him). ASMA then assigns each one of these resulting morphemes a POS tag. For an explanation of Arabic morphology, cf. section 2.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "The most successful approaches to date that render this level of morphological segmentation (addressing both inflectional as well as agglutinative boundaries) typically rely on employing a morphological analyzer in the process . We show that it is possible to efficiently perform full morpho-syntactic disambiguation employing language-independent methods that are not based on a morphological analyzer. Our motivation is that dependence on a morphological analyzer comes at the cost of development since such an analyzer is generally based on manually written rules and an extensive lexicon.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "ASMA performs both inflectional morpheme segmentation and agglutinative clitic segmentation, as well as fine grained POS tagging of Modern Standard Arabic (MSA). In ASMA, a segment is a stem, an inflectional affix, or a clitic. ASMA does not handle morphotactic boundaries, thereby potentially deriving stems which may not be smoothed into correct lexemic forms for the POS process. An example of the result of the segmentation in ASMA is as follows: the surface form (AlwilaAyaAt, Eng. 'the states') is segmented into + + (Al+wilaAy+aAt) where wilaAy is a stem, Al is a clitic, and At is an affixival inflectional suffix. It should be noted that wilaAy is not a valid Arabic lexeme. For ASMA to convert it into a lexeme, it would have to process the morphotactics on the stem and render it as (wilaAyap) restoring the lexeme/lemma final p.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "The remainder of the paper is structured as follows: Section 2 describes the pertinent facts about Arabic morphology. Section 3 describes related work, namely on AMIRA 2.1 and MADA+TOKAN 3.2. In section 4, we describe ASMA, the overall system, in section 5, we report results on the segmentation task, and in section 6 on the POS tagging task. In section 7 we provide an error analysis, and conclude in section 9.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "Arabic exhibits derivational, inflectional, and agglutinative morphology. Derivational morphology is mostly templatic where a word is made up of a root and a pattern, along with some idiosyncratic information. For example, a root such as (k t b) if combined with the pattern 1a2a3, where the numbers [1, 2, 3] designate the root radicals, respectively, it results in the derivational form (katab, Eng. 'to write'). Likewise for the same root when it combines with the pattern 1A2a3, it result in the word (kaAtab, Eng. 'to correspond'). All derivation forms undergo inflection reflecting various types of functional features such as voice, number, aspect, gender, grammatical case, tense, etc. The resulting word is known as a lexeme. Therefore a lexeme such as (katabat, Eng. 'she wrote') reflects feminine [gender] , singular [number], past [tense], perfective [aspect], 3rd [person] inflections for the verb. Typically, one of the fully inflected lexemes is considered a citation form, and it is known as the lemma. The choice of a specific lexeme as a citation form is a convention, and it is typically the 3rd person masculine singular perfective form for verbs and the 3rd person singular form for nouns. Hence in this case the lemma is (kataba, Eng. 'he wrote'). Arabic words often undergo clitic agglutination to form surface words. For example, the lexeme (kAtabat, Eng. 'she corresponded') could have an enclitic/suffixal pronoun as follows:", | |
| "cite_spans": [ | |
| { | |
| "start": 300, | |
| "end": 303, | |
| "text": "[1,", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 304, | |
| "end": 306, | |
| "text": "2,", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 307, | |
| "end": 309, | |
| "text": "3]", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 808, | |
| "end": 816, | |
| "text": "[gender]", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 863, | |
| "end": 885, | |
| "text": "[aspect], 3rd [person]", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Arabic Morphology", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "(kAtabathum, Eng. 'she corresponded with them'). The agglutination process results in morphotactic variations at the morpheme boundaries where the orthography is changed for the underlying lexeme. For example, in a noun such as (wabiHasanathim, Eng. 'and by their virtue'), the underlying lexeme (same as lemma in this case) is the noun (Hasanap), where the lexeme final Taa-Marbuta (p) is changed into a regular (t) when followed by a pronominal clitic. Accordingly, segmenting off agglutinative clitics without handling boundary morphotactics to restore the underlying lexeme form results in stems.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Arabic Morphology", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "3 Related Work AMIRA 2.1 Diab, 2009) is a supervised SVM-based machine learning algorithm for processing MSA, including clitic tokenization and normalization, POS tagging, and base phrase chunking. Diab. et al. adopt the inside-outside-beginning (IOB) chunk approach (Ramshaw and Marcus, 1995) for clitic tokenization, i.e., each letter in a word is labeled as being at the beginning (B), the inside (I), or the outside (O) of a chunk. Note that the tokenization by Diab et al. does not split off inflectional morphology. For example, while ASMA would segment (wbHsnAthm) into w+b+Hsn+At+hm, AMIRA 2.1 would output w+b+HsnAt+hm, i.e., it does not split off the number and gender inflectional suffix", | |
| "cite_spans": [ | |
| { | |
| "start": 25, | |
| "end": 36, | |
| "text": "Diab, 2009)", | |
| "ref_id": "BIBREF8" | |
| }, | |
| { | |
| "start": 267, | |
| "end": 293, | |
| "text": "(Ramshaw and Marcus, 1995)", | |
| "ref_id": "BIBREF16" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Arabic Morphology", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "At from the stem (Hsn).", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Arabic Morphology", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "One advantage of ASMA over AMIRA 2.1 is thus that ASMA identifies inflectional morpheme boundaries. Similar to AMIRA 2.1, ASMA employs an IOB chunking approach on the character level for segmentation of words into morphemic chunks (clitics, stems, and inflectional affixes). AMIRA 2.1 achieves an F-measure of 99.15% for the entire word being segmented correctly. AMIRA 2.1 also performs POS tagging. It uses multiple POS tagsets ranging from a basic 24 tagset called Reduced TagSet (RTS) to an enriched tagset (ERTS) of 75 tags. AMIRA 2.1. achieves an accuracy of 96.6% for RTS and 96.13% for ERTS. ASMA, in contrast, uses a fuller tagset of 139 POS tags, which includes morphological information, e.g., on gender and number. MADA+TOKAN 3.2 Habash et al. Habash and Rambow (2005; developed MADA, a system for the morphological disambiguation of MSA. MADA relies on the output of the Buckwalter Arabic Morphological Analyzer (BAMA) (Buckwalter, 2004) and uses 14 individual SVM classifiers for learning individual fea-tures, which makes it computationally costly compared to ASMA which uses a single classifier for each of the two tasks of segmentation and morphological disambiguation. TOKAN, a separate tool, performs tokenization on the output of MADA. For tokenization, Habash et al. report 98.85% word level accuracy and for POS tagging, 96.1% accuracy. MADA+TOKAN 3.2 perform segmentation similar to ASMA. However, MADA+TOKAN 3.2 depend on the underlying morphological analyzer. In contrast to ASMA, MADA+TOKAN 3.2 perform POS tagging yielding the fully specified morphological analysis in the ATB, which comprises 440 unique tags.", | |
| "cite_spans": [ | |
| { | |
| "start": 756, | |
| "end": 780, | |
| "text": "Habash and Rambow (2005;", | |
| "ref_id": "BIBREF9" | |
| }, | |
| { | |
| "start": 932, | |
| "end": 950, | |
| "text": "(Buckwalter, 2004)", | |
| "ref_id": "BIBREF1" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Arabic Morphology", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "For both segmentation and POS tagging, we use memory-based learning (MBL) (Aha et al., 1991) classifiers. MBL is a lazy learning method that does not abstract rules from the data, but rather keeps all training data. During training, the learner stores the training instances without abstraction. Given a new instance, the classifier finds the k nearest neighbors in the training set and chooses their most frequent class for the new instance. MBL has been shown to have a suitable bias for NLP problems (Daelemans et al., 1999 ; Daelemans and van den Bosch, 2005) since it does not abstract over irregularities or subregularities. For each of the two classification tasks (i.e., segmentation and POS tagging), we use MBT (Daelemans et al., 1996) , a memory-based POS tagger that has access to previous tagging decisions in addition to an expressive feature set.", | |
| "cite_spans": [ | |
| { | |
| "start": 74, | |
| "end": 92, | |
| "text": "(Aha et al., 1991)", | |
| "ref_id": "BIBREF0" | |
| }, | |
| { | |
| "start": 503, | |
| "end": 526, | |
| "text": "(Daelemans et al., 1999", | |
| "ref_id": "BIBREF4" | |
| }, | |
| { | |
| "start": 721, | |
| "end": 745, | |
| "text": "(Daelemans et al., 1996)", | |
| "ref_id": "BIBREF3" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "ASMA 4.1 Method: Memory-Based Learning", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "We use segmentation and POS data from the Penn Arabic Treebank (PATB) (Maamouri et al., 2004) , specifically, we use the following parts: ATB1V4, ATB2V3, ATB3V3.1 and ATB3V3.2 with different splits as described below. The textual basis of the treebank consists of newswire articles covering political, economic, cultural, sports, etc. topics. Table 1 presents for each part the number of words, the number of tokens (i.e., only clitics are split off), the number of segments (i.e., clitic and inflectional morphology is split off), the number of news reports, and the source of the reports (i.e., the news agency) 2 . As mentioned above, Arabic is generally written without diacritics. While the ATB does have a version with diacritics restored, for our experiments, we use the version without diacritics, for both segmentation and POS tagging.", | |
| "cite_spans": [ | |
| { | |
| "start": 70, | |
| "end": 93, | |
| "text": "(Maamouri et al., 2004)", | |
| "ref_id": "BIBREF14" | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 343, | |
| "end": 350, | |
| "text": "Table 1", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Data Sets and Splits", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "For a fair comparison of ASMA to both AMIRA and MADA, we adopt two different data splits, AMIRA-SPLIT and MADA-SPLIT, with each split corresponding to the data splits used in the evaluations of these systems. The same splits are used both for segmentation and POS tagging. For the AMIRA-SPLIT, we follow the procedure by Diab et al. (2004) , but we use more recent releases of the ATB than Diab et al. We split each of the first three parts into 10% development data (DEV), 80% training data (TRAIN), and 10% test data (TEST). We then concatenate the respective splits from each part. For example, to create a single DEV set from the three parts, we concatenate the 10% DEV data from ATB1V4, ATB2V3, and ATB3V3.2, etc. For MADA-SPLIT, we follow the MADA manual 3 . For this split, ATB1V4 and ATB2V3 and the first 80% of ATB3V3.1 are used as the TRAIN set, the last 20% of ATB3V3.1 are divided into two halves, i.e. DEV and TEST (each making up 10% of ATB3V3.1) respectively. The development sets are used for parameter and feature optimization.", | |
| "cite_spans": [ | |
| { | |
| "start": 321, | |
| "end": 339, | |
| "text": "Diab et al. (2004)", | |
| "ref_id": "BIBREF5" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data Sets and Splits", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "We define segmentation as an IOB classification task, where each letter in a word is tagged with a label indicating its place in a segment. The tagset is {B-SEG, I-SEG, O}, where B is a tag assigned to the beginning of a segment, I denotes the inside of a segment, and O spaces between surface form words.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Setup", | |
| "sec_num": "5.1" | |
| }, | |
| { | |
| "text": "Procedure: We performed a non-exhaustive search for optimal settings for the following MBT parameters: the MBL algorithm, the similarity metric, the feature weighting method, and the value of the k nearest neighbors. The best setting used the IB1 algorithm with weighted overlap as the similarity metric, gain ratio (GR) as a feature weighting method, and a value of k = 1.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Setup", | |
| "sec_num": "5.1" | |
| }, | |
| { | |
| "text": "Data set # words # tokens # segments # texts Source ATB1V4 145,386 167,280 209,187 734 AFP ATB2V3 144,199 169,319 221,001 501 UMMAH ATB3V3.1 340,281 402,291 551,171 600 An Nahar ATB3V3.2 339,710 402,291 512,932 599 An Nahar Table 1 : Data statistics and sources", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 45, | |
| "end": 217, | |
| "text": "Source ATB1V4 145,386 167,280 209,187 734 AFP ATB2V3 144,199 169,319 221,001 501 UMMAH ATB3V3.1 340,281 402,291 551,171 600 An Nahar ATB3V3.2 339,710 402,291", | |
| "ref_id": "TABREF2" | |
| }, | |
| { | |
| "start": 239, | |
| "end": 246, | |
| "text": "Table 1", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Setup", | |
| "sec_num": "5.1" | |
| }, | |
| { | |
| "text": "Our complete feature set comprises the six preceding characters, the previous tag decisions of all the six preceding characters except the character immediately preceding the focus character, the focus character itself and its ambiguity tag (henceforth, ambitag), and the seven following characters. For features, we tested (1) left only, right only, and left and right contexts across various window sizes and (2) different types of information, e.g., feature sets with/without previous tag decisions for left context, feature sets with/without ambitags of right context. An ambitag is a combination of all tags of the ambiguity set of a word.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Setup", | |
| "sec_num": "5.1" | |
| }, | |
| { | |
| "text": "Evaluation: We evaluate segmentation in terms of character-based accuracy, word level accuracy, and precision, recall, and F-measure for segments. For example, the word (AlwlAyAt, Eng.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Setup", | |
| "sec_num": "5.1" | |
| }, | |
| { | |
| "text": "'the states') has the correct segmentation + + Al+wlAy+At and comprises 8 characters. If it is segmented incorrectly as + (Al-wlAy+At), one of the 8 characters, the 'w' is incorrectly classified as I as opposed to B, and consequently, we have a character based accuracy of 7/8, a word level based accuracy of 0/8. On the segment level, precision is 50%, recall 33.33%, and the F-measure 41.65. Table 2 shows the results for segmentation on the two data splits, AMIRA-SPLIT and MADA-SPLIT. For both data splits, the best features are the six preceding characters, the previous tag decisions of all the six preceding characters except the character immediately preceding the focus character, the focus character itself and its ambitag, and the seven following characters.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 394, | |
| "end": 401, | |
| "text": "Table 2", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Setup", | |
| "sec_num": "5.1" | |
| }, | |
| { | |
| "text": "AMIRA-SPLIT: On the TEST data for this split, we reach an accuracy of 99.53%, a precision of 97.97%, a recall of 98.04% and an F of 98.01%. The segmentation accuracy is at 98.34% for words.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Segmentation Results", | |
| "sec_num": "5.2" | |
| }, | |
| { | |
| "text": "MADA-SPLIT: For this data set, we achieve an accuracy of 99.49%, precision of 97.72%, a recall of 97.85% and an F of 97.79%. Segmentation accuracy for words is at 98.10%", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Segmentation Results", | |
| "sec_num": "5.2" | |
| }, | |
| { | |
| "text": "These experiments show that on the segmentation level, the MADA split is slightly more complex than the AMIRA split.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Segmentation Results", | |
| "sec_num": "5.2" | |
| }, | |
| { | |
| "text": "Our segmentation results are not fully comparable to the tokenization performance of AMIRA (Diab et al., 2004 ) since AMIRA does not split off inflectional morphology. MADA (Habash and Rambow, 2005; , in contrast, does perform segmentation, but it is based on a morphological analyzer. ASMA, without the use of any external resources, achieved a word accuracy of 98.10% on the MADA-SPLIT, which is only slightly lower than MADA's 98.85% word accuracy.", | |
| "cite_spans": [ | |
| { | |
| "start": 91, | |
| "end": 109, | |
| "text": "(Diab et al., 2004", | |
| "ref_id": "BIBREF5" | |
| }, | |
| { | |
| "start": 173, | |
| "end": 198, | |
| "text": "(Habash and Rambow, 2005;", | |
| "ref_id": "BIBREF9" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Segmentation Results", | |
| "sec_num": "5.2" | |
| }, | |
| { | |
| "text": "POS tagging is defined here so that each segment, rather than a full word (as in (K\u00fcbler and Mohamed, 2012) ) or a token (as in (Diab et al., 2004) ), is assigned a POS tag. For the experiments reported here, we modify the ATB tagset such that case and mood tags are removed since those are syntactic features that cannot be determined based on a local context. While AMIRA, similar to ASMA, does not predict case and mood, MADA does at the cost of some performance loss. The remaining tagset comprises 139 segment-based tags. The input for the POS tagger consists of gold segmented data. The reasons for this decision are mainly to allow us to compare our system to AMIRA, which also uses gold segmentation.", | |
| "cite_spans": [ | |
| { | |
| "start": 81, | |
| "end": 107, | |
| "text": "(K\u00fcbler and Mohamed, 2012)", | |
| "ref_id": "BIBREF13" | |
| }, | |
| { | |
| "start": 128, | |
| "end": 147, | |
| "text": "(Diab et al., 2004)", | |
| "ref_id": "BIBREF5" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "POS Tagging", | |
| "sec_num": "6" | |
| }, | |
| { | |
| "text": "Procedure: We performed a non-exhaustive search for the best parameters described in section 5. We use the IGTREE algorithm. We identified the modified value difference metric (MVDM) as similarity metric, gain ratio (GR) as a feature weighting method, and k = 1 for known words Table 2 : Segmentation results and k = 30 for unknown words as optimal parameters. For both data splits, the following feature sets give optimal results on the DEV set: For known segments, the best feature set uses the focus segment, its ambitag, two previous segments, and the predicted tag of three previous segments. For unknown segments, the feature set consists of the five previous segments and their predicted tags, the focus segment itself and its ambitag, the first five characters and the last three characters of the focus segment, and six following segments and their ambitags.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 278, | |
| "end": 285, | |
| "text": "Table 2", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Setup", | |
| "sec_num": "6.1" | |
| }, | |
| { | |
| "text": "Evaluation: We evaluate based on segments, i.e. on the units which were used for POS tagging, rather than on full words. We report overall accuracy as well as accuracy on known segments and on unknown segments. Table 3 shows the results for POS tagging on the two data sets given the settings and the feature set described above. AMIRA-SPLIT: Using the feature set described above, we reach an accuracy of 96.61% on known words and 74.46% on unknown words, averaging 96.26% on all words.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 211, | |
| "end": 218, | |
| "text": "Table 3", | |
| "ref_id": "TABREF2" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Setup", | |
| "sec_num": "6.1" | |
| }, | |
| { | |
| "text": "MADA-SPLIT: We reach an accuracy of 94.61% on known words and of 86.00% on unknown words, averaging 94.67% on all words. In comparison, the results for unknown words are much higher. This is due to the fact that in the MADA split, we only have 593 unknown words while the AMIRA split has more than twice as many (i.e. 1261 unknown words).", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "POS Tagging Results", | |
| "sec_num": "6.2" | |
| }, | |
| { | |
| "text": "These experiments show that for POS tagging, the MADA split is considerably more challenging than the AMIRA split. This means that even if results reported for MSA are based on the same sub-word analysis, the data splits have to be taken into account in a comparison as well.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "POS Tagging Results", | |
| "sec_num": "6.2" | |
| }, | |
| { | |
| "text": "Our POS tagging results are not directly comparable to AMIRA, because of the differences in segmentation and because of the different POS tagsets. They are comparable to those obtained with MADA using tokenization by TOKAN. Roth et al. (2008) report 94.7% accuracy on predicting 10 morphological types of features, the closest setting to our tagset. This is very close to the 94.67% we report using the MADA-SPLIT. Roth et al. report a slight improvement for an extended system using diacritic markers as additional input, but as K\u00fcbler and Mohamed (2012) have shown, automatic diacritization must be extremely accurate in order to be useful for POS tagging.", | |
| "cite_spans": [ | |
| { | |
| "start": 224, | |
| "end": 242, | |
| "text": "Roth et al. (2008)", | |
| "ref_id": "BIBREF17" | |
| }, | |
| { | |
| "start": 530, | |
| "end": 555, | |
| "text": "K\u00fcbler and Mohamed (2012)", | |
| "ref_id": "BIBREF13" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "POS Tagging Results", | |
| "sec_num": "6.2" | |
| }, | |
| { | |
| "text": "We also ran experiments with two other tagsets, the standard RTS tagset, which is composed of 25 tags, and the CATiB tagset , which comprises only 6 tags, in order to investigate the effect of using different levels of morphological and morpho-syntactic information in the tagset. The full tagset, as mentioned above, includes all morphological information, except for case and mood markers. The RTS tagset is a reduced version, resulting in a tagset that is similar to the English Penn Treebank tagset (Santorini, 1990) . Using the RTS tagset also allows us to make our results more comparable to AMIRA. The CATiB tagset represents only the major word classes, such as noun or verb. We used CATiB because its tagset corresponds to traditional notions in Arabic grammar and because it was used in the Columbia Arabic Treebank .", | |
| "cite_spans": [ | |
| { | |
| "start": 503, | |
| "end": 520, | |
| "text": "(Santorini, 1990)", | |
| "ref_id": "BIBREF18" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Experimenting with Other Tagsets", | |
| "sec_num": "6.3" | |
| }, | |
| { | |
| "text": "For this set of experiments, we use the same parameters and feature settings as described in section 6.2 above. Thus, the results reported on this set of experiments are potentially suboptimal. In the future, we plan to tune the performance of ASMA with each of these tagsets. Table 4 shows the results of these experiments.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 277, | |
| "end": 284, | |
| "text": "Table 4", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Experimenting with Other Tagsets", | |
| "sec_num": "6.3" | |
| }, | |
| { | |
| "text": "AMIRA-SPLIT: Using RTS, we reach an accuracy of 96.28%. This is very slightly higher than our results for the full POS tagset (96.26%), and it is very close to AMIRA's results when using Table 4 : POS tagging results with the RTS and CATiB tagsets the RTS. But note that AMIRA uses tokenization rather than segmentation; thus the results are not directly comparable. We also notice that ASMA's performance on unknown words improves by almost 3 percent points to 77.79%, as opposed to 74.46% using the full tagset. This is to be expected since guessing the morphological information for an unknown word is more difficult than guessing only the main category in RTS. MADA-SPLIT: Here, ASMA reaches an overall accuracy of 94.06%. This is slightly lower than for the full tagset (94.67%), due to a drop in accuracy on unknown words, from 86.00% to 84.99% and a slight drop in accuracy on known words from 94.80% to 94.20%.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 187, | |
| "end": 194, | |
| "text": "Table 4", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "RTS", | |
| "sec_num": "6.3.1" | |
| }, | |
| { | |
| "text": "The results for the RTS on both data splits show that ASMA reaches state-of-the-art results, without using morphological analysis and while using a classifier not optimized for sequence handling, but which has access to previous classification decisions. The results also show that, in general, using the reduced tagset does not significantly change the difficulty of the task. In other words, giving up morphological information in the tagset in this specific case does not lead to higher tagging accuracy.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "RTS", | |
| "sec_num": "6.3.1" | |
| }, | |
| { | |
| "text": "AMIRA-SPLIT: With the CATiB tagset, ASMA reaches an overall accuracy of 97.59%, showing that an extreme reduction of the tagset to one completely devoid of morphological information increases tagging accuracy.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "CATiB", | |
| "sec_num": "6.3.2" | |
| }, | |
| { | |
| "text": "MADA Both sets of experiments show that the amount of morphological and morpho-syntactic information present in the POS tagset has an influence on the difficulty of the POS tagging step, even though the connection is not always a direct one. Thus, if ASMA is used as a preprocessing system for upstream modules, it is necessary to choose the tagset with regard to the upstream task.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "CATiB", | |
| "sec_num": "6.3.2" | |
| }, | |
| { | |
| "text": "We performed an error analysis to see which types of errors ASMA makes. Table 5 presents a confusion matrix for the ATB tagset we used in section 6.2. We provide results only with the AMIRA split, as the results for the MADA split are similar. The table is sorted based on the contribution the confusion pair makes towards the overall error rate.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 72, | |
| "end": 79, | |
| "text": "Table 5", | |
| "ref_id": "TABREF4" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Error Analysis", | |
| "sec_num": "7" | |
| }, | |
| { | |
| "text": "The table shows that because of the high number of POS labels, each confusion case contributes only marginally to the overall error rate. The most likely errors involve nouns (NOUN), proper nouns (NOUN PROP), and adjectives (ADJ). These errors can be explained via the characteristics of Arabic: Proper nouns in Arabic are generally standard nouns used as names. Thus, the same word can be used as either noun or proper noun, depending on the context. Additionally, unlike English, Arabic proper nouns are not marked by capitalization or other orthographic means. The nounadjective distinction is not clear in Arabic: Adjectives can be used as nouns, and they share the same morphological patterns as nouns.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Error Analysis", | |
| "sec_num": "7" | |
| }, | |
| { | |
| "text": "The next set concerns the POS tags PV PASS, IV PASS, and ADJ.VN. With the lack of diacritics, the classifier is prone to erring with regard to cases where diacritics play a crucial factor in carrying the grammatical function. Since passivization is marked using diacritics in Arabic, passive verbs also suffer from the lack of diacritics, both in the perfective (i.e., PV PASS) and imperfective (i.e., IV PASS) cases, and hence the misclassification and high percent of confusion between passive and active verbs in the data. Adjectival verbal nouns -i.e., ADJ.VN as in (mu'lin, Eng. 'announcing') -are also confused with adjectives as these two parts of speech have very similar contexts, especially given the lack of diacritic nunation 4 characteristic of the adjectival verbal noun.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Error Analysis", | |
| "sec_num": "7" | |
| }, | |
| { | |
| "text": "As described above, ASMA performs both inflectional morpheme segmentation and agglutinative clitic segmentation, as well as fine grained POS tagging of Modern Standard Arabic (MSA). Compared to AMIRA, ASMA performs more fine grained morphological disambiguation due to ASMA's identification of inflectional morpheme boundaries. Compared to MADA, ASMA performs the same tasks, however without using a morphological analyzer. Given that restriction, it still achieves state-of-the-art results, only minimally lower than MADA's. One major advantage of ASMA is the high speed with which it oper-ates: On a PowerPC 970 machine, with a Darwin Kernel Version 8.11.0 and 2GB memory, it takes ASMA about 5 minutes to process 100 000 words. Although we have not had the chance to compare ASMA and MADA in terms of the speed with which each operates, we believe that ASMA is significantly faster than MADA. After all, whereas MADA employs 14 individual SVM classifiers to learn individual features, ASMA employs a single classifier per task, segmentation and morpho-syntactic disambiguation. AMIRA is observably slower than ASMA. In addition, while the MBL framework in ASMA uses virtually no time to train, SVMs (which AMIRA and MADA use) are known for long training times. Its speed makes ASMA valuable especially for real-world tasks, such as information retrieval and extraction, and tasks depending on big data processing.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "ASMA in Comparison", | |
| "sec_num": "8" | |
| }, | |
| { | |
| "text": "ASMA is flexible in terms of the granularity of its output as it renders morphological disambiguation with three different tagsets (i.e., the full ATB 139 tagset, the RTS, and the reduced CATiB tagset). As such, ASMA can be customized to different NLP tasks depending on the specific needs of each task. Both AMIRA and MADA also employ different tagsets. In the context of our introduction of ASMA, we have shown how it is that performance varies according to the size of the tagset used. To the best of our knowledge, this is the first report exploiting the CATiB tagset.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "ASMA in Comparison", | |
| "sec_num": "8" | |
| }, | |
| { | |
| "text": "In this paper, we have presented ASMA, a system for automatic segmentation and morphosyntactic disambiguation of Modern Standard Arabic (MSA). We compared ASMA to the two most popular Arabic processing suites, AMIRA and MADA, and showed ASMA's advantages. ASMA has the advantages of speed as well as non-dependence on an external morphological analyzer (unlike MADA). It also identifies morpheme boundaries at a level more fine grained than AMIRA. Moreover, ASMA performs POS tagging with different degrees of granularity and hence can be customized according to an upstream task if used as a preprocessing system. For the future, we plan to investigate the utility of using a conditional random fields classifier either to complement or replace ASMA's current memory-based classifier. In addition, we will attempt to improve ASMA's performance based on our error analysis.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusion and Future Work", | |
| "sec_num": "9" | |
| }, | |
| { | |
| "text": "The information is based on the LDC documentation at http://www.ldc.upenn.edu/Catalog/ docs/ * .3 http://www1.ccls.columbia.edu/MADA", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Nunation indicates indefiniteness and refers to wordfinal diacritics occuring as a short vowel followed by an unwritten /n/ sound.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| } | |
| ], | |
| "back_matter": [], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "title": "Instance-based learning algorithms", | |
| "authors": [ | |
| { | |
| "first": "David", | |
| "middle": [], | |
| "last": "Aha", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Dennis", | |
| "middle": [], | |
| "last": "Kibler", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Marc", | |
| "middle": [ | |
| "K" | |
| ], | |
| "last": "Albert", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1991, | |
| "venue": "Machine Learning", | |
| "volume": "6", | |
| "issue": "", | |
| "pages": "37--66", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "David Aha, Dennis Kibler, and Marc K. Albert. 1991. Instance-based learning algorithms. Machine Learning, 6:37-66.", | |
| "links": null | |
| }, | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "title": "Arabic morphological analyzer version 2.0. Linguistic Data Consortium", | |
| "authors": [ | |
| { | |
| "first": "Tim", | |
| "middle": [], | |
| "last": "Buckwalter", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2004, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Tim Buckwalter. 2004. Arabic morphological ana- lyzer version 2.0. Linguistic Data Consortium.", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "Memory Based Language Processing", | |
| "authors": [ | |
| { | |
| "first": "Walter", | |
| "middle": [], | |
| "last": "Daelemans", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Antal", | |
| "middle": [], | |
| "last": "Van Den", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Bosch", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2005, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Walter Daelemans and Antal van den Bosch. 2005. Memory Based Language Processing. Cambridge University Press, Cambridge, UK.", | |
| "links": null | |
| }, | |
| "BIBREF3": { | |
| "ref_id": "b3", | |
| "title": "MBT: A memory-based part of speech tagger-generator", | |
| "authors": [ | |
| { | |
| "first": "Walter", | |
| "middle": [], | |
| "last": "Daelemans", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Jakub", | |
| "middle": [], | |
| "last": "Zavrel", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Peter", | |
| "middle": [], | |
| "last": "Berck", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Steven", | |
| "middle": [], | |
| "last": "Gillis", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1996, | |
| "venue": "Proceedings of the 4th Workshop on Very Large Corpora", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "14--27", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Walter Daelemans, Jakub Zavrel, Peter Berck, and Steven Gillis. 1996. MBT: A memory-based part of speech tagger-generator. In Proceedings of the 4th Workshop on Very Large Corpora, pages 14-27, Copenhagen, Denmark.", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "title": "Forgetting exceptions is harmful in language learning", | |
| "authors": [ | |
| { | |
| "first": "Walter", | |
| "middle": [], | |
| "last": "Daelemans", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1999, | |
| "venue": "Machine Learning", | |
| "volume": "34", | |
| "issue": "", | |
| "pages": "11--43", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Walter Daelemans, Antal van den Bosch, and Jakub Za- vrel. 1999. Forgetting exceptions is harmful in lan- guage learning. Machine Learning, 34(1-3):11-43. Special Issue on Natural Language Learning.", | |
| "links": null | |
| }, | |
| "BIBREF5": { | |
| "ref_id": "b5", | |
| "title": "Automatic tagging of Arabic text: From raw text to base phrase chunks", | |
| "authors": [ | |
| { | |
| "first": "Mona", | |
| "middle": [], | |
| "last": "Diab", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Kadri", | |
| "middle": [], | |
| "last": "Hacioglu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Daniel", | |
| "middle": [], | |
| "last": "Jurafsky", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2004, | |
| "venue": "Proceedings of the 5th Meeting of the North American Chapter of the Association for Computational Linguistics and Human Language Technologies Conference (HLT-NAACL)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "149--152", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Mona Diab, Kadri Hacioglu, and Daniel Jurafsky. 2004. Automatic tagging of Arabic text: From raw text to base phrase chunks. In Proceedings of the 5th Meeting of the North American Chapter of the As- sociation for Computational Linguistics and Human Language Technologies Conference (HLT-NAACL), pages 149-152, Boston, MA.", | |
| "links": null | |
| }, | |
| "BIBREF6": { | |
| "ref_id": "b6", | |
| "title": "Automatic processing of Modern Standard Arabic text", | |
| "authors": [ | |
| { | |
| "first": "Mona", | |
| "middle": [], | |
| "last": "Diab", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Kadri", | |
| "middle": [], | |
| "last": "Hacioglu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Dan", | |
| "middle": [], | |
| "last": "Jurafsky", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2007, | |
| "venue": "Arabic Computational Morphology", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "159--179", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Mona Diab, Kadri Hacioglu, and Dan Jurafsky. 2007. Automatic processing of Modern Standard Arabic text. Arabic Computational Morphology, pages 159-179.", | |
| "links": null | |
| }, | |
| "BIBREF7": { | |
| "ref_id": "b7", | |
| "title": "Towards an optimal POS tag set for Modern Standard Arabic processing", | |
| "authors": [ | |
| { | |
| "first": "Mona", | |
| "middle": [], | |
| "last": "Diab", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2007, | |
| "venue": "Proceedings of Recent Advances in Natural Language Processing (RANLP)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Mona Diab. 2007. Towards an optimal POS tag set for Modern Standard Arabic processing. In Proceed- ings of Recent Advances in Natural Language Pro- cessing (RANLP).", | |
| "links": null | |
| }, | |
| "BIBREF8": { | |
| "ref_id": "b8", | |
| "title": "Second generation AMIRA tools for Arabic processing: Fast and robust tokenization, POS tagging, and base phrase chunking", | |
| "authors": [ | |
| { | |
| "first": "Mona", | |
| "middle": [], | |
| "last": "Diab", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2009, | |
| "venue": "2nd International Conference on Arabic Language Resources and Tools", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Mona Diab. 2009. Second generation AMIRA tools for Arabic processing: Fast and robust tokenization, POS tagging, and base phrase chunking. In 2nd International Conference on Arabic Language Re- sources and Tools, Cairo, Egypt.", | |
| "links": null | |
| }, | |
| "BIBREF9": { | |
| "ref_id": "b9", | |
| "title": "Arabic tokenization, part-of-speech tagging and morphological disambiguation in one fell swoop", | |
| "authors": [ | |
| { | |
| "first": "Nizar", | |
| "middle": [], | |
| "last": "Habash", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Owen", | |
| "middle": [], | |
| "last": "Rambow", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2005, | |
| "venue": "Proceedings of the 43rd Annual Meeting of the Association for Computational Linguistics (ACL)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "573--580", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Nizar Habash and Owen Rambow. 2005. Arabic tok- enization, part-of-speech tagging and morphological disambiguation in one fell swoop. In Proceedings of the 43rd Annual Meeting of the Association for Computational Linguistics (ACL), pages 573-580, Ann Arbor, MI.", | |
| "links": null | |
| }, | |
| "BIBREF10": { | |
| "ref_id": "b10", | |
| "title": "CATiB: The Columbia Arabic Treebank", | |
| "authors": [ | |
| { | |
| "first": "Nizar", | |
| "middle": [], | |
| "last": "Habash", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Ryan", | |
| "middle": [], | |
| "last": "Roth", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2009, | |
| "venue": "Proceedings of the ACL-IJCNLP 2009 Conference Short Papers", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "221--224", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Nizar Habash and Ryan Roth. 2009. CATiB: The Columbia Arabic Treebank. In Proceedings of the ACL-IJCNLP 2009 Conference Short Papers, pages 221-224, Suntec, Singapore.", | |
| "links": null | |
| }, | |
| "BIBREF11": { | |
| "ref_id": "b11", | |
| "title": "MADA+TOKAN: A toolkit for Arabic tokenization, diacritization, morphological disambiguation, POS tagging, stemming and lemmatization", | |
| "authors": [ | |
| { | |
| "first": "Nizar", | |
| "middle": [], | |
| "last": "Habash", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Owen", | |
| "middle": [], | |
| "last": "Rambow", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Ryan", | |
| "middle": [], | |
| "last": "Roth", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2009, | |
| "venue": "Proceedings of the Second International Conference on Arabic Language Resources and Tools", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "102--109", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Nizar Habash, Owen Rambow, and Ryan Roth. 2009. MADA+TOKAN: A toolkit for Arabic tokenization, diacritization, morphological disambiguation, POS tagging, stemming and lemmatization. In Proceed- ings of the Second International Conference on Ara- bic Language Resources and Tools, pages 102-109, Cairo, Egypt.", | |
| "links": null | |
| }, | |
| "BIBREF12": { | |
| "ref_id": "b12", | |
| "title": "Morphological tagging: Data vs. dictionaries", | |
| "authors": [], | |
| "year": 2000, | |
| "venue": "Proceedings of the1st conference of the North American Chapter of the Association for Computational Linguistics", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "94--101", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Jan Haji\u010d. 2000. Morphological tagging: Data vs. dictionaries. In Proceedings of the1st conference of the North American Chapter of the Association for Computational Linguistics, pages 94-101.", | |
| "links": null | |
| }, | |
| "BIBREF13": { | |
| "ref_id": "b13", | |
| "title": "Part of speech tagging for Arabic", | |
| "authors": [ | |
| { | |
| "first": "Sandra", | |
| "middle": [], | |
| "last": "K\u00fcbler", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Emad", | |
| "middle": [], | |
| "last": "Mohamed", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2012, | |
| "venue": "Natural Language Engineering", | |
| "volume": "18", | |
| "issue": "4", | |
| "pages": "521--548", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Sandra K\u00fcbler and Emad Mohamed. 2012. Part of speech tagging for Arabic. Natural Language Engi- neering, 18(4):521-548.", | |
| "links": null | |
| }, | |
| "BIBREF14": { | |
| "ref_id": "b14", | |
| "title": "The Penn Arabic Treebank: Building a large-scale annotated Arabic corpus", | |
| "authors": [ | |
| { | |
| "first": "Mohamed", | |
| "middle": [], | |
| "last": "Maamouri", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Anne", | |
| "middle": [], | |
| "last": "Bies", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Tim", | |
| "middle": [], | |
| "last": "Buckwalter", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "W", | |
| "middle": [], | |
| "last": "Mekki", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2004, | |
| "venue": "NEMLAR Conference on Arabic Language Resources and Tools", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "102--109", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Mohamed Maamouri, Anne Bies, Tim Buckwalter, and W. Mekki. 2004. The Penn Arabic Tree- bank: Building a large-scale annotated Arabic cor- pus. In NEMLAR Conference on Arabic Language Resources and Tools, pages 102-109.", | |
| "links": null | |
| }, | |
| "BIBREF15": { | |
| "ref_id": "b15", | |
| "title": "Is arabic part of speech tagging feasible without word segmentation?", | |
| "authors": [ | |
| { | |
| "first": "Emad", | |
| "middle": [], | |
| "last": "Mohamed", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Sandra", | |
| "middle": [], | |
| "last": "K\u00fcbler", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2010, | |
| "venue": "Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the Association for Computational Linguistics", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "705--708", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Emad Mohamed and Sandra K\u00fcbler. 2010. Is ara- bic part of speech tagging feasible without word segmentation? In Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the Association for Computational Lin- guistics, pages 705-708. Association for Computa- tional Linguistics.", | |
| "links": null | |
| }, | |
| "BIBREF16": { | |
| "ref_id": "b16", | |
| "title": "Text chunking using transformation-based learning", | |
| "authors": [ | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Lance", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Mitchell", | |
| "middle": [ | |
| "P" | |
| ], | |
| "last": "Ramshaw", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Marcus", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1995, | |
| "venue": "Proceedings of the ACL 3rd Workshop on Very Large Corpora", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "82--94", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Lance A. Ramshaw and Mitchell P. Marcus. 1995. Text chunking using transformation-based learning. In Proceedings of the ACL 3rd Workshop on Very Large Corpora, pages 82-94, Cambridge, MA.", | |
| "links": null | |
| }, | |
| "BIBREF17": { | |
| "ref_id": "b17", | |
| "title": "Arabic morphological tagging, diacritization, and lemmatization using lexeme models and feature ranking", | |
| "authors": [ | |
| { | |
| "first": "Ryan", | |
| "middle": [], | |
| "last": "Roth", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Owen", | |
| "middle": [], | |
| "last": "Rambow", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Nizar", | |
| "middle": [], | |
| "last": "Habash", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Mona", | |
| "middle": [], | |
| "last": "Diab", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Cynthia", | |
| "middle": [], | |
| "last": "Rudin", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2008, | |
| "venue": "Proceedings of ACL-08: HLT, Short Papers", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "117--120", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Ryan Roth, Owen Rambow, Nizar Habash, Mona Diab, and Cynthia Rudin. 2008. Arabic morphologi- cal tagging, diacritization, and lemmatization using lexeme models and feature ranking. In Proceed- ings of ACL-08: HLT, Short Papers, pages 117-120, Columbus, Ohio.", | |
| "links": null | |
| }, | |
| "BIBREF18": { | |
| "ref_id": "b18", | |
| "title": "Part-of-speech tagging guidelines for the Penn Treebank Project. Department of Computer and Information Science", | |
| "authors": [ | |
| { | |
| "first": "Beatrice", | |
| "middle": [], | |
| "last": "Santorini", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1990, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Beatrice Santorini. 1990. Part-of-speech tagging guidelines for the Penn Treebank Project. Depart- ment of Computer and Information Science, Univer- sity of Pennsylvania, 3rd Revision, 2nd Printing.", | |
| "links": null | |
| }, | |
| "BIBREF19": { | |
| "ref_id": "b19", | |
| "title": "Statistical parsing of morphologically rich languages (SPMRL): What, how and whither", | |
| "authors": [ | |
| { | |
| "first": "Reut", | |
| "middle": [], | |
| "last": "Tsarfaty", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Djam\u00e9", | |
| "middle": [], | |
| "last": "Seddah", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Yoav", | |
| "middle": [], | |
| "last": "Goldberg", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Sandra", | |
| "middle": [], | |
| "last": "K\u00fcbler", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Marie", | |
| "middle": [], | |
| "last": "Candito", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Jennifer", | |
| "middle": [], | |
| "last": "Foster", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Yannick", | |
| "middle": [], | |
| "last": "Versley", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Ines", | |
| "middle": [], | |
| "last": "Rehbein", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Lamia", | |
| "middle": [], | |
| "last": "Tounsi", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2010, | |
| "venue": "Proceedings of the NAACL Workshop on Statistical Parsing of Morphologically Rich Languages", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Reut Tsarfaty, Djam\u00e9 Seddah, Yoav Goldberg, San- dra K\u00fcbler, Marie Candito, Jennifer Foster, Yan- nick Versley, Ines Rehbein, and Lamia Tounsi. 2010. Statistical parsing of morphologically rich languages (SPMRL): What, how and whither. In Proceedings of the NAACL Workshop on Statistical Parsing of Morphologically Rich Languages, Los Angeles, CA.", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "TABREF2": { | |
| "content": "<table><tr><td colspan=\"2\">Tagset System</td><td>Split</td><td colspan=\"3\">Acc: known Acc: unknown Acc: all</td></tr><tr><td>RTS</td><td>ASMA</td><td>AMIRA-SPLIT</td><td>96.56</td><td>77.79</td><td>96.28</td></tr><tr><td/><td>ASMA</td><td>MADA-SPLIT</td><td>94.20</td><td>84.99</td><td>94.06</td></tr><tr><td/><td>AMIRA 2.1</td><td/><td/><td/><td>96.60</td></tr><tr><td colspan=\"2\">CATiB ASMA</td><td>AMIRA-SPLIT</td><td>97.88</td><td>79.27</td><td>97.59</td></tr><tr><td/><td>ASMA</td><td>MADA-SPLIT</td><td>96.04</td><td>88.36</td><td>95.92</td></tr></table>", | |
| "html": null, | |
| "type_str": "table", | |
| "num": null, | |
| "text": "POS tagging results" | |
| }, | |
| "TABREF4": { | |
| "content": "<table><tr><td>: Example results per POS category and</td></tr><tr><td>their respective confusable modified ATB POS tag</td></tr><tr><td>with this split, ASMA reaches an overall accuracy</td></tr><tr><td>of 95.92%.</td></tr></table>", | |
| "html": null, | |
| "type_str": "table", | |
| "num": null, | |
| "text": "" | |
| } | |
| } | |
| } | |
| } |