| { | |
| "paper_id": "S01-1039", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T15:35:35.597537Z" | |
| }, | |
| "title": "Automatic WSD: Does it make sense of Estonian?", | |
| "authors": [ | |
| { | |
| "first": "Kadri", | |
| "middle": [], | |
| "last": "Vider", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "University of Tartu", | |
| "location": { | |
| "postCode": "50410", | |
| "settlement": "Tartu", | |
| "country": "Estonia" | |
| } | |
| }, | |
| "email": "[email protected]" | |
| }, | |
| { | |
| "first": "Kaarel", | |
| "middle": [], | |
| "last": "Kaljurand", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "University of Tartu", | |
| "location": { | |
| "postCode": "50410", | |
| "settlement": "Tartu", | |
| "country": "Estonia" | |
| } | |
| }, | |
| "email": "[email protected]" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "This paper describes a fully automatic Estonian word sense disambiguation system called semyhe which is based on Estonian WordNet (EstWN) hyponymjhypernym hierarchies and meant to disambiguate both nouns and verbs.", | |
| "pdf_parse": { | |
| "paper_id": "S01-1039", | |
| "_pdf_hash": "", | |
| "abstract": [ | |
| { | |
| "text": "This paper describes a fully automatic Estonian word sense disambiguation system called semyhe which is based on Estonian WordNet (EstWN) hyponymjhypernym hierarchies and meant to disambiguate both nouns and verbs.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Abstract", | |
| "sec_num": null | |
| } | |
| ], | |
| "body_text": [ | |
| { | |
| "text": "The main inspiration for our system is Agirre and Rigau (1996) similar system that disambiguates the English noun senses based on WordNet hyponymjhypernym hierarchy, taking into consideration the distances between the nodes corresponding to the word senses in the WordNet tree as well as the density of the tree. They have also experimented with using meronyms/holonyms in addition to hyponyms/hypernyms but report that it does not improve the results.", | |
| "cite_spans": [ | |
| { | |
| "start": 39, | |
| "end": 62, | |
| "text": "Agirre and Rigau (1996)", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Short description of the system", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "Our main object was not to focus on the homonymous words only (lexical sample), but to try to disambiguate all nouns and verbs in the text. The Estonian WordNet (EstWN) also contains adjectives but they are not linked by hyponym/hypernym relations. The word sense disambiguation could also try to describe a unique sense for adverbs but in our case such words have not yet been included in the thesaurus.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Short description of the system", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "As far as we know this is the first attempt on automatic Estonian word sense disambiguation.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Short description of the system", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "The input text for our system must be morphologically analyzed, meaning that each word is provided with its lemma and morphological reading. Taking those two into account we can localize the senses that correspond to the word in EstWN hyponym/hypernym tree (Vider et aL 1999) . It must be mentioned that although the morphological description in the input can be quite detailed, we only use the information on whether the word is a noun or a verb.", | |
| "cite_spans": [ | |
| { | |
| "start": 257, | |
| "end": 275, | |
| "text": "(Vider et aL 1999)", | |
| "ref_id": "BIBREF4" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Input", | |
| "sec_num": "1.1" | |
| }, | |
| { | |
| "text": "A simple morphological analysis that only looks at the word-form and not its context can result in very ambiguous output. On average 45% of the words are morphologically ambiguous in Estonian texts (Kaalep, 1997) . The ambiguity can be greatly reduced by also applying the Estonian morphological disambiguator (Kaalep and Vaino, 1998) to the text before the word sense disambiguation. Since even then the words can in principle stay morphologically ambiguous, our system doesn't require each word to have exactly one morphological reading assigned to it in the input text.", | |
| "cite_spans": [ | |
| { | |
| "start": 198, | |
| "end": 212, | |
| "text": "(Kaalep, 1997)", | |
| "ref_id": "BIBREF2" | |
| }, | |
| { | |
| "start": 310, | |
| "end": 334, | |
| "text": "(Kaalep and Vaino, 1998)", | |
| "ref_id": "BIBREF1" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Input", | |
| "sec_num": "1.1" | |
| }, | |
| { | |
| "text": "Similarly to the morphological analysis, we do not try to provide each word with exactly one sense. In case two (or more) senses have equal evaluation results then both of those prevail in the output.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Output", | |
| "sec_num": "1.2" | |
| }, | |
| { | |
| "text": "We apply the exact same algorithm for both nouns and verbs. Nouns and verbs cannot be compared with each other since in terms of hyponym/hypernym hierarchy they are located in different trees in the thesaurus. So the disambiguation is carried out in 2 runs, first nouns are disambiguated, then verbs, or vice versa.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Sense disambiguation algorithm", | |
| "sec_num": "1.3" | |
| }, | |
| { | |
| "text": "A window is shifted on the text and as a word moves through the window its senses are compared with the senses of other words in the window. The context is either made out of nouns or verbs depending on which part of speech is being disambiguated.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Sense disambiguation algorithm", | |
| "sec_num": "1.3" | |
| }, | |
| { | |
| "text": "The basis of the comparison is the similarity between the senses which is defined through the notion of conceptual distance, the distance between the nodes corresponding to the senses in Est WN tree. \\Vinners are the senses that minimize\u2022 the total distance between the word senses in the window, all the rest are removed from the list of candidates for the correct reading. semyhe leaves the word ambiguous when there are more than one senses with equal result. This usually happens when the senses of the context vmrds are located in different hierarchies and hence can not be compared. Currently there are 108 different top nodes in EstVVN, 29 corresponding to nouns and 79 to verbs.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Sense disambiguation algorithm", | |
| "sec_num": "1.3" | |
| }, | |
| { | |
| "text": "In addition, the work of the system can be modified via several options in the configuration file:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Sense disambiguation algorithm", | |
| "sec_num": "1.3" | |
| }, | |
| { | |
| "text": "\u2022 The window-size can be changed, increasing it makes the output less ambiguous since there is a higher possibility that the comparable senses end up in one window.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Sense disambiguation algorithm", | |
| "sec_num": "1.3" | |
| }, | |
| { | |
| "text": "On the other hand a bigger window may span across several sentences making the compared words possibly irrelevant to each other. For the moment we have used window of 5 words.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Sense disambiguation algorithm", | |
| "sec_num": "1.3" | |
| }, | |
| { | |
| "text": "\u2022 Since we use no syntactic analysis before word sense disambiguation, the context of any word under observation is unstructured, the only syntactic information that we can use is therefore only the distance of the words from each other in terms of running texL A set of weights can be defined that is mapped to the distances, so that the similarity of the senses of the words that are far away from each other is less relevant for the total score.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Sense disambiguation algorithm", | |
| "sec_num": "1.3" | |
| }, | |
| { | |
| "text": "\u2022 We can also take into account the average depth of the compared nodes in the treethe bigger the depth the more reliable the score.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Sense disambiguation algorithm", | |
| "sec_num": "1.3" | |
| }, | |
| { | |
| "text": "So far we haven't yet experimented with any of those options much.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Sense disambiguation algorithm", | |
| "sec_num": "1.3" | |
| }, | |
| { | |
| "text": "For the purposes of analyzing the quality of disambiguation, tests were made with 12 manually sense tagged texts. These text samples were mainly from fiction, in a part also from newspapers and they contained approximately 160 10,000 tokens that corresponded to either nouns or verbs.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Analysis of the results", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "Manual tagging naturally had to remove all the morphological ambiguity, therefore the results obtained on those texts should be better than on the texts that have only been treated automatically before the word sense disambiguation. Words that occurred in the text but were not present in Est WN were marked as having 0 senses, approximately 30% of such words are proper names.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Analysis of the results", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "Manual tagging also recognized multi-word units, which in our case are mostly noncontiguous verbal phrases that are hard to detect automatically even if we had a complete list of such units.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Analysis of the results", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "Using semyhe, we set out to disambiguate all the nouns and verbs contained in the texts. Since semyhe can leave a word ambiguous, it makes sense to evaluate its work in terms of recall and precision. Table 1 lists semyhe results when the window of context words has size 5. The table also shows the results obtained with a random method which chooses exactly one sense for every word randomly (in this case recall and precision have equal values).", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 200, | |
| "end": 207, | |
| "text": "Table 1", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Analysis of the results", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "The row groups of the table refer respectively to the results with polysemous words and the overall results. Note that the words which were manually marked as having 0 senses were considered monosemous and so they are always correctly analyzed, with unique sense selected for every word. Table 1 : semyhe results with 10,000 nouns and verbs Figure 1 shows the distribution of words between cliff:'erent number of senses according to those texts. This shows the ambiguity that any automatic analysis has to cope with.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 288, | |
| "end": 295, | |
| "text": "Table 1", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 341, | |
| "end": 349, | |
| "text": "Figure 1", | |
| "ref_id": "FIGREF0" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Analysis of the results", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "No,te that there is an unusually large number of words with 9 different senses. The main reason for this is the frequent word 'olema' (to be, to have). Fortunately the distribution of its senses is highly skewed, meaning that mostly this word is used in one or two senses. Including the sense frequency information in the disambiguation process could considerably improve the results.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "POS recall", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Several problems have been discovered concerning the relatively simple approach described above.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Proble1ns and solutions", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "The output of the morphological analyzer often contains valuable information for word sense disambiguation which we have currently ignored.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Proble1ns and solutions", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "\u2022 in some cases the word-form used in the text can uniquely specify the sense of the word, although its lemma is ambiguous, e.g. the word 'palk' can either mean salary or log, tree trunk, but its genitive form is different in each meaning (either 'palga' or 'palgi'). By using only the lemma we ignore this distinction that can be explicitly present in the text. The number of words behaving this way, though, is not very large.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Proble1ns and solutions", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "\u2022 the modal verbs are explicitly marked in the output of the morphological disambiguator, when a verb is marked as such, then the senses that don't correspond to the modal senses could be removed and the winning sense should be chosen from the prevailing ones, e.g. the word 'saama' has all together 12 senses in the thesaurus, but only 2 of them correspond to the modal use of the word (either can or may).", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Proble1ns and solutions", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "Right now the frequency information of the senses has not been used. Most probably theresults that could be obtained with the \"commonest\" baseline (Kilgariff and Rosenzweig, 2000) would beat the results of semyhe. We think that even the frequency information that could be calculated using the 10,000 manually sensetagged words can be very useful for disambiguating purposes.", | |
| "cite_spans": [ | |
| { | |
| "start": 147, | |
| "end": 179, | |
| "text": "(Kilgariff and Rosenzweig, 2000)", | |
| "ref_id": "BIBREF3" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Proble1ns and solutions", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "At the moment the input text contains no information about its syntactic structure, most importantly the verbal phrases and other multiword units are not marked as such, therefore the analyzer tries to disambiguate all the components of a multi-word unit separately, this of course results in an incorrect analysis. Also, having the information about the syntactic structure of the sentences could help to reduce the number of possible senses to choose from. For example the word 'olema' that was already mentioned above has five more frequent senses:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Proble1ns and solutions", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "1. be -copula, used with an adjective or a predicate noun 2. exist -have an existence, be extant 3. stay in place, be stationary or spend a certain length of time 4. be somewhere, occupy a certain area, occupy a certain position 5. have, have got, hold -have or possess, either in a concrete or an abstract sense", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Proble1ns and solutions", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "The first sense is present in complementary clauses; senses 2, 3 and 4 appear in existential sentences and the last one in possessive sentences. If the information about the nature of the sentence was present in the input text it would certainly help the disambiguation process.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Proble1ns and solutions", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "The output of semyhe stays often very ambiguous. This either happens when the sense-nodes of the context words are located in different trees so that their similarity cannot be calculated; or when different nodes of one word have the same parent node and are equally distant from the rest of the sense-nodes so that the similarity measure for them will be equal. The second reason may not be a big problem considering WordNet's fine-grainedness and the fact that for some applications a detailed sense distinction is not needed. The disambiguation result in this case can be simply seen as the union of the prevailed senses. Often, though, this approach does not hold, e.g. it is crucial for translation that the senses of the word 'naine' which can either stand for woman, wife or generally female person, are fully disambiguated, although the senses stand for more or less the same thing.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Proble1ns and solutions", | |
| "sec_num": "3" | |
| } | |
| ], | |
| "back_matter": [], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "title": "\\Vord Sense Disambiguation using Conceptual Density", | |
| "authors": [ | |
| { | |
| "first": "E", | |
| "middle": [], | |
| "last": "Agirre", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [ | |
| "R" | |
| ], | |
| "last": "", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1996, | |
| "venue": "COLING-96", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "E. Agirre and G. R.igau. 1996. \\Vord Sense Disambiguation using Conceptual Density. In COLING-96.", | |
| "links": null | |
| }, | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "title": "Kas vale rneetodiga oiged tulemused? Statistikale tuginev eesti keele morfoloogiline iihestamine", | |
| "authors": [ | |
| { | |
| "first": "H.-J", | |
| "middle": [], | |
| "last": "Kaalep", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Vaino", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1998, | |
| "venue": "Estonian. English title: Getting right result with a wrong method? Statistical morphological disambiguation of Estonian", | |
| "volume": "1", | |
| "issue": "", | |
| "pages": "30--36", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "H.-J. Kaalep and T. Vaino. 1998. Kas vale rneetodiga oiged tulemused? Statistikale tug- inev eesti keele morfoloogiline iihestamine. Keel ja Kirjandus, 1:30-36. In Estonian. En- glish title: Getting right result with a wrong method? Statistical morphological disam- biguation of Estonian.", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "An Estonian morphological analyser and the impact of a corpus on its development", | |
| "authors": [ | |
| { | |
| "first": "H.-J", | |
| "middle": [], | |
| "last": "Kaalep", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1997, | |
| "venue": "Computers and the Humanities", | |
| "volume": "31", | |
| "issue": "", | |
| "pages": "115--133", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "H.-J. Kaalep. 1997. An Estonian morphologi- cal analyser and the impact of a corpus on its development. Computers and the Human- ities, 31:115-133.", | |
| "links": null | |
| }, | |
| "BIBREF3": { | |
| "ref_id": "b3", | |
| "title": "Framework and Results for English SENSEVAL. Computers and the Hmnanitz:es", | |
| "authors": [ | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Kilgariff", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Rosenzweig", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2000, | |
| "venue": "", | |
| "volume": "34", | |
| "issue": "", | |
| "pages": "15--48", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "A. Kilgariff and J. Rosenzweig. 2000. Frame- work and Results for English SENSEVAL. Computers and the Hmnanitz:es, 34:15-48.", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "title": "The Estonian Worclnet", | |
| "authors": [ | |
| { | |
| "first": "K", | |
| "middle": [], | |
| "last": "Vider", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Paldre", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Orav", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Oim", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1999, | |
| "venue": "Deliverable 2D014", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "K. Vider, L. Paldre, H. Orav, and H. Oim. 1999. The Estonian Worclnet. In C. Kunze, edi- tor, Final Wordnets for German. French, Es- tonian and Czech. EuroWorclNet (LE-8328), Deliverable 2D014.", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "FIGREF0": { | |
| "text": "distribution of words in running text", | |
| "num": null, | |
| "uris": null, | |
| "type_str": "figure" | |
| } | |
| } | |
| } | |
| } |