Datasets:

WINGNUS
/

ACL-OCL

	{
	"paper_id": "O00-3001",
	"header": {
	"generated_with": "S2ORC 1.0.0",
	"date_generated": "2023-01-19T07:59:01.405211Z"
	},
	"title": "Adaptive Word Sense Disambiguation Using Lexical Knowledge in a Machine-readable Dictionary",
	"authors": [
	{
	"first": "Jen",
	"middle": [
	"Nan"
	],
	"last": "Chen",
	"suffix": "",
	"affiliation": {
	"laboratory": "",
	"institution": "Ming Chuan University",
	"location": {
	"addrLine": "Shih-lin",
	"country": "Taiwan, R.O.C"
	}
	},
	"email": ""
	}
	],
	"year": "",
	"venue": null,
	"identifiers": {},
	"abstract": "This paper describes a general framework for adaptive conceptual word sense disambiguation. The proposed system begins with knowledge acquisition from machine-readable dictionaries. Central to the approach is the adaptive step that enriches the initial knowledge base with knowledge gleaned from the partial disambiguated text. Once the knowledge base is adjusted to suit the text at hand, it is applied to the text again to finalize the disambiguation decision. Definitions and example sentences from the Longman Dictionary of Contemporary English are employed as training materials for word sense disambiguation, while passages from the Brown corpus and Wall Street Journal (WSJ) articles are used for testing. An experiment showed that adaptation did significantly improve the success rate. For thirteen highly ambiguous words, the proposed method disambiguated with an average precision rate of 70.5% for the Brown corpus and 77.3% for the WSJ articles.",
	"pdf_parse": {
	"paper_id": "O00-3001",
	"_pdf_hash": "",
	"abstract": [
	{
	"text": "This paper describes a general framework for adaptive conceptual word sense disambiguation. The proposed system begins with knowledge acquisition from machine-readable dictionaries. Central to the approach is the adaptive step that enriches the initial knowledge base with knowledge gleaned from the partial disambiguated text. Once the knowledge base is adjusted to suit the text at hand, it is applied to the text again to finalize the disambiguation decision. Definitions and example sentences from the Longman Dictionary of Contemporary English are employed as training materials for word sense disambiguation, while passages from the Brown corpus and Wall Street Journal (WSJ) articles are used for testing. An experiment showed that adaptation did significantly improve the success rate. For thirteen highly ambiguous words, the proposed method disambiguated with an average precision rate of 70.5% for the Brown corpus and 77.3% for the WSJ articles.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Abstract",
	"sec_num": null
	}
	],
	"body_text": [
	{
	"text": "Word sense disambiguation is a long-standing problem in natural language understanding. It seems to be very difficult to statistically acquire enough word-based knowledge about a language to build a robust system capable of automatically disambiguating senses in unrestricted text. For such a system to be effective, a large number of balanced materials must be assembled in order to cover many idiosyncratic aspects of the language. There exist three issues in a lexicalized statistical word sense disambiguation (WSD) model: data sparseness, the lack of abstraction, and static learning. First, a word-based model has a plethora of parameters that are difficult to estimate reliably even with a very large corpus. Under-trained models lead to low precision. Second, word-based models lack a degree of abstraction that is crucial for a broad coverage system. Third, a static WSD model is unlikely to be robust and portable, since it is very difficult to build a single model relevant to a wide variety of unrestricted texts. Several WSD systems have been developed that apply word-based models to a specific or genre domain to disambiguate senses appearing in generally easy context that has a large number of typically salient words. In the case of unrestricted text, however, the context tends to be very diverse and difficult to capture with a lexicalized model; therefore, a corpus-trained system is unlikely to transfer well to a new domain.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Introduction",
	"sec_num": "1."
	},
	{
	"text": "Generality and adaptability are, therefore, keys to a robust and portable WSD system. A concept-based model for WSD requires fewer parameters and has an element of generality built in. Conceptual classes make it possible to generalize from word-specific context in order to disambiguate word senses appearing in an unfamiliar context in terms of word recurrences. An adaptive system, armed with an initial lexical and conceptual knowledge base extracted from machine-readable dictionaries (MRD), has two strong advantages over static lexicalized models trained on a corpus. First, the initial knowledge is rich and unbiased enough for a substantial portion of text to be disambiguated correctly. Second, based on the result of initial disambiguation, an adaptation step can then be performed to make the knowledge base more relevant to the task at hand, thus resulting in broader and more precise WSD.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Introduction",
	"sec_num": "1."
	},
	{
	"text": "In this paper we explore in some depth the question of whether conceptual knowledge in the MRD is effective enough to provide a general solution for disambiguating contexts of unrestricted texts, such as the Brown and Wall Street Journal (WSJ) corpora. Major emphasis has previously been placed on self-adaptation [Chen and Chang 1998a] . This approach is based on the hypothesis that a substantial part of a given text is easy or prototypical and, therefore, susceptible to interpretation based on general knowledge derived from the MRD. By adapting the contextual representation of word senses to those in the easy context, we hope to be better equipped to interpret the other part, which is usually considered a hard context. Adaptation results in gaps in the general knowledge being filled in or domain specific information being added to the initial knowledge base. Either way, adaptation makes the knowledge base more relevant to the text and, therefore, more effective for WSD in a hard context. We will give experimental results showing the effectiveness of this adaptive WSD approach based on initial knowledge base acquired from the MRD. Although our adaptive approach requires virtually no domain-specific training, it nevertheless achieves high precision rates for WSD of unrestricted text rivaling those of static methods that demand very lengthy training using a very large corpus. Figure 1 lays out the general framework for the adaptive conceptual WSD approach which this research employed. The learning process described here begins with a step involving knowledge acquisition from MRDs. With this acquired knowledge, the input text is read and a trial disambiguation step is carried out. An adaptation step follows which combines the initial knowledge base with knowledge gleaned from the partially disambiguated text. Once the knowledge base is adjusted to suit the text at hand, it is then applied to the text again to finalize the disambiguation result. For instance, the Adptive 3 initial contextual representation (CR) extracted from the Longman Dictionary of Contempory English [Proctor 1978, LDOCE] for the bank-GEO sense contains both lexical and conceptual information: {land, river, lake, \u2022\u2022\u2022} \u222a {GEO, MOTION, \u2022\u2022\u2022}. The initial CR is informative enough to disambiguate a passage containing \"a deer near the river bank\" in the input text. The trial disambiguation step produces sense tagging of deer/ANIMAL and bank/GEO, but certain instances of bank are left untagged due to the lack of WSD knowledge. We observe that the bank-GEO sense in the context of vole is unresolved since there is no link between ANIMAL and GEOGRAPHY. Subsequently, the adaptation step adds deer and ANIMAL to the contextual representation for bank-GEO. The adapted CR is now enriched with information capable of disambiguating the instance of bank in the context of vole to produce the final disambiguation result.",
	"cite_spans": [
	{
	"start": 314,
	"end": 336,
	"text": "[Chen and Chang 1998a]",
	"ref_id": "BIBREF6"
	},
	{
	"start": 2102,
	"end": 2123,
	"text": "[Proctor 1978, LDOCE]",
	"ref_id": null
	}
	],
	"ref_spans": [
	{
	"start": 1396,
	"end": 1404,
	"text": "Figure 1",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "Introduction",
	"sec_num": "1."
	},
	{
	"text": "The rest of this paper is organized as follows. First of all, we will peresent how easy contexts are interpreted and ambiguous words are labeled in the initial disambiguation step using general knowledge derived from MRD. Next, we describe the adaptation step that uses the sense labels assigned to polysemous words. After that, we will describe the strategy of using the adapted knowledge base and defaults. Next, we will give a detailed account of experiments conducted to assess the effectiveness of the adaptive approach, including the experiment setup, results and evaluation. Following that, we will review the recent WSD literature from the perspective of various types of contextual knowledge and different representation schemes. Finally, we will draw conclusions.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Introduction",
	"sec_num": "1."
	},
	{
	"text": "General framework for adaptive WSD using MRD.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Figure 1",
	"sec_num": null
	},
	{
	"text": "In this section, we will describe how the conceptual characterization technique is applied to MRD definitions and give examples of acquiring WSD knowledge. First, we will show word level definitions based on a lexical CR and then a conceptual CR. Next, we will show the advantage of including information gained from an example sentence. Finally, we will combine these techniques to perform adaptative WSD computation.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Acquisition of Disambiguation Knowledge using MRD",
	"sec_num": "2."
	},
	{
	"text": "A word-level contextual representation from MRD definitions can be derived almost effortlessly. Let CR(W, S) denote the contextual representation of the sense S of headword W. Intuitively, it is composed of the content words in the definition with a specific sense. Thus, CR(W, S) can be represented symbolically as { x \| x\u2208DEF 1 W, S and x is not a function word }. To illustrate, the CRs for the nine nominal senses of bank in LDOCE listed below are shown in Table 1: bank.1.n.1 land along the side of a river, lake, etc.; bank.1.n.2 earth which is heaped up in a field or garden, often making a border or division; bank.1.n.3 a mass of snow, clouds, mud, etc.; bank.1.n.4 a slope made at bends in a road or race-track, so that they are safer for cars to go round; bank.1.n.5 a high underwater of bank in a river, harbour, etc.; bank.3.n.1 a row, esp. of OAR in an ancient boat or KEY on a TYPEWRITER; bank.4.n.1 a place in which money is kept and paid out on demand, and where related activities go on; bank.4.n.2 a place where something is held ready for use, esp. ORGANIC products of human origin for medical use; bank.4.n.3 a person who keeps a supply of money or pieces for payment or use in a game of chance. ",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 461,
	"end": 469,
	"text": "Table 1:",
	"ref_id": "TABREF0"
	}
	],
	"eq_spans": [],
	"section": "Lexicalized Contextual Representation",
	"sec_num": "2.1.1"
	},
	{
	"text": "The word-based CR from MRD definitions is highly precise and effective but not broad enough to work alone effectively. Word-based sense representation is hampered by the difficulty of providing estimates for a very large parameter space leading to limited coverage in WSD. Certainly, there are many situations that call for a conceptual generalization of a word-based representation of word sense from an example sentence. For instance, the RIVER sense of bank in Example (1c) can be correctly interpreted by an MRD-based CR, but only when the contextual word river in the CR is generalized to all words related to RIVER, including the word stream:",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Conceptualized Contextual Representation",
	"sec_num": "2.1.2"
	},
	{
	"text": "(1) a. a ribbon of mist along the river bank; b. a small excavation in the river bank; c. the left bank of the stream.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Conceptualized Contextual Representation",
	"sec_num": "2.1.2"
	},
	{
	"text": "There are many possible approaches to making such a generalization and deriving a conceptualized CR (CCR) of word sense. Chen and Chang [1998b] described one such approach based on thesaurus topics. The CCR for each MRD sense can be viewed as relating to words listed under some Longman Lexicon of Contemporary English [McArthur 1992, LLOCE] topics. By linking MRD senses to thesaurus senses and by classifying senses according to linked senses, we can derive the CCR for a sense definition. Table 2 shows the topical CCR for the senses of bank in LDOCE. Each sense in MRD is given a list of weighted LLOCE topics. The weights in the CCR are normalized to a sum of unity for the obvious reason. Table 3 shows the lists of words listed in LLOCE under the topics relevant to bank senses. We sum up the above description and outline the procedure as Algorithm 1 for creating a CCR for a word W of sense S with definition D.",
	"cite_spans": [
	{
	"start": 121,
	"end": 143,
	"text": "Chen and Chang [1998b]",
	"ref_id": "BIBREF7"
	},
	{
	"start": 319,
	"end": 341,
	"text": "[McArthur 1992, LLOCE]",
	"ref_id": null
	}
	],
	"ref_spans": [
	{
	"start": 492,
	"end": 500,
	"text": "Table 2",
	"ref_id": "TABREF1"
	},
	{
	"start": 696,
	"end": 703,
	"text": "Table 3",
	"ref_id": "TABREF2"
	}
	],
	"eq_spans": [],
	"section": "Conceptualized Contextual Representation",
	"sec_num": "2.1.2"
	},
	{
	"text": "Step 1: Run the TopSense algorithm described by Chen and Chang [1998b] to map D to SC(D), a set of semantic categories in a thesaurus.",
	"cite_spans": [
	{
	"start": 48,
	"end": 70,
	"text": "Chen and Chang [1998b]",
	"ref_id": "BIBREF7"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Algorithm 1: Creating a conceptualized contextual representation CCR(D w, s ) for a word W of sense S with definition D.",
	"sec_num": null
	},
	{
	"text": "2 See Appendix A for more details.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Algorithm 1: Creating a conceptualized contextual representation CCR(D w, s ) for a word W of sense S with definition D.",
	"sec_num": null
	},
	{
	"text": "Step 2: Create a conceptualized contextual representation CCR(D w, s ) for sense S with definition D:",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Adptive 7",
	"sec_num": null
	},
	{
	"text": "CCR(D w, s ) = \u2211 \u2208 ) ( SC ) ( D T T WORD 3 ,",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Adptive 7",
	"sec_num": null
	},
	{
	"text": "where WORD(T) is a set of related words in semantic category T. In the following, we demonstrate how Algorithm 1 works. Given the sense definition of bank.4.n.1 shown in Section 2.1.1, the CCR(bank.4.n.1) can be acquired by means of Algorithm 1.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Adptive 7",
	"sec_num": null
	},
	{
	"text": "Step 1: After running the TopSense algorithm, we have SC(bank.4.n.1) = {Je, Jf, Jd}.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Adptive 7",
	"sec_num": null
	},
	{
	"text": "Step 2: Next, we expand each of three topics in SC(bank.4.n.1) to a cluster of words. Thus, we have WORD (Je) ={money, pay, cash, capital, account, charge, ...}, WORD (Jf) ={pay, bond, bill, charge, ...} and WORD (Jd) ={money, cash, fund, check, ...}.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Adptive 7",
	"sec_num": null
	},
	{
	"text": "Finally, the CCR (bank.4.n.1) =WORD (Je) pay, cash, capital, ..., pay, bond, bill, charge, ..., money, cash, fund, check, . .. }.",
	"cite_spans": [
	{
	"start": 41,
	"end": 123,
	"text": "pay, cash, capital, ..., pay, bond, bill, charge, ..., money, cash, fund, check, .",
	"ref_id": null
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Adptive 7",
	"sec_num": null
	},
	{
	"text": "+ WORD (Jf) + WORD (Jd) ={money,",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Adptive 7",
	"sec_num": null
	},
	{
	"text": "Dictionary examples are intended to show typical use of words in context. Therefore, MRD examples provide rich information supplementary to definitions. In this section, we will describe a method for tagging bilingual sentences with sense labels based on dictionary definitions and translations in a bilingual MRD.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Contextual Representation from an Example Sentence",
	"sec_num": "2.2"
	},
	{
	"text": "However, the sense for each word in an example is not explicitly marked except for the word being defined. That limits the potential for using dictionary examples as knowledge sources for WSD. Gale, Church and Yarowsky [1992b] first pointed out that the strong constraint of one-sense-per-translation can be exploited to tag a bilingual corpus for training a statistical WSD model. Building on their idea, we describe a new method for tagging bilingual sentences, in the MRD or elsewhere, for automatic acquisition of the CR of senses.",
	"cite_spans": [
	{
	"start": 193,
	"end": 226,
	"text": "Gale, Church and Yarowsky [1992b]",
	"ref_id": "BIBREF13"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Contextual Representation from an Example Sentence",
	"sec_num": "2.2"
	},
	{
	"text": "Now we are ready to propose a heuristic algorithm for tagging bilingual sentences with sense labels. First, the translation morphemes of an MRD definition are added to the CR so that not only the English context, but also the translation (in Chinese for the particular implementation of LDOCE/E-C we will be describing) is considered. For instance, the representation of MONEY-bank contains not only FINANCE words such as money, pay, cash, capital, account, charge, etc., but also the morphemes \"\u9280\" Adptive 9",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Sense Tagging Based on Conceptual Context Representation and Translations",
	"sec_num": "2.2.1"
	},
	{
	"text": "and \"\ufa08\" in the translation of the definition. (See Table 4 for some examples of bilingual context representations for the bank senses in LDOCE/E-C.) Subsequently, each CR for a polysemous word is compared with the bilingual sentences. The polysemous word is tagged in favor of the relevant CR that has the most overlap with the bilingual sentences. For instance, consider the case of tagging the instance of bank in Example (2) extracted from LDOCE/E-C:",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 51,
	"end": 58,
	"text": "Table 4",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "Sense Tagging Based on Conceptual Context Representation and Translations",
	"sec_num": "2.2.1"
	},
	{
	"text": "(2) a. the interest in my bank account accrued over the years; b. \u6211\u9280\ufa08\u5e33\u6236\u7684\uf9dd\u606f\u9010\uf98e\u6709\u6240\u589e\u52a0\u3002 Under the assumption of the one-sense-per-translation constraint, the morphemes \"\u9280\" and \"\ufa08\" in the translation are sufficient evidence for tagging the instance of bank as MONEY-bank. Even if such telling evidence is not present, there nevertheless is a great chance that the sentence contains enough words related to a relevant topic for correct sense tagging to happen. For instance, the FINANCE words, such as interest and account, in Example (2) lead to the correct sense label MONEY-bank for this instance of bank, even when it is not translated as \"\u9280\ufa08.\" The contextual representation derived from the MRD definition also acts as a safety net when the one-translation-per-sense constraint does not hold. For instance, based on the one-translation-per-sense constraint, the instance of star in Example (3) can not be labeled as ENTERTAINMENT-star because both the ENTERTAINMENT and HEAVENLY-BODY senses of star are translated as \"\u661f\":",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Sense Tagging Based on Conceptual Context Representation and Translations",
	"sec_num": "2.2.1"
	},
	{
	"text": "(3) a. she is a star with the theatre company; b. \u5979\u662f\u5287\u5718\u7684\u7d05\u661f\u3002",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Sense Tagging Based on Conceptual Context Representation and Translations",
	"sec_num": "2.2.1"
	},
	{
	"text": "In such an event, the ENTERTAINMENT words, such as theatre and company, nonetheless result in the correct sense label: ENTERTAINMENT-star.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Sense Tagging Based on Conceptual Context Representation and Translations",
	"sec_num": "2.2.1"
	},
	{
	"text": "A bilingual example in the MRD, or text in a bilingual corpus, can be tagged in the way described above, word by word and sentence by sentence. Unambiguous words with only one sense label are tagged as such. Tagging is done only for content words within the scope of this work. Function words can be treated similarly [Chang, Hsu and Chen 1996] . Sentences in English tagged as training materials can facilitate acquisition of WSD knowledge. The method for tagging a bilingual training corpus is summarized as Algorithm 2. Table 5 shows the result of applying Algorithm 2 to some LDOCE/E-C examples. Step 1: Form contextual representation CR(W, S) of sense S of word W with definition D and translation T as follows:",
	"cite_spans": [
	{
	"start": 318,
	"end": 344,
	"text": "[Chang, Hsu and Chen 1996]",
	"ref_id": null
	}
	],
	"ref_spans": [
	{
	"start": 523,
	"end": 530,
	"text": "Table 5",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "Sense Tagging Based on Conceptual Context Representation and Translations",
	"sec_num": "2.2.1"
	},
	{
	"text": "CR(W, S) = LCR(D w, s ) + CCR(D w, s ) + LCR(T w, s ).",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Sense Tagging Based on Conceptual Context Representation and Translations",
	"sec_num": "2.2.1"
	},
	{
	"text": "Step 2: For each word W in an example sentence E, compute the similarity between its context and translation, C E , and each of the contextual representations CR(W,S) based on the Dice Coefficient:",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Adptive 11",
	"sec_num": null
	},
	{
	"text": "Sim (C E , CR(W, S)) = \u2211 \u00d7 \u2208 E \| ) , ( \| + \| \| ) ) , ( , In( 2 E C c S W CR C S W CR C ,",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Adptive 11",
	"sec_num": null
	},
	{
	"text": "where In(a, B) = the weight of a in B, if a \u2208 B and 0, otherwise.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Adptive 11",
	"sec_num": null
	},
	{
	"text": "Step 3:",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Adptive 11",
	"sec_num": null
	},
	{
	"text": "Label W in E with S * such that Sim (C E , CR(W,S * )) is maximized; Sim (C E , CR(W,S * )) = L Max Sim (C E , CR(W,L))",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Adptive 11",
	"sec_num": null
	},
	{
	"text": "and is greater than a certain threshold.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Adptive 11",
	"sec_num": null
	},
	{
	"text": "Lexicalized and conceptualized CR can be constructed from tagged MRD examples in a fashion similar to that described in Section 2.1 for MRD definitions. Given an ambiguous word W labeled with sense S in a set of example sentences E w, s , every content word appearing in E is gathered to form LCR(E w, s ), shown as follows:",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Acquiring Contextual Representations for Example Sentences",
	"sec_num": "2.2.2"
	},
	{
	"text": "LCR(E w, s ) = { x \| x \u2208 E w, s and",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Acquiring Contextual Representations for Example Sentences",
	"sec_num": "2.2.2"
	},
	{
	"text": "x is not a function word }. Table 6 shows some of the contextual words in the LDOCE examples that appear in the context of each of eight bank senses. Notice that the entry for MONEY-bank contains many strong collocates, such as (rob, bank), (bank, account), etc. These collocates are potentially very helpful for WSD. Although some of the contextual words merely repeat information in the definition-based representation, LCR(D w, s ) + CCR(D w, s ), many do provide new information. For instance, fifteen instances of river reaffirm the defining word river as an important collocate for RIVER-bank, while contextual words such as north, east, deer, and vole provide additional, richer context. ",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 28,
	"end": 35,
	"text": "Table 6",
	"ref_id": "TABREF4"
	}
	],
	"eq_spans": [],
	"section": "Acquiring Contextual Representations for Example Sentences",
	"sec_num": "2.2.2"
	},
	{
	"text": "(2), north(2), stream(2), east(1), air(1), deer(1), south(1), sea(1), vole(1), ... EARTH build(2), earth(2), flood(1), rise(1), water(1), ... PILE cloud(2), dark(2), heavy(1), storm(1), ... ROAD moss(2), wood(2), rest(1), sit(1), ... ROW - MEDICINE - GAMBLE -",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Acquiring Contextual Representations for Example Sentences",
	"sec_num": "2.2.2"
	},
	{
	"text": "In the previous section, we showed that these contextual words are neither frequent nor necessarily likely to recur. However, when viewed as representing a typical topic or concept, they certainly are recurring. For instance, although there is only one instance of north bank in LDOCE examples, there are quite a few south bank, and right bank instances, all of which signal a recurring context of the DIRECTION concept. Therefore, it is a good idea to derive a conceptualized contextual representation from the set of examples E relevant to a sense label S. For instance, representing the co-occurring concept of the DIRECTION with RIVER-bank, CCR(E bank, river ) would contain such words as east, west, south, north, left, and right, etc.: CCR(E bank, river ) = { east, west, south, north, left, right, \u2026 }.",
	"cite_spans": [
	{
	"start": 699,
	"end": 741,
	"text": "west, south, north, left, and right, etc.:",
	"ref_id": null
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Acquiring Contextual Representations for Example Sentences",
	"sec_num": "2.2.2"
	},
	{
	"text": "For this purpose, we again turn to the information retrieval (IR) technique. Since the LDOCE in general strictly uses words in the controlled vocabulary for both definitions and examples, the same method described by Chen and Chang [1998b] for forming conceptual characterization of MRD definitions also works for MRD examples. Table 7 shows a list of topical words that characterize the context of each of the eight bank senses based on sense tagged LDOCE examples. The results obtained using an IR-based method seem to characterize the context in a general way that can be very useful for WSD. ",
	"cite_spans": [
	{
	"start": 217,
	"end": 239,
	"text": "Chen and Chang [1998b]",
	"ref_id": "BIBREF7"
	}
	],
	"ref_spans": [
	{
	"start": 328,
	"end": 335,
	"text": "Table 7",
	"ref_id": "TABREF5"
	}
	],
	"eq_spans": [],
	"section": "Acquiring Contextual Representations for Example Sentences",
	"sec_num": "2.2.2"
	},
	{
	"text": "Definition-based and example-based CR as described in Sections 2.1 and 2.2 can be put together to form a combined CR for acquiring word sense. For simplicity, we merge the two to produce the final MRD-based CR. For a polysemous word W and a relevant word sense S, with the definition D of sense Adptive 13 S and the set of examples E containing an instance of S, the contextual representation CR(W, S) can be represented as follows:",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Combining Definition-based and Example-based CR",
	"sec_num": "2.3"
	},
	{
	"text": "WORD(W, S) = LCR(D w, s ) + CCR(D w, s ) + LCR(E w, s ) + CCR(E w, s ),",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Combining Definition-based and Example-based CR",
	"sec_num": "2.3"
	},
	{
	"text": "where To take into account the significance of each contextual word in CR(W, S), the IR technique for weighting index terms for relevancy can be applied here to good effect. Using the IR analogy, the collective context of each word sense is viewed as a document, and the relevance of a contextual word t to a sense S of word W depends on its term frequency tf and inverse document frequency idf. The term frequency tf is the number of instances of t in WORD(W, S), and idf is the percentage of CRs in which an instance of t appears. The relevancy of a contextual word is estimated using the commonly used scheme: tf \u00d7idf. Experiments show that the simple scheme tends to give a high weight to strong collocations, such as (rob, MONEY-bank) and (river, RIVER-bank), thus leading to a representation that is potentially very effective for WSD.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Combining Definition-based and Example-based CR",
	"sec_num": "2.3"
	},
	{
	"text": "LCR(D w,",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Combining Definition-based and Example-based CR",
	"sec_num": "2.3"
	},
	{
	"text": "We sum up the above descriptions and outline the procedure as Algorithm 3. The algorithm combines definition-based and example-based CR into an integrated contextual representation CR(W, S) for the sense S of the polysemous word W.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Combining Definition-based and Example-based CR",
	"sec_num": "2.3"
	},
	{
	"text": "Step 1: Given a polysemous word W, one of its senses S and a collection of bilingual examples C, run Algorithms 1 and 2 to obtain",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Algorithm 3: Combining definition-based CR",
	"sec_num": null
	},
	{
	"text": "LCR(D w, s ), CCR(D w, s ), LCR(E w, s ) and CCR(E w, s )",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Algorithm 3: Combining definition-based CR",
	"sec_num": null
	},
	{
	"text": ", where E is a set of examples that each contain an instance of S.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Algorithm 3: Combining definition-based CR",
	"sec_num": null
	},
	{
	"text": "Step 2: Merge the following word list for W and S:",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Algorithm 3: Combining definition-based CR",
	"sec_num": null
	},
	{
	"text": "WORD(W, S) = LCR(D w, s ) + CCR(D w, s ) + LCR(E w, s ) + CCR(E w, s ).",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Algorithm 3: Combining definition-based CR",
	"sec_num": null
	},
	{
	"text": "Step 3: For each WORD(W, S), compute a list of distinct words X with weight W X,S as follows:",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Algorithm 3: Combining definition-based CR",
	"sec_num": null
	},
	{
	"text": "CR(W, S) = { X (W X, S ) \| X is a distinct word in WORD(W, S)}, where tf X, S = the frequency of X in WORD(W, S), idf X = 1/the percentage of senses S such that X\u2208WORD(W, S), W X, S = tf X, S \u00d7 idf X.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Algorithm 3: Combining definition-based CR",
	"sec_num": null
	},
	{
	"text": "Step 4: The weights W X, S in CR(W, S) for each word sense S are normalized to a sum of 100.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Algorithm 3: Combining definition-based CR",
	"sec_num": null
	},
	{
	"text": "In the following, we will demonstrate how Algorithm 3 works. Given a MONEY-bank sense, the integrated CR(bank, MONEY) can be acquired by doing the following (where the numbers in parentheses following collocates denote the frequency):",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "An Illustrative Example",
	"sec_num": null
	},
	{
	"text": "Step 1: After running Algorithms 1 and 3, we obtain the following: Step 2:",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "An Illustrative Example",
	"sec_num": null
	},
	{
	"text": "LCR(D bank.4.n.1 ) =",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "An Illustrative Example",
	"sec_num": null
	},
	{
	"text": "WORD(bank, MONEY) = LCR(D bank.4.n.1 ) + CCR(D bank.4.n.1 ) + LCR(E bank.4.n.1 ) + CCR(E bank.4.n.1 ).",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "An Illustrative Example",
	"sec_num": null
	},
	{
	"text": "Similar calculations can be performed for other senses of bank to obtain WORD(bank, RIVER), WORD(bank, EARTH), etc.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "An Illustrative Example",
	"sec_num": null
	},
	{
	"text": "Step 3: Compute tf and idf for each distinct word in WORD(bank, S Step 4: The weight W X, S in CR(W, S) for each word sense S is normalized to a sum of 100. For instance, the total of the weights CR(bank, MONEY) = 6274.5; therefore, the normalized weight W account, MONEY = 2.04. Table 8 shows more details about the contextual words and normalized weights in CR(bank, S) for all bank senses S. The ten top-weighted context words from the CRs of the bank senses listed in Table 8 seem to be very relevant to each sense and to have strong collocates listed in BBI [Benson, Benson and Ilson 1993] . These weighted context words form the general CCR knowledge for senses of bank. In the next section, we will show that this knowledge is effective for applying WSD to unrestricted text. ",
	"cite_spans": [
	{
	"start": 563,
	"end": 594,
	"text": "[Benson, Benson and Ilson 1993]",
	"ref_id": "BIBREF0"
	}
	],
	"ref_spans": [
	{
	"start": 280,
	"end": 287,
	"text": "Table 8",
	"ref_id": "TABREF8"
	},
	{
	"start": 472,
	"end": 479,
	"text": "Table 8",
	"ref_id": "TABREF8"
	}
	],
	"eq_spans": [],
	"section": "An Illustrative Example",
	"sec_num": null
	},
	{
	"text": "Among the recently proposed WSD systems, almost all have the property that the knowledge obtained is fixed when the system completes the training phase. This means that the acquired knowledge can not be enriched during the course of disambiguation. Such fixed knowledge is referred to as static knowledge. We believe that this property limits WSD performance. We propose lifting this limitation by adjusting the initial acquired knowledge to suit the text at hand. Alternatively, such expanded knowledge is referred to as adaptive knowledge. In this section, we will show how to distinguish between senses of text using adaptive disambiguation techniques. First, we will start with disambiguation of polysemous words in easy (trivial) contexts by using the fundamental knowledge previously acquired from MRD. Next, we will expand the acquired knowledge based on these disambiguated contexts. Finally, we will resolve the senses in the remaining contexts, called hard contexts.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Word Sense Disambiguating Algorithms",
	"sec_num": "3."
	},
	{
	"text": "The proposed WSD method starts with a simple disambiguation step using the topical CR described in a previous section. For instance, to disambiguate the word bank in Examples (4) through (6), the content words in its context are extracted, lemmatized and matched against the contextual representation of each of bank's word senses. Each instance of bank is given a sense label in favor of a CR most similar to the context in question. A sense label is assigned only when the match is strong enough and the runner-up sense is sufficiently weak. In the following subsections, we will describe how to distinguish between strong and weak signals. For instance, there is enough overlap between the CR for the instance of MONEY-bank in Examples (4) and (6) to warrant a sense label of MONEY-bank for the two instances of bank, but the match is not strong enough for the instance in Example (4). We call Examples (4) and (6) easy 5 contexts, while Example (5) is a hard 6 context.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Disambiguating Polysemous Words in Easy Contexts",
	"sec_num": "3.1"
	},
	{
	"text": "(4) \u2026 Participation loans are those made jointly by the SBA and banks or other private lending institutions ...",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Disambiguating Polysemous Words in Easy Contexts",
	"sec_num": "3.1"
	},
	{
	"text": "(5) \u2026 individual action by every nation in position to help, we must squarely face this titanic challenge \u2026 (6) \u2026 from investment firms all over the nation, all of them wanting a part of shares that would be sold (185,000 to the public at $12.50 with another 5,000 reserved for Morton Foods employers at $11.50 a share) there was even a cable in French from a bank in Switzerland that had somehow \u2026 In addition, the contextual words closer to an ambiguous word may have greater influence on the sense of a word. For instance, consider Example 7, where the intended sense of bank is MONEY. We observe that there are two salient words, mortgage and river, around an ambiguous word bank. The word mortgage favors a MONEY sense, while the word river favors a RIVER sense. Intuitively, the MONEY sense should be given more favorable consideration since mortgage is nearer to the ambiguous word than river is. There are various representations for distance-based weights. Here we adopt the metric proposed by Hawking and Thistlewaite [1995] to weigh the relevance of salient words in a text. (7) \u2026 and an effort to get this religious center out of its rut of wild worship into a modern church organization. He emphasized to the Presiding Elder the plan of giving up the old church and moving across the river. The Presiding Elder was sure that that would be impossible. But he told Wilson to \"go ahead and try\". And Wilson tried. It did seem impossible. The bank which held the mortgage on the old church declared that the interest was considerably in arrears, and the real estate people said flatly that the land across the river was being held for an eventual development for white working people who were coming in, and that none would be sold to colored folk. When it was proposed to rebuild the church, Wilson found that the terms for \u2026",
	"cite_spans": [
	{
	"start": 1003,
	"end": 1034,
	"text": "Hawking and Thistlewaite [1995]",
	"ref_id": "BIBREF16"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Disambiguating Polysemous Words in Easy Contexts",
	"sec_num": "3.1"
	},
	{
	"text": "To sum up, we outline a general WSD method using MRD-based contextual representation as Algorithm 4 for labeling an instance of a polysemous word W in a particular context CON(W).",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Disambiguating Polysemous Words in Easy Contexts",
	"sec_num": "3.1"
	},
	{
	"text": "Step 1: Preprocess the context and produce a list of lemmatized content words CON(W) in W's context.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Algorithm 4: (StaticSense) WSD using MRD-based contextual representation",
	"sec_num": null
	},
	{
	"text": "Step 2: For each sense S of W, compute the similarity between the context representation CR(W, S) and topical context CON(W).",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Algorithm 4: (StaticSense) WSD using MRD-based contextual representation",
	"sec_num": null
	},
	{
	"text": "Sim (CR(W,S), CON(W)) = \u2211 \u2211 + + \u2211 \u2208 \u2208 \u2208 ) , ( t ) ( s t, t s t, ) ( S W CR t W CON t M t W W W W , where M = CR(W,S) \u2229 CON(W), s t, W",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Algorithm 4: (StaticSense) WSD using MRD-based contextual representation",
	"sec_num": null
	},
	{
	"text": "= the weight of a contextual word t with sense S in CR(W), t W = the weight of t in CON(W ",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 81,
	"end": 89,
	"text": "in CON(W",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "Algorithm 4: (StaticSense) WSD using MRD-based contextual representation",
	"sec_num": null
	},
	{
	"text": ") = t 1 X , X t = the distance from t to W in number of words, S(W, CON(W)) = s max arg Sim (CR(W,S), CON(W)), S\"(W, CON(W)) = s max arg {Sim (CR(W,S), CON(W)) \| Sim (CR(W,S), CON(W)) < S(W, CON(W))}, TSCORE(W, CON(W)) = )) ( , ( S\" )) ( , ( * S W CON W W CON W , RANK-S(W, CON(W)) = the rank of S(W, CON(W)) among all S(X, CON(X))",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Algorithm 4: (StaticSense) WSD using MRD-based contextual representation",
	"sec_num": null
	},
	{
	"text": "for all n instances of polysemous word X and context CON(X), RANK-T(W, CON(W)) = the rank of TSCORE(W, CON(W)) among TSCORE(X, CON(X)) for all n instances and context of polysemous word X.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Algorithm 4: (StaticSense) WSD using MRD-based contextual representation",
	"sec_num": null
	},
	{
	"text": "Step 3: Construct the set of the triples T, where",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Algorithm 4: (StaticSense) WSD using MRD-based contextual representation",
	"sec_num": null
	},
	{
	"text": "T = { (W, S, CON(W)) \| S = S*(W, CON(W)) such that RANK-S(W, CON(W)) \u2264 n/c and RANK-T(W, CON(W)) \u2264 n/c, where the constant c \u2265 1 } 7 .",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Algorithm 4: (StaticSense) WSD using MRD-based contextual representation",
	"sec_num": null
	},
	{
	"text": "Step 4: DEFAULT(W)= S such that the count of (W, S, CON(W))\u03b5T is the largest among all the senses of W.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Algorithm 4: (StaticSense) WSD using MRD-based contextual representation",
	"sec_num": null
	},
	{
	"text": "Step 5: Assign (W, CON(W)) as the relevant sense S if (W, S, CON(W)) is in T, and assign DEFAULT(W) otherwise.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Algorithm 4: (StaticSense) WSD using MRD-based contextual representation",
	"sec_num": null
	},
	{
	"text": "The adaptive approach to WSD hinges on two assumptions. First, we assume that it is possible to build an initial general knowledge base so that a substantial portion of disambiguated text can be used to adapt the knowledge base to fit the text itself. The second condition for the adaptive approach to be feasible is that there is indeed new and effective information to be gained from the partially disambiguated text. In this section, we will first show the kinds of contexts in the Brown corpus and WSJ articles in which word sense ambiguity can be confidently resolved by using an MRD-based knowledge base. In these contexts, one will find an abundance of rich task-specific information not easily covered in a general or static knowledge base. We will also justify the use of contextual information and a task-specific default for WSD.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Adapting the Knowledge Base to Fit the Text",
	"sec_num": "3.2"
	},
	{
	"text": "There is indeed an abundance of new and useful contextual information for word sense to be gained from typical, easy contexts. Such information can be extracted as long as ambiguity in these typical contexts can be interpreted successfully. For instance, the Brown corpus passage reproduced here as Example 8is obviously very typical of MONEY-bank with salient words such as accounts, stocks and property in its context. Without a doubt, this instance of MONEY-bank can be resolved successfully using the kind of MRD-based knowledge base described in Section 3.1. Even though the overall context of this instance of MONEY-bank is a general one, it nevertheless contains many words, such as law and state, not in the MRD-based knowledge base. Such words might very well be incidental and have no intrinsic relation with the sense. For instance, the word law might just as likely be associated with RIVER-bank as MONEY-bank. Without much stretching of the imagination, it is possible to think of a likely event where the state of Texas passes a law to declare an outer bank off limits to commercial development. However, more often than not, these unexpected words will indicate real recurring contexts of word sense, either generally or in a task-specific way. Therefore, adapting the knowledge base to fit such a context is beneficial for WSD. For instance, the instances of tree and camping in the context of RIVER-bank in Example (9) seem to be reasonable additions to CR(bank, RIVER) in the sense that tree and camping are, in general, more strongly associated with RIVER-bank than with MONEY-bank. Even if that assertion generally does not hold, adding tree and camping to CR(bank, RIVER) as a way of adapting the knowledge base is still beneficial since it is likely to be valid in the very text where this association is discovered. The same argument holds for the local cue of through in the context of PILE-bank in Example (10), and for the instances of donor and transfusion in the context of MEDICINE-bank in Example (11). (See Table 9 for further details.) (8) \u2026 63 million dollars at the end of the current fiscal year next Aug. 31. He told the committee the measure would merely provide means of enforcing the escheat law which has been on the books \"since Texas was a republic\". It permits the state to take over bank accounts, stocks and other personal property of persons missing for seven years or more. The bill, which Daniel said he drafted personally, would force banks, insurance firms, pipeline companies and other \u2026 (9) \u2026 On shooting preserves? Ask Sammy Shooter. WE WERE CAMPING a few weeks ago on Cape Hatteras Campground in that land of pirates, seagulls and bluefish on North Carolina's famed outer banks. This beach campground with no trees or hills presents a constant camping show with all manner of equipment in actual use. With the whole camp exposed to view we could see the variety of canvas shelters in which Americans are camping now. There were \u2026 (10) \u2026 to let down through the overcast and see the ground before it hit him. Bob Fogg didn't have today's advantages of Instrument Flight and Ground Control Approach systems. At the end of the calculated time he'd nose the Waco down through the cloud bank and hope to break through where some feature of the winter landscape would be recognizable. Usually back in Concord by noon, there was just time to get partially thawed out, refuel, and grab a bit of Mrs. Fogg's \u2026 (11) \u2026 agreed, but explained that it would be necessary first to check Fred's blood to ascertain whether or not it was of the same type as Papa's. To give a patient the wrong type of blood, said the doctor, would likely kill him. That was in the days before blood banks, of course, and transfusions had to be given directly from donor to patient. One had to find a donor, and usually very quickly, whose blood corresponded with the patient's. And then it took considerably ",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 2038,
	"end": 2045,
	"text": "Table 9",
	"ref_id": "TABREF9"
	}
	],
	"eq_spans": [],
	"section": "Discovering Task-specific Contextual Information",
	"sec_num": "3.2.1"
	},
	{
	"text": "The distribution of senses of a word might not follow Zipf's law because their rank-frequency plot does not follow the power-law well, and it is often quite skewed even in a balanced corpus. In the Brown corpus, 60% of the instances of twelve polysemous words are the top-ranking sense of the word, according to an experimental report by Luk [1995] .",
	"cite_spans": [
	{
	"start": 338,
	"end": 348,
	"text": "Luk [1995]",
	"ref_id": "BIBREF20"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Using the Default Sense",
	"sec_num": "3.2.2"
	},
	{
	"text": "Generally, the top-ranking sense of a word is corpus-dependent. Table 10 presents some statistics about the distribution of senses in different corpora. For instance, we find that CURIOSITY-interest is favored over MONEY-interest 194 to 49 in the Brown corpus, while preference is reversed with counts of 53 and 122 in the WSJ corpus. On the other case, GRAMMAR-sentence is favored over JUDGEMENT-sentence 22 and 10 in the Brown corpus while preference is reversed with counts of 1 to 11 in the WSJ corpus. Using a fixed default would be disastrous for interest or sentence in at least one of these corpora. The adaptive method alternatively uses a set of disambiguated samples from the text in question to estimate the default. ",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 64,
	"end": 72,
	"text": "Table 10",
	"ref_id": "TABREF0"
	}
	],
	"eq_spans": [],
	"section": "Using the Default Sense",
	"sec_num": "3.2.2"
	},
	{
	"text": "We are now ready to present a new adaptive approach to WSD based on the fundamental knowledge base acquired from MRD. Previous sections have already shown how such a knowledge base can be built and described its advantages. We will show one way of using a MRD-based knowledge base for WSD. Although the knowledge base does not guarantee high precision and 100% coverage, a substantial portion, say 50%, can be disambiguated at a high precision rate. In this section, we will show how such a level of coverage and high precision can be put to use in an adaptive way to maintain the same high precision rate at 100% coverage. We will first describe the adaptive algorithm. Examples will be given in Section 3.4 to illustrate how the algorithm works and to give some idea of the potential effectiveness of adaptation.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The Adaptive WSD Algorithm",
	"sec_num": "3.3"
	},
	{
	"text": "The algorithm starts with an initial disambiguation step using the knowledge base derived from the MRD. An adaptation step follows which produces a knowledge base from the partially disambiguated text. Finally, the undisambiguated part is disambiguated according to the adapted knowledge base. Algorithm 5 gives a formal and detailed description.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "The Adaptive WSD Algorithm",
	"sec_num": "3.3"
	},
	{
	"text": "Step 1: Run Algorithm 4 to obtain triples T 1 of word, word sense and context.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Algorithm 5: (AdaptSense) Adaptive WSD",
	"sec_num": null
	},
	{
	"text": "Step 2: From the selected triples (W, S, CON(W))\u2208T 1 , compute a new set of contextual representations: WORD(W,S) = { u \| u\u2208CON(W)and (W, S, CON(W))\u2208T 1 }.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Adptive 21",
	"sec_num": null
	},
	{
	"text": "Step 3: Build the contextual representation CR(W,S) of sense S of word W from WORD(W, S) according to Algorithm 3. DEFAULT(W) = S such that the count of (W, S, CON(W))\u2208T 1 is the highest among all the senses of W.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Adptive 21",
	"sec_num": null
	},
	{
	"text": "Step W = the weight of a contextual word t with sense s in CR(W,S), t W = the weight of t in CON(W) = Step 5: Construct the set of triples T 2 , where ",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Adptive 21",
	"sec_num": null
	},
	{
	"text": "T 2 = { (W,",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Adptive 21",
	"sec_num": null
	},
	{
	"text": "To show how Algorithm 5 works in an adaptive fashion, we will consider the case of disambiguating the Brown corpus, focusing on the polysemous word bank. For this purpose, we will describe step by step how the algorithm operates on the two following passages in the Brown corpus containing an instance of bank. The two passages are reproduced here as Examples 12and 13, showing a context window of 50 words before and after the polysemous word which is used in the algorithm for disambiguation. Huntley and her husband also will be questioned about \u2026 (13) \u2026 Of cattle in a pasture without throwin' 'em together for the purpose was called a \"pasture count\". The counters rode through the pasture countin' each bunch of grazin' cattle, and drifted it back so that it didn't get mixed with the uncounted cattle ahead. This method of countin' was usually done at the request, and in the presence, of a representative of the bank that held the papers against the herd. The notes and mortgages were spoken of as \"cattle paper\". A \"book count\" was the sellin' of cattle by the books, commonly resorted to in the early days, sometimes much to the profit of the seller. This led to the famous sayin' in the Northwest of the \"books won't freeze\". This became a common byword durin' the \u2026 Step1: Identifying an easy context This step corresponds to five substeps of Algorithm 4. First, only the salient words that are in CR(bank, S) for S in {MONEY, RIVER, EARTH, PILE, ROW, ROAD, MEDICINE, GAMBLE} are of interest; all other words are thrown out for now. To calculate similarity values, the weights for these words with respect to relevant senses are pulled out from the initial knowledge base. Tables 11 (a) and (b) show these words, their position relative to bank, and their weights according to a knowledge base extracted from LDOCE. The context of Example (12) resembles the CR of MONEY-bank the most. Table 11 (a) indicates clearly that very salient words in CR(bank, MONEY-bank), such as robbery, branch, and charge, occur in close proximity to the word bank. Although words related to other senses, such as drive and report, do occur, they are fewer and are located at quite a longer distance. It is not surprising that the similarity of this context with MONEY-bank and the t-score ranks high enough for this instance to be included in T 1 .",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 1897,
	"end": 1905,
	"text": "Table 11",
	"ref_id": "TABREF0"
	}
	],
	"eq_spans": [],
	"section": "An Illustrative Example",
	"sec_num": "3.4"
	},
	{
	"text": "On the other hand, Example (13) does not resemble the CR of any particular sense of bank more than it does those of other senses. That is evident from the weights shown in Table 11(b). The only indicative word representative is not enough to enable interpretation of the intended sense of MONEY-bank. All other words are either not in any CRs or ambivalent (hold, note and paper), indicating a number of senses competing with MONEY-bank. Hence, the similarity of this context with MONEY-bank and the t-score do not rank high enough for this instance to be included in T 1 . From the triples T 1 , a list WORD(S) of contextual words for each sense S of word bank and the most frequent sense DEFAULT(bank) are calculated. Therefore, the contextual words in the triple from Example (12) will be lemmatized. With stop words removed, we obtain a list like the following: charge, assault, robbery, portland, detectives, say, Friday, mrs, lavaughn, huntley, accuse, drive, getaway, car, use, robbery, woodyard, bros, grocery, burnside, st, april, husband, sentence, year, federal, prison, mcneil, island, last, april, robbery, hillsdale, branch, multnomah, charge, store, holdup, secret, grand, jury, indictment, return, against, pair, last, week, detective, murray, logan, report, phoenix, arrest, culminate, year, investigation, detective, william, taylor, officer, taylor, say, mrs, huntley, husband, question, \u2026 } Step 3: Assigning weight to the contextual representation From the word lists for all senses of bank, the new set of CRs can be derived. The CR(bank, S) for the word sense S basically consists of every word in WORD(S) associated with a weight. Weights are assigned in favor of contextual words frequently occurring in the context of a particular word sense and that of a smaller number of other senses. For instance, the word cooperative occurs very frequently and only in the context of MONEY-bank in the part of the Brown corpus resolved in Step 1. According to our experiment, there are quite a number of bank instances in the Brown corpus that are very typical and can be reliably resolved using LDOCE-based contextual representation. Those instances are predominately resolved as MONEY-bank. Therefore, we have DEFAULT(bank) = MONEY-bank. Table 13 .",
	"cite_spans": [
	{
	"start": 866,
	"end": 1410,
	"text": "charge, assault, robbery, portland, detectives, say, Friday, mrs, lavaughn, huntley, accuse, drive, getaway, car, use, robbery, woodyard, bros, grocery, burnside, st, april, husband, sentence, year, federal, prison, mcneil, island, last, april, robbery, hillsdale, branch, multnomah, charge, store, holdup, secret, grand, jury, indictment, return, against, pair, last, week, detective, murray, logan, report, phoenix, arrest, culminate, year, investigation, detective, william, taylor, officer, taylor, say, mrs, huntley, husband, question, \u2026 }",
	"ref_id": null
	}
	],
	"ref_spans": [
	{
	"start": 2255,
	"end": 2263,
	"text": "Table 13",
	"ref_id": "TABREF0"
	}
	],
	"eq_spans": [],
	"section": "Adptive 23",
	"sec_num": null
	},
	{
	"text": "Adptive 25 WORD(MONEY-bank) = {face,",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Adptive 23",
	"sec_num": null
	},
	{
	"text": "Step 1, where the MRD-based CR is used, there are now more words in the context that are indicative of the intended sense. From the perspective of the new CRs, the words method, usual, request, paper, note, and book all point to the sense of MONEY-bank and not to any other sense. These words either do not exist or ambivalent with respect to the MRD-based CR. As a whole, these words provide enough evidence to reverse the previous inconclusive situation leading to the expected sense of MONEY-bank. In the event that the maximal similarity is lower than a threshold value, the default sense of MONEY-bank is used. In this particular case, the default happens to be correct. ",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Contrary to the situation in",
	"sec_num": null
	},
	{
	"text": "The experimental setup can be described in a number of steps as follows. (1) A set of 13 polysemous words was selected as the target for disambiguation and evaluation. (2) For each of the polysemous words, a sense division was established based on the LDOCE treatment of relevant nominal senses. The LDOCE's sense division was used largely as is, with only a couple of closely related senses merged. 3Two sets of text from corpora were gathered as the test sets. (4) Two human judges were asked to assign a sense label to each nominal instance of these 13 words in the two test sets. (5) Two WSD programs were written to disambiguate nominal instances of these polysemous words in the test sets. 6The results of running the two programs on both test sets were compared against those of human assessors. The number of test instances and that of correctly disambiguated ones in these four experiments were tallied to produce a precision rate for each experiment. In the following, we describe each step in turn. Table 13 . Weights for salient words in Example (13) after adaptation. (1) Test words We limited our experiment and evaluation to a set of thirteen words with higher than usual ambiguity. That is due mainly to the fact that the process of evaluation is a difficult and expensive one. It is often difficult to pin down the number of senses allowed for a word in the experiment. For the purpose of comparing results with other approaches, we stick to words that have been studied in various experiments reported in the literature on computational linguistics. These words include bank, bass, bow, cone, duty, galley, interest, issue, mole, sentence, slug, star, and taste. (2) Sense division The sense division for each of these test words was very crucial in the WSD experiment. We used a sense division based on LDOCE's treatment of the nominal senses of these words. The division is somehow more fine-grained than those used in other WSD studies. This level of sense division is very close to the kind of granularity required for machine translation. For most cases, a word sense has a unique Chinese translation.",
	"cite_spans": [
	{
	"start": 1594,
	"end": 1680,
	"text": "bass, bow, cone, duty, galley, interest, issue, mole, sentence, slug, star, and taste.",
	"ref_id": null
	}
	],
	"ref_spans": [
	{
	"start": 1010,
	"end": 1018,
	"text": "Table 13",
	"ref_id": "TABREF0"
	}
	],
	"eq_spans": [],
	"section": "Experiment",
	"sec_num": "4.1"
	},
	{
	"text": "Word X t W t W t, s in CR(bank, S) S MONEY S RIVER S EARTH S PILE S ROW S ROAD S MEDICINE S",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Experiment",
	"sec_num": "4.1"
	},
	{
	"text": "(",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Experiment",
	"sec_num": "4.1"
	},
	{
	"text": "We aimed to determine the effectiveness of the proposed approach for unrestricted text and to find out how domain and genre affect WSD. Therefore, we used the Brown corpus and a collection of WSJ articles from October 30 to November 2, 1989 as the test sets. Passages of 100 words centered at an instance of the test words in the two corpora were extracted using a SED program. It is in general not hard to write a regular expression in the SED program to exclude verbal instances, so only a small number of verb cases were extracted. These verbal instances were excluded from the experiment according to the marks made by human judges. For these thirteen words under investigation, we had 846 and 903 passages of nominal senses from the Brown corpus and WSJ test sets, respectively.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "3) Test corpora",
	"sec_num": null
	},
	{
	"text": "(4) Judgement To be as subjective as possible, we asked two human judges to assign a sense label to each nominal instance of these thirteen words in the two test sets. There were also cases which fell out of the scope of our sense division. Most of these cases used proper nouns, so they bore none of the meaning represented in our sense division. Cases judged to be verbal uses or proper names were removed from the test cases. For instance, the word bow in a Brown corpus passage, reproduced here as Example (14), was an instance of a proper name and, therefore, was excluded from the test cases.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "3) Test corpora",
	"sec_num": null
	},
	{
	"text": "(14) \u2026 The announcement that the secrets of the Dreadnought had been stolen was made in Bow St. police court here at the end of a three day hearing \u2026 (5) Static vs. Adaptive WSD In the previous sections, we argued in favor of using an MRD-derived knowledge base because we believe that the fundamental information in an MRD can be very helpful for WSD. Despite our belief in the effectiveness of the MRD-derived knowledge base, we also expected that adaptation could improve its effectiveness a bit further. Therefore, we implemented programs for both Algorithms 4 and 5. These two programs were executed in order to disambiguate the test cases in the Brown and WSJ corpora.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "3) Test corpora",
	"sec_num": null
	},
	{
	"text": "The results of running the two programs on both test sets were compared against those of human assessors. The number of test instances and that of correct assignments in these four experiments were tallied to calculate the precision rate for each experiment. All results were based on 100% applicability 8 . Statistics for the experimental results are summarized in Tables 14 and 15 . Several observations can be made based on the results. First, evidently, the MRD-based knowledge base was reasonably helpful for WSD. The results shown in Tables 14 and 15 indicate that without adaptation, the knowledge extracted from LDOCE and LLOCE could be used to deliver a precision rate of 65.2% for the Brown corpus and 76.6% for the WSJ articles. Second, adaptation indeed helped boost the precision rate by over 5% for the Brown corpus. As for the WSJ test set (see Table 15 ), adaptation only marginally increased the average precision rate. Closer examination of the results for this test set shows that three words, bank, interest, and issue, dominated the experiment and evaluation results. The precision rate for bank was over 95%, which left adaptation with very little room for improvement. The other two words, interest and issue, were very general and difficult to disambiguate.",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 366,
	"end": 382,
	"text": "Tables 14 and 15",
	"ref_id": "TABREF0"
	},
	{
	"start": 540,
	"end": 556,
	"text": "Tables 14 and 15",
	"ref_id": "TABREF0"
	},
	{
	"start": 860,
	"end": 868,
	"text": "Table 15",
	"ref_id": "TABREF0"
	}
	],
	"eq_spans": [],
	"section": "Evaluation",
	"sec_num": "4.2"
	},
	{
	"text": "Although it is often difficult to compare results from experiments based on different domains, genres and setups, the experimental results presented here seem to compare favorably with the experimental results reported in previous WSD research. Our adaptive approach could disambiguate with an average precision rate of 71.2% for these thirteen words in Brown and of 76.5% for these words in WSJ. For the Brown corpus, Luk [1995] experimented with the same words we used except for the word bank and reported that there were totally 616 instances of these words (slightly less than the 749 instances we found).",
	"cite_spans": [
	{
	"start": 405,
	"end": 429,
	"text": "Brown corpus, Luk [1995]",
	"ref_id": null
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Discussion",
	"sec_num": "5."
	},
	{
	"text": "The precision rate for all instances was 60%. Leacock, Towell and Voorhees [1993] reported a precision rate of 76% for disambiguating the word line in a sample of WSJ articles. Besides the precision rate, a number of interesting features of this approach are also important. First, the proposed disambiguation system is robust and portable, since absolutely no corpus-specific knowledge is needed in the disambiguation procedure. It can be applied readily to test data in a variety of domains and genres with performance rivaling that of methods requiring a substantial training corpus. Second, the proposed approach is considerably more time efficient when compared to other learning strategies. Although the bootstrap approach proposed by Yarowsky [1995] has an element of adaptation to it, his method still requires a long training process to derive a static knowledge base for WSD. The differences between our method and his lie in the initial knowledge, the level of abstraction, and the learning cycle. We propose to exploit rich conceptualized knowledge from MRD at the outset, while the bootstrap method uses merely a couple of word collocations for each sense to start the learning process. Since the bootstrap method aims to derive a word-based conceptual representation with a large parameter space, a very large training corpus is required. The thesaurus used in the proposed approach provides an appropriate level of abstraction and, thus, alleviates the need for a very large corpus. The time required for learning in the two approaches is also quite different. The adaptive approach requires a single round of adaptation for effective WSD, while the bootstrap method needs many rounds of learning. Speedy adaptation is the consequence of using rich conceptualized knowledge to start the learning process. To show that this is truly the case, we have revised Algorithm 5 by adding a second and a third adaptation step and by applying the new CR to a reserved batch of low-ranking instances instead of using defaults. The results obtained using more adaptation steps are shown in Figure 2 . The precision rates show that the additional adaptation steps have only a marginal effect.",
	"cite_spans": [
	{
	"start": 46,
	"end": 81,
	"text": "Leacock, Towell and Voorhees [1993]",
	"ref_id": "BIBREF18"
	},
	{
	"start": 741,
	"end": 756,
	"text": "Yarowsky [1995]",
	"ref_id": "BIBREF33"
	}
	],
	"ref_spans": [
	{
	"start": 2093,
	"end": 2101,
	"text": "Figure 2",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "Discussion",
	"sec_num": "5."
	},
	{
	"text": "One of the limiting factors of this approach is the quality of sense definition in the MRD. Short and vague definitions tend to lead to inclusion of inappropriate topics in the contextual representation. With such inferior CRs, it is not possible to produce enough precise samples in the initial step for subsequent adaptation. For instance, it is difficult to derive appropriate contextual knowledge for the LDOCE senses in (15) since their definitions mainly consist of either function words or very common words:",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Figure 2 Average precision rates with and without adaptation.",
	"sec_num": null
	},
	{
	"text": "(15)interest.1.n.1 a readiness to give attention issue.1.n.1 the act of coming out issue.1.n.2 an example of this issue.1.n.3 something which comes or is given out issue.1.n.4",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Figure 2 Average precision rates with and without adaptation.",
	"sec_num": null
	},
	{
	"text": "the act of bringing out something in a new form",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Figure 2 Average precision rates with and without adaptation.",
	"sec_num": null
	},
	{
	"text": "The experiment and evaluation results show that adaptation is most effective when a high-frequency word with contrasting senses is involved. For low-frequency senses, such as EARTH, ROW, and ROAD senses of bank, the approach does not seem to be very effective. That is not a problem specific to the adaptive approach, and all other approaches in the literature suffer from the same problem of data sparseness. Even with static knowledge acquired from a very large corpus, these senses were disambiguated at a considerably lower precision rate than other senses. There has been increasing interest in using a machine to identify the intended sense of a polysemous word in a given context. Recently, various approaches to WSD have been proposed in the natural language processing literature, and old ideas have been superseded by newer ones at a rapid rate. Central to these development efforts are the kind of contextual knowledge encoded and the way this knowledge is represented and acquired. In this section, we review the recent literature on WSD from the perspectives of different types of contextual knowledge and their representational schemes.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Figure 2 Average precision rates with and without adaptation.",
	"sec_num": null
	},
	{
	"text": "Any kind of scheme for acquiring contextual information of word sense must begin with a way of identifying the word sense since word sense is an abstract concept not clear on the surface. Once this is done, we can use the surrounding words to build a contextual representation of the word sense for WSD. There are three approaches to the chicken-and-egg problem of dividing word senses. First, one can resort to human intervention to get a hand-tagged corpus of word senses. Most early WSD works used this approach and went to the trouble of hand-tagging the intended sense of each polysemous word in the training corpus [Kelly and Stone 1975; Hearst 1991] . Second, one can take the numbered sense entries readily available in a machine-readable dictionary and treat their definitions and examples as contextual information [Lesk 1986; Veronis and Ide 1990; Wilks et al. 1990; Guthrie et al. 1991] . The third way of identifying word sense exploits linguistic constraints. For instance, three linguistic constraints can be exploited for successful sense tagging and WSD.",
	"cite_spans": [
	{
	"start": 621,
	"end": 643,
	"text": "[Kelly and Stone 1975;",
	"ref_id": "BIBREF17"
	},
	{
	"start": 644,
	"end": 656,
	"text": "Hearst 1991]",
	"ref_id": "BIBREF15"
	},
	{
	"start": 825,
	"end": 836,
	"text": "[Lesk 1986;",
	"ref_id": "BIBREF19"
	},
	{
	"start": 837,
	"end": 858,
	"text": "Veronis and Ide 1990;",
	"ref_id": "BIBREF28"
	},
	{
	"start": 859,
	"end": 877,
	"text": "Wilks et al. 1990;",
	"ref_id": "BIBREF31"
	},
	{
	"start": 878,
	"end": 898,
	"text": "Guthrie et al. 1991]",
	"ref_id": "BIBREF14"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Lexicalized vs. Conceptual Encoding of Context",
	"sec_num": "6.1"
	},
	{
	"text": "\u2022 One sense per discourse The senses of all instances of a polysemous word are highly consistent within any given document.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Lexicalized vs. Conceptual Encoding of Context",
	"sec_num": "6.1"
	},
	{
	"text": "\u2022One sense per collocation Nearby words provide strong and consistent clues to the sense of a target word, conditional on the relative distance, order, and syntactic relationship.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Lexicalized vs. Conceptual Encoding of Context",
	"sec_num": "6.1"
	},
	{
	"text": "\u2022One sense per translation Translations in a bilingual corpus can be used to represent the senses of words.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Lexicalized vs. Conceptual Encoding of Context",
	"sec_num": "6.1"
	},
	{
	"text": "As an example of the first constraint, consider the word suit. The constraint captures the intuition that if the first occurrence of suit is a LAWSUIT sense, then later occurrences in the same discourse are also likely to refer to LAWSUIT [Gale, Church and Yarowsky 1992a] . The second constraint indicates that most works on statistical disambiguation have made the basic assumption that word sense is strongly correlated with certain contextual features, like occurrence of particular words in a window around the ambiguous word. However, Yarowsky [1995] proposed an approach in which strong collocations were identified for WSD. If a bilingual corpus was available, differences in translations of the polysemous word allowed one to delineate the intended sense, particularly in the case of contrasting polysemy. Gale, Church and Yarowsky [1992b] used French translations in parallel texts to disambiguate some polysemous words in English. For instance, the senses of duty were usually translated as two different French words, droit and devoir, respectively, representing the senses tax and obligation. Thus, a number of tax sense instances of duty could be collected by extracting instances of duty that were translated as droit, and the same could be done for obligation sense instances of duty.",
	"cite_spans": [
	{
	"start": 239,
	"end": 272,
	"text": "[Gale, Church and Yarowsky 1992a]",
	"ref_id": "BIBREF12"
	},
	{
	"start": 541,
	"end": 556,
	"text": "Yarowsky [1995]",
	"ref_id": "BIBREF33"
	},
	{
	"start": 815,
	"end": 848,
	"text": "Gale, Church and Yarowsky [1992b]",
	"ref_id": "BIBREF13"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Lexicalized vs. Conceptual Encoding of Context",
	"sec_num": "6.1"
	},
	{
	"text": "Once word senses are identified in one way or another, the context of a particular word sense can then be acquired and encoded in some way for use in the subsequent disambiguation step. There are at least two ways of encoding contextual knowledge. The obvious way, the lexicalized representation, is a surface scheme that keeps a weighted list of words appearing in the context of a particular sense. On the other hand, the conceptual representation encodes the classes of words that might appear in the context.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Lexicalized vs. Conceptual Encoding of Context",
	"sec_num": "6.1"
	},
	{
	"text": "Dictionary Definitions as Context Lesk [1986] described a word-sense disambiguation technique based on the number of overlaps between words in a dictionary definition and the fixed-size window of words surrounding the target. The author reported WSD performance ranging from 50% to 70% when the method was applied to a sample of ambiguous words. Lesk's method had failed to determine the correct senses of words when two or more senses of a word had the same number of overlaps with the context. Veronis and Ide [1990] constructed an artificial neural network from sense definitions, representing each word in the definition text as a node in the network. Different senses of each word competed with each other through the mechanism of spreading activation initiated at the nodes of contextual words. White [1988] , Guthrie et al. [1991] , and Slator [1991] used measures of words in context overlapping with dictionary definitions. One major problem of these earlier approaches was their lack of abstraction. The rich semantic information in the definition, such as the genus term, differentia, and implicit topics, was not exploited to the fullest. Gale, Church and Yarowsky [1992b] indicated that translation in a bilingual corpus could be used to provide tagged material for supervised learning of WSD knowledge. In their experiment, French translations were, in effect, used to represent the senses of some English words under the assumption of one-sense-per-translation. The Bayesian model was used to represent the contextual words in terms of their probabilities of occurrence. They reported a 90% accuracy rate in discriminating between two constrasting senses of six ambiguous nouns in the Canadian Hansards: duty, drug, land, language, position, and sentence. The weaknesses of this approach include the dreaded problem of data sparseness. Even when a very large corpus is available, it is still difficult to guarantee that each word sense will have enough contextual samples to avoid running into the problem of zero frequency, namely, the difficulty of assigning appropriate probabilistic values to words that do not appear in these contextual samples. Yarowsky [1992] improved on the WSD method proposed by Gale, Church and Yarowsky [1992b] by smoothing the concurrence probability via predefined semantic classification. Basically, that was done by lumping the probabilities related to all the senses in a thesaurus category to smooth the zero frequency cases. For instance, the contextual information of bird and other animals was used to build a contextual representation for all the senses in the animal category in Roget's Thesaurus [1987] . His experiment showed in a close test using Grolier's Encyclopedia that instances of twelve words, bass, bow, cone, duty, galley, interest, issue, mole, sentence, slug, star , and taste, could be disambiguated with an average precision rate of 92%. However, a very large corpus is required to train such a lexicalized contextual model, and clearly this kind of static model has a portability problem.",
	"cite_spans": [
	{
	"start": 34,
	"end": 45,
	"text": "Lesk [1986]",
	"ref_id": "BIBREF19"
	},
	{
	"start": 496,
	"end": 518,
	"text": "Veronis and Ide [1990]",
	"ref_id": "BIBREF28"
	},
	{
	"start": 801,
	"end": 813,
	"text": "White [1988]",
	"ref_id": "BIBREF30"
	},
	{
	"start": 816,
	"end": 837,
	"text": "Guthrie et al. [1991]",
	"ref_id": "BIBREF14"
	},
	{
	"start": 844,
	"end": 857,
	"text": "Slator [1991]",
	"ref_id": "BIBREF27"
	},
	{
	"start": 1151,
	"end": 1184,
	"text": "Gale, Church and Yarowsky [1992b]",
	"ref_id": "BIBREF13"
	},
	{
	"start": 2166,
	"end": 2181,
	"text": "Yarowsky [1992]",
	"ref_id": "BIBREF32"
	},
	{
	"start": 2221,
	"end": 2254,
	"text": "Gale, Church and Yarowsky [1992b]",
	"ref_id": "BIBREF13"
	},
	{
	"start": 2652,
	"end": 2658,
	"text": "[1987]",
	"ref_id": null
	},
	{
	"start": 2753,
	"end": 2834,
	"text": "words, bass, bow, cone, duty, galley, interest, issue, mole, sentence, slug, star",
	"ref_id": null
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Lexicalized Representation of Context",
	"sec_num": "6.1.1"
	},
	{
	"text": "Context as Definition-Based Conceptual Co-occurrence Luk [1995] advocated using defining words in the MRD for the contextual representation of word sense. Reminiscent of an earlier work by Wilks et al. [1990] , Luk proposed a definition-based concept co-occurrence model (DBCC) for WSD. With the model, the context of each word sense is represented using a vector of LDOCE defining words in the sense definition. The author argued that by using a fixed, relatively small number of concepts, a small corpus could provide enough concept co-occurrence data for statistical sense disambiguation. In a close test, the DBCC model trained on the Brown corpus was found to be capable of disambiguating 60% 9 of the instances of the same twelve ambiguous words used in Yarowsky's experiment.",
	"cite_spans": [
	{
	"start": 53,
	"end": 63,
	"text": "Luk [1995]",
	"ref_id": "BIBREF20"
	},
	{
	"start": 189,
	"end": 208,
	"text": "Wilks et al. [1990]",
	"ref_id": "BIBREF31"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Conceptual Representation of Context",
	"sec_num": "6.1.2"
	},
	{
	"text": "Many researchers have exploited the semantic categories in a thesaurus, such as Roget's and LLOCE, or the subject information in a dictionary for context representation and WSD. Walker and Amsler [1986] applied subject codes in LDOCE as semantic representation for WSD. Black [1988] reported an accuracy rate of around 50% when Walker and Amsler's algorithm was applied to a sample of five ambiguous words: interest, point, power, state, and term. Pure conceptual representation is the most economical kind of WSD model since it requires the smallest parameter space and requires no substantial texts for training. Chen et al. [1996] proposed a mixed representational scheme for context based on contextual words as well as LLOCE topics. With a contextual representation acquired from example sentences in LDOCE/E-C, the authors reported that the method could disambiguate around 70% of the instances of thirteen polysemous words in the Brown corpus.",
	"cite_spans": [
	{
	"start": 178,
	"end": 202,
	"text": "Walker and Amsler [1986]",
	"ref_id": "BIBREF29"
	},
	{
	"start": 270,
	"end": 282,
	"text": "Black [1988]",
	"ref_id": "BIBREF1"
	},
	{
	"start": 627,
	"end": 633,
	"text": "[1996]",
	"ref_id": null
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Context as Thesaurus Categories",
	"sec_num": null
	},
	{
	"text": "In almost all the studies described in Section 2.1, topical context was used in WSD. In a number of research works related to machine translation, researchers have used local context to solve a problem closely related to WSD, namely, the lexical choice problem. We will examine these two different kinds of contextual information in this section.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Topical vs. Local Representation of Context",
	"sec_num": "6.2"
	},
	{
	"text": "With topical representation of context, the context of a given sense of a target word is a bag of words without any structure. Information in topical context is generally quite helpful for WSD. For instance, consider Examples 16and 17extracted from the Brown corpus, each containing an instance of the ambiguous word bass.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Topical Context",
	"sec_num": "6.2.1"
	},
	{
	"text": "(16) \u2026 for scintillating flights of meaningless improvisations, and he has a quiet way of getting back and restating the melody after the improvising is over. In this he is sticking with tradition, however far removed from it he may seem to be. SHEARING TAKES OVER George Shearing took over with his well disciplined group, a sextet consisting of vibes, guitar, bass, drums, Shearing's piano and a bongo drummer. He met with enthusiastic audience approval, especially when he swung from jazz to Latin American things like the Mambo. Shearing, himself, seemed to me to be playing better piano than in his recent Newport appearances. A very casual, pleasant program-one of those easy-going things that make Newport's afternoon programs such a \u2026 (17) \u2026 Breakfast was at the Palace Hotel, luncheon was somewhere in the mountain forest, and dinner was either at Boulder Creek or at Santa Cruz. Gazing too long at the scenery could be tiring, so halts were contrived between meals. Then the Chinese hostler, who rode with Vernon on the box, would break open a hamper and produce filets of smoked bass or sturgeon, sandwiches, pickled eggs, and a rum sangaree to be heated over a spirit lamp. In spring and in autumn the run was made for a group of botanists which included an old friend of mine. They gathered roots, bulbs, odd ferns, leaves, and bits of resin from the rare Santa Lucia fir, which exists only on a forty-five mile strip on the westerly side of these mountains. In the Spanish \u2026 Intuitively, the first instance of bass can be disambiguated as INSTRUMENT-bass since guitar, drum, piano, jazz, etc. are likely to appear in the topical context of INSTRUMENT-bass. Similarly, the second instance can be disambiguated as FISH-bass since meal, sandwiches, egg, etc. are often found in the topical context of FISH-bass. Generally, the sense representation of topical context is acquired from a very large corpus. Gale, Church and Yarowsky [1992b] experimented on acquiring topical context from a substantial bilingual training corpus and reported good results.",
	"cite_spans": [
	{
	"start": 1916,
	"end": 1949,
	"text": "Gale, Church and Yarowsky [1992b]",
	"ref_id": "BIBREF13"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Topical Context",
	"sec_num": "6.2.1"
	},
	{
	"text": "Local context includes structured information about word order, distance, and syntactic features. For instance, the local context of a line from does not suggest the same sense for the word line as a line for does. Brown et al. [1990] used the trigram model as a way of resolving sense ambiguity for lexical selection in statistical machine translation. This model makes the assumption that only the previous two words have any effect on the translation, and thus, the word sense of the next word.",
	"cite_spans": [
	{
	"start": 215,
	"end": 234,
	"text": "Brown et al. [1990]",
	"ref_id": "BIBREF3"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Local Context",
	"sec_num": "6.2.2"
	},
	{
	"text": "The model was used to attack the problem of lexical ambiguity and produced satisfactory results, under some strong assumptions. For instance, the authors showed that the French sentence Je vais prendre la decision could be correctly translated as I will make the decision using this model. Although in isolation, take was more likely than make to translate as prendre, the trigram language reversed the decision in favor of make. A major problem with the trigram model is long distance dependency. For instance, the model incorrectly rendered the French sentence Je vais prendre ma propre decision as I will take my own decision. The language model did not consider make my own decision more probable since prendre and decision did not fall within a window of three words.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Trigram as Local Context",
	"sec_num": null
	},
	{
	"text": "Lexical Relation Dagan, Itai and Schwall [1991] and Dagan and Itai [1994] made use of translations of different senses from a Hebrew/English bilingual dictionary to disambiguate contexts. Local context in the form of lexical relations was analyzed in a foreign corpus. The basic idea of the algorithm is best explained with an example. Given two Hebrew words hoze and shalom, hoze has two translations in English: contract and treaty, while shalom is often translated into English as peace.",
	"cite_spans": [
	{
	"start": 17,
	"end": 47,
	"text": "Dagan, Itai and Schwall [1991]",
	"ref_id": "BIBREF11"
	},
	{
	"start": 52,
	"end": 73,
	"text": "Dagan and Itai [1994]",
	"ref_id": "BIBREF10"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Trigram as Local Context",
	"sec_num": null
	},
	{
	"text": "Their experiment showed that all instances of peace appear before treaty and none before contract in the corpus of English language. Therefore, the authors concluded that this instance of hoze in the phrase hoze shalom was best translated as treaty. The authors experimented on lexical choice with 105 Hebrew words and 54 German words from news articles. The precision rates achieved ranged from 75% to 92% for coverage rates between 59% and 70%. Brown et al. [1991] described a statistical algorithm for partitioning the senses of a word into two groups. The authors used mutual information to find a local contextual feature that most reliably indicated which of the senses of the French ambiguous word was used. For instance, for the verb prendre, the object was a good indicator: prendre une measure translated as to take a measure, and prendre une decision as to make a decision. Therefore, words (any word, first verb or first noun) immediately to the left or right of the word were evaluated for their effectiveness as good indicators for WSD and lexical choice. The authors reported 20% improvement in the performance of a machine translation system (from 37 to 45 sentences correct out of 100) when the words were first disambiguated in this way.",
	"cite_spans": [
	{
	"start": 447,
	"end": 466,
	"text": "Brown et al. [1991]",
	"ref_id": "BIBREF4"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Trigram as Local Context",
	"sec_num": null
	},
	{
	"text": "By DEF S, W , we mean the definition of sense S of headword W in MRD.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "",
	"sec_num": null
	},
	{
	"text": "WORD(T) is a bag of words rather than a set. By summation of bags, we mean collecting all the word instances in the bags and keeping track of counts.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "",
	"sec_num": null
	},
	{
	"text": "Weights for all contextual words of a sense are normalized to a sum of 100.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "",
	"sec_num": null
	},
	{
	"text": "The algorithm for identifying easy contexts is Algorithm 4.6 The algorithm for resolving hard contexts is Algorithm 5.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "",
	"sec_num": null
	},
	{
	"text": "We use the WSD results of the top-ranking c'th instances in S* as well as TSCORE values which are more reliable.For instance, setting c to 2 amounts to taking the top-ranking 25% quantile of the test cases.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "",
	"sec_num": null
	},
	{
	"text": "Applicability (coverage) denotes the proportion of cases in which the WSD model performed disambiguation.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "",
	"sec_num": null
	},
	{
	"text": "The originally reported value, 77%, was based on the average of the precision rates for all twelve words. This form of evaluation is sensitive to the outcome of a handful of test samples since the precision rate of a word with a couple of samples could have an overly strong impact on the average. In this paper, we use the average rate of precision calculated over all instances.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "",
	"sec_num": null
	}
	],
	"back_matter": [
	{
	"text": "We have described an adaptive approach to word sense disambiguation. Under this new learning strategy, a contextual representation for each sense discriminator is first built based on the sense definition and example sentence in MRD and represented as a weighted-vector of concepts represented by word lists in a thesaurus. This knowledge representation acquired through MRD is based on a limited number of concepts; thus, the dreaded problem of data sparseness is avoided. Conceptual knowledge also offers the additional advantages of reduced storage requirements and increased efficiency due to reduced dimensionality. Also, we can correctly identify at least 50% of the word senses in unrestricted texts. In addition, these disambiguated texts can be used to adjust the fundamental knowledge in an adaptive fashion so to improve disambiguation precision. We have demonstrated that this approach can outperform established static approaches based on direct comparison of results obtained for the same words. This level of performance is achieved without lengthy training or the use of a very large training corpus.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Conclusions",
	"sec_num": "7"
	},
	{
	"text": "Here, we list 129 topics found in LLOCE. The column labeled \"Topic\" shows a set of two-character symbols representing the topics in LLOCE. Each topic is giving a gloss. ",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Appendix A A Glossary of LLOCE Topics",
	"sec_num": null
	}
	],
	"bib_entries": {
	"BIBREF0": {
	"ref_id": "b0",
	"title": "The BBI Combinatory Dictionary of English",
	"authors": [
	{
	"first": "M",
	"middle": [],
	"last": "Benson",
	"suffix": ""
	},
	{
	"first": "E",
	"middle": [],
	"last": "Benson",
	"suffix": ""
	},
	{
	"first": "R",
	"middle": [],
	"last": "Ilson",
	"suffix": ""
	}
	],
	"year": 1993,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Benson, M., E. Benson and R. Ilson. The BBI Combinatory Dictionary of English, John Benjamins Publishing Company. Amstersam/Philadelphia. 1993.",
	"links": null
	},
	"BIBREF1": {
	"ref_id": "b1",
	"title": "An Experiment in Computational Discrimination of English Word Sense",
	"authors": [
	{
	"first": "E",
	"middle": [],
	"last": "Black",
	"suffix": ""
	}
	],
	"year": 1988,
	"venue": "IBM Journal of Research and Developmen",
	"volume": "32",
	"issue": "",
	"pages": "185--194",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Black, E. \"An Experiment in Computational Discrimination of English Word Sense. \" IBM Journal of Research and Developmen. Vol. 32. pp. 185-194. 1988.",
	"links": null
	},
	"BIBREF2": {
	"ref_id": "b2",
	"title": "A Corpus-Based Approach to Language Learning",
	"authors": [
	{
	"first": "E",
	"middle": [],
	"last": "Brill",
	"suffix": ""
	}
	],
	"year": 1993,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Brill, E. A Corpus-Based Approach to Language Learning, PH. D. thesis. Department of Computer and Information Science. University of Pennsylvania. 1993.",
	"links": null
	},
	"BIBREF3": {
	"ref_id": "b3",
	"title": "A Statistical Approach to Machine Translation",
	"authors": [
	{
	"first": "P",
	"middle": [
	"F"
	],
	"last": "Brown",
	"suffix": ""
	},
	{
	"first": "J",
	"middle": [],
	"last": "Cocke",
	"suffix": ""
	},
	{
	"first": "S",
	"middle": [
	"A"
	],
	"last": "Della Pietra",
	"suffix": ""
	},
	{
	"first": "V",
	"middle": [
	"J"
	],
	"last": "Della Pietra",
	"suffix": ""
	},
	{
	"first": "F",
	"middle": [],
	"last": "Jelinek",
	"suffix": ""
	},
	{
	"first": "J",
	"middle": [
	"D"
	],
	"last": "Lafferty",
	"suffix": ""
	},
	{
	"first": "R",
	"middle": [
	"L"
	],
	"last": "Mercer",
	"suffix": ""
	},
	{
	"first": "P",
	"middle": [
	"S"
	],
	"last": "Roosin",
	"suffix": ""
	}
	],
	"year": 1990,
	"venue": "Computational Linguistics",
	"volume": "16",
	"issue": "2",
	"pages": "79--85",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Brown, P. F., J. Cocke, S. A. Della Pietra, V. J. Della Pietra, F. Jelinek, J. D. Lafferty, R. L. Mercer and P. S. Roosin. \"A Statistical Approach to Machine Translation.\" Computational Linguistics,.16(2). pp. 79-85. 1990.",
	"links": null
	},
	"BIBREF4": {
	"ref_id": "b4",
	"title": "Word-Sense Disambiguation using Statistical Methods",
	"authors": [
	{
	"first": "P",
	"middle": [
	"F"
	],
	"last": "Brown",
	"suffix": ""
	},
	{
	"first": "S",
	"middle": [
	"A"
	],
	"last": "Della Pietra",
	"suffix": ""
	},
	{
	"first": "V",
	"middle": [
	"J"
	],
	"last": "Della Pietra",
	"suffix": ""
	},
	{
	"first": "R",
	"middle": [
	"L"
	],
	"last": "Mercer",
	"suffix": ""
	}
	],
	"year": 1991,
	"venue": "Proceedings of the 29th Annual Meeting of the Association for Computational Linguistic",
	"volume": "",
	"issue": "",
	"pages": "264--270",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Brown, P. F., S. A. Della Pietra, V. J. Della Pietra and R. L. Mercer. \"Word-Sense Disambiguation using Statistical Methods\" In Proceedings of the 29th Annual Meeting of the Association for Computational Linguistic. pp. 264-270. 1991.",
	"links": null
	},
	"BIBREF5": {
	"ref_id": "b5",
	"title": "Automatic Extraction Rules on Preposition Phrase",
	"authors": [
	{
	"first": "J",
	"middle": [
	"S"
	],
	"last": "Chang",
	"suffix": ""
	},
	{
	"first": "R",
	"middle": [
	"H"
	],
	"last": "Shu",
	"suffix": ""
	},
	{
	"first": "M",
	"middle": [
	"H"
	],
	"last": "Chen",
	"suffix": ""
	}
	],
	"year": null,
	"venue": "Proceedings of R.O.C. Computational Linguistics Conference IX (ROCLING-IX)",
	"volume": "",
	"issue": "",
	"pages": "295--320",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Chang, J. S., R. H. Shu and M. H. Chen. \"Automatic Extraction Rules on Preposition Phrase.\" In Proceedings of R.O.C. Computational Linguistics Conference IX (ROCLING-IX). pp. 295-320.",
	"links": null
	},
	"BIBREF6": {
	"ref_id": "b6",
	"title": "A Concept-based Adaptive Approach to Word Sense Disambiguation",
	"authors": [
	{
	"first": "J",
	"middle": [
	"N"
	],
	"last": "Chen",
	"suffix": ""
	},
	{
	"first": "J",
	"middle": [
	"S"
	],
	"last": "Chang",
	"suffix": ""
	}
	],
	"year": 1998,
	"venue": "Proceedings of the 36th Annual Meeting of the Association for Computational Linguistics",
	"volume": "",
	"issue": "",
	"pages": "237--244",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Chen, J. N. and J. S. Chang, \"A Concept-based Adaptive Approach to Word Sense Disambiguation. \" In Proceedings of the 36th Annual Meeting of the Association for Computational Linguistics. pp. 237-244. Montreal. Canada. 1998a.",
	"links": null
	},
	"BIBREF7": {
	"ref_id": "b7",
	"title": "TopSense: A Topical Sense Clustering Method based on Information Retrieval Techniques on Machine Readable Resources",
	"authors": [
	{
	"first": "J",
	"middle": [
	"N"
	],
	"last": "Chen",
	"suffix": ""
	},
	{
	"first": "J",
	"middle": [
	"S"
	],
	"last": "Chang",
	"suffix": ""
	}
	],
	"year": 1998,
	"venue": "Special Issue on Word Sense Disambiguation. Computational Linguistics",
	"volume": "24",
	"issue": "1",
	"pages": "61--95",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Chen, J. N. and J. S. Chang. \"TopSense: A Topical Sense Clustering Method based on Information Retrieval Techniques on Machine Readable Resources.\" Special Issue on Word Sense Disambiguation. Computational Linguistics. 24(1). pp. 61-95. 1998b.",
	"links": null
	},
	"BIBREF8": {
	"ref_id": "b8",
	"title": "Combining Machine Readable lexical Resources and Bilingual Corpora for Broad Word Sense Disambiguation",
	"authors": [
	{
	"first": "J",
	"middle": [
	"N"
	],
	"last": "Chen",
	"suffix": ""
	},
	{
	"first": "J",
	"middle": [
	"S"
	],
	"last": "Chang",
	"suffix": ""
	},
	{
	"first": "H",
	"middle": [
	"H"
	],
	"last": "Sheng",
	"suffix": ""
	},
	{
	"first": "S",
	"middle": [
	"J"
	],
	"last": "Ker",
	"suffix": ""
	}
	],
	"year": null,
	"venue": "Proceedings of the Second Adptive",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Chen, J. N., J. S. Chang, H. H. Sheng and S. J. Ker. \"Combining Machine Readable lexical Resources and Bilingual Corpora for Broad Word Sense Disambiguation.\" In Proceedings of the Second Adptive 41",
	"links": null
	},
	"BIBREF9": {
	"ref_id": "b9",
	"title": "Conference of the Association for Machine Translation",
	"authors": [],
	"year": null,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "115--124",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Conference of the Association for Machine Translation. pp. 115-124. Montreal. Quebec. Canada.",
	"links": null
	},
	"BIBREF10": {
	"ref_id": "b10",
	"title": "Word Sense Disambiguation Using a Second Language Monolingual Corpus",
	"authors": [
	{
	"first": "I",
	"middle": [],
	"last": "Dagan",
	"suffix": ""
	},
	{
	"first": "A",
	"middle": [],
	"last": "Itai",
	"suffix": ""
	}
	],
	"year": 1994,
	"venue": "Computational Linguistic",
	"volume": "20",
	"issue": "4",
	"pages": "563--596",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Dagan, I. and A. Itai. \"Word Sense Disambiguation Using a Second Language Monolingual Corpus.\" Computational Linguistic. 20(4). pp. 563-596. 1994.",
	"links": null
	},
	"BIBREF11": {
	"ref_id": "b11",
	"title": "Two Languages are More Informative than One",
	"authors": [
	{
	"first": "I",
	"middle": [],
	"last": "Dagan",
	"suffix": ""
	},
	{
	"first": "A",
	"middle": [],
	"last": "Itai",
	"suffix": ""
	},
	{
	"first": "U",
	"middle": [],
	"last": "Schwall",
	"suffix": ""
	}
	],
	"year": 1991,
	"venue": "Proceedings of the 29th Annual Meeting of the Association for Computational Linguistics",
	"volume": "",
	"issue": "",
	"pages": "130--137",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Dagan, I., A. Itai, and U. Schwall. \"Two Languages are More Informative than One.\" In Proceedings of the 29th Annual Meeting of the Association for Computational Linguistics. pp. 130-137. 1991.",
	"links": null
	},
	"BIBREF12": {
	"ref_id": "b12",
	"title": "One Sense Per Discourse",
	"authors": [
	{
	"first": "W",
	"middle": [
	"A"
	],
	"last": "Gale",
	"suffix": ""
	},
	{
	"first": "K",
	"middle": [
	"W"
	],
	"last": "Church",
	"suffix": ""
	},
	{
	"first": "D",
	"middle": [],
	"last": "Yarowsky",
	"suffix": ""
	}
	],
	"year": 1992,
	"venue": "Proceedings of the Speech and Natural Language Workshop",
	"volume": "",
	"issue": "",
	"pages": "233--237",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Gale, W. A., K. W. Church and D. Yarowsky. \"One Sense Per Discourse,\" In Proceedings of the Speech and Natural Language Workshop. pp. 233-237. 1992a.",
	"links": null
	},
	"BIBREF13": {
	"ref_id": "b13",
	"title": "Using Bilingual Materials to Develop Word Sense Disambiguation Methods",
	"authors": [
	{
	"first": "W",
	"middle": [
	"A"
	],
	"last": "Gale",
	"suffix": ""
	},
	{
	"first": "K",
	"middle": [
	"W"
	],
	"last": "Church",
	"suffix": ""
	},
	{
	"first": "D",
	"middle": [],
	"last": "Yarowsky",
	"suffix": ""
	}
	],
	"year": 1992,
	"venue": "Proceedings of the 4th International Conference on Theoretical and Methodological Issues in Machine Translation",
	"volume": "",
	"issue": "",
	"pages": "101--112",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Gale, W. A., K. W. Church, and D. Yarowsky. \"Using Bilingual Materials to Develop Word Sense Disambiguation Methods.\" In Proceedings of the 4th International Conference on Theoretical and Methodological Issues in Machine Translation. pp. 101-112. 1992b.",
	"links": null
	},
	"BIBREF14": {
	"ref_id": "b14",
	"title": "Subject-dependent Co-occurrence and Word Sense Disambiguation",
	"authors": [
	{
	"first": "J",
	"middle": [],
	"last": "Guthrie",
	"suffix": ""
	},
	{
	"first": "L",
	"middle": [],
	"last": "Guthrie",
	"suffix": ""
	},
	{
	"first": "Y",
	"middle": [],
	"last": "Wilks",
	"suffix": ""
	},
	{
	"first": "H",
	"middle": [],
	"last": "Aidinejad",
	"suffix": ""
	}
	],
	"year": 1991,
	"venue": "Proceedings of the 29th Annual Meeting of the Association for Computational Linguistics",
	"volume": "",
	"issue": "",
	"pages": "146--152",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Guthrie, J., L. Guthrie, Y. Wilks and H. Aidinejad. \"Subject-dependent Co-occurrence and Word Sense Disambiguation.\" In Proceedings of the 29th Annual Meeting of the Association for Computational Linguistics. pp. 146-152. 1991.",
	"links": null
	},
	"BIBREF15": {
	"ref_id": "b15",
	"title": "Noun Homonym Disambiguation using Local Context in Large Text Corpora",
	"authors": [
	{
	"first": "M",
	"middle": [],
	"last": "Hearst",
	"suffix": ""
	}
	],
	"year": 1991,
	"venue": "Proceedings of the 7th International Conference on of UW Centre for the New OED and Text Research: Using Corpora",
	"volume": "",
	"issue": "",
	"pages": "1--22",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Hearst, M. \"Noun Homonym Disambiguation using Local Context in Large Text Corpora.\" In Proceedings of the 7th International Conference on of UW Centre for the New OED and Text Research: Using Corpora. pp. 1-22. 1991.",
	"links": null
	},
	"BIBREF16": {
	"ref_id": "b16",
	"title": "Proximity Operators -So Near and So Far",
	"authors": [
	{
	"first": "D",
	"middle": [],
	"last": "Hawking",
	"suffix": ""
	},
	{
	"first": "P",
	"middle": [],
	"last": "Thistlewaite",
	"suffix": ""
	}
	],
	"year": 1995,
	"venue": "Proceedings of the fourth Text REtrieval Conference (TREC-4)",
	"volume": "",
	"issue": "",
	"pages": "1--13",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Hawking, D. and P. Thistlewaite. \"Proximity Operators -So Near and So Far,\" In Proceedings of the fourth Text REtrieval Conference (TREC-4). pp. 1-13. 1995.",
	"links": null
	},
	"BIBREF17": {
	"ref_id": "b17",
	"title": "Computer Recognition of English Word Senses",
	"authors": [
	{
	"first": "E",
	"middle": [],
	"last": "Kelly",
	"suffix": ""
	},
	{
	"first": "P",
	"middle": [],
	"last": "Stone",
	"suffix": ""
	}
	],
	"year": 1975,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Kelly, E. and P. Stone. Computer Recognition of English Word Senses. North-Holland. Amsterdam. 1975.",
	"links": null
	},
	"BIBREF18": {
	"ref_id": "b18",
	"title": "Corpus-based Statistical Sense Resolution",
	"authors": [
	{
	"first": "C",
	"middle": [],
	"last": "Leacock",
	"suffix": ""
	},
	{
	"first": "G",
	"middle": [],
	"last": "Towell",
	"suffix": ""
	},
	{
	"first": "E",
	"middle": [
	"M"
	],
	"last": "Voorhees",
	"suffix": ""
	}
	],
	"year": 1993,
	"venue": "Proceedings of the ARPA Workshop on Human Language Technology",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Leacock, C., G. Towell and E. M. Voorhees. \"Corpus-based Statistical Sense Resolution.\" In Proceedings of the ARPA Workshop on Human Language Technology. 1993.",
	"links": null
	},
	"BIBREF19": {
	"ref_id": "b19",
	"title": "Automatic Sense Disambiguation using Machine Readable Dictionaries: How to Tell a Pine Cone from an Ice Cream Cone",
	"authors": [
	{
	"first": "M",
	"middle": [
	"E"
	],
	"last": "Lesk",
	"suffix": ""
	}
	],
	"year": 1986,
	"venue": "Proceedings of the ACM SIGDOC Conference",
	"volume": "",
	"issue": "",
	"pages": "24--26",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Lesk, M. E. \"Automatic Sense Disambiguation using Machine Readable Dictionaries: How to Tell a Pine Cone from an Ice Cream Cone,\" In Proceedings of the ACM SIGDOC Conference. pp. 24-26, Toronto. Ontario. 1986.",
	"links": null
	},
	"BIBREF20": {
	"ref_id": "b20",
	"title": "Statistical Sense Disambiguation with Relatively Small Corpora using Dictionary Definitions",
	"authors": [
	{
	"first": "A",
	"middle": [
	"K"
	],
	"last": "Luk",
	"suffix": ""
	}
	],
	"year": 1995,
	"venue": "Proceedings of the 33rd Annual Meeting of the Association for Computational Linguistics",
	"volume": "",
	"issue": "",
	"pages": "181--188",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Luk, A. K. \"Statistical Sense Disambiguation with Relatively Small Corpora using Dictionary Definitions.\" In Proceedings of the 33rd Annual Meeting of the Association for Computational Linguistics. pp. 181-188. 1995.",
	"links": null
	},
	"BIBREF23": {
	"ref_id": "b23",
	"title": "Longman Dictionary of Contemporary English. Harlow: Longman Group",
	"authors": [
	{
	"first": "P",
	"middle": [],
	"last": "Proctor",
	"suffix": ""
	}
	],
	"year": 1978,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Proctor, P. (ed.) Longman Dictionary of Contemporary English. Harlow: Longman Group. 1978.",
	"links": null
	},
	"BIBREF26": {
	"ref_id": "b26",
	"title": "Roget's Thesaurus of English words and Phrases. Longman Group UK Limited",
	"authors": [],
	"year": 1987,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Roget's Thesaurus of English words and Phrases. Longman Group UK Limited. 1987.",
	"links": null
	},
	"BIBREF27": {
	"ref_id": "b27",
	"title": "Using Context for Sense Preference",
	"authors": [
	{
	"first": "B",
	"middle": [],
	"last": "Slator",
	"suffix": ""
	}
	],
	"year": 1991,
	"venue": "Zernik (ed.) Lexical Acquisition: Exploiting On-line Resources to Build a Lexicon. Lawrence Erlbaum. Hillsdale. NJ",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Slator, B. \"Using Context for Sense Preference.\" In Zernik (ed.) Lexical Acquisition: Exploiting On-line Resources to Build a Lexicon. Lawrence Erlbaum. Hillsdale. NJ. 1991.",
	"links": null
	},
	"BIBREF28": {
	"ref_id": "b28",
	"title": "Word Sense Disambiguation with Very Large Neural Networks Extracted from Machine Readable Dictionaries",
	"authors": [
	{
	"first": "J",
	"middle": [],
	"last": "Veronis",
	"suffix": ""
	},
	{
	"first": "N",
	"middle": [],
	"last": "Ide",
	"suffix": ""
	}
	],
	"year": 1990,
	"venue": "Proceedings of the 13th International Conference on Computational Linguistics",
	"volume": "",
	"issue": "",
	"pages": "389--394",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Veronis, J. and N. Ide. \"Word Sense Disambiguation with Very Large Neural Networks Extracted from Machine Readable Dictionaries.\" In Proceedings of the 13th International Conference on Computational Linguistics. pp. 389-394. 1990.",
	"links": null
	},
	"BIBREF29": {
	"ref_id": "b29",
	"title": "The use of Machine-Readable Dictionaries in Sublanguage Analysis",
	"authors": [
	{
	"first": "D",
	"middle": [],
	"last": "Walker",
	"suffix": ""
	},
	{
	"first": "R",
	"middle": [],
	"last": "Amsler",
	"suffix": ""
	}
	],
	"year": 1984,
	"venue": "Proceedings of Workshop on Sublanguage Analysis",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Walker, D. and R. Amsler. \"The use of Machine-Readable Dictionaries in Sublanguage Analysis.\" In Analyzing Language in Restricted Domains. Grishman. R. and R. Kittredge (eds.). Lawrence Erlbaum Associates. Hillsdale. New Jersey. 1986. (also available in R. Kittredge (ed.), Proceedings of Workshop on Sublanguage Analysis; New York 1984).",
	"links": null
	},
	"BIBREF30": {
	"ref_id": "b30",
	"title": "Determination of Lexical-Semantic Relations for Multi-Lingual Terminology Structures",
	"authors": [
	{
	"first": "J",
	"middle": [],
	"last": "White",
	"suffix": ""
	}
	],
	"year": 1988,
	"venue": "Relational Models of the Lexicon",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "White, J. S. \"Determination of Lexical-Semantic Relations for Multi-Lingual Terminology Structures.\" In Relational Models of the Lexicon. Cambridge University Press, Cambridge. UK. 1988.",
	"links": null
	},
	"BIBREF31": {
	"ref_id": "b31",
	"title": "Providing Tractable Dictionaty Tools",
	"authors": [
	{
	"first": "Y",
	"middle": [
	"A"
	],
	"last": "Wilks",
	"suffix": ""
	},
	{
	"first": "D",
	"middle": [
	"C"
	],
	"last": "Fass",
	"suffix": ""
	},
	{
	"first": "C",
	"middle": [
	"M"
	],
	"last": "Guo",
	"suffix": ""
	},
	{
	"first": "J",
	"middle": [
	"E"
	],
	"last": "Mcdonald",
	"suffix": ""
	},
	{
	"first": "T",
	"middle": [],
	"last": "Plate",
	"suffix": ""
	},
	{
	"first": "B",
	"middle": [
	"M"
	],
	"last": "Slator",
	"suffix": ""
	}
	],
	"year": 1990,
	"venue": "Machine Translation. 5",
	"volume": "",
	"issue": "",
	"pages": "99--154",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Wilks, Y. A., D. C. Fass, C. M. Guo, J. E. McDonald, T. Plate and B. M. Slator. \"Providing Tractable Dictionaty Tools.\" Machine Translation. 5. pp. 99-154. 1990.",
	"links": null
	},
	"BIBREF32": {
	"ref_id": "b32",
	"title": "Word-Sense Disambiguation using Statistical Models of Roget's Categories Trained on Large Corpora",
	"authors": [
	{
	"first": "D",
	"middle": [],
	"last": "Yarowsky",
	"suffix": ""
	}
	],
	"year": 1992,
	"venue": "Proceedings of the 14th International Conference on Computational Linguistics",
	"volume": "",
	"issue": "",
	"pages": "454--460",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Yarowsky, D. \"Word-Sense Disambiguation using Statistical Models of Roget's Categories Trained on Large Corpora.\" In Proceedings of the 14th International Conference on Computational Linguistics. pp. 454-460. Nantes. France. 1992.",
	"links": null
	},
	"BIBREF33": {
	"ref_id": "b33",
	"title": "Unsupervised Word Sense Disambiguation Rivaling Supervised Methods",
	"authors": [
	{
	"first": "D",
	"middle": [],
	"last": "Yarowsky",
	"suffix": ""
	}
	],
	"year": 1995,
	"venue": "Proceedings of the 33rd Annual Meeting of the Association for Computational Linguistics",
	"volume": "",
	"issue": "",
	"pages": "189--196",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Yarowsky, D. \"Unsupervised Word Sense Disambiguation Rivaling Supervised Methods.\" In Proceedings of the 33rd Annual Meeting of the Association for Computational Linguistics. pp. 189-196. 1995.",
	"links": null
	},
	"BIBREF34": {
	"ref_id": "b34",
	"title": "Human Behavior and the Principle of Least Eeffect",
	"authors": [
	{
	"first": "G",
	"middle": [],
	"last": "Zipf",
	"suffix": ""
	}
	],
	"year": 1994,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Zipf, G. Human Behavior and the Principle of Least Eeffect. Hafner. New York. 1994.",
	"links": null
	}
	},
	"ref_entries": {
	"FIGREF0": {
	"type_str": "figure",
	"text": "s ) is the lexical contextual representation derived from definition D, CCR(D w, s ) is the conceptual contextual representation derived from definition D, LCR(E w, s ) is the lexical contextual representation derived from the set of examples E, and CCR(E w, s ) is the conceptual contextual representation derived from the set of examples E.",
	"num": null,
	"uris": null
	},
	"FIGREF1": {
	"type_str": "figure",
	"text": "the distance from t to W in number of words. CR(W,S), CON(W)) \| Sim (CR(W,S), CON(W)) < S(W, CON(W)) }, (W, CON(W)) = the rank of S(W, CON(W)) among all S*(X, CON(X)) for all n instances of polysemous word X and context CON(X), RANK-T(W, CON(W)) = the rank of TSCORE(W, CON(W)) among TSCORE(X, CON(X)) for all n instances and context of polysemous word X.",
	"num": null,
	"uris": null
	},
	"TABREF0": {
	"type_str": "table",
	"html": null,
	"text": "Lexical contextual representations for bank senses.",
	"num": null,
	"content": "<table><tr><td>Sense ID</td><td>Sense Label S</td><td>Lexical Context Representation LCR(D bank, s )</td></tr><tr><td>bank.4.n.1</td><td>MONEY</td><td>{place, money, keep, pay, demand, activity}</td></tr><tr><td>bank.1.n.1</td><td>RIVER</td><td>{land, lake, river}</td></tr><tr><td>bank.1.n.5</td><td>SANDBANK</td><td>{underwater, sand, harbour}</td></tr><tr><td>bank.1.n.2</td><td>EARTH</td><td>{earth, heap, field, garden, boarder, division}</td></tr><tr><td>bank.1.n.3</td><td>PILE</td><td>{mass, snow, cloud, mud}</td></tr><tr><td>bank.1.n.4</td><td>ROAD</td><td>{car, aircraft, move, side, turn}</td></tr><tr><td>bank.3.n.1</td><td>ROW</td><td>{row, oar, boat, key, typewriter}</td></tr><tr><td>bank.4.n.2</td><td>MEDICINE</td><td>{place, hold, use, organic, product, human, origin, medical}</td></tr><tr><td>bank.4.n.3</td><td>GAMBLE</td><td>{person, keep, supply, money, payment, game, chance}</td></tr></table>"
	},
	"TABREF1": {
	"type_str": "table",
	"html": null,
	"text": "Conceptual context representations for a definition D of bank senses.",
	"num": null,
	"content": "<table><tr><td>Sense Label S</td><td>Topics 2 (with weights) on CCR(D bank, s )</td></tr><tr><td>MONEY</td><td>Je(0.45), Jf(0.33), Jd(0.22)</td></tr><tr><td>RIVER</td><td>Ld(0.45), Mf(0.26), Me(0.14), Hc(0.07), Af(0.05), Ad(0.04)</td></tr><tr><td>EARTH</td><td>La(0.36), Ld(0.24), Eg(0.20), Me(0.12), Ie(0.08)</td></tr><tr><td>PILE</td><td>Lc(0.59), Db(0.13), Hc(0.09), La(0.09), Md(0.09)</td></tr><tr><td>ROAD</td><td>Md(0.45), Me(0.38), Ld( 0.17)</td></tr><tr><td>ROW</td><td>Md(0.49), Gd(0.18), Mc(0.16), Kb(0.12), Me(0.06)</td></tr><tr><td>MEDICINE</td><td>Bd(0.70), Bj( 0.30)</td></tr><tr><td>GAMBLE</td><td>Ke(0.35), Kh(0.28), Kf(0.23), Cn(0.14)</td></tr></table>"
	},
	"TABREF2": {
	"type_str": "table",
	"html": null,
	"text": "Definition-based conceptual context representation and related word lists for bank senses",
	"num": null,
	"content": "<table><tr><td>Sense Division S</td><td>Topics</td><td>Word List on CCR(D bank, s )</td></tr><tr><td>MONEY</td><td>Je(0.45)</td><td>{money, pay, cash, capital, account, charge, ...</td></tr><tr><td/><td>Jf(0.33)</td><td>pay, bond, bill, charge, ...</td></tr><tr><td/><td>Jd(0.22)</td><td>money, cash, fund, check, ... }</td></tr><tr><td>RIVER</td><td>Ld(0.45)</td><td/></tr><tr><td/><td>Mf(0.26)</td><td/></tr><tr><td/><td>Me(0.14)</td><td/></tr><tr><td/><td>Hc(0.07)</td><td/></tr><tr><td/><td>Af(0.05) Ad(0.04)</td><td>.}</td></tr><tr><td>EARTH</td><td>La(0.36)</td><td/></tr><tr><td/><td>Ld(0.24)</td><td/></tr><tr><td/><td>Eg(0.20)</td><td/></tr><tr><td/><td>Me(0.12) Ie(0.08)</td><td>.}</td></tr><tr><td>PILE</td><td>Lc(0.59) Db(0.13)</td><td>{weather, climate, sky, cloud, fog, steam, ... roof, ceiling, wall, door, ground, ...</td></tr><tr><td/><td>Hc(0.09)</td><td>rock, stone, clay, soil, ...</td></tr><tr><td/><td>La(0.09) Md(0.09)</td><td>universe, space, planet, constellation, ... transport, vehicle, car, motorcar, transit, ...}</td></tr><tr><td>ROAD</td><td>Md(0.45)</td><td>{transport, vehicle, car, motorcar, transit, ...</td></tr><tr><td/><td>Me(0.38)</td><td>place, edge, road, border, ...</td></tr><tr><td/><td>Ld(0.17)</td><td>lake, land, river, shore, stream, beach, ...}</td></tr><tr><td>ROW</td><td>Md(0.49) Gd(0.18)</td><td>{transport, vehicle, car, motorcar, transit, ... printing, sign, letter, code, \u2026</td></tr><tr><td/><td>Mc( 0.16)</td><td>sail, caravan, itinerary, \u2026</td></tr><tr><td/><td>Kb(0.12) Me(0.06)</td><td>song, melody, dance, \u2026 road, street, \u2026}</td></tr><tr><td>MEDICINE</td><td>Bd(0.70)</td><td>{blood, trunk, breast, back, buttock, waist, ...</td></tr><tr><td/><td>Bj(0.30)</td><td>patient, examine, diagnose, soothe, ...}</td></tr><tr><td>GAMBLE</td><td>Ke(0.35)</td><td>{athletics, run, jump, ride, game, round, ...</td></tr><tr><td/><td>Kh(0.28)</td><td>ball game, shoot, golf, pitch, football, ...</td></tr></table>"
	},
	"TABREF3": {
	"type_str": "table",
	"html": null,
	"text": "Bilingual contextual representations for bank senses based on conceptual context representations from definition and dictionary translation. \u5bb6, athletics, run, jump, ride, game, round, ..., ball game, shoot, golf, pitch, football, ..., cards, pack, suit, heart, club, poker, dice, ... war, warfare, conflict, fight, battleground, ... } Results of sense tagging.",
	"num": null,
	"content": "<table><tr><td>Sense</td><td/><td colspan=\"2\">Context on CCR(D bank, s )</td></tr><tr><td>Division S</td><td/><td/></tr><tr><td colspan=\"4\">MONEY {\u9280, \ufa08, money, pay, cash, capital, account, charge ..., pay, bond, bill, charge ..., money,</td></tr><tr><td/><td/><td>cash, fund, check, ... }</td></tr><tr><td colspan=\"4\">RIVER {\u5cb8, \u5824, \u6c99, \u6d32,lake, land, river, shore, stream, beach, ..., boat, ship, craft, port, ..., place,</td></tr><tr><td/><td colspan=\"3\">edge, road, border, ..., rock, stone, clay, soil, ..., fish, crab, coral, shell, fur, ..., chicken, duct,</td></tr><tr><td/><td/><td>goose, seabird, ... }</td></tr><tr><td colspan=\"4\">EARTH {\u7530, \u57c2, universe, space, planet, constellation, ..., lake, land, river, shore, stream, beach, ...,</td></tr><tr><td/><td colspan=\"3\">farming, field, crop, stock, productive, ..., place, edge, road, border, ..., playgroup, school,</td></tr><tr><td/><td/><td colspan=\"2\">college, classroom, ... }</td></tr><tr><td colspan=\"4\">PILE {\u4e00, \u584a, \u4e00, \u5718, weather, climate, sky, cloud, fog, steam, ..., roof, chimney, ceiling, wall,</td></tr><tr><td/><td colspan=\"3\">door, ground, ..., rock, stone, clay, soil, ..., universe, space, planet, constellation, \u2026, transport,</td></tr><tr><td/><td/><td colspan=\"2\">vehicle, car, motorcar, transit, ... }</td></tr><tr><td colspan=\"4\">ROAD {\u908a, \u5761, transport, vehicle, car, motorcar, transit, ..., place, edge, road, border, ..., lake, land,</td></tr><tr><td/><td/><td colspan=\"2\">river, shore, stream, beach, ... }</td></tr><tr><td>ROW</td><td colspan=\"3\">{\u4e00, \u6392, transport, vehicle, car, motorcar, transit, \u2026 , printing, sign, letter, code, \u2026, sail,</td></tr><tr><td/><td/><td colspan=\"2\">caravan, itinerary,\u2026, song, melody, dance, \u2026, road, street, \u2026\u2026... }</td></tr><tr><td>MEDICIN</td><td colspan=\"3\">{\u8840, \u5eab, blood, trunk, breast, back, buttock, waist, ..., patient, examine, diagnose,</td></tr><tr><td>E</td><td/><td>soothe, ... }</td></tr><tr><td colspan=\"2\">GAMBLE {\u838a, Example</td><td colspan=\"2\">The interest in my bank account accrued over the years.</td></tr><tr><td colspan=\"2\">Translation</td><td colspan=\"2\">\u6211\u9280\ufa08\u5e33\u6236\u7684\uf9dd\u606f\u9010\uf98e\u6709\u6240\u589e\u52a0\u3002</td></tr><tr><td colspan=\"2\">Tagged Keywords</td><td colspan=\"2\">interest/Je, bank/Je, account/Je, accrue/Nd</td></tr><tr><td colspan=\"2\">Gloss for Topics</td><td>Je</td><td>Banking</td></tr><tr><td/><td/><td>Nd</td><td>Size</td></tr><tr><td colspan=\"3\">Algorithm 2: Labeling bilingual training corpus</td></tr></table>"
	},
	"TABREF4": {
	"type_str": "table",
	"html": null,
	"text": "Lexicalized contextual representations for bank from the set of LDOCE examples E.",
	"num": null,
	"content": "<table><tr><td>Sense Label S</td><td>Context (with frequency) in LCR(E bank, s )</td></tr><tr><td>MONEY</td><td>rob(23), account(15), money(8), criminal(6), interest(5), keep(5), paper(4), police(4),</td></tr><tr><td/><td>robber(4), thief(4), cheque(3), ...</td></tr><tr><td>RIVER</td><td>river(15), city</td></tr></table>"
	},
	"TABREF5": {
	"type_str": "table",
	"html": null,
	"text": "Conceptualized contextual representation for bank from the set of LDOCE examples E.",
	"num": null,
	"content": "<table><tr><td>Sense Division S</td><td>Related Topics</td><td>Context on CCR(E bank, s )</td></tr><tr><td>MONEY</td><td>Je(0.45), Jf(0.33), Jd(0.22)</td><td>{officer, cop, detective, guard, protect,</td></tr><tr><td/><td/><td>gangster, hoodlum, larceny, hijacking,</td></tr><tr><td/><td/><td>burglar, steal, fraud, swindle, \u2026}</td></tr><tr><td>RIVER</td><td>Ld(0.45), Mf(0.26), Me(0.14), Hc(0.07),</td><td>{east, west, north, south, up, down, erode, elk,</td></tr><tr><td/><td>Af( 0.05), Ad(0.04)</td><td>moose, rat, mouse, rabbit, hare, \u2026 }</td></tr><tr><td>EARTH</td><td>La(0.36), Ld(0.24), Eg(0.20), Me(0.12),</td><td>{tide, ebb, current, spate, \u2026}</td></tr><tr><td/><td>Ie(0.08)</td><td/></tr><tr><td>PILE</td><td>Lc(0.59), Db(0.13), Hc(0.09), La(0.09),</td><td>{fog, steam, haze, dew, mist, \u2026}</td></tr><tr><td/><td>Md(0.09)</td><td/></tr><tr><td>ROAD</td><td>Md(0.45), Me(0.38), Ld( 0.17)</td><td>{forest, jungle, hole, crack, \u2026}</td></tr><tr><td>ROW</td><td>Md(0.49), Gd(0.18), Mc(0.16), Kb(0.12),</td><td>-</td></tr><tr><td/><td>Me(0.06)</td><td/></tr><tr><td>MEDICINE</td><td>Bd(0.70), Bj( 0.30)</td><td>-</td></tr><tr><td>GAMBLE</td><td>Ke(0.35), Kh(0.28), Kf(0.23), Cn(0.14)</td><td>-</td></tr></table>"
	},
	"TABREF6": {
	"type_str": "table",
	"html": null,
	"text": "{place, money, keep, pay, demand, activity }, CCR(D bank.4.n.1 ) = {money, pay, cash, capital, account, charge, ... pay, bond, bill, charge, ... money, cash, fund, check, ... } ,",
	"num": null,
	"content": "<table><tr><td>LCR(E bank.4.n.1 ) = {rob(23), account(15), money(8), criminal(6), interest(5),</td></tr><tr><td>keep(5), paper(4), police(4), robber(4), thief(4),</td></tr><tr><td>cheque(3), ...}, and</td></tr><tr><td>CCR(E bank.4.n.1 ) = {officer, cop, detective, guard, protect, gangster, hoodlum,</td></tr><tr><td>larceny, hijacking, burglar, steal, fraud, swindle, \u2026}.</td></tr></table>"
	},
	"TABREF7": {
	"type_str": "table",
	"html": null,
	"text": "). For instance, there are 16 instances of account in WORD(bank, MONEY) and no other word list WORD(bank, S), S \u2260 MONEY, contains account. Thus, we have tf account, MONEY = 16 and idf account = 8. Thus, the weight for account in CR(bank, MONEY) is W account, MONEY = 16 * 8 = 128.",
	"num": null,
	"content": "<table/>"
	},
	"TABREF8": {
	"type_str": "table",
	"html": null,
	"text": "Top-ranked words in combined contextual representation based on definitions and examples of bank senses.",
	"num": null,
	"content": "<table><tr><td>Sense S</td><td>Context(with weights 4 ) on CR(bank,S)</td></tr><tr><td colspan=\"2\">MONEY rob(6.17), money(2.17), account(2.04), criminal(1.61), interest(1.40), keep(1.39), pay(1.18),</td></tr><tr><td/><td>police(1.07), robber(1.07), thief(1.07), \u2026</td></tr><tr><td>RIVER</td><td>river(5.54), leave(2.18), towards(2.18), ship(1.23), city(1.09), dangerous(1.09), deer(1.09),</td></tr><tr><td/><td>descend(1.09), excavation(1.09), north(0.73), fish(0.56), \u2026</td></tr><tr><td colspan=\"2\">EARTH build(3.92), vole(3.92), earth(1.42), rise(0.98), flood(0.73), water(0.65), agricultural(0.20),</td></tr><tr><td/><td>barn(0.20), farm(0.20), garden(0.20), \u2026</td></tr><tr><td>PILE</td><td>cloud(2.26), dark(1.97), heavy (0.99), storm(0.64), hall(0.36), shower(0.36),</td></tr><tr><td/><td>atmosphere(0.18), blizzard(0.18), blow(0.06), breeze(0.06), \u2026</td></tr><tr><td colspan=\"2\">ROAD moss(4.38), sit (2.19), wood (1.14), rest (1.09), gradient (0.18), junction (0.18), subway(0.08),</td></tr><tr><td/><td>tunnel(0.08), accelerator(0.06), accident(0.06), \u2026</td></tr><tr><td>ROW</td><td>call(0.19), page(0.19), classical(0.16), compose(0.16), composition(0.16), leader(0.16),</td></tr><tr><td/><td>caravan(0.12), porter(0.12), bell(0.11), horn(0.11), \u2026</td></tr><tr><td>MEDICI</td><td>crutch(0.48), gut(0.48), abdomen(0.34), abdominal(0.34), ankle(0.34), anal(0.34), anus(0.34),</td></tr><tr><td>NE</td><td>aorta(0.34), pendicities(0.34), armpit(0.34), \u2026</td></tr><tr><td>GAMBLE</td><td>club(0.43), cup(0.32), loser(0.24), win(0.24), defense(0.20), bet(0.17), champion(0.17),</td></tr><tr><td/><td>competition(0.17), gamble(0.17), games(0.17), \u2026</td></tr></table>"
	},
	"TABREF9": {
	"type_str": "table",
	"html": null,
	"text": "Samples of disambiguated topical contexts of bank in the Brown corpus. That was in the days before blood banks, of course, and transfusions had to be given directly from donor to",
	"num": null,
	"content": "<table><tr><td>Sense</td><td>Example</td><td>Typical, Easy Context of Various Senses of bank</td><td>General</td><td>Task-specif</td></tr><tr><td/><td>No.</td><td/><td>Topical</td><td>ic Context</td></tr><tr><td/><td/><td/><td>Context</td><td/></tr><tr><td>MONEY</td><td>(9)</td><td>\u2026 It permits the state to take over bank accounts, stocks</td><td>Account</td><td>law</td></tr><tr><td/><td/><td>and other personal property of persons missing for seven</td><td>Stock</td><td>bill</td></tr><tr><td/><td/><td>years or more. \u2026</td><td>Property</td><td/></tr><tr><td>RIVER</td><td>(10)</td><td>\u2026 WE WERE CAMPING a few weeks ago on Cape</td><td>Seagull</td><td>tree</td></tr><tr><td/><td/><td>Hatteras Campground in that land of pirates, seagulls and</td><td>Bluefish</td><td>camping</td></tr><tr><td/><td/><td>bluefish on North Carolina's famed outer banks. \u2026</td><td>Hill</td><td/></tr><tr><td>PILE</td><td>(11)</td><td>\u2026 At the end of the calculated time he'd nose the Waco</td><td>Cloud</td><td>flight</td></tr><tr><td/><td/><td>down through the cloud bank and hope to breakthrough</td><td/><td>through</td></tr><tr><td/><td/><td>where some feature of the winter landscape would be</td><td/><td/></tr><tr><td/><td/><td>recognizable. \u2026</td><td/><td/></tr><tr><td colspan=\"3\">MEDICINE (12) \u2026 patient. \u2026</td><td>blood patient doctor</td><td>donor transfusion</td></tr></table>"
	},
	"TABREF10": {
	"type_str": "table",
	"html": null,
	"text": "Skewed sense distribution is corpus dependent",
	"num": null,
	"content": "<table><tr><td>Word</td><td>Sense</td><td>Brown</td><td>WSJ</td></tr><tr><td>Interest</td><td>MONEY</td><td>49</td><td>122</td></tr><tr><td/><td>CURIOSITY</td><td>194</td><td>53</td></tr><tr><td>Sentence</td><td>GAMMAR</td><td>22</td><td>1</td></tr><tr><td/><td>JUDGEMENT</td><td>10</td><td>11</td></tr><tr><td>Bass</td><td>MUSIC</td><td>15</td><td>2</td></tr><tr><td/><td>FISH</td><td>1</td><td>0</td></tr></table>"
	},
	"TABREF14": {
	"type_str": "table",
	"html": null,
	"text": "Weights for salient words in Example (12) for bank in the initial WSD stage.",
	"num": null,
	"content": "<table><tr><td/><td/><td/><td/><td/><td/><td colspan=\"3\">W t, s in CR (bank, S)</td></tr><tr><td/><td>X t</td><td>W t</td><td>S MONE</td><td>S RIVER</td><td>S EARTH</td><td>S PILE</td><td colspan=\"3\">S ROW S ROAD S MEDICINE</td><td>S GAMBLE</td></tr><tr><td/><td/><td/><td>Y</td><td/><td/><td/><td/><td/></tr><tr><td colspan=\"3\">drive -46 0.15</td><td>-</td><td>0.81</td><td>0.02</td><td>0.02</td><td>0.07</td><td>0.01</td><td>-</td><td>0.04</td></tr><tr><td colspan=\"4\">getaway -44 0.15 1.37</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>car</td><td colspan=\"3\">-43 0.15 0.94</td><td>-</td><td>-</td><td>0.81</td><td>0.12</td><td>0.07</td><td>-</td><td>-</td></tr><tr><td colspan=\"4\">robbery -39 0.16 0.81</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td colspan=\"4\">husband -24 0.20 1.07</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>year</td><td colspan=\"3\">-18 0.24 0.81</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td colspan=\"4\">prison -14 0.27 2.04</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td colspan=\"4\">robbery -7 0.38 2.31</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td colspan=\"4\">branch -3 0.58 3.81</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>bank</td><td>0</td><td/><td/><td/><td/><td/><td/><td/></tr><tr><td>charge</td><td colspan=\"3\">3 0.58 3.58</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td colspan=\"4\">return 13 0.28 1.77</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>week</td><td colspan=\"3\">18 0.24 1.34</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td colspan=\"3\">report 22 0.21</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>0.90</td><td>-</td><td>-</td></tr><tr><td>year</td><td colspan=\"3\">30 0.18 0.81</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td colspan=\"4\">husband 45 0.15 1.07</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr></table>"
	},
	"TABREF15": {
	"type_str": "table",
	"html": null,
	"text": "Weights for salient words in Example (13) for bank in the initial WSD stage.With the instance bank in Example (12) resolved to MONEY-bank, the following triple is created and added to T 1 .( bank, MONEY-bank, \"to face charges of assault and robbery, Portland detectives said Friday. Mrs. Lavaughn Huntley is accused of driving the getaway car used in a robbery of the Woodyard Bros' Grocery, 2825 E. Burnside St., in April of 1959. Her husband, who was sentenced to 15 years in the federal prison at McNeil Island last april for robbery of the Hillsdale branch of Multnomah Bank, also was charged with the store holdup. Secret Grand Jury indictments were returned against the pair last week, Detective Murray Logan reported. The Phoenix arrest culminates more than a year's investigation by Detective William Taylor and other officers. Taylor said Mrs. Huntley and her husband also will be questioned about\" )",
	"num": null,
	"content": "<table><tr><td>Word</td><td/><td/><td/><td/><td colspan=\"3\">W t, s in CR (bank, S)</td><td/><td/></tr><tr><td/><td>X t W t</td><td>S MONE</td><td colspan=\"3\">S RIVER S EARTH S PILE</td><td>S ROW</td><td>S ROAD</td><td colspan=\"2\">S MEDICINE S GAMBLE</td></tr><tr><td/><td/><td>Y</td><td/><td/><td/><td/><td/><td/><td/></tr><tr><td>call</td><td>-50 0.14</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>0.99</td><td>-</td><td>-</td></tr><tr><td>pasture</td><td>-48 0.14</td><td>-</td><td>-</td><td>1.00</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>count</td><td colspan=\"2\">-47 0.15 0.82</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>0.86</td></tr><tr><td>counter</td><td>-45 0.15</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>0.92</td></tr><tr><td>ride</td><td>-44 0.15</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>0.97</td></tr><tr><td>pasture</td><td>-41 0.16</td><td>-</td><td>-</td><td>1.00</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>ahead</td><td>-20 0.22</td><td>-</td><td>0.94</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td colspan=\"3\">representative -3 0.58 3.04</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>bank</td><td>0</td><td/><td/><td/><td/><td/><td/><td/><td/></tr><tr><td>hold</td><td>2 0.71</td><td>-</td><td>-</td><td>3.65</td><td>-</td><td>-</td><td>3.03</td><td>-</td><td>3.04</td></tr><tr><td>paper</td><td colspan=\"3\">4 0.50 1.07 0.81</td><td>-</td><td>0.81</td><td>-</td><td>0.82</td><td>-</td><td>-</td></tr><tr><td>note</td><td colspan=\"2\">9 0.33 1.79</td><td>-</td><td>-</td><td>-</td><td>-</td><td>1.58</td><td>-</td><td>-</td></tr><tr><td>paper</td><td colspan=\"3\">17 0.24 1.07 0.81</td><td>-</td><td>0.81</td><td>-</td><td>0.82</td><td>-</td><td>-</td></tr><tr><td>book</td><td colspan=\"2\">19 0.23 0.93</td><td>-</td><td>-</td><td>-</td><td>-</td><td>0.89</td><td>-</td><td>-</td></tr><tr><td>count</td><td colspan=\"2\">20 0.22 0.82</td><td>-</td><td>-</td><td>-</td><td>-</td><td/><td>-</td><td>0.86</td></tr><tr><td>book</td><td colspan=\"2\">28 0.19 0.93</td><td>-</td><td>-</td><td>-</td><td>-</td><td>0.89</td><td>-</td><td>-</td></tr><tr><td>profit</td><td colspan=\"2\">40 0.16 0.84</td><td>-</td><td>-</td><td>-</td><td>-</td><td/><td>-</td><td>-</td></tr><tr><td>lead</td><td>45 0.15</td><td>-</td><td>0.81</td><td>-</td><td>0.82</td><td>0.92</td><td>0.89</td><td>-</td><td>-</td></tr><tr><td>Step 2:</td><td/><td/><td/><td/><td/><td/><td/><td/><td/></tr></table>"
	},
	"TABREF16": {
	"type_str": "table",
	"html": null,
	"text": "",
	"num": null,
	"content": "<table/>"
	},
	"TABREF17": {
	"type_str": "table",
	"html": null,
	"text": "Selected sample of initial knowledge for bank senses.Armed with the new CRs, the instances that do not pass the test in Step 1 are re-evaluated again in Step 4. The similarity for each of those instances, including Example (13), is re-calculated for all possible word senses. The new weights for contextual words in Example (13) are shown in",
	"num": null,
	"content": "<table><tr><td>Sense</td><td>Top-ranking Contextual Words (with weights)</td></tr><tr><td>MONEY</td><td>rob(6.17), money(2.17), account(2.04), interest(1.40), pay(1.17), robber(1.07),</td></tr><tr><td/><td>month(0.80), robbery(0.81), prison(0.81), year(0.81), charge(0.58), \u2026</td></tr><tr><td>RIVER</td><td>river(5.54), ship(1.23), deer(1.09), hunter(1.09), drive(0.81), fish(0.55), air(0.54),</td></tr><tr><td/><td>hill(0.44), east(0.36), south(0.36), boat(0.13), boatman(0.13), \u2026</td></tr><tr><td>EARTH</td><td>build(3.91), water(0.65), sky(0.29), plant(0.19), west(0.17), north(0.11), south(0.11),</td></tr><tr><td/><td>sidewalk(0.02), street(0.02), drive(0.02) , bridge(0.01), \u2026</td></tr><tr><td>PILE</td><td>wet(0.29), basin(0.06), window(0.06), table(0.03), clay(0.02), cloth(0.02),</td></tr><tr><td/><td>drive(0.02), \u2026</td></tr><tr><td>ROAD</td><td>moss(4.38), sit(2.19), wood(1.14), subway(0.18), car(0.07), drive(0.01), \u2026</td></tr><tr><td>ROW</td><td>car(0.12), letter(0.09), write(0.09), visit(0.08), drive(0.07), column(0.04),</td></tr><tr><td/><td>story(0.04), \u2026</td></tr><tr><td>MEDICINE</td><td>blood(0.33), body(0.33), shoulder(0.33), patient(0.14), doctor(0.07), course(0.03), \u2026</td></tr><tr><td>GAMBLE</td><td>box(0.11), pocket(0.11), play(0.08), check(0.05), point(0.05), drive(0.04), \u2026</td></tr></table>"
	},
	"TABREF18": {
	"type_str": "table",
	"html": null,
	"text": "Selected sample of adaptive knowledge for bank senses.",
	"num": null,
	"content": "<table><tr><td>Sense</td><td>Selected Contextual Words (with weights)</td></tr><tr><td>MONEY</td><td>cooperatives(1.50), department(1.37), affairs(1.07), export-import(0.69), federal(0.60),</td></tr><tr><td/><td>government(0.54), short-term(0.43), cooperative(0.41), administration(0.38),</td></tr><tr><td/><td>firm(0.35), sponsor(0.35), \u2026</td></tr><tr><td>RIVER</td><td>church(0.80), soldier(0.68), dill(0.66), camping(0.54), fame(0.52), outer(0.52),</td></tr><tr><td/><td>rhine(0.52), motel(0.40), sight(0.40), tree(0.40), camp(0.33), \u2026</td></tr><tr><td>EARTH</td><td>manchester(4.35), company(3.77), telegraph(3.77), goodwin(3.11), power(1.88),</td></tr><tr><td/><td>light(1.69), door(1.64), cemetery(1.31), commercial(1.31), dwelling(1.31),</td></tr><tr><td/><td>electric(1.31), business(1.23), construction(1.23), \u2026</td></tr><tr><td>PILE</td><td>tiber(3.69), fold(2.83), moonlight(1.84), thick(1.84), anatomy(0.98), bedside(0.98),</td></tr><tr><td/><td>buckle(0.98), damn(0.98), dancer(0.98), dark(0.58), \u2026</td></tr><tr><td>ROAD</td><td>-</td></tr><tr><td>ROW</td><td>feel(8.51), error(5.37), correct(3.58), shareholder(3.47), people(3.36), data(0.89),</td></tr><tr><td/><td>fund(0.89), funds(0.89), \u2026</td></tr><tr><td>MEDICINE</td><td>stumbled(4.51), transfusions(3.46), donor(2.25), frail(2.25), child(1.20),</td></tr><tr><td/><td>laboratory(1.20), neck(1.20), night(1.20), sample(1.20), \u2026</td></tr><tr><td>GAMBLE</td><td>fraud(4.10), drink(2.85), grade(2.67), stare(2.67), chief(1.42), collusion(1.42),</td></tr><tr><td/><td>conclusive(1.42), death(1.42), \u2026</td></tr></table>"
	},
	"TABREF20": {
	"type_str": "table",
	"html": null,
	"text": "Disambiguation results for thirteen ambiguous words in the Brown corpus.",
	"num": null,
	"content": "<table><tr><td colspan=\"3\">Word Sample Sizes StaticSense</td><td/><td>AdaptSense</td></tr><tr><td/><td/><td># of correct</td><td># of correct in 1 st run</td><td># of correct in 2 nd run</td></tr><tr><td>bank</td><td>97</td><td>68</td><td>71</td><td>71</td></tr><tr><td>bass</td><td>16</td><td>16</td><td>16</td><td>16</td></tr><tr><td>bow</td><td>12</td><td>3</td><td>3</td><td>2</td></tr><tr><td>cone</td><td>14</td><td>14</td><td>14</td><td>14</td></tr><tr><td>duty</td><td>75</td><td>67</td><td>69</td><td>69</td></tr><tr><td>galley</td><td>4</td><td>4</td><td>4</td><td>4</td></tr><tr><td>interest</td><td>346</td><td>213</td><td>228</td><td>226</td></tr><tr><td>issue</td><td>141</td><td>67</td><td>88</td><td>97</td></tr><tr><td>mole</td><td>4</td><td>2</td><td>2</td><td>2</td></tr><tr><td>sentence</td><td>32</td><td>30</td><td>30</td><td>30</td></tr><tr><td>slug</td><td>8</td><td>4</td><td>6</td><td>6</td></tr><tr><td>star</td><td>46</td><td>28</td><td>29</td><td>29</td></tr><tr><td>taste</td><td>51</td><td>36</td><td>36</td><td>36</td></tr><tr><td>Total</td><td>846</td><td>552</td><td>596</td><td>602</td></tr><tr><td colspan=\"2\">precision</td><td>65.2%</td><td>70.5%</td><td>71.2%</td></tr></table>"
	},
	"TABREF21": {
	"type_str": "table",
	"html": null,
	"text": "Disambiguation results for thirteen ambiguous words in the WSJ test set.",
	"num": null,
	"content": "<table><tr><td>Word</td><td>Sample Sizes</td><td>StaticSense</td><td colspan=\"2\">AdaptSense</td></tr><tr><td/><td/><td># of correct</td><td># of correct in 1 st run</td><td># of correct in 2 nd run</td></tr><tr><td>Bank</td><td>370</td><td>350</td><td>353</td><td>353</td></tr><tr><td>Bass</td><td>2</td><td>2</td><td>2</td><td>2</td></tr><tr><td>Bow</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>Cone</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>duty</td><td>25</td><td>19</td><td>22</td><td>22</td></tr><tr><td>galley</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>interest</td><td>221</td><td>123</td><td>127</td><td>122</td></tr><tr><td>issue</td><td>260</td><td>181</td><td>177</td><td>175</td></tr><tr><td>mole</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>sentence</td><td>12</td><td>11</td><td>12</td><td>12</td></tr><tr><td>slug</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>star</td><td>7</td><td>3</td><td>2</td><td>2</td></tr><tr><td>taste</td><td>6</td><td>3</td><td>3</td><td>3</td></tr><tr><td>Total</td><td>903</td><td>692</td><td>698</td><td>691</td></tr><tr><td colspan=\"2\">precision</td><td>76.6%</td><td>77.3%</td><td>76.5%</td></tr></table>"
	}
	}
	}
	}