{ "paper_id": "O06-4004", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T08:07:11.512670Z" }, "title": "A Probe into Ambiguities of Determinative-Measure Compounds", "authors": [ { "first": "Shih-Min", "middle": [], "last": "Li", "suffix": "", "affiliation": { "laboratory": "", "institution": "Academia Sinica", "location": { "country": "Taipei" } }, "email": "" }, { "first": "Su-Chu", "middle": [], "last": "Lin", "suffix": "", "affiliation": { "laboratory": "", "institution": "Academia Sinica", "location": { "country": "Taipei" } }, "email": "" }, { "first": "Chia-Hung", "middle": [], "last": "Tai", "suffix": "", "affiliation": { "laboratory": "", "institution": "Academia Sinica", "location": { "country": "Taipei" } }, "email": "" }, { "first": "Keh-Jiann", "middle": [], "last": "Chen", "suffix": "", "affiliation": { "laboratory": "", "institution": "Academia Sinica", "location": { "country": "Taipei" } }, "email": "kchen@iis.sinica.edu.tw" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "This paper aims to further probe into the problems of ambiguities for automatic identification of determinative-measure compounds (DMs) in Chinese and to develop sets of rules to identify DMs and their parts of speech. It is known that Chinese DMs are identifiable by regular expressions. DM rule matching helps one solve word segmentation ambiguities, and parts of speech help one improve sense recognition and part-of-speech tagging. In this paper, a deep analysis based on corpus data was studied. With analyses of error identification and disambiguation of DM compounds, the authors classified three types of ambiguities, i.e. word segmentation, sense, and pos ambiguities. DM rules are necessary complements to dictionaries and helpful to resolve word segmentation ambiguities by applying resolution principles and segmentation models. Sense and pos ambiguities are also expected to be resolved by different approaches during postprocessing.", "pdf_parse": { "paper_id": "O06-4004", "_pdf_hash": "", "abstract": [ { "text": "This paper aims to further probe into the problems of ambiguities for automatic identification of determinative-measure compounds (DMs) in Chinese and to develop sets of rules to identify DMs and their parts of speech. It is known that Chinese DMs are identifiable by regular expressions. DM rule matching helps one solve word segmentation ambiguities, and parts of speech help one improve sense recognition and part-of-speech tagging. In this paper, a deep analysis based on corpus data was studied. With analyses of error identification and disambiguation of DM compounds, the authors classified three types of ambiguities, i.e. word segmentation, sense, and pos ambiguities. DM rules are necessary complements to dictionaries and helpful to resolve word segmentation ambiguities by applying resolution principles and segmentation models. Sense and pos ambiguities are also expected to be resolved by different approaches during postprocessing.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "To a speaker of English, one of the most striking features of the Mandarin noun phrase is the classifier. A classifier is a word that must occur with a number and/or a demonstrative, or certain quantifiers before the noun [Li and Thompson 1981] . Furthermore, Li and Thompson [1981] assert that any measure word can be a classifier, so the combination of demonstrative and/or number or quantifier plus a classifier or a measure is defined as a classifier phrase or a measure phrase. For example, san ge in san ge ren 'three people' (\u4e09\u500b\u4eba), zhe zhan in zhe zhan deng 'this lamp' (\u9019\u76de\u71c8), ji jian in ji jian yifu 'a few / how many garments' (\u5e7e\u4ef6\u8863 \u670d), liu li in liu li lu 'six miles of road' (\uf9d1\uf9e9\uf937), na jin in na jin yangrou 'that tael of lamb' (\u90a3\u65a4\u7f8a\u8089) and ji gang in ji gang cu 'a few / how many vats of vinegar' (\u5e7e\u7f38\u918b) are classifier phrases / measure phrases, which are regarded as Determinative-Measure (DM) compounds in Chao [1968] . A determinative (D) and a measure normally make a compound with unlimited versatility and form a transient word of no lexical import [Chao 1968] . Although the demonstratives, numerals, and measures may be listed exhaustedly, their combination is inexhaustible. It is impossible to list thoroughly all combinations of DMs in dictionaries. Therefore, it requires a representational model to express DM compounds in Chinese NLP.", "cite_spans": [ { "start": 222, "end": 244, "text": "[Li and Thompson 1981]", "ref_id": "BIBREF9" }, { "start": 260, "end": 282, "text": "Li and Thompson [1981]", "ref_id": "BIBREF9" }, { "start": 915, "end": 926, "text": "Chao [1968]", "ref_id": "BIBREF6" }, { "start": 1062, "end": 1073, "text": "[Chao 1968]", "ref_id": "BIBREF6" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "In Chinese, word segmentation, sense, and pos (i.e. part-of-speech) ambiguities commonly occur in certain constructions of DMs or DM-like structures. For examples:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "( The DM compound sange in example (1) modifies moon-cake stores as well as months. The string sangeyuebingpu can be segmented into either sange yuebingpu or sangeyue bingpu, which has word segmentation ambiguity. In example (2), ming functions either as a measure or as a noun. Although example (2) has two meanings and is sense ambiguous, the roles assigned to ciming in (2a) and (2b) are both the goal. In example (3), ershiwunian is a time referent, and it can either be a time point specifying the event-time of the verb or denote the period of time delimitating the time length of the event. In Sinica Treebank, no matter whether ershiwunian behaves as a time point or a time length, ershiwunian is tagged as a unit. When ershiwunian behaves as a time point, its pos is Ndaad and semantic role is time 1 . Furthermore, when ershiwunian behaves as a time length, its pos is DM and semantic role is duration. However, according to CKIP's word segmentation standard of Sinica Corpus, the temporal and locative DM structures are combined together when the meaning of the structure is not obtained from the composition of the components of the structure. Therefore, the temporal DM ershiwunian in (3a) is combined as a unit and tagged as Nd in Sinica Corpus while that in (3b) is segmented into two units, i.e. ershiwu and nian, and tagged as Neu and Nf individually. Therefore, example (3) has pos ambiguity. The different degrees of ambiguities are shown in examples (1) to (3).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "In this paper, we examine and analyze Mandarin Chinese DMs in Sinica Corpus and Sinica Treebank. In section 3, we introduce the regular expression approach to identify DMs and their poses. In section 4, we make a study of the structures and ambiguities of DMs, and then try to analyze and disambiguate these DMs. Section 5 is for implementation and evaluation.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "To deal with DMs, first one must give a proper definition to DMs. Thus, one can delimit the scope of the discussion. There are numerous discussions on determinatives as well as measures, especially on the types of measures. 2 The classification of measures is beyond the scope of this paper. To avoid confusion between classifiers and measures, one must pay attention to the relation between them. Tai [1994] asserts that in the literature on general grammar as well as Chinese grammar, classifiers and measures words are often treated together under one single framework of analysis. Chao [1968] treats classifiers as one kind of measure. In his definition, a measure is a bound morpheme which forms a DM compound with one of the determinatives enumerated above [Chao 1968 ]. Classifiers are defined as 'individual measures', which is one of the nine kinds of measures. As was mentioned in the section of introduction, Chao considers that determinatives are listable and measures are largely listable so D and M can be defined by enumeration, and that DM compounds have unlimited versatility. However, Li and Thompson [1981] blend classifiers with measures. They conclude that, not only does a measure word generally not take a classifier, but also any measure word can be a classifier. In Tai's opinion [1944] , in order to better understand the nature of categorization in a classifier system, it is not only desirable but also necessary to differentiate classifiers from measure words. These studies on the distinction between classifiers and measures are not very clear-cut. In this paper, we discuss ambiguities of DMs in NLP as well as adopt the CKIP DM rules and symbols of poses, and therefore inherit the definition of determinative-measure compounds (DMs) in Mo et al. [1991] . Mo et al. define a DM as the composition of one or more determinatives together with an optional measure. The definition of Mo et al. is used to apply to NLP and somewhat different from traditional linguistics definition.", "cite_spans": [ { "start": 224, "end": 225, "text": "2", "ref_id": null }, { "start": 398, "end": 408, "text": "Tai [1994]", "ref_id": "BIBREF14" }, { "start": 585, "end": 596, "text": "Chao [1968]", "ref_id": "BIBREF6" }, { "start": 763, "end": 773, "text": "[Chao 1968", "ref_id": "BIBREF6" }, { "start": 1103, "end": 1125, "text": "Li and Thompson [1981]", "ref_id": "BIBREF9" }, { "start": 1291, "end": 1311, "text": "Tai's opinion [1944]", "ref_id": null }, { "start": 1770, "end": 1786, "text": "Mo et al. [1991]", "ref_id": "BIBREF11" } ], "ref_spans": [], "eq_spans": [], "section": "Literature Review", "sec_num": "2." }, { "text": "As for ambiguity, Crystal [1991] specifies that the general sense of ambiguity is a word or sentence which expresses more than one meaning. The most widely discussed type of ambiguity in recent years has been grammatical (or structural) ambiguity. In the structure new houses and shops, it could be analysed either as new [houses and shops] (i.e. both are new) or [new houses] and shops (i.e. only the houses are new). Furthermore, according to Crystal's assertion, ambiguity which does not arise from the grammatical analysis of a sentence, but is due solely to the alternative meanings of an individual lexical item, is referred to as lexical ambiguity, e.g. I found the table fascinating (= 'object of furniture' or 'table of figures'). Moreover, the definition of structural and lexical ambiguities can be referred to Prins [2005] . Prins mentions if one restricts his or her attention to the syntax in texts, then one may focus on ambiguity in two forms. The first is lexical ambiguity, the second is structural ambiguity. Lexical ambiguity arises when one word can have several meanings. Structural ambiguity arises when parts of a sentence can be syntactically combined in more than one way. Prins believes humans can resolve most ambiguity, of both types, without even being consciously aware of the alternatives. The remaining ambiguity, of which we are aware, is resolved when knowledge about the world is used in combination with what is known about the linguistic context of the ambiguity to arrive at the most likely analysis. Jurafsky and Martin [2000] define ambiguity as a sentence or words which can have more than one parse. Deciding which category a word belongs to can be solved by part-of-speech tagging. Deciding what sense a word has can be solved by word sense disambiguation. Resolutions of part-of-speech and word sense ambiguities are two important kinds of lexical disambiguation [Jurafsky and Martin 2000] . Furthermore, structural ambiguity occurs when the grammar assigns more than one possible parse to a sentence. Structural disambiguation / syntactic disambiguation can be addressed by probabilistic parsing. Three particularly common kinds of structural ambiguity are attachment ambiguity, coordination ambiguity, and noun-phrase bracketing ambiguity [Jurafsky and Martin 2000] . In the following analysis, we find that Prins' division of ambiguity into structural ambiguity and lexical ambiguity is not enough to deal with ambiguities in NLP. One must apply Jurafsky and Martin's classification to obtain more detailed discussion on ambiguities of DMs in NLP. Word segmentation ambiguity caused by different segmentation of words is a kind of structural ambiguity. With the same word segmentation, that string of words may still be ambiguous because the string may either have more than one meaning or have different parts of speech, semantic roles and functions. Therefore, weone have sense ambiguity and pos ambiguity, which are the two subtypes of lexical ambiguities of DMs.", "cite_spans": [ { "start": 26, "end": 32, "text": "[1991]", "ref_id": null }, { "start": 822, "end": 834, "text": "Prins [2005]", "ref_id": "BIBREF12" }, { "start": 1540, "end": 1566, "text": "Jurafsky and Martin [2000]", "ref_id": "BIBREF8" }, { "start": 1908, "end": 1934, "text": "[Jurafsky and Martin 2000]", "ref_id": "BIBREF8" }, { "start": 2286, "end": 2312, "text": "[Jurafsky and Martin 2000]", "ref_id": "BIBREF8" } ], "ref_spans": [], "eq_spans": [], "section": "Literature Review", "sec_num": "2." }, { "text": "In this section, we introduce the regular expression approach to identify different types of DMs, their representational rules and their poses. Since this paper focuses on the DM defined in Mo et al., which is the composition of one or more determinatives together with an optional measure, the DM structure includes prototypical DMs, variant forms of DMs (e.g. the ellipsis of the determinative and the insertion of an adjective into a DM 3 ), reduplicative forms of DMs (e.g. the reduplication of 'M', 'DM' or 'AM'), and forms of the numeral yi preceding the reduplicative measures (e.g. 'yiMM' (\u4e00 MM) and 'yiAM' (\u4e00 AM)). Due to the infinite of the number of possible DMs, Mo et al. [1991] proposed identification of DMs by regular expression before parsing as part of their morphological module in NLP. For example, when the DM compound is the composition of one determinative, e.g. numerals in (4), rules (5a), (5b) or (5c) will first apply, and then rules (5d), 5(e) or (5f) will apply to compose complex numeral structures, and finally rules (5g) will apply to generate the pos Neu of numeral structures. From the processes of regular expression, the numerals 534 and 319 in (4) are auto-tagged as Neu. Regular expression approach also applies in dealing with ordinal numbers, decimals, and fractional numbers. The ordinal number diyi in (6) applies rules (9) and (5g) so that it is regarded as a unit and tagged as Neu. Rules (10) and (5g) apply to decimals such as sandianyi in (7). Therefore, decimals are viewed as one unit and tagged as Neu. Depending upon the forms of fractional numbers, rules (11a) or (11b) apply to fractional numbers like sanfenzhiyi in (8) and are treated as single units. Then, after application of rules (11a) or (11b), rule (11c) applies to the fractional numbers. Thus the fractional numbers are tagged as Neqa. DM rules will generate/identify ambiguous DM compounds, e.g. (12). If nian and ban are regarded as measures, wunianyiban is segmented into two DMs, i.e. wunian 'five years' and yiban 'one run', like (12a). Therefore, when identifying DMs treated as classes, one has to apply the resolution principles and two DM rules (13a) and (13b). Then, classes in schools such as (12b) are viewed as a single unit, and the noun phrase wunianyiban is restricted to be a unit with the pos Ncb. The word segmentation algorithm will reduce semantic anomalies resulting from possible parses of sentences.", "cite_spans": [ { "start": 190, "end": 200, "text": "Mo et al.,", "ref_id": null }, { "start": 675, "end": 691, "text": "Mo et al. [1991]", "ref_id": "BIBREF11" } ], "ref_spans": [], "eq_spans": [], "section": "Regular Expression Approach for Identifying DMs", "sec_num": "3." }, { "text": "When dealing with addresses, especially indicating the floor, number, alley, lane, section and neighbourhood, we also adopt a regular expression approach for identification. The following instances show the same forms with different segmentation between DMs and addresses. ", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Regular Expression Approach for Identifying DMs", "sec_num": "3." }, { "text": "(16) a. NC1 -> IN1 {\u9130,\u5df7,\uf943,\u6bb5,\u865f,\uf94c}; b. NC2 -> IN1 {\uf94c,\u865f} {\u4e4b,-} IN1 ; c. NC3 -> IN1 {\u4e4b,-} IN1 {\u865f} ;", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Regular Expression Approach for Identifying DMs", "sec_num": "3." }, { "text": "Normally, DMs such as sanlouzhiyi and yizhiyilingerhao are segmented into several units. The former is segmented into three units, i.e. sanlou 'the third floor', zhi 'DE' and yi 'one' while the latter is segmented into three units, i.e. yi 'one', zhi 'DE' and yilingerhao 'No. 102'. However, according to CKIP Technical Report 96-01 (1996: 50), the determinative measure structures expressing time and location will be combined together as a unit. The reason why the locative DMs are combined is because the first joint principle of segmentation stipulates that when the meaning of a string of words is not obtained from the composition of these components, this string should be segmented as a unit. Consequently, in 14, DM rule (16a) applies to sanduan and qihao, and DM rule (16b) applies to sanlouzhiyi. In 15, yizhiyilingerhao applies DM rule (16c). Then sanduan, qihao, sanlouzhiyi and yizhiyilingerhao are all processed by the application of DM rule (13b). Thus sanduan, qihao and sanlouzhiyi in (14) and yizhiyilingerhao in (15) are all segmented as a single unit and tagged as Ncb, not DM, in Sinica Treebank.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Regular Expression Approach for Identifying DMs", "sec_num": "3." }, { "text": "To deal with the reduplicative measures, e.g. 'yiMM' (\u4e00 MM) and 'MM' (MM), we also adopt a context-sensitive regular expression to identify them. For example, zhongzhong in (17) and yizhangzhang in (18) are regarded as a single unit. First, the context-sensitive DM rule (19) is applied, where two measure words in MM are restricted to be equal, and then rule (11c), so the form of reduplicative measures with a preceding optional yi is tagged as Neqa. From the examples and rules illustrated above, one knows that the regular expression approach helps people identify certain DMs. However, DMs still have ambiguities.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Regular Expression Approach for Identifying DMs", "sec_num": "3." }, { "text": "The adoption of DMs rules really improves the recall of recognition, but one still has to resolve segmentation, sense, and pos ambiguities of DMs as shown in the preceding examples in Section 1.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Ambiguities of DMs", "sec_num": "4." }, { "text": "There are two types of word segmentation ambiguities, i.e. covering ambiguities and overlapping ambiguities. Covering ambiguities are always associated with sense ambiguities, since different segmentations result in different sense interpretations. Goh et al. [2005] mention that the covering ambiguity is defined as follows: For a string w = xy, x \u2208 W, y \u2208 W, and w \u2208 W. As almost any single character in Chinese can be considered a word, the above definition reflects only those cases where both word boundaries .../xy/... and .../x/y/... can be found in sentences. Example 20is ambiguous in its meaning and has two different segmentations.", "cite_spans": [ { "start": 249, "end": 266, "text": "Goh et al. [2005]", "ref_id": "BIBREF7" } ], "ref_spans": [], "eq_spans": [], "section": "Word Segmentation Ambiguities of DMs", "sec_num": "4.1" }, { "text": "When fu functions as a verb, fuyao is segmented as a unit, and the meaning of (20) is (20a).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Word Segmentation Ambiguities of DMs", "sec_num": "4.1" }, { "text": "When fu functions as a measure, yifu is tagged as DM, and the meaning of (20) is (20b).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Word Segmentation Ambiguities of DMs", "sec_num": "4.1" }, { "text": "Because the combinations of determinatives and measures are countless, DMs such as yifu won't be listed in the CKIP dictionary. Different word segmentation will bring structural ambiguities forth. Another segmentation ambiguity exists in the ellipsis of the determinatives. In (21), kou is sense ambiguous in that it functions as a noun or as a measure. When kou behaves as a noun in (21a), the sentence is semantically anomalous and syntactically ungrammatical. Only when kou functions as a measure, will the sentence (21b) be well-processed. The first group dichotomized into three parts. Goh et al. [2005] state that overlapping the ambiguity is defined as follows: For a string w = xyz, both w1 = xy \u2208 W and w2 = yz \u2208 W hold. Mo et al. [1991] list a resolution principle to reduce word segmentation ambiguity. The principle asserts if ambiguous word breaks occur between the words in the lexicon and the DMs, the words in the lexicon should have higher priority to get the shared characters. Therefore, (22a), (23a) and (24a) have higher priority to (22b), (23b) and (24b), individually.", "cite_spans": [ { "start": 591, "end": 608, "text": "Goh et al. [2005]", "ref_id": "BIBREF7" }, { "start": 730, "end": 746, "text": "Mo et al. [1991]", "ref_id": "BIBREF11" } ], "ref_spans": [], "eq_spans": [], "section": "Word Segmentation Ambiguities of DMs", "sec_num": "4.1" }, { "text": "According to the above discussions, to resolve word segmentation ambiguities, we propose the following resolution principles which were implemented in the word segmentation system of Ma and Chen [2003] . a) DM compounds are expressed and matched by regular expressions. b) Lexical words have higher precedence than DM compounds (cf. 22, 23, 24) . c) Longest matching principle (i.e. 9, 10, 11, 13, 16, 19): a long DM has higher precedence than a short DM (cf. 6, 7, 8, 12 14, 15, 17, 18) . The word segmentation ambiguity is caused by different possible segmentations. Although the above examples have ambiguities, after the application of the resolution principles, the ambiguous segmentation is resolved and the correct segmentation has higher priority.", "cite_spans": [ { "start": 183, "end": 201, "text": "Ma and Chen [2003]", "ref_id": "BIBREF10" }, { "start": 328, "end": 344, "text": "(cf. 22, 23, 24)", "ref_id": null }, { "start": 452, "end": 462, "text": "DM (cf. 6,", "ref_id": null }, { "start": 463, "end": 465, "text": "7,", "ref_id": null }, { "start": 466, "end": 468, "text": "8,", "ref_id": null }, { "start": 469, "end": 475, "text": "12 14,", "ref_id": null }, { "start": 476, "end": 479, "text": "15,", "ref_id": null }, { "start": 480, "end": 483, "text": "17,", "ref_id": null }, { "start": 484, "end": 487, "text": "18)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Word Segmentation Ambiguities of DMs", "sec_num": "4.1" }, { "text": "Senses and semantic functions of DMs are sometimes related to the types of measures. The sense of certain types of DMs can be identified by types of measures. However, as usual, some DMs have ambiguous senses. Their ambiguity resolution is almost equivalent to word sense disambiguation. Therefore, context sensitive rules and collocation bi-grams are suitable information for resolving sense ambiguities. Methods for word sense disambiguation are also applicable for DMs.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Sense Ambiguities of DMs", "sec_num": "4.2" }, { "text": "As mentioned in Section 3, we have adopted a regular expression approach for identifying structures denoting addresses. For example, DM rule (16a) is applied to segment yiduan, bahao, and yilou in (25) as single units, and each of them is tagged as Ncb. However, in (26), hao and duan are both measures specifying the fixed amount or quantity of the road, so yiyiqihao and yiduan are both tagged as DM. Examples 25and 26have sense ambiguities during the processes of DM recognition, so they will be automatically segmented as (25a) and (25b) as well as (26a) and (26b), individually. According to CKIP Technical Report 96-01 (1996) , locative DMs are combined so that one can disambiguate (25) and (26). Then, the correct segmentations (25b) and (26b) are resumed according to their contexts. belong to a part of the 117th road Similar to dealing with addresses, sense ambiguities occur when one refers to percentages. One can adopt two forms to represent percentage, i.e. Chinese characters in (8) and mathematical symbols in (27). Regular expression approach can help one identify word segmentation ambiguity existing in (8). The form of the mathematical symbols has sense ambiguity, which can refer to either percentage like (27) or a time point like (28). Without any context, the symbol \"\uff11\uff10\uff0f\uff12\uff11\" can be identified either as a fractional number and read as \"ershiyifenzhishi\" (\u4e8c\u5341\u4e00\u5206\u4e4b\u5341 'ten over twenty one') with the pos Neqa or as a time point and read as \"shiyueershiyiri\" (10 \u6708 21 \u65e5 'Oct. 21') with the pos Ndabd whose semantic role is time. Besides, the form of mathematical symbols is also used to refer to another kind of time point, such as (29). To reduce such kinds of sense ambiguities caused by mathematical symbols, DM rules (30a), (30b), (31a) and (31b) exist. DM rules (30a) and (30b) apply to the forms of mathematic symbols like (27) tagged as Neqa (numbers), while DM rules (31a) and (31b) apply to forms like (29) tagged as Nd (time point). Although DM rules (30) and 31help disambiguate sense ambiguities between forms of mathematic symbols denoting percentage and time points such as (27) and 29, one still has to have context to make (27) and 28 In the autumn in the year of thirty, the anti-smuggling team is rectified to the tax policemen team. Examples (32) and (33) have the same temporal phrase sanshinian, but their senses, functions and roles are different. The temporal phrase in (32) denotes a time length and is tagged as DM. The semantic role of sanshinian in (32) is duration. However, sanshinian in (33) indicates a time point and is tagged as Ndaad, whose semantic role is time. Even though example (33) omits either a Chinese reign title or Mingguo (\u6c11\u570b) preceding sanshinian, we still know sanshinian is a specific time, not a period, from the context. When the measures nian 'year' (\uf98e) and ri 'day' (\u65e5) are preceded by numerals, the temporal phrases always have sense ambiguities. We have two tricks to differentiate between time points and time lengths. If DMs denote time points, they are usually preceded by key words of Mingguo, the Christian era like Gongyuan (\u516c\u5143) and Xiyuan (\u897f\u5143), or a Chinese reign title such as Guanxu (\u5149\u7dd2), Qianlong (\u4e7e\uf9dc), Tianbao (\u5929\u5bf6), Jiajing (\u5609\u9756) and so on. When DMs are preceded by these key words, they are tagged as Ndaad (i.e. a time point). Another trick that helps one to recognize DMs as time points is their neighbouring temporal phrases. Time points, not time lengths, are usually combined with two or three co-occurring temporal phrases, e.g. erlinglingwunian liuyue (2005 \uf98e 6 \u6708 'June 2005'), liuyue sanshiri (6 \u6708 30 \u65e5 'June 30'), erlinglingwunian liuyue sanshiri (2005 \uf98e 6 \u6708 30 \u65e5 'June 30, 2005'), etc. The two tricks mentioned above and context will help one reduce some ambiguities of phrases and mathematical symbols specifying time.", "cite_spans": [ { "start": 619, "end": 631, "text": "96-01 (1996)", "ref_id": null }, { "start": 2157, "end": 2161, "text": "(27)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Sense Ambiguities of DMs", "sec_num": "4.2" }, { "text": "In conclusion, sense ambiguity resolution of DMs is almost equivalent to word sense disambiguation. Therefore, context sensitive rules and collocation bi-grams are suitable information for resolving sense ambiguities. Methods for word sense disambiguation are also applicable here.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Sense Ambiguities of DMs", "sec_num": "4.2" }, { "text": "Here, we first discuss ambiguities about temporal adverbs to illustrate pos ambiguities and possible resolution methods. The temporal phrases in (34) and (35) are regarded either as time points or as time lengths. These temporal phrases have the same strings and word segmentation but have different parts of speech, semantic roles and functions. As is known, in Chinese a folktale, a woman called Wang Baochuan (\u738b\u5bf6\u91e7) went through 18 years of hardship for her husband's turning back home. In Mainland China, the Tiananmen Square massacre occurred in 1989. Therefore, the semantic roles of the temporal phrases shibanian in (34) and bajiunian in (35) will be labelled as duration and time individually. The pos of the former is DM while that of the latter is Ndaad. The reason for making different assignments of semantic roles is concerned with logical interpretation of sense collocations according to common sense and the real world knowledge. When detecting DMs in Sinica Corpus and Sinica Treebank, one finds some interesting examples. The verb phrases (36) and (37) have the same lexical items except for their linear word order. The pos of the DM structure in (36) is DM whose semantic role is duration while that in (37) is Ndaad whose semantic role is time. It seems that different positions of temporal DMs will affect the meanings of sentences. Therefore, we briefly reviewed the data in Sinica Treebank. The totality of the semantic role time of NPs and of PPs following a verb is close to that of the semantic role duration. But the totality of the semantic role time of NPs and of PPs preceding a verb is much more than that of duration. The statistics indicate that temporal DMs preceding verbs mostly function as time. Another problem with assignment of semantic role to a similar structure is illustrated by (38) and (39). The DM structure in (38) is tagged as DM and assigned the semantic role duration while, in (39), it is tagged as Ndaad and assigned the semantic role time. This kind of pos ambiguity has relation to situation types. The situation type of fuxing in (38) is an Activity while that of panxing in (39) is an Achievement. The feature [\u00b1Durative] 4 of the events causes differences. As for the pos ambiguity of yixia, yixia in (40) means 'for a while', which is tagged as Nddc and assigned the role of duration. However, yixia in (41) means 'once', which is tagged as DM and assigned the role of frequency. Nevertheless, yixia in (42) is POS ambiguous and has two senses. One is tagged as Nddc and means 'for a while' while another is tagged as DM and means 'once'. The former is labelled as duration and the latter is assigned as frequency. Equal to the cause of differences between (38) and (39), the ambiguities in (42) are due to situation types. Semantic role assignment is not an easy task, since it requires world knowledge as well as linguistic knowledge. In You and Chen [2004] , they identify parameters of determining semantic roles and propose an instance-based approach to resolve ambiguities. They adopt dependency decision making and example-based approaches. Semantic roles are determined by four parameters, including syntactic and semantic categories of the target word, case markers, phrasal head, and sub-categorization frame and its syntactic patterns. The refinements of features extraction, canonical representation for certain classes of words and dependency decisions improve role assignment. To assign the semantic roles of DMs, the above parameters are further refined as the features of relative positions and situation types.", "cite_spans": [ { "start": 2900, "end": 2919, "text": "You and Chen [2004]", "ref_id": "BIBREF15" } ], "ref_spans": [], "eq_spans": [], "section": "POS Ambiguities of DMs", "sec_num": "4.3" }, { "text": "The examples above show that ambiguities are unavoidable when one deals with DMs. In addition to the typical DMs, some related structures like reduplicative DMs, numerals, the ellipsis of measures, etc. are also topics for discussion. During DM processing, certain DMs are ambiguous to automatic identification in word segmentations, senses as well as poses. Here, yi dian (\u4e00\u9ede) is taken as an example. The phrase yidian functions as a pronoun and tagged as DM in (43), functions as a quantitative determinative modifying xinyi and tagged as Neqa in (44), functions as a time noun and tagged as Ndabe in (45), and functions as a post-verb adverb of degree and tagged as Dfb in (46). While in (47), yidian has sense ambiguities depending upon context. In addition, yidian in (47a) and (48) is pos ambiguous. For another example, qi (\u8d77) functions as a measure in both siqi anjian (\u56db\u8d77\u6848\u4ef6) and yiqi mingan (\u4e00\u8d77\u547d\u6848). In fongyun siqi (\u98a8\u96f2\u56db\u8d77) and yiqi sikao (\u4e00\u8d77\u601d\u8003), siqi and yiqi are tagged as VA11 and Dh individually. However, in yiqi fanan (\u4e00\u8d77\u72af\u6848) and fanan siqi (\u72af\u6848\u56db\u8d77), the DMs yiqi and siqi are ambiguous. It is obvious that the ambiguities of DMs are complex and that a DM compound can have more than one classification of ambiguities.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "POS Ambiguities of DMs", "sec_num": "4.3" }, { "text": "No matter whether the ambiguity is from word segmentation, sense, or pos, the prescription of resolution principles and DM rules are helpful in disambiguating DMs. Besides, the neighbouring morphemes and context are other tricks in reducing ambiguities. Furthermore, pos ambiguities are concerned with common sense, and the resolution features also include positions of temporal DMs and situation types. Such ambiguities have to be reduced by the application of parameters of context vector models.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "POS Ambiguities of DMs", "sec_num": "4.3" }, { "text": "We randomly chose 2035 sentences (11697 word tokens) from Sinica Treebank as our development set. In total, 545 tokens of the development data are processed by the revised DM rules (as shown in the appendix). Among the 545 tokens, 504 tokens are correctly segmented, and 443 tokens are correctly pos tagged. The segmentation accuracy of the development data is 92.5%, the tag accuracy of the development data is 81.3%, and the tag accuracy with the correct segmentation of the development data is 88.0%. Contrastively, the segmentation accuracy and tag accuracy of the development set processed by the original DM rules are both lower than those applied the revised DM rules. The segmentation accuracy is 84.2%, and the tag accuracy is 71.0%. Then, to test the accuracy of the revised DM rules, we randomly chose 2111 sentences (12209 word tokens), which have no overlap with the development set, from Sinica Treebank as the testing set. In total 564 tokens of the testing data were processed using the revised DM rules. Among those 564 tokens, 508 tokens were correctly segmented, and 424 tokens were correctly pos tagged. By application of the revised DM rules, the segmentation accuracy of the testing data is 90.1%, the tag accuracy of the testing data is 75.2%, and the tag accuracy with the correct segmentation of the testing data is 83.5%. Contrastively, processed by the original DM rules, the segmentation accuracy and the tag accuracy of the testing data is 77.8% and 60.3% individually. Table 1 is the evaluation result. Table 1 shows that both segmentation accuracy and tag accuracy of development set and of testing set processed by the revised DM rules are higher than those processed by the original DM rules, although the segmentation accuracy, tag accuracy and tag accuracy with correct segmentation of the testing set are a little bit lower than those of the development set.", "cite_spans": [], "ref_spans": [ { "start": 1499, "end": 1506, "text": "Table 1", "ref_id": "TABREF9" }, { "start": 1533, "end": 1540, "text": "Table 1", "ref_id": "TABREF9" } ], "eq_spans": [], "section": "Implementation and Evaluation", "sec_num": "5." }, { "text": "After data analysis, we found that there were several reasons for inaccurate results. The most crucial factor resulting in inaccuracy is ambiguity, including word segmentation ambiguities, sense ambiguities and pos ambiguities. The word segmentation ambiguities, sense ambiguities, and pos ambiguities caused 11%, 32.4% and 27.2% of errors, respectively. For example, (49a) and (50a) are correct sentences. The contrastive sentences (49b) and 50bhave errors in word segmentation ambiguities. Sentences (51b) and (52b) have errors in sense ambiguities, which are contrastive to the correct ones (51a) and (52a). Sentences 53b Other than errors in ambiguities, there are four kinds of errors bringing about inaccuracy. The first kind of errors is a result of the segmentation model (a Hidden Markov Model). In HMM, there are several possible paths, including the correct one. However, the result chosen was not the correct one. For example, (55a) and (55b) are the possible paths in HMM. For the result, the inaccurate one (55b) was chosen. The error of (56b) is also due to HMM. The percentage of errors made by HMM is 10.3%. The second kind of errors is because different contexts cause different tagging. In Sinica Treebank, one or more determinatives together with an optional measure will constitute a DM. However, certain determinatives and measures are tagged differently than usual because of the context. In (57a), liangsanmiaozhong (\uf978\u4e09\u79d2\u9418 'two and three seconds') is composed of liangmiaozhong 'two seconds' and sanmiaozhong 'three seconds'. The measure miaozhong 'second' is shared by determinatives liang 'two' and san 'three'. What is in (57c) is the tree structure of (57a). Both liang and san are tagged as Neu, and miaozhong is tagged as Nf. In (58a), qibayue (\u4e03\u516b\u6708 'July and August') is composed of qiyue 'July' and bayue 'August'. The diagram (58c) is the bracketed tree diagram of (58a). From the context, one knows that qi in (58a) is not the numeral 'seven' but a month 'July' so qi is tagged as Nd not Neu. The The third kind of errors occurs when there is only one measure without any other determinatives, e.g. (59b) and (60b). The percentage of errors in one measure is 8.1%. The error in wrong tagging of one measure is because of the training data from Sinica Corpus. A sole measure is tagged as Nf in Sinica Corpus, but in Sinica Treebank it is viewed as a DM structure and tagged as DM. Therefore, a sole measure always has incorrect pos. This kind of error has to be dealt with during postprocessing. In our implementation and evaluation, the accuracy of segmentation, of tag, and of tag with correct segmentation of the development set processed by the revised DM rules are higher than those processed by the original DM rules. Through application of the revised DM rules, the segmentation accuracy, tag accuracy and tag accuracy with correct segmentation of the testing set are a little bit lower than those of the development set. The percentage of ambiguities causing inaccuracy is 70.6% while the total percentage of other factors is 29.4%.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Implementation and Evaluation", "sec_num": "5." }, { "text": "The high proportion of ambiguity shows that, although a regular expression approach was used in applying DM rules to deal with DMs, eventually, ambiguity is the most crucial issue one must confront. Therefore, the application of resolution principles, of DM rules, of context sensitive rules, of collocation bi-grams and of parameters of context vector models are necessary to help one disambiguate. Language reflects the human view of the world. Differing personal world knowledge may result in different explanations of sentences. Some reduction of ambiguities of DMs depends upon the human's common sense knowledge.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Implementation and Evaluation", "sec_num": "5." }, { "text": "DMs are not a closed set, so one has to apply DM rules during the process of automatic identification of DMs. By observing Sinica Treebank, we had developed a set of regular expression rules to identify DMs and their parts of speech. Thus, all DM candidates can be matched and classified by regular expression rules. However, due to segmentation, pos and sense ambiguities of DMs, DM rules are necessary complements to dictionaries and helpful to resolve ambiguities by applying resolution principles and segmentation models. Sense and pos ambiguities are also expected to be resolved by different approaches during post-processing by applying context sensitive rules, collocation bi-grams and parameters of context vector models. ", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusion", "sec_num": "6." }, { "text": "All the symbols such as Ndaad will be defined in the appendix of this paper.2 Chao [1968] andLi and Thompson [1981] detect measures and classifiers. He [2000] traces the diachronic names of measures and mentions related literature on measures. The dictionary of measures pressed by Mandarin Daily News Association and CKIP [1997] lists all the possible measures in Mandarin Chinese.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "The insertion of an adjective into a DM has the form of 'yiAM'.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null } ], "back_matter": [], "bib_entries": { "BIBREF0": { "ref_id": "b0", "title": "\u6280\u8853\u5831 \u544a 96-01, \u4e2d\u592e\u7814\u7a76\u9662,\u53f0\uf963", "authors": [ { "first": "\"", "middle": [], "last": "\u4e2d\u592e\u7814\u7a76\u9662\u8a5e\u5eab\u5c0f\u7d44(ckip)", "suffix": "" }, { "first": "", "middle": [], "last": "\u300c\u641c\u300d\u6587\u89e3\u5b57\u2500\u4e2d\u6587\u8a5e\u754c\u7814\u7a76\u8207\u8cc7\u8a0a\u7528\u5206\u8a5e\u6a19\u6e96", "suffix": "" } ], "year": 1996, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "\u4e2d\u592e\u7814\u7a76\u9662\u8a5e\u5eab\u5c0f\u7d44(CKIP), \"\u300c\u641c\u300d\u6587\u89e3\u5b57\u2500\u4e2d\u6587\u8a5e\u754c\u7814\u7a76\u8207\u8cc7\u8a0a\u7528\u5206\u8a5e\u6a19\u6e96,\" \u6280\u8853\u5831 \u544a 96-01, \u4e2d\u592e\u7814\u7a76\u9662,\u53f0\uf963, 1996.", "links": null }, "BIBREF3": { "ref_id": "b3", "title": "CKIP word segmentation system", "authors": [ { "first": "Academia", "middle": [], "last": "Sinica", "suffix": "" } ], "year": null, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Academia Sinica, CKIP word segmentation system, http://ckipsvr.iis.sinica.edu.tw/.", "links": null }, "BIBREF4": { "ref_id": "b4", "title": "Sinica Corpus", "authors": [ { "first": "Academia", "middle": [], "last": "Sinica", "suffix": "" } ], "year": 2001, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Academia Sinica, Sinica Corpus, version 4.0, http://www.sinica.edu.tw/SinicaCorpus/, 2001.", "links": null }, "BIBREF6": { "ref_id": "b6", "title": "A Dictionary of Linguistics and Phonetics", "authors": [ { "first": "Y.-R", "middle": [], "last": "Chao", "suffix": "" }, { "first": "D", "middle": [], "last": "Crystal", "suffix": "" } ], "year": 1968, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Chao, Y.-R., A Grammar of Spoken Chinese, University of California Press, Berkeley, 1968. Crystal, D., A Dictionary of Linguistics and Phonetics, Blackwell, Massachusetts, Cambridge, 1991.", "links": null }, "BIBREF7": { "ref_id": "b7", "title": "Chinese Word Segmentation by Classification of Characters", "authors": [ { "first": "C.-L", "middle": [], "last": "Goh", "suffix": "" }, { "first": "M", "middle": [], "last": "Asahara", "suffix": "" }, { "first": "Y", "middle": [], "last": "Matsumoto", "suffix": "" } ], "year": 2005, "venue": "Computational Linguistics and Chinese Language Processing", "volume": "10", "issue": "3", "pages": "381--96", "other_ids": {}, "num": null, "urls": [], "raw_text": "Goh, C.-L., M. Asahara and Y. Matsumoto, \"Chinese Word Segmentation by Classification of Characters,\" Computational Linguistics and Chinese Language Processing, 10(3), 2005, pp. 381-96.", "links": null }, "BIBREF8": { "ref_id": "b8", "title": "Speech and Language Processing: An Introduction to natural Language Processing, Computational Linguistics, and Speech Recognition", "authors": [ { "first": "D", "middle": [], "last": "Jurafsky", "suffix": "" }, { "first": "J", "middle": [ "H" ], "last": "Martin", "suffix": "" } ], "year": 2000, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Jurafsky, D. and J. H. Martin, Speech and Language Processing: An Introduction to natural Language Processing, Computational Linguistics, and Speech Recognition, Prentice Hall, Upper Saddle River, N.J., 2000.", "links": null }, "BIBREF9": { "ref_id": "b9", "title": "Mandarin Chinese: A Functional Reference Grammar", "authors": [ { "first": "C", "middle": [ "N" ], "last": "Li", "suffix": "" }, { "first": "S", "middle": [ "A" ], "last": "Thompson", "suffix": "" } ], "year": 1981, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Li, C. N. and S. A. Thompson, Mandarin Chinese: A Functional Reference Grammar, University of California Press, Berkeley, 1981.", "links": null }, "BIBREF10": { "ref_id": "b10", "title": "Introduction to CKIP Chinese Word Segmentation System for the First International Chinese Word Segmentation Bakeoff", "authors": [ { "first": "W.-Y", "middle": [], "last": "Ma", "suffix": "" }, { "first": "K.-J", "middle": [], "last": "Chen", "suffix": "" } ], "year": 2003, "venue": "Proceedings of ACL, Second SIGHAN Workshop on Chinese Language Processing", "volume": "", "issue": "", "pages": "168--171", "other_ids": {}, "num": null, "urls": [], "raw_text": "Ma, W.-Y. and K.-J. Chen, \"Introduction to CKIP Chinese Word Segmentation System for the First International Chinese Word Segmentation Bakeoff,\" In Proceedings of ACL, Second SIGHAN Workshop on Chinese Language Processing, 2003, Sapporo, Japan, pp.168-171.", "links": null }, "BIBREF11": { "ref_id": "b11", "title": "Determinative-Measure Compounds in Mandarin Chinese: Their Formation Rules and Parser Implementation", "authors": [ { "first": "R.-P", "middle": [ "J" ], "last": "Mo", "suffix": "" }, { "first": "Y.-J", "middle": [], "last": "Yang", "suffix": "" }, { "first": "K.-J", "middle": [], "last": "Chen", "suffix": "" }, { "first": "C.-R", "middle": [], "last": "Huang", "suffix": "" } ], "year": 1991, "venue": "Proceedings of ROCLING IV (R.O.C. Computational linguistics Conference)", "volume": "", "issue": "", "pages": "111--134", "other_ids": {}, "num": null, "urls": [], "raw_text": "Mo, R.-P. J., Y.-J. Yang, K.-J. Chen and C.-R. Huang, \"Determinative-Measure Compounds in Mandarin Chinese: Their Formation Rules and Parser Implementation,\" In Proceedings of ROCLING IV (R.O.C. Computational linguistics Conference), 1991, National Chiao-Tung University, Hsinchu, Taiwan, pp. 111-134.", "links": null }, "BIBREF12": { "ref_id": "b12", "title": "Finite-State Pre-Processing for Natural Language Analysis", "authors": [ { "first": "R", "middle": [ "P" ], "last": "Prins", "suffix": "" } ], "year": 2005, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Prins, R. P., Finite-State Pre-Processing for Natural Language Analysis, Art Dissertation, 2005", "links": null }, "BIBREF13": { "ref_id": "b13", "title": "The Parameter of Aspect", "authors": [ { "first": "C", "middle": [ "S" ], "last": "Smith", "suffix": "" } ], "year": 1991, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Smith, C. S., The Parameter of Aspect, Kluwer Academic Publishers, Dordrecht, 1991.", "links": null }, "BIBREF14": { "ref_id": "b14", "title": "Chinese classifier systems and human categorization", "authors": [ { "first": "J", "middle": [], "last": "Tai", "suffix": "" }, { "first": "", "middle": [], "last": "H-Y", "suffix": "" } ], "year": 1994, "venue": "Interdisciplinary Studies on Language and Language Change", "volume": "", "issue": "", "pages": "479--494", "other_ids": {}, "num": null, "urls": [], "raw_text": "Tai, J. H-Y, \"Chinese classifier systems and human categorization,\" In Honor of William S-Y. Wang: Interdisciplinary Studies on Language and Language Change, ed. by M. Y. Chen and O J.-L. Tzeng, Pyramid Press, Taipei, 1994, pp. 479-494.", "links": null }, "BIBREF15": { "ref_id": "b15", "title": "Automatic Semantic Role Assignment for a Tree Structure", "authors": [ { "first": "J.-M", "middle": [], "last": "You", "suffix": "" }, { "first": "K.-J", "middle": [], "last": "Chen", "suffix": "" } ], "year": 2004, "venue": "Proceedings of 3rd ACL SIGHAN Workshop", "volume": "", "issue": "", "pages": "109--115", "other_ids": {}, "num": null, "urls": [], "raw_text": "You, J.-M. and K.-J. Chen, \"Automatic Semantic Role Assignment for a Tree Structure,\" In Proceedings of 3rd ACL SIGHAN Workshop, 2004, Barcelona, Spain, pp. 109-115.", "links": null } }, "ref_entries": { "FIGREF0": { "num": null, "text": "", "uris": null, "type_str": "figure" }, "FIGREF2": { "num": null, "text": "then take effect Every time when he takes medicine, the illness is completely cured. b. yifu yao jiu jianxiao one-M medicine then take effect One dose is effective. (21) \u4ed6\u5011\u5bb6\u56db\u53e3\u4eba *a. tamen jia si kou ren they family four mouth person b. tamen jia sikou ren they family four-M personTheir family has four members.", "uris": null, "type_str": "figure" }, "FIGREF3": { "num": null, "text": "d) Covering ambiguities are resolved by collocation context (cf. 20, 21).", "uris": null, "type_str": "figure" }, "FIGREF5": { "num": null, "text": "and (54b) have errors in pos ambiguities. The total percentage of errors caused by ambiguities is about 70.6%. (49) a. \u6c7d\uf902(Na) \u53ef(D) \uf967(D) \u662f(SHI) \u4e00\u500b(DM) \u4eba(Na) \u6240(automobile should NEG SHI one-M person that which invent DE Automobiles are not invented by one person. *b. \u6c7d\uf902(Na) \u53ef(D) \uf967(D) \u662f(SHI) \u4e00(Neu) \u500b\u4eba(Nh) \u6240(D) \u767c\u660e (VC) \u7684(DE) have Sec. 1 heart-stirring DE story (52) a. \u4f46(Cbb) \u4e5f(D) \u4ed8\u51fa(VC) \uf969\u4f4d(DM) \u5275\u696d(Nv) \u5148\u9032(Na) \u5bf6\u8cb4(VH) \u7684(DE) \u751f\u547d(Na) \u53ca(Caa) \u640d\u5931(VJ) \uff11\uff14\u67b6(DM) \u98db\u6a5f(Na) \u7684(DE) \u6158\u75db(VH) \u4ee3\u50f9(Na) dan ye fuchu shuwei chuangye xianjin baoguei de shengming but also pay several-M pioneering precursor valued DE life ji sunshi shisijia feiji de cantong daijia and lose 14-M airplane DE cruel cost but also pay several pioneering precursors' valued lives and suffer the cruel costs of losing fourteen airplanes *b.\u4f46(Cbb) \u4e5f(D) \u4ed8\u51fa(VC) \uf969\u4f4d(A) \u5275\u696d(VA) \u5148\u9032(VH) \u5bf6\u8cb4(VH) \u7684(DE) \u751f\u547d(Na) \u53ca(Caa) \u640d\u5931(Na) \uff11\uff14\u67b6(DM) \u98db\u6a5f(Na) \u7684 (DE) \u6158\u75db(VH) \u4ee3\u50f9(Na) dan ye fuchu shuwei chuangye xianjin baoguei de shengming but also pay several-M digital precursor valued DE life a. \u5c0d\u65bc(P) \u4e03\u5341\u4e5d\uf98e(Nd) \uf98e\u7d42(Nd) \u734e\uf90a(Na) dueiyu qishijiunian nianzhong jiangjin about the year of 79 year-end bonus about the year-end bonus in the year of 79 *b. \u5c0d\u65bc(P) \u4e03\u5341\u4e5d\uf98e(DM) \uf98e\u7d42(Nd) \u734e\uf90a(Na) dueiyu qishijiunian nianzhong jiangjin about 79-M year-end bonus", "uris": null, "type_str": "figure" }, "FIGREF6": { "num": null, "text": "55) a. \u56de\u60f3(VE) \u8d77(Di) \u4e8c\u5341\uf98e(DM) \u524d(Ng) \u7684(DE) \u5f80\u4e8b(Na) hueixiang qi ershinian qian de wangshi recall ASP 20-M before DE past recall the past twenty years ago *b. \u56de\u60f3\u8d77(VE) \u4e8c\u5341(Neu) \uf98e(Nf) \u524d(Ng) \u7684(DE) \u5f80\u4e8b(Na)", "uris": null, "type_str": "figure" }, "FIGREF7": { "num": null, "text": "percentage of errors resulting from contexts is 8.8%. (57) a. \u9019(Nep) \u59ff\u52e2(Na) \u4fdd\u6301(VJ) \uf978(Neu) \u4e09(Neu) \u79d2\u9418(Nf) zhe zishi baochi liang san miaozhong this pose keep two three M This pose is keeping for two and three seconds.*b. \u9019(Nep) \u59ff\u52e2(Na) \u4fdd\u6301(VJ) \uf978\u4e09\u79d2\u9418(DM) zhe zishi baochi liangsanmiaozhong this pose keep two-three-M c. S(theme:NP(quantifier:Nep:\u9019|Head:Nac:\u59ff\u52e2)|Head:VJ1:\u4fdd\u6301 |duration:DM(Head:Neu(Head:Neu:\uf978|Head:Neu:\u4e09)|Head:Nfg:\u79d2\u9418)) (58) a. \uff17(Nd) \u3001(Caa) \uff18\u6708(Nd) \u70ba(VG) \u4e0b\u5348(Nd) \uff14\u9ede(Nd) \u81f3afternoon four o'clock to four o'clock thirty minutes In July and August, it is from four to four-thirty o'clock. *b. \uff17(Neu) \u3001(PAUSECATEGORY) \uff18\u6708(Nd) \u70ba(P) \u4e0b\u5348(Nd) \uff14\u9ede (DM) \u81f3(Caa) \uff14\u9ede(DM) \uff13\uff10\u5206(DM) qi bayue wei xiawu sidian zhi sidian sanshifen seven August is afternoon four-M to four-M thirty minutes c. S(theme:NP(Head:Ndabc(DUMMY1:Ndabc:\uff17|Head:Caa:\u3001 |DUMMY2:Ndabc:8\u6708))|Head:VG2:\u70ba|range:NP(property:Ndabe:\u4e0b\u5348 |Head:NP(DUMMY1:NP(Head:Ndabe: \uff14\u9ede)|Head:Caa:\u81f3 |DUMMY2:NP(property:Ndabe:\uff14\u9ede|Head:Ndabe:\uff13\uff10\u5206))))", "uris": null, "type_str": "figure" }, "FIGREF8": { "num": null, "text": "59) a. \u6700\u8fd1(Nd) \u807d\u5230(VE) \uf9ba(Di) \u500b(DM) \u99ed\u4eba\u807d\u805e(VH) \u7684(DE) \u6545\u4e8b(Na) zueijin tingdao le ge hairentingwen de gushi recent hear LE M shocking DE story recently hear one shocking story *b. \u6700\u8fd1(Nd) \u807d\u5230(VE) \uf9ba(Di) \u500b(Nf) \u99ed\u4eba\u807d\u805e(VH) \u7684(DE) \u6545\u4e8b(Na) zueijin tingdao le ge hairentingwen de gushi recent hear LE M shocking DE story (60) a. \u6709(V_2) \u9846(DM) \u5584\uf97c(VH) \u7684(DE) \u5fc3(Na) \u6709(V_2) \u9846(Nf) \u5584\uf97c(VH) \u7684(DE) \u5fc3(Na) you ke shanliang de xin have M kindhearted DE heartThe last kind of error is because of unknown word identification such as (61) and (62). The unknown words in (61b) and (62b) are not identified correctly, so errors occur. The percentage of errors in unknown words is 2.2%.(61) a. \u8ab0(Nh) \u8a00(VE) \u5bf8(DM) \u8349(Na) \u5fc3(Na)", "uris": null, "type_str": "figure" }, "FIGREF9": { "num": null, "text": "\uf96d,\u5dde,\u7e23,\u9109,\u6751,\u93ae,\u9130,\uf9e9,\u90e1,\u5340,\u7ad9,\u5df7,\uf943,\u6bb5,\u865f,\u5730,\u8857,\uf94c,\u8856,\u5e02,\u6d32,\u90e8,\u53f8,\u8ab2,\u9662, \u79d1,\u7cfb,\u7d1a,\u80a1,\u5ba4,\u5ef3} ; Ndabe = {\u6e05\u6668,\uf955\u6668,\u65e9\u6668,\u65e9\u4e0a,\u665a\u4e0a,\u4e0a\u5348,\u4e2d\u5348,\u4e0b\u5348,\u6668\u9593,\u5348\u9593,\u665a\u9593,\u534a\u591c,\u5348\u591c,\u6668,\u5348,\u665a, \u508d\u665a,\u6df1\u591c,\u6651\u5348,\u5b50\u6642,\u4e11\u6642,\u5bc5\u6642,\u536f\u6642,\uf971\u6642,\u5df3\u6642,\u5348\u6642,\u672a\u6642,\u7533\u6642,\u9149\u6642,\u620c\u6642,\u4ea5 \u6642} ; ND = {\u5fae\u79d2,\u91d0\u79d2,\u79d2,\u79d2\u9418,\u5206,\u5206\u9418,\u523b,\u523b\u9418,\u9ede,\u9ede\u9418,\u9ede\u591a\u9418,\uf901,\u65ec,\u7d00,\uf9d7,\u5929,\u65e5,\u9031,\u5468,\uf9b6 \u62dc,\u5b63,\uf98e,\uf98e\u4efd,\u8f09,\u865f,\u665a,\u5bbf,\u9031\uf98e,\u5468\uf98e,\u5468\u6b72,\u6703,\u6703\u5152,\u9663,\u4e16,\u8f29,\uf98e\u671f} ; ND2 = {\u6642,\u4e16\u7d00,\uf98e\ufa01,\u6708,\u6708\u4efd,\u9663\u5b50,\u5b78\u671f,\u5b78\uf98e,\uf98e\u4ee3,\u4e0b\u5b50} ; DESC = {\u5927,\u5c0f,\u6574} ; PNM = {\u591a,\u9918,\u534a,\u51fa\u982d,\u597d\u5e7e,\u958b\u5916,\u6574,\u6b63,\u8a31,\u8db3,\u4e4b\u591a} ;Nfa = {\u672c,\u628a,\u74e3,\u90e8,\u67c4,\u5e8a,\u8655,\u671f,\u9f63,\u5834,\u6735,\u9802,\u5835,\u9053,\u9813,\u9320,\u95cb,\u68df,\u5e62,\u6a94\u5b50,\u6a94,\u5c01,\u5e45,\u767c,\u5206, \u4efd,\u670d,\u500b,\u6839,\u6839\u5152,\u7ba1,\ufa08,\u6236,\u4ef6,\u5bb6,\u67b6,\u5377,\u5177,\u7bc0,\uf906,\u5c46,\u7b87,\u6372,\u5291,\u96bb,\u5c0a,\u76de,\u5f35,\u679d,\u652f, \u6a01,\u5e40,\u53ea,\u682a,\u6298,\u70b7,\u8ef8,\u53e3,\u68f5,\u6b3e,\u5ba2,\u8f1b,\uf9f9,\u7247,\uf9d7,\u679a,\u9762,\u9580,\u5e55,\u5339,\u6240,\u8258,\u6247,\u9996,\u4e58,\u8972, \u982d,\u689d,\u9577\u689d,\u53f0,\u81fa,\u633a,\u5802,\u5e16,\u9846,\u5ea7,\u5247,\u518a,\u4efb,\u5c3e,\u5473,\u4f4d,\u9801,\uf96e,\u623f,\u5f4e,\u73ed,\u54e1,\u4ecb,\u4e38,\u540d, \u9805,\u8d77,\u9593,\u7bc7,\u984c,\u76ee,\u62db,\u80a1,\u56de} ; Nfb = {\u901a,\u53e3,\u9813,\u76e4,\u5c40,\u756a} ; Nfc = {\u5c0d,\u96d9,\u5b97,\u756a,\u7566,\u9910,\ufa08,\u8eab,\uf99c,\u9577\uf99c,\u7cfb\uf99c,\u6392,\u9577\u6392,\u526f,\u4ed8,\u5957,\u7b46,\uf905,\u9577\uf905,\u639b,\u5e6b,\u623f,\u6279, \u7d44,\u7aa9,\u7d91,\u6346,\u7fa4,\u80ce,\u684c,\u561f\u5695,\u90e8,\u7a2e,\uf9d0,\u6a23,\u6a23\u5152,\u6d3e,\uf937,\uf942,\uf918,\u4f19,\u5925,\u675f,\u7c07,\u5e2d,\u758a,\u7d2e,\u8272, \u7968,\u53e2,\u968a,\u6524,\u5f0f,\u84ec,\u9805} ; Nfd = {\u4e9b,\u5206,\u5206\u5152,\u5718,\u5806,\u6ce1,\u7db9,\u64ae,\u628a,\u80a1,\u7058,\u6c6a,\u9663,\u53e3,\u53e3\u5152,\u62b9,\u584a,\u6ef4,\uf91d,\u6367,\u62b1,\u5c64,\u91cd,\u5e36,\u622a, \u9577\u622a,\u9577\u622a\u5152,\u622a\u5152,\u7bc0,\u7bc0\u5152,\u9577\u7bc0,\u9577\u7bc0\u5152,\u6bb5,\u6bb5\u5152,\u9577\u6bb5,\u9577\u6bb5\u5152,\u7d72,\u7d72\u5152,\u9ede,\u9ede\u5152, \u7247,\uf950,\u90e8\u4efd,\u90e8\u5206,\u5768,\u5339,\u758b,\u968e,\u6294,\u6ce2,\u9053} ; Nfe = {\u76d2,\u76d2\u5b50,\u5323,\u5323\u5b50,\u7bb1,\u7bb1\u5b50,\u6ac3\u5b50,\u6ae5,\u6ae5\u5b50,\u7c43,\u7c43\u5b50,\u7c0d,\u7c0d\u5b50,\uf932\u5b50,\u5305,\u5305\u5152,\u888b,\u888b\u5152,\u6c60 \u5b50,\u74f6,\u6876,\u807d,\u7f50,\u7f48,\u76c6,\u934b,\uf944,\u76e4,\u7897,\u676f,\u52fa,\u52fa\u5b50,\u5319,\u6e6f\u5319,\u7b52,\u64d4,\u74f6\u5b50,\u6876\u5b50,\u7f50\u5b50,\u7f48\u5b50, \u76c6\u5b50,\u934b\u5b50,\uf944\u5b50,\u76e4\u5b50,\u676f\u5b50,\u7b52\u5b50,\u64d4\u5b50,\u7c6e\u7b50,\u6753,\u6753\u5b50,\uf9fe\u5319,\u58fa,\u76c5,\u7b50,\u74e2,\u936c,\u7f38} ; Nff = {\u8eab,\u982d,\u81c9,\u9f3b\u5b50,\u5634,\u809a\u5b50,\u624b,\u8173,\u684c\u5b50,\u9662\u5b50,\u5730,\u5c4b\u5b50,\u6c60,\u8154,\u5bb6\u5b50} ; Nfg = {\u516c\u5398,\u516c\u5bf8,\u516c\u5206,\u516c\u5c3a,\u516c\u4e08,\u516c\u5f15,\u516c\uf9e9,\u5e02\u5c3a,\u516c\u91d0,\u71df\u9020\u5c3a,\u53f0\u5c3a,\u540b,\u544e,\u78bc,\u54e9,\u6d6c,\u6d77 \u6d6c,\u5eb9,\u565a,\u5c3a,\uf9e9,\u91d0,\u5bf8,\u4e08,\u7c73,\u5398,\u5398\u7c73,\u6d77\u54e9,\u6d77\uf9e9,\u82f1\u5c3a,\u82f1\uf9e9,\u82f1\u544e,\u82f1\u5bf8,\u7c73\u7a81,\u7c73\u5c3a, \u5fae\u7c73,\u6beb\u7c73,\u82f1\u540b,\u82f1\u54e9,\u5149\uf98e,\u516c\u755d,\u516c\u9803,\u5e02\u755d,\u71df\u9020\u755d,\u576a,\u755d,\u5206,\u7532,\u9803,\u5e73\u65b9\u516c\uf9e9,\u5e73 \u65b9\u516c\u5c3a,\u5e73\u65b9\u5c3a,\u5e73\u65b9\u516c\u5206,\u5e73\u65b9\u82f1\u54e9,\u82f1\u755d,\u516c\u514b,\u516c\u65a4,\u516c\u5678,\u5e02\u65a4,\u53f0\uf978,\u53f0\u65a4,\u65e5\u65a4, \u76ce\u53f8,\u76ce\u65af,\u78c5,\u516c\u64d4,\u516c\u8861,\u516c\uf978,\u514b\uf925,\u65a4,\uf978,\u9322,\u5678,\u514b,\u82f1\u78c5,\u82f1\uf978,\u516c\u9322,\u6beb\u514b,\u6beb\u5206,\u516c \u6beb,\u4edf\u514b,\u516c\u64ae,\u516c\u5347,\u5e02\u5347,\u71df\u9020\u5347,\u53f0\u5347,\u65e5\u5347,\u76ce\u53f8,\u54c1\u812b,\u52a0\u4f96,\u84b2\u5f0f\u8033,\u516c\u6597,\u516c\u77f3, \u516c\u79c9,\u516c\u5408,\u516c\u52fa,\u6597,\u6beb\u5347,\u5938,\u5938\u7279,\u5938\u723e,\uf9f7\u65b9\u7c73,\uf9f7\u65b9\u5398\u7c73,\uf9f7\u65b9\u516c\u5206,\uf9f7\u65b9\u516c\u5bf8,\uf9f7 \u65b9\u516c\u5c3a,\uf9f7\u65b9\u516c\uf9e9,\uf9f7\u65b9\u82f1\u5c3a,\u77f3,\u659b,\u897f\u897f,\u89d2,\u6bdb,\u5143,\u5713,\u584a,\u5148\uf9a8,\uf933\u6bd4,\u6cd5\u90ce,\u6cd5\uf929,\u8fa8 \u58eb,\u99ac\u514b,\u938a,\u82f1\u938a,\u7f8e\u5143,\uf965\u58eb,\uf9e9\uf925,\u65e5\u5143,\u65e5\u5713,\u5200,\u6253,\uf9a8,\u7db8,\u7c6e,\u5927\u7c6e,\u7126\u8033,\u5343\u5361,\u4edf\u5361, \u71ed\u5149,\u4edf\u74e6,\u5343\u74e6,\u4f0f\u7279,\u99ac\uf98a,\u723e\u683c,\u74e6\u7279,\u74e6,\u5361\uf937\uf9e9,\u5361,\u99ac\u514b,\u4edf\u8d6b,\u5343\u8d6b,\u5146\u8d6b,\u8d6b,\u8d6b \u8332,\u4f4d\u5143,\u83ab\u8033,\u6b50\u59c6,\u6cd5\uf925\u7b2c,\u5b89\u57f9,\u5206\u8c9d,\u5c45\uf9e9,\u5fae\u5c45\uf9e9,\u6beb\u5c45\uf9e9,\u6beb\u5b89\u57f9,\u6beb\u7c73,\u6beb\u5df4,\u9054 \u56e0,\u725b\u9813,\u5468\u6ce2,\u6b72,\u2103} ; Nfh = {\u7a0b,\u4f5c,\u5206,\u5398,\u6beb,\u7d72,\u570d,\u6307,\u8c61\u9650,\ufa01,\u958b,\u958b\uf90a,\uf997,\u5e2b,\uf983,\u5718,\u71df,\u4f0d,\u73ed,\u6392,\uf99a,\u7403,\u6ce2,\u56de\u5408, \u6298,\u968e,\u647a,\u7b49,\u7968,\uf9ca,\u68d2,\u8072,\u6b21,\u80a1} ; Nfi = {\ufa01,\uf9d7,\u56de,\u6b21,\u904d,\u8d9f,\u4e0b,\u4e0b\u5152,\u906d,\u756a,\u8072,\u8072\u5152,\u97ff,\u5708,\u5708\u5152,\u6b65,\u628a,\u4ed7,\u89ba,\u9813,\u95dc,\u624b,\u624b\u5152, \u8173,\u638c,\u5df4\u638c,\u62f3,\u62f3\u982d,\u773c,\u53e3,\u5200,\u69cc,\u69cc\u5b50,\u677f,\u677f\u5b50,\u97ad,\u97ad\u5b50,\u68d2,\u68cd,\u68cd\u5b50,\u9663,\u91dd,\u7bad,\u69cd,\u69cd \u77db,\u7832,\u5834,\u5468,\u66f2,\u8de4,\u8a18,\u56de\u5408,\u7968} ; {\u570b,\uf96d,\u5dde,\u7e23,\u9109,\u6751,\u93ae,\u9130,\uf9e9,\u90e1,\u5340,\u7ad9,\u5df7,\uf943,\u6bb5,\u865f,\u5730,\uf94c,\u8856,\u5e02,\u6d32} ; {WQP,DQP,PQP,NOP,NOP2} ;", "uris": null, "type_str": "figure" }, "TABREF9": { "num": null, "type_str": "table", "html": null, "content": "
datadevelopment settesting set
2035 sentences2111 sentences
accuracy(11697 word tokens)(12209 word tokens)
original DM rulessegmentation accuracy84.2%77.8%
tag accuracy71.0%60.3%
segmentation accuracy92.5%90.1%
revisedtag accuracy81.3%75.2%
DM rulestag accuracy
with correct88.0%83.5%
segmentation
", "text": "" } } } }