Datasets:

pdf2dataset

/

edd8453fa5ee53fbb5b51e0a4c5b4c30

like 0

SALMA: Arabic Sense-Annotated Corpus and WSD Benchmarks Mustafa Jarrar, Sanad Malaysha, Tymaa Hammouda, Mohammed Khalilia Birzeit University, Palestine {mjarrar, smalaysha, thammouda, mkhalilia}@birzeit. edu Abstract SALMA, the first Arabic sense-annotated cor-pus, consists of ~34K tokens, which are all sense-annotated. The corpus is annotated using two dif-ferent sense inventories simultaneously (Modern and Ghani). SALMA novelty lies in how tokens and senses are associated. Instead of linking a token to only one intended sense, SALMA links a token to multiple senses and provides a score to each sense. A smart web-based annotation tool was developed to support scoring multiple senses against a given word. In addition to sense annota-tions, we also annotated the corpus using six types of named entities. The quality of our annotations was assessed using various metrics (Kappa, Lin-ear Weighted Kappa, Quadratic Weighted Kappa, Mean Average Error, and Root Mean Square Er-ror), which show very high inter-annotator agree-ment. To establish a Word Sense Disambiguation baseline using our SALMA corpus, we developed an end-to-end Word Sense Disambiguation sys-tem using Target Sense Verification. We used this system to evaluate three Target Sense Verification models available in the literature. Our best model achieved an accuracy with 84. 2% using Modern and 78. 7% using Ghani. The full corpus and the annotation tool are open-source and publicly avail-able at https://sina. birzeit. edu/salma/. 1 Introduction WSD aims to determine a word's intended mean-ing (sense) in a given context. WSD is underdevel-oped in Arabic due to the lack of sense-annotated datasets. This is in addition to the challenging nature of the WSD task due to the semantic poly-semy of the words (Al-Hajj and Jarrar, 2021). For instance, the Arabic word ( á «,ayn) has sixteen meanings in the Contemporary Arabic Dictionary (Omar, 2008). In the context ( á ªË@ ø @P éJK @Pr-aytuh r-ay¯al,ayn), word ( á «,ayn) refers to eye, while in ( ZAÖÏ@ á « á ÓIK. Qå…šribt min,ayn¯alm¯a-), it refers to water spring. Similarly, the English word book as a noun has ten different senses in Princeton Word Net (Miller et al., 1990), such as (a written work or composition that has been published), or (number of pages bound together). WSD has been consid-ered a challenging task for many years (Weaver,1949/1955), but it has recently gained more atten-tion due to the advances in learning contextualized word representations from language models, such as BERT (Devlin et al., 2019) and GPT (Radford et al., 2018). As glosses are short descriptions of senses (Jar-rar, 2006, 2005), recent research has demonstrated promising results in WSD task by framing the prob-lem as a sentence-pair (context-gloss) binary clas-sification task, referred to as Target Sense Verifi-cation (TSV), where the context is a sentence con-taining the ambiguous word (Huang et al., 2019; Yap et al., 2020; Blevins and Zettlemoyer, 2020). Al-Hajj and Jarrar (2021) proposed an approach for Arabic WSD (using TSV) based on context-gloss pairs extracted from the Arabic Ontology and lexicons and they achieved 84% accuracy, but this evaluation was done on a TSV dataset rather than a WSD evaluation using a sense-annotated corpus. Additionally, Al-Hajj and Jarrar (2021) presented an attempt for Arabic Word-in-Context (Wi C) disambiguation using the dataset provided by the Sem Eval shared task (Martelli et al., 2021). This article presents SALMA, the first sense-annotated Arabic corpus consisting of about 34K tokens, which are manually annotated with senses. Since there are no available sense inventories for Arabic, We used two Arabic lexicons as sense inventories: Contemporary Arabic Dictio-nary (è Qå•AªÖ Ï@é J K. QªË@é ªÊË@¯all˙gh¯al,rbyh ¯alm,¯as. rh), here-after we refer to as Modern (Omar, 2008), and Al-Ghani Al-Zaher ( Që@ QË@ ú æ ªË@¯al˙gny¯alz¯ahr), hereafter we refer to as Ghani (Abul-Azm, 2014). These two lexicons are part of the lexicon digitization project and lexicographic database at Sina Lab1(Jarrar and Amayreh, 2019; Alhafi et al., 2019; Amayreh et al., 2019; Ghanem et al., 2023; Jarrar et al., 2021). We introduce a novel sense-annotation framework (Section 3), in which all candidate senses, from both lexicons, are scored to indicate their semantic 1https://sina. birzeit. edu/

JMHK23.pdf

relatedness to a token appearing within a context. The higher the score, the more semantically related the sense is. For better coverage, we annotated each token in our corpus using both lexicons inde-pendently and in parallel. The scores assigned to senses of the Modern do not influence the scoring of the Ghani senses. In addition, we also annotated our corpus using six types of named entities: per-son (PERS), organization (ORG), geopolitical en-tity (GPE), location (LOC), facility (FAC), and cur-rency (CURR). The corpus was annotated by three linguists and we assessed the inter-annotator agree-ment (IAA) using 2. 6% of the annotated words in the corpus. To establish a baseline for WSD in Ara-bic, we developed an end-to-end WSD system, in which we benchmarked three available TSV mod-els, with different settings. The best model resulted in 84. 2% accuracy using Modern and 78. 7% using Ghani. The main contributions of this paper are: Sense-annotated corpus, annotated with two sense inventories independently, and six named entities; and most importantly, each word is linked with all of its senses, and each sense is given a score. Web-based sense-annotation framework to score all senses of a given word. End-to-end WSD system, implemented and evaluated using three different TSV models. WSD baseline for Arabic, with different set-tings. The remainder of the article is organized as fol-lows: Section 2 highlights the related work, Section 3 presents the corpus, Section 4 describes the inter-annotator agreement, Sections 5 and 6 present how the baselines are produced, we conclude in Section 7 and outline the limitations and future work in Section 8. 2 Related Work We will first review related sense-annotated cor-pora, then we will review related sense inventories. One of the known English sense-annotated cor-pora is Sem Cor (Miller et al., 1993), which is an-notated using the Princeton Word Net (Miller et al., 1990). It contains about 200K sense annotations for around 700K words, but not all words are sense-annotated in the Sem Cor corpus, especially multi-word expressions, articles, and prepositions. The An Cora corpus for Spanish and Catalan languages (Taulé et al., 2008) was collected from newspapers and consists of 500K words, but only 200K noun words are semantically annotated using the Span-ish Word Net. An Cora also includes morphological, semantic, and syntactic annotations. Tu Ba-D/Z is a German annotated corpus, manually collected from newspapers and annotated using the German-Net senses (Telljohann et al., 2004). Tu Ba-D/Z was later used as a gold standard for the WSD task by (Petrolito and Bond, 2014). The Italian Syntactic-Semantic Treebank (ISST) is a corpus built for the Italian language with 89,941 sense-annotated words (Montemagni and Venturi, 2003). The ISST annotations cover five levels that are related to lexico-semantics such as orthographic, morpho-syntactic, semantic, and syntactic aspects. The NTU-MC corpus (Tan and Bond, 2012) cov-ers eight languages including Thai, Vietnamese, Arabic, Korean, Indonesian, Japanese, Mandarin Chinese, and English. However, the Arabic version is not publicly available. This corpus was collected from short stories, essays, and tourism articles re-sulting in a total of 116K words, but only 63K words are annotated. KPWr, a Polish corpus, con-tains text from multiple domains including science, law, religion, and press (Broda et al., 2012) with a total of 438,327 words, but only 9,157 words are annotated using the Polish Word Net (Maziarz et al., 2012). For Arabic, the focus of research has been pri-marily on developing corpora for morphological and syntactic tagging (Darwish et al., 2021) rather than semantic and sense annotation, as noted by Elayeb (2019) and Naser-Karajah et al. (2021). For instance, part of the Onto Notes corpus (Weischedel et al., 2013) covers limited semantic annotations for Arabic using a small sense inventory of size 261 senses (150 verbs and 111 nouns). Additionally, AQMAR corpus (Schneider et al., 2012) is anno-tated with 25 super-sense labels representing broad semantic fields such as ARTIFACT and PERSON, which can be considered as general types of named entities, rather than word-sense annotations. They annotated ~22K nouns out of 65K tokens corpus. Table 1 compares our proposed corpus and related Arabic resources. In addition to the lack of sense-annotated cor-pora, Arabic lacks reliable sense inventories. Al-though there are some available semantic resources, they are not mature enough to be used as sense

JMHK23.pdf

Corpus Unique Senses Annotation Type Corpus Size(tokens)Annotations Nouns Verbs Func. Words Punc. + Digits Total AQMAR25 semantic fields (closer to named entities)selected words each one sense65K ~22K---~22K Onto Notes5261 semantic fields (high-level grouped senses)selected words each one sense300K 8,700 4,300--13K SALMA (ours)4,151 word senses (from each sense inventory) 6 types of named entitiesall senses of all words34K 19,030 2,763 7,116 5,344 34,253 Table 1: Overview of related Arabic sense-annotated corpora. inventories. For example, the Arabic Word Net (Black et al., 2006) contains about 10K senses, and the Arabic Ontology (Jarrar, 2021, 2011) contains about 18K synsets. However, both resources cannot be used as sense inventories as they do not provide a complete set of senses for a given lemma (i. e., lex-icon entry). The lexicographic database developed at Birzeit University contains about 150 Arabic lexicons (Jarrar and Amayreh, 2019; Jarrar et al., 2019), but these lexicons are not well-structured or suitable to be used as sense inventories (Jarrar and Amayreh, 2019). Due to the lack of depend-able Arabic sense inventory, we decided to obtain a license to digitize and use two Arabic lexicons as sense inventories, namely, Modern (Omar, 2008) and Ghani (Abul-Azm, 2014). 3 Corpus Construction and Annotation 3. 1 Corpus Collection Our SALMA corpus is part of the Wojood corpus (Jarrar et al., 2022), and was collected from 33 online media sources written in Modern Standard Arabic (MSA) and covering general topics. Some of those sources include mipa. institute, sanaacen-ter. org, hrw. org, diplomatie. ma, sa. usembassy. gov, eeas. europa. eu, crisisgroup. org, and mofaic. gov. ae. The corpus was then segmented into sentences and tokenized, resulting in 1439 sentences and ~34K tokens, with an average of 23. 8 tokens per sentence. 3. 2 Annotation Framework This section presents a novel sense annotation framework, where instead of linking a word to one sense, we propose to score all semantically related senses to the word. The score ranges between 1-100% and a sense with a score ≥60%is considered a correct sense of the word. The ranking scale is divided into six categories: Explicate /è Qå…AJ. Ó(100%): direct and explicate semantics (ém' Qå•ðéj J m•é ËBX). General /ÐA« ú æªÓ(80%): correct but implicate semantics (è Qå…AJ. ÓQ «éj J m•é ËBX). Referral /é K ñ ªËé ËBX(60%): generally correct semantics, but is referred to another lemma (É«A ¯ Õæ…@,PY’Ó ÉJÓ@Yg. é ÓA« áºËðéj J m•). For example, the word drinker and its gloss ( active partici-ple of drink ). Related /é¯C« H@ X(40%): weak semantics (AJ ËBX AîD k @, ¡® ¯é ÓAªË@é ËBYË@ ú ¯é»Q‚Ó). For example, the term (é» Qå„Ë@éƒAJ ƒsy¯ash¯alšrkh ) /company's policy, is related to the sense ( the policy used to collect taxes ) which is not a sense of the lemma (éƒAJ ƒsy¯ash), but semantically related. Root semantics /P Yg. é ËBX(20%): share root se-mantics ( Aê ÊÒm' ú æ Ë@è XQj. ÖÏ@é ËBYË@ ú ¯ ¼Q‚ áºËðé ®ÊJ m×é ËBX é K PAj. ÖÏ@é ËBYË@ ÉJÓ, P Ym. Ì'@). In Arabic lexical seman-tics, all words with the same root share part of the semantics of this root (Ryding, 2014; Boudelaa and Marslen-Wilson, 2004; Boude-laa et al., 2010). For example, all senses of the lemma (éƒAJ ƒsy¯ash), such as politics and policies share an abstract meaning (e. g., issues related to governing and acting). Different /é ®ÊJ m×(1%): unrelated semantics (AÓAÖßé ®ÊJ m×é ËBX). This framework serves several purposes. First, in case of underdeveloped sense inventories (such as the Modern and Ghani lexicons), in which glosses might be vague, redundant, or overlapping, our framework allows the annotators to score each sense. In this paper, we linked every word in the corpus with all semantically related senses in Modern and Ghani, thus we were able to compare and evaluate the lexical coverage in both lexicons (see Section 3. 5). Another advantage of using this framework (i. e., scoring all senses) is that our corpus can be used to benchmark ranking-based WSD methods (Conia and Navigli, 2021; Yap et al.,

JMHK23.pdf

Figure 1: Screenshot of our web-based annotation tool. 2020), which is not possible in the case of one-sense annotated corpora. 3. 3 Annotation Tool We developed a web-based tool optimized for our sense annotation framework and methodology. On the right side of Figure 1, the linguist selects a word to be annotated (such as "éƒAJ ‚Ë@¯alsy¯ash"). The tool will then retrieve all sentences (i. e. contexts) in the corpus containing the selected word. The tool will also automatically fetch the lemma of the selected word, and the linguist has the ability to search for the lemma manually. After selecting a lemma, the tool retrieves senses associated with the lemma from both lexicons, Modern and Ghani. The linguist can then select the score category for each sense according to our guideline and apply these scores to all selected words (in contexts) as shown in Figure 1. The scores are selected from a Combo Box of the six categories (See Section 3. 2), however, the tool internally stores their correspond-ing numeric values. 3. 4 Annotation Process The annotation was carried out in three phases:Phase 1 (training) : we recruited three undergradu-ate students majoring in linguistics. The students were trained in three steps in order to produce con-sistent annotations. We first assigned 50 words to each linguist and trained them to conduct the annotation jointly. Second, we assigned the same 150 words to each student separately, then asked them to compare and consolidate their annotations, which helps in calibrating their scoring. Third, we repeated the second phase, but using 300 words and again we asked them to compare their annotations. Phase 2 (annotation ): out of ~34K tokens, ex-cluding digits and punctuations, we assigned about 9. 6K words to each of the three linguists. Each linguist was asked to annotate all occurrences of each word in the corpus-resulting in about ~29K annotations for the whole words. Phase 3 (validation ): after finishing the annota-tions, we used the tool to automatically validate the annotations and flag those that violated the fol-lowing cases: (i) a word is annotated with more than one Explicit or General sense in the same lex-icon, which is an indication of either a mistake or redundant or overlapping senses in the lexicon. (ii) a word is missing either an Explicit or a General

JMHK23.pdf

sense; this is an indication of a mistake or the lexi-con is missing this sense. (iii) if the selected sense is a proper noun, then all other senses should be ranked as Different. The linguists were asked to review these flagged annotations and revise them if necessary. The linguists were encouraged to discuss among themselves and take joint decisions when facing difficulties, especially in the case of vague glosses or contexts. In addition, as will be discussed in Section 3. 5, missing lemmas and senses are manu-ally added to the lexicons. Table 2 provides general statistics about the annotations. It is worth noting that sense annotations are typically costly and time-consuming. The linguists spent about 600 working days (i. e., 4800 working hours) to carry out the three phases described above. Term Noun Verb Func. Words Punc+ Digits Total Tokens 19,030 2,763 7,116 5,344 34,253 Unique Tokens6,670 1,593 322 175 8,760 Unique Lemmas2,904 677 119 175 3,875 Unique Senses3,151 792 206 24,151 Table 2: Statistics of the SALMA corpus. Term Modern Ghani Lemmas 80% (2,788/3,522) 78% (2,724/3,522) Senses (Without Proper nouns)83% (3,430/4,151) 78% (3,226/4,151) Proper Nouns Senses4% (9/213) 14% (30/213) Table 3: Coverage of Modern and Ghani lexicons. 3. 5 Discussion and Lexical Coverage We evaluated the coverage of both lexicons based on the sense-annotated tokens. As Table 3 shows, Modern has higher coverage of lemmas (80%) com-pared to Ghani's coverage (78%), and has higher sense coverage (83%) compared to Ghani (78%). Moreover, glosses in Modern are more precise, less ambiguous and well-formulated as discussed in Section 4. 1. The proper nouns are the main reason for the missing lemmas and senses, as the Modern and Ghani cover 4% and 14% of proper nouns in SALMA corpus, respectively. Lemmas and senses that are not covered by any of the two lexicons were added manually by the linguists. All numerical val-ues are annotated with the same "digit" sense thatcovers ordinal and nominal numbers, and similarly, punctuation marks are all annotated with "Punc". 3. 6 Named Entity Annotations Named-entity annotations are important in sense-annotated corpora because sense inventories do not typically cover names of organizations, towns, people, landmarks, and others. Tag Description PERS Person names: first, middle, last, nickname... ORG Organizations: company, team, government... GPE Geopolitical entities: country, city, state... LOC Geographical locations: river, sea, mountain... FAC facilities: landmark, road, building, airport... CURR Currency names or symbols. Table 4: Types of named entities. In addition to word-sense annotations, we anno-tated our corpus using six types of named entities listed in Table 4. As our corpus is a part of the Wo-jood, which is annotated with 21 types of nested named-entities (Jarrar et al., 2022), in this article we annotated SALMA with six flat entities only. We used the IOB2 tagging scheme (Sang and Veen-stra, 1999), where B indicates the beginning of the entity mention, I the inside token, and O outside token. Tag Named Entity Mentions Tokens in the Entity Mentions PERS 294 568 ORG 1,123 2,108 GPE 1,086 1,295 LOC 166 318 FAC 22 59 CURR 37 41 Total 2,728 4,389 Table 5: Statistics of named entities in SALMA corpus. We applied the NER guidelines that were used to annotate the Onto Notes5 corpus (Weischedel et al., 2011). Table 5 presents statistics about all named entities in the SALMA corpus, which shows that 4389 (about 15%) of the tokens are part of an entity mention. 4 Inter-Annotation Agreement (IAA) To evaluate our annotations, we selected 250 annotated words from each annotator A∈ {A1, A2, A3}, and assigned them to a different annotator to perform double annotations. This yielded a total of 750 words (2. 6% of the anno-tated words) divided among three pairs of anno-tators, {(A1, A2),(A1, A3),(A2, A3)}. Because

JMHK23.pdf

our sense annotations contain scores (i. e., not dis-crete values), computing IAA is not straightfor-ward. We chose to use various evaluation met-rics especially those that take ranking into con-sideration. The IAA metrics used are: (i) Kappa, (ii) Linear Weighted Kappa (LWK), (ii) Quadratic Weighted Kappa (QWK), (iv) Mean Average Error (MAE), and (v) Root Mean Square Error (RMSE). Kappa is usually used when the data is nominal (Eugenio and Glass, 2004), so we set a threshold on the score ( ≥60%) in the six categories to be able to calculate Cohen's Kappa. The senses with scores above or equal this threshold carry the intended meanings that map with the context of the targeted word (See section 3. 2). Nonetheless, a more suit-able metric for ranked labels is either the LWK or QWK, as specified in the following equations, which we adopt from (Vanbelle, 2016): QWK = 1-KP i,j=1(yi-yj)2 (K-1)2. foij KP i,j=1(yi-yj)2 (K-1)2. feij(1) LWK = 1-KP i,j=1|yi-yj)| (K-1). foij KP i,j=1|yi-yj)| (K-1). feij(2) where foijis the observed frequency of the cat-egories ( iandj) per the annotators selection, feij is the expected frequency for both annotators' se-lected categories, (yi-yjx)denotes the distance between the categories, and Kis number of cate-gories. Both LWK and QWK take the distance between categories into consideration, where the distance is defined as the number of categories separating the two annotators' selection. The difference is that LWK calculates the distance linearly while QWK calculates it quadratically. For measuring the ranking error deviation among annotators we used MAE and RMSE. 4. 1 IAA Results Table 6 summarizes the result of the inter-annotator-agreement, the value in parenthesis is the standard deviation among pairs of annotators. Overall, we see higher agreement among the annotators for the Modern. The higher agreement is clear from all IAA metrics and the standard deviation. We see less confidence in the Ghani annotations as the IAAMetric Lexicons Average (STD) Kappa Modern Ghani90. 48 (±2. 97) 78. 68 (±8. 49) LWK Modern Ghani88. 29 (±5. 37) 79. 56 (±9. 35) QWK Modern Ghani91. 94 (±3. 42) 86. 03 (±5. 41) RMSE Modern Ghani13. 44 (±3. 08) 19. 12 (±3. 06) MAE Modern Ghani4. 46 (±2. 04) 8. 27 (±3. 52) Table 6: Inter-Annotator Agreement (IAA) average among the three linguists using different metrics. dropped across all metrics with higher variability among annotators, presented in higher standard de-viation. Kappa was affected the most with a drop of 11. 8% when measured on the Ghani, followed by LWK with a drop of 8. 73%. QWK has the small-est drop of 5. 91% and also has the least variability among annotators. We believe the reason for the higher IAA on Modern is because Modern has bet-ter quality glosses compared to the Ghani, which has shorter glosses and in many cases are ambigu-ous. However, regardless of the lexicon used, we observed higher agreement among annotators as measured by LWK and QWK since they take ad-vantage of the scores assigned to each gloss, while Kappa ignores the scoring information. Figure 2: BERT-based TSV Architecture. We reach similar conclusions for RMSE and MAE. Both metrics are lower for Modern com-pared to Ghani. The Average RMSE among all annotator pairs on the Modern is 13. 44 compared

JMHK23.pdf

Rank 1 20. 6 0. 4 0. 7 0. 3[SEP] رؤﯾﺔ اﻟﻰ اﻟﻣﺳﺗﻧدة اﻷﻣرﯾﻛﯾﺔ\token>اﻟﺳﯾﺎﺳﺔtoken> ﺳﺎھﻣت ﻛﯾف [CLS] [SEP] ﺷُؤُوﻧِﮭﺎ وَﺗَدْﺑﯾرُ واﻟﺧﺎرِﺟِﯾﱠﺔِ اﻟدﱠاﺧِﻠِﯾﱠﺔِ أَﻋْﻣﺎﻟِﮭﺎ وَﺗَﺳْﯾِﯾرُ أُﻣورِھﺎ، ﺗَوَﻟﱢﻲ ': اﻟﺑِﻼدِ ﺳﯾﺎﺳَﺔُ' اﻟﺳﯾﺎﺳﺔ ﺳﺎھﻣت ﻛﯾف رؤﯾﺔ إﻟﻰ اﻟﻣﺳﺗﻧدة اﻷﻣرﯾﻛﯾﺔ. واﻗِﻊٌ ھُوَ ﺑِﻣﺎ اﻟﺗﱠﺳْﻠﯾمُ أَي': اﻟواﻗِﻊِ اﻷَﻣْرِ ﺳِﯾﺎﺳَﺔُ' Generate context-gloss pairs Candidate Glosses TSV Model (See Figure 2) واﻟﺧﺎرِﺟِﯾﱠﺔِ اﻟدﱠاﺧِﻠِﯾﱠﺔِ أَﻋْﻣﺎﻟِﮭﺎ وَﺗَﺳْﯾِﯾرُ أُﻣورِھﺎ، ﺗَوَﻟﱢﻲ ' : اﻟﺑِﻼدِ ﺳﯾﺎﺳَﺔُ' . ﺷُؤُوﻧِﮭﺎ وَﺗَدْﺑﯾرُ g1 g2 p1 p2 True False [SEP] رؤﯾﺔ اﻟﻰ اﻟﻣﺳﺗﻧدة اﻷﻣرﯾﻛﯾﺔ \token>اﻟﺳﯾﺎﺳﺔtoken> ﺳﺎھﻣت ﻛﯾف[CLS] [SEP] واﻗِﻊٌ ھُوَ ﺑِﻣﺎ اﻟﺗﱠﺳْﻠﯾمُ أَي': اﻟواﻗِﻊِ اﻷَﻣْرِ ﺳِﯾﺎﺳَﺔُ' Rank glosses based on the True scores Lookup Glosses (ﺳِﯾﺎﺳَﺔ) lexicon ﺳِﯾﺎﺳَﺔLemmatize (اﻟﺳﯾﺎﺳﺔ) Softmax Scores Score 0. 7 0. 6Gloss g2 g1Figure 3: An end-to-end WSD using the TSV model (SALMA system). to 19. 12 for Ghani, while the average MAE for the Modern is 4. 46 compared to 8. 27 on the Ghani. 5 Computing WSD Baselines using SALMA In this section, we present the baseline for Arabic WSD using our SALMA corpus. To the best of our knowledge, there are no available Arabic WSD systems to evaluate. The only available Arabic models are TSV, which are related, but not the same as WSD. In what follows, we explain the difference between WSD and TSV tasks, and propose an end-to-end WSD system using TSV. 5. 1 The TSV Task The TSV task is a binary classification task used to determine whether a pair of sentences (context and gloss) are True or False (see Figure 2). In other words, given a context ccontainig the target word w, and a gloss gi, TSV aims to classify the context-gloss pair (c, gi)as True or False. It is True if the gloss giis the intended sense of winc, otherwise, it is False (Breit et al., 2020). It is important to note that TSV is different from WSD, which determines which gloss, among a set of glosses, is the intended meaning for the target word. There are three available Arabic TSV models with the same architecture: (1) the Razzaz model, trained using 31K context-gloss pairs extracted from Modern (El-Razzaz et al., 2021); (2) the Arab Gloss BERT model, trained on a larger dataset (167K context-gloss pairs) extracted from sev-eral Arabic lexicons (Al-Hajj and Jarrar, 2021); and (3) the Aug-Arab Gloss BERT (D9) model, trained on an augmented data, generated using back-translation of the Arab Gloss BERT dataset (Malaysha et al., 2023). In what follows, we propose to develop an end-to-end WSD system using TSV (called SALMA system) and in Section 6, we benchmark our pro-posed system using the SALMA corpus. 5. 2 Building WSD System Using TSV In this section, we propose an end-to-end solution for WSD using TSV. The solution consists of the following phases (Figure 3): 1) candidate glosses lookup, 2) target sense verification, and 3) gloss ranking. 1. Candidate Glosses Lookup : given a target word win a context c, we first lemmatize w(i. e., deter-mine its lemma l), where we use our own in-house lemmatizer, then retrieve the set of ncandidate glosses, G={g1, g2,..., g n}, oflfrom the lexicon (i. e., sense inventory). Example : the word w(éƒAJ ‚Ë@¯alsy¯ash ) in c (é K ð P ú Í@è Y J‚ÖÏ@é JºKQÓ B@éƒAJ‚Ë@ IÒë Aƒ ­J») has the lemma (éƒAJ ƒsiya¯asatun ) with two corresponding glosses ({g1, g2}) in the Ghani, as shown in Figure 3.

JMHK23.pdf

2. TSV : once we have the set of ncandidate glosses, we input to the TSV model a set of n context-gloss pairs, P={(c, gi)|∀gi∈G}, as illustrated with (p1, p2)in Figure 3. The target word wincis wrapped with special tokens "<token> w</token>", to emphasize the target word during training and testing of the TSV models. For each context-gloss pair, the TSV model returns confidence scores for the True and False labels, but the TSV model does not compare or rank glosses in this phase. 3. Gloss Ranking : we determine the intended meaning by ranking the glosses based on their True confidence scores calculated in the previous step. The gloss with the highest score is selected as the intended gloss for w. 6 Experiments and Results 6. 1 Experimental Setup To evaluate the three available Arabic TSV mod-els using our SALMA corpus, we implemented three instances of the WSD system depicted in Fig-ure 3, each with a different TSV model. For each word in each context in the SALMA corpus, we generated context-gloss pairs similar to the exam-ple shown in Figure 3. Because our corpus was sense-annotated using two lexicons (i. e., two sense inventories), we generated two sets of context-gloss pairs. In this way, we compute a separate baseline for each of the Modern and Ghani. We neither in-cluded annotations of digits and punctuations, nor the named-entity annotations presented in Section 3. 6. The length of the contexts may impact the WSD accuracy, so in addition to using the full context around w, we also experimented with different con-text sizes, s∈ {3,5,7,9,11}. For example, the context size s= 5means that there are two tokens before and two tokens after w. As will be discussed in the next subsection, we evaluated three TSV models: Razzaz2, Arab Gloss-BERT3, and Aug-Arab Gloss BERT(D9)4. We used context size s= 11, which gave the best results. Following the authors of these models, we did not 2We reproduced the TSV model using the code and data available at https://github. com/MElrazzaz/Arabic-word-sense-disambiguation-bench-mark 3Arab Gloss BERT fine-tuned model Version 1 (CC-BY-4. 0) at https://huggingface. co/Sina Lab/Arab Gloss BERT/tree/main 4Fine-tuned model D9 (CC-BY-4. 0) at https://huggingface. co/Sina Lab/Arab Gloss BERT/tree/Augmentuse any signal to mark up target words in the case of the Razzaz and Aug-Arab Gloss BERT(D9); how-ever, we used UNUSED0 for Arab Gloss BERT. The experiments have been implemented in Python, specifically using the Transformers library provided by Huggig Face5, which is used to load and test the models. To speed-up the models evalu-ation, we have run the codes using a GPU (SVGA II) instance, where each run took around 20 hours. TSV Model Lexicons Accuracy Razzaz Modern Ghani66. 0% 68. 4% Arab Gloss BERT Modern Ghani84. 2% 77. 6% Aug-Arab Gloss BERT(D9) Modern Ghani82. 6% 78. 7% Table 7: WSD baselines for three TSV models, with context length = 11. 6. 2 Baselines and Discussion Table 7 presents our evaluation of the three TSV models using both Modern and Ghani with context sizes= 11. As shown in this table, the Arab Gloss-BERT is the best-performing model(84. 2%), which most probably because it was trained on a larger and higher quality dataset of lexicon definitions. The accuracy was calculated for nouns and verbs. We excluded the functional words as they mostly do not carry semantics. Window Lexicon Accuracy Target Sense Rank Accuracy (Top1) per POS Top1 Top2 Top3 Noun Verb Func. All Modern 82. 8 94. 2 97. 4 83. 5 77. 9 41. 2 Ghani 77. 0 89. 3 94. 1 78. 5 66. 0 36. 0 11Modern 84. 2 95. 1 98. 1 85. 4 76. 1 37. 9 Ghani 77. 6 90. 1 94. 9 79. 4 61. 7 31. 8 9Modern 83. 5 95. 0 97. 9 84. 4 78. 3 37. 7 GHani 77. 3 90. 1 94. 8 79 63. 7 32. 2 7Modern 83. 8 95. 1 97. 9 84. 8 77. 4 38. 9 Ghani 77. 3 90. 0 94. 9 79. 1 62. 9 31. 8 5Modern 84. 0 95. 1 98. 1 85. 3 75. 6 40. 0 Ghani 77. 6 90. 1 94. 9 79. 5 61. 6 31. 7 3Modern 82. 8 94. 4 97. 6 84. 4 71. 8 42. 1 Ghani 77. 4 90. 0 94. 8 79. 4 59. 7 32. 1 Table 8: Baselines-evaluation of Arab Gloss BERT on two sense inventories, with different context windows and sense orderings. Table 8 presents further evaluation of Arab Gloss-BERT, which illustrates the following: (i) using Modern is better than using Ghani in all experi-ments. This might be because of the better quality 5https://huggingface. co/docs/transformers/index

JMHK23.pdf

of Modern glosses (refer to IAA in Section 4); (ii) While window 11 and 5 have the highest WSD ac-curacy, the use of context windows does not make major difference (only 1. 4% for Modern and 0. 6% for Ghani); (iii) the ranking of the intended sense among the top 1, 2, and 3 senses illustrates a con-sistent and reasonable increase in the WSD accu-racy; and (iv) when evaluating the model accuracy for noun and verb, the accuracy of nouns is about 8. 5% better than verbs for Modern, which might be because verbs are typically more ambiguous (Malaysha et al., 2023). The WSD accuracy for functional words is very low with both lexicons. This is because functional words are highly poly-semous and their glosses describe their functions rather than semantics. 7 Conclusion We presented SALMA, the first sense-annotated Arabic corpus. The novelty of SALMA lies in utilizing two sense inventories and named entity annotations. In addition, instead of linking a word to one intended sense, we scored all semantically related senses of each token in the corpus. The qual-ity of the annotations was assessed using various inter-annotator agreement metrics (Kappa, LWK, QWK, MAE, and RSME). To compute a WSD baseline using our corpus, we proposed to build an end-to-end WSD system using TSV, and evalu-ated this system using three different TSV models. The full corpus, annotations, and the tool, are open source and publicly available on Git Hub. 8 Limitations and Future Work Although Modern provides a better quality of glosses compared with the Ghani, some of Mod-ern's glosses are referrals, i. e., referred to another related lemma. At this stage, we annotated these referrals as senses. Nevertheless, in order to use the Modern as a general sense inventory, these referrals need to be treated differently. We plan to replace all referral glosses with the senses they refer to, which can be done semi-automatically. For miss-ing lemmas in Modern, we plan to map between the lemmas in both lexicons and then import missing lemmas and their senses from Ghani to Modern. In this way, we expect to have a richer Arabic sense inventory. Additionally, our sense annotations are limited to the senses of a single-word lemma. We plan to annotate the corpus with multiword expres-sions (Jarrar et al., 2018). Furthermore, the corpuswe presented in this article is limited to MSA. To extend this corpus with dialectal text, plan to sense-annotate portions of the available corpora Curras (Haff et al., 2022; Jarrar et al., 2017), Baladi (Haff et al., 2022), Nabra (Nayouf et al., 2023) and Lisan (Jarrar et al., 2023). Acknowledgment We would like to thank Shimaa Hamayel, Tamara Qaimari, Raghad Aburahma, Hiba Zayed, and Rwaa Assi for helping us in the corpus annotation. References Abdul-Ghani Abul-Azm. 2014. Al-ghani al-zaher dic-tionary. Rabat: Al-Ghani Publishing Institution. Moustafa Al-Hajj and Mustafa Jarrar. 2021. Arabgloss-bert: Fine-tuning bert on context-gloss pairs for wsd. In Proceedings of the International Conference on Recent Advances in Natural Language Processing (RANLP 2021), pages 40-48, Online. INCOMA Ltd. Diana Alhafi, Anton Deik, and Mustafa Jarrar. 2019. Usability evaluation of lexicographic e-services. In The 2019 IEEE/ACS 16th International Conference on Computer Systems and Applications (AICCSA), pages 1-7. IEE. Hamzeh Amayreh, Mohammad Dwaikat, and Mustafa Jarrar. 2019. Lexicon digitization-a framework for structuring, normalizing and cleaning lexical entries. Technical Report, Birzeit University. William Black, Sabri Elkateb, Horacio Rodriguez, Musa Alkhalifa, Piek V ossen, Adam Pease, Christiane Fell-baum, et al. 2006. Introducing the arabic wordnet project. In Proceedings of the third international Word Net conference, pages 295-300. Jeju Korea. Terra Blevins and Luke Zettlemoyer. 2020. Mov-ing down the long tail of word sense disambigua-tion with gloss-informed biencoders. ar Xiv preprint ar Xiv:2005. 02590. Sami Boudelaa and William D Marslen-Wilson. 2004. Abstract morphemes and lexical representation: The cv-skeleton in arabic. Cognition, 92(3):271-303. Sami Boudelaa, Friedemann Pulvermüller, Olaf Hauk, Yury Shtyrov, and William Marslen-Wilson. 2010. Arabic morphology in the neural language system. Journal of cognitive neuroscience, 22(5):998-1010. Anna Breit, Artem Revenko, Kiamehr Rezaee, Moham-mad Taher Pilehvar, and Jose Camacho-Collados. 2020. Wic-tsv: An evaluation benchmark for tar-get sense verification of words in context. In Pro-ceedings of the 16th Conference of the European Chapter of the Association for Computational Lin-guistics: Main Volume, EACL 2021, Online, April 19-23, 2021, pages 1635-1645.

JMHK23.pdf

Bartosz Broda, Michał Marci ´nczuk, Marek Maziarz, Adam Radziszewski, and Adam Wardy ´nski. 2012. KPWr: Towards a free corpus of Polish. In Proceed-ings of the Eighth International Conference on Lan-guage Resources and Evaluation (LREC'12), pages 3218-3222, Istanbul, Turkey. European Language Resources Association (ELRA). Simone Conia and Roberto Navigli. 2021. Framing word sense disambiguation as a multi-label problem for model-agnostic knowledge integration. In Pro-ceedings of the 16th Conference of the European Chapter of the Association for Computational Lin-guistics: Main Volume, EACL 2021, Online, April 19-23, 2021, pages 3269-3275. Association for Com-putational Linguistics. Kareem Darwish, Nizar Habash, Mourad Abbas, Hend Al-Khalifa, Huseein T. Al-Natsheh, Houda Bouamor, Karim Bouzoubaa, Violetta Cavalli-Sforza, Samhaa R. El-Beltagy, Wassim El-Hajj, Mustafa Jar-rar, and Hamdy Mubarak. 2021. A panoramic survey of natural language processing in the arab worlds. Commun. ACM, 64(4):72-81. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: pre-training of deep bidirectional transformers for language under-standing. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Tech-nologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 4171-4186. Association for Computational Linguistics. Mohammed El-Razzaz, Mohamed Waleed Fakhr, and Fahima A Maghraby. 2021. Arabic gloss wsd using bert. Applied Sciences, 11(6):2567. Bilel Elayeb. 2019. Arabic word sense disambiguation: a review. Artif. Intell. Rev., 52(4):2475-2532. Barbara Di Eugenio and Michael Glass. 2004. The kappa statistic: A second look. Computational lin-guistics, 30(1):95-101. Sana Ghanem, Mustafa Jarrar, Radi Jarrar, and Ibrahim Bounhas. 2023. A benchmark and scoring algorithm for enriching arabic synonyms. In Proceedings of the 12th International Global Wordnet Conference (GWC2023), pages 215-222. Global Wordnet Asso-ciation. Karim El Haff, Mustafa Jarrar, Tymaa Hammouda, and Fadi Zaraket. 2022. Curras + baladi: Towards a levantine corpus. In Proceedings of the International Conference on Language Resources and Evaluation (LREC 2022), Marseille, France. Luyao Huang, Chi Sun, Xipeng Qiu, and Xuanjing Huang. 2019. Glossbert: BERT for word sense dis-ambiguation with gloss knowledge. In Proceedings of the 2019 Conference on Empirical Methods in Nat-ural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP 2019, Hong Kong, China, Novem-ber 3-7, 2019, pages 3507-3512. Association for Computational Linguistics. Mustafa Jarrar. 2005. Towards Methodological Princi-ples for Ontology Engineering. Ph. D. thesis, Vrije Universiteit Brussel. Mustafa Jarrar. 2006. Towards the notion of gloss, and the adoption of linguistic resources in formal ontol-ogy engineering. In Proceedings of the 15th interna-tional conference on World Wide Web (WWW2006), pages 497-503. ACM Press, New York, NY. Mustafa Jarrar. 2011. Building a formal arabic ontology (invited paper). In Proceedings of the Experts Meet-ing on Arabic Ontologies and Semantic Networks. ALECSO, Arab League. Mustafa Jarrar. 2021. The arabic ontology-an arabic wordnet with ontologically clean content. Applied Ontology Journal, 16(1):1-26. Mustafa Jarrar and Hamzeh Amayreh. 2019. An arabic-multilingual database with a lexicographic search engine. In The 24th International Conference on Ap-plications of Natural Language to Information Sys-tems (NLDB 2019), volume 11608 of LNCS, pages 234-246. Springer. Mustafa Jarrar, Hamzeh Amayreh, and John P. Mc Crae. 2019. Representing arabic lexicons in lemon-a preliminary study. In The 2nd Conference on Lan-guage, Data and Knowledge (LDK 2019), volume 2402, pages 29-33. CEUR Workshop Proceedings. Mustafa Jarrar, Nizar Habash, Faeq Alrimawi, Diyam Akra, and Nasser Zalmout. 2017. Curras: An anno-tated corpus for the palestinian arabic dialect. Jour-nal Language Resources and Evaluation, 51(3):745-775. Mustafa Jarrar, Eman Karajah, Muhammad Khalifa, and Khaled Shaalan. 2021. Extracting synonyms from bilingual dictionaries. In Proceedings of the 11th In-ternational Global Wordnet Conference (GWC2021), pages 215-222. Global Wordnet Association. Mustafa Jarrar, Mohammed Khalilia, and Sana Ghanem. 2022. Wojood: Nested arabic named entity corpus and recognition using bert. In Proceedings of the International Conference on Language Resources and Evaluation (LREC 2022), Marseille, France. Mustafa Jarrar, Fadi Zaraket, Rami Asia, and Hamzeh Amayreh. 2018. Diacritic-based matching of arabic words. ACM Asian and Low-Resource Language Information Processing, 18(2):10:1-10:21. Mustafa Jarrar, Fadi Zaraket, Tymaa Hammouda, Daan-ish Masood Alavi, and Martin Waahlisch. 2023. Lisan: Yemeni, irqi, libyan, and sudanese arabic di-alect copora with morphological annotations. In The 20th IEEE/ACS International Conference on Com-puter Systems and Applications (AICCSA). IEEE.

JMHK23.pdf

Sanad Malaysha, Mustafa Jarrar, and Mohammad Khalilia. 2023. Context-gloss augmentation for im-proving arabic target sense verification. In Proceed-ings of the 12th International Global Wordnet Con-ference (GWC2023). Global Wordnet Association. Federico Martelli, Najla Kalach, Gabriele Tola, and Roberto Navigli. 2021. Semeval-2021 task 2: Mul-tilingual and cross-lingual word-in-context disam-biguation (mcl-wic). In Proceedings of the 15th In-ternational Workshop on Semantic Evaluation, Se-m Eval@ACL/IJCNLP 2021, Virtual Event / Bangkok, Thailand, August 5-6, 2021, pages 24-36. Associa-tion for Computational Linguistics. Marek Maziarz, Maciej Piasecki, and Stan Szpakowicz. 2012. Approaching plwordnet 2. 0. In Proceedings of 6th International Global Wordnet Conference, The Global Word Net Association, pages 189-196. George A Miller, Richard Beckwith, Christiane Fell-baum, Derek Gross, and Katherine J Miller. 1990. Introduction to wordnet: An on-line lexical database. International journal of lexicography, 3(4):235-244. George A Miller, Claudia Leacock, Randee Tengi, and Ross T Bunker. 1993. A semantic concordance. In Human Language Technology: Proceedings of a Workshop Held at Plainsboro, New Jersey, March 21-24, 1993. Simonetta Montemagni and Guglielmo Venturi. 2003. Building sense-tagged corpora for all: The itec project. In Proceedings of the Fourth International Conference on Language Resources and Evaluation (LREC'04). Eman Naser-Karajah, Nabil Arman, and Mustafa Jarrar. 2021. Current trends and approaches in synonyms extraction: Potential adaptation to arabic. In Pro-ceedings of the 2021 International Conference on Information Technology (ICIT), pages 428-434, Am-man, Jordan. IEEE. Amal Nayouf, Mustafa Jarrar, Fadi zaraket, Tymaa Ham-mouda, and Mohamad-Bassam Kurdy. 2023. Nâbra: Syrian arabic dialects with morphological annota-tions. In Proceedings of the 1st Arabic Natural Lan-guage Processing Conference (Arabic NLP), Part of the EMNLP 2023. ACL. Ahmed Mukhtar Omar. 2008. Contemporary arabic dic-tionary. (i1). World of Books, Cairo, Egypt. Retrieval Date, 14(8):2020. Tommaso Petrolito and Francis Bond. 2014. A survey of wordnet annotated corpora. In Proceedings of the Seventh Global Wordnet Conference, GWC 2014, Tartu, Estonia, January 25-29, 2014, pages 236-245. University of Tartu Press. Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. 2018. Improving language under-standing by generative pre-training. Open AI Techni-cal Report, 1(8). Karin C Ryding. 2014. Arabic: A linguistic introduc-tion. Cambridge University Press. Erik F. Tjong Kim Sang and Jorn Veenstra. 1999. Repre-senting text chunks. In EACL 1999, 9th Conference of the European Chapter of the Association for Com-putational Linguistics, June 8-12, 1999, University of Bergen, Bergen, Norway, pages 173-179. The As-sociation for Computer Linguistics. Nathan Schneider, Behrang Mohit, Kemal Oflazer, and Noah A. Smith. 2012. Coarse lexical semantic anno-tation with supersenses: An arabic case study. In The 50th Annual Meeting of the Association for Compu-tational Linguistics, Proceedings of the Conference, July 8-14, 2012, Jeju Island, Korea-Volume 2: Short Papers, pages 253-258. The Association for Com-puter Linguistics. Liling Tan and Francis Bond. 2012. Building and an-notating the linguistically diverse NTU-MC (NTU-multilingual corpus). Int. J. Asian Lang. Process., 22(4):161-174. Mariona Taulé, M. Antònia Martí, and Marta Recasens. 2008. An Cora: Multilevel annotated corpora for Catalan and Spanish. In Proceedings of the Sixth International Conference on Language Resources and Evaluation (LREC'08), Marrakech, Morocco. European Language Resources Association (ELRA). Heike Telljohann, Erhard Hinrichs, and Ra Ubler. 2004. The tüba-d/z treebank: Annotating german with a context-free backbone. Proceedings of the Fourth International Conference on Language Resources and Evaluation (LREC). Sophie Vanbelle. 2016. A new interpretation of the weighted kappa coefficients. Psychometrika, 81(2):399-410. Warren Weaver. 1949/1955. Translation. In William N. Locke and A. Donald Boothe, editors, Machine Translation of Languages, pages 15-23. MIT Press, Cambridge, MA. Reprinted from a memorandum written by Weaver in 1949. Ralph Weischedel, Martha Palmer, Mitchell Marcus, Eduard Hovy, Sameer Pradhan, Lance Ramshaw, Ni-anwen Xue, Ann Taylor, Jeff Kaufman, Mohammed El-Bachouti, Robert Belvin, and Ann Houston. 2013. Ontonotes release 5. 0 ldc2013t19. Philadelphia: Lin-guistic Data Consortium. Ralph Weischedel, Sameer Pradhan, Lance Ramshaw, Martha Palmer, Nianwen Xue, Mitchell Marcus, Ann Taylor, Craig Greenberg, Eduard Hovy, Robert Belvin, et al. 2011. Ontonotes release 4. 0. LDC2011T03, Philadelphia, Penn. : Linguistic Data Consortium. Boon Peng Yap, Andrew Koh, and Eng Siong Chng. 2020. Adapting bert for word sense disambiguation with gloss selection objective and example sentences. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 41-46.

JMHK23.pdf