I. Mapping ambiguity in the Portuguese DELAF

1The Portuguese DELAF (v. 2), built by the LabEL team2, has been publicly available since 2002. It contains 936,215 entries, consisting of 889,986 different inflected forms. There are 43,127 different ambiguous inflected forms, that is, word forms belonging to different categories and / or lemmas, which correspond to 89,356 DELAF ambiguous entries3. For example, the inflected word capital (capital) is ambiguous because it has three entries in the DELAF4 :

2It is possible to obtain the full list of different, inflected, ambiguous forms with the grammatical categories associated to them5. The three entries above are thus factorized as :

capital, A_N_N

3About 10 % of Portuguese DELAF entries are ambiguous words, which correspond to 4.85 % of the DELAF different inflected forms.

1. Ambiguity class (AC)

4One can consider A_N_N an ambiguity class (AC). There are 137 ambiguity classes in the DELAF. The number of ambiguous word forms per ambiguity class is very uneven (Table 1) :

Table 1 : Distribution of ambiguity classes (AC) per number of ambiguous word forms (WF)

5There are 56 ambiguity classes (40.88 %) with only one word form. The quantitatively more important ambiguity classes are (Table 2) :

Table 2 : Quantitatively more important ambiguity classes (AC)*

6The 7 largest ambiguity classes represent 97.15 % of the ambiguous word forms of the DELAF ; the A_N ambiguity class alone contains more than 56 % of the total of the ambiguous word forms.

2. Ambiguity level (Amb)

7Each ambiguity class corresponds to a certain ambiguity level (Amb). For instance, for the A_N class Amb = 2, while for A_N_V class Amb = 3, and so on. Distribution of ambiguity classes by the ambiguity level of their word forms is shown below (Table 3) :

Table 3 : Distribution of ambiguity classes (AC) by ambiguity level (Amb)*

8The majority of the ambiguous word forms show Amb = 2 (93.51 %), or Amb = 3 (5.87 %), but the number of ambiguity classes with Amb = 2 is slightly less (40) than that with Amb = 3 (50). The most ambiguous word forms present Amb = 6, but there are only two words of this kind, each one belonging to a different ambiguity class. The distribution of the most important ambiguity classes with Amb = 2 is shown below (Table 4).

9Five ambiguity classes with Amb = 2 stand out as the most prominent : N_N, V_V, N_V, A_V and A_N. Together, these five classes constitute about 92.5 % of the ambiguous words of DELAF, with A_N ambiguity class alone representing more than half.

Table 4 : Ambiguity classes (AC) with ambiguity level Amb = 2*

3. Ambiguity types (AT)

10It is also possible to consider different ambiguity types, i.e., ntuples of categories involved in different ambiguity classes. For example, in ambiguity class A_N_V there are three (binary) ambiguity types : A/N, A/V and N/V, besides the (ternary) A/N/V ambiguity type. An ambiguity type involves different ambiguity classes. For example, here are the 25 ambiguity classes where the ambiguity type A/V may be found (Table 5).

11As we can see, the ambiguity class A_V alone contains most of the word forms presenting this ambiguity type (82.6 %), followed by class A_N_V (15 %). The remaining classes are quantitatively less important. On the other hand, the A/V ambiguity type involves ambiguity classes were all other categories are represented.

Table 5 : Ambiguity type A/V*

4. Using ambiguity information with DELAF

12The three types of information on ambiguity – Amb (ambiguity level), AC (ambiguity class) and AT (ambiguity type) – can be added to the DELAF : it is possible to build a priority DELAF of ambiguous words where this information is given, as in the following entries of capital :

capital, capital. A+Amb=3+AC=A_N_N+AT=A/N+AT=N/N : ms : f s
capital, capital. N+Amb=3+AC=A_N_N+AT=A/N+AT=N/N : fs
capital, capital. N+Amb=3+AC=A_N_N+AT=A/N+AT=N/N : ms

13This information could then be used on disambiguation grammars, as we will see below.

II. Measuring the impact of ambiguity in a Portuguese corpus

1. Preliminaries

14In order to see how ambiguity affects real texts, we have applied the Portuguese DELAF to a large corpus of Portuguese journalistic text, the CETEMPúblico corpus6, using INTEX (v. 4.31)7. Here are raw results from lexical analysis of this text (Table 6).

15The corpus contains about 12.8 million tokens (9.7 million simple words). About 8.96 million simple words were identified by the DELAF (92.324 % of the text words). The simple word dictionary of the text (DLF) contains 103,045 entries and 83,690 different inflected forms (46.67 % of the corpus different simple words).

Table 6 : Lexical analysis of CETEMPúblico (Part 01) corpus using DELAF

16A finite-state transducer (FST)8, allows the (tentative) tagging of proper nouns (N+Nprop ; Silberztein 2000 :123), i.e. simple words beginning in capital letter, followed by another letter, and not previously identified by the DELAF (Fig.1) :

Fig. 1 : Npr+.fst

17In this FST, <U> stands for any uppercase letter from the alphabet, while <W> stands for any letter. There are 57,275 different tokens that may be (tentatively) tagged in this way (31.963 % of the corpus different simple words), otherwise, they would not be recognised and would then be considered unknown tokens (ERR). The remaining 15,175 different words (8.468 % of the corpus different simple words) were left as unknown (ERR). For the purposes of this paper, the list of unknown tokens, ERR, and the list of tokens tagged by the Npr+.fst were not taken into consideration. We will thus focus on the remaining words that form the dictionary of simple words of the text (DLF).

2. The DLF of the CetemPúblico (Part 01) corpus

18The distribution of the DLF lexical entries by ambiguity level is shown below (Table 7) :

Table 7 : Distribution of the simple word dictionary of the CETEMPúblico (Part 01) corpus by ambiguity level (Amb).

19In the first place, it is interesting to verify that almost 80 % of the DLF different inflected forms (DIF) are non-ambiguous and represent 64,57 % of DLF entries9.

20As it is now possible to associate the DIF list to the frequency list of the corpus’ tokens, we can see that non-ambiguous words represent about 44.52 % of the tokens. This constitutes a first approximation to the degree of ambiguity of the corpus’ words, for it means that more than a half (55.48 %) of the corpus different tokens is ambiguous.

21However, we proposed (Baptista e Faísca 2001) another measure for evaluating the impact of ambiguity in a text. The weight of an ambiguous word to the global ambiguity of a text may be viewed as the product of its ambiguity level (Amb) by its frequency (FRQ). In the absence of a better name, we plainly called it product (PROD). This measure seems more adequate than mere frequency, for it is related with the number of analysis that must removed from the text in the process of reducing its ambiguity. Any gain in ambiguity reduction can be calculated as the difference between the total PROD (17,691,627) and the remaining PROD after the disambiguation process has been carried out. The base-line would be a figure corresponding to the total disambiguation of the texts words, that is, the sum of their frequencies (8,960,421).

22Considering this PROD measure, it is now clear that even if non-ambiguous words are a substantial part of a text’s tokens, ambiguous words contribute significantly (77.452 %) to the global ambiguity of a corpus such as this. Here we find something that can be viewed as a surprise : while frequency values associated to each ambiguity level are quite different, the impact of ambiguity classes with Amb = 3 to 5 is relatively similar, even if ambiguity classes with Amb = 2 reflect a greater impact.

3. The dlf of the tokens’ frequency list of the CetemPúblico corpus

23We tried to verify if there were significant changes when, instead of just Part 1 (with less than 10 million words), we would consider the ambiguity of the entire CetemPúblico corpus (that contains about 200 million words). For this, we used the frequency list of the corpus10. The distribution of tokens by ambiguity level is shown in Table 8.

Table 8 : Distribution of the simple word dictionary of the CETEMPúblico corpus by ambiguity level (Amb)

24It the first place, while the number of tokens of the entire corpus is 17.8 times larger than that of Part 1, the number of different word forms and the number of lexical entries is just about twice as large. Still, even with such a large corpus, DIF only constitute 17,752 % of the entire DELAF different inflected forms, while DE represents 19,893 % of DELAF lexical entries Secondly, when we compared the difference between the percentages for DIF, LE, FRQ and PROD by ambiguity level among the two corpora (Table 9), we find that they were, for the most part, minimal :

Table 9 : Difference between the percentages of DIF, LE, FRQ and PROD by ambiguity level in the entire CETEMPúblico (CP) and only in Part 1 (P1)

25The only significant differences were : the percentage of non-ambiguous different inflected forms (DIF) and of lexical entries (LE) in the DLF of the entire CETEMPúblico is slightly larger than that of Part 1. However, the percentage of FRQ and of PROD is smaller, therefore, the impact on the corpora global ambiguity of ambiguous words with Amb = 2 or 3 is slightly larger (2.8 %), especially for words with Amb = 3 (1.91 %).

26It is thus reasonable to assume that most observations regarding ambiguity in Part 1 will also hold for the entire CETEMPúblico corpus.

III. Disambiguation grammar of past participle in compound tenses with ter (to have) : a case study

27We begin by explaining that in Portuguese, the past participle, noted <V : K> in the DELAF and abbreviated as K in this paper, is an invariable form of the verb, only used in compound tenses with temporal auxiliary ter (to have) and (more rarely, and quickly becoming obsolete) haver (to there be)11 as in :

O João cumprimentou a Maria
John greeted Mary
O João (tinha +havia) cumprimentado a Maria
John has greeted Mary

28There are 11,374 K in the Portuguese DELAF. In many cases (5,301), K is ambiguous with words of other categories, mainly with adjectives. In fact, the ambiguity type A/V :k affects 5,182 word forms, which represents 97.755 % of K ambiguous forms and constitutes 59,061 % of all ambiguity classes with ambiguity type A/V.

29In view of identifying the compound tense, it is necessary to disambiguate K. Usually, general-purpose disambiguation grammars, such as the one shown below (Fig.2) only use the grammatical categories available in the DELAF.

30This grammar can be viewed as a first approach, obviously very incipient, to the compound past tense disambiguation problem.

31When we apply this grammar to the CETEMPúblico (Part 1), we obtain 27,327 matches from the 145,998 occurrences of ambiguous K word forms. Naturally, the next step is to evaluate the results of this grammar. In this sense, we think that ambiguity information can be useful to identify the cases where the grammar could be improved or to detect problems in the dictionary. Results are shown in Table 10, sorted by ambiguity level.

Fig. 2 : Disambiguation grammar Ter_Vk.fst

Table 10 : Ambiguous <V : K> in the DELAF

32In order to evaluate the performance of the grammar, we added to the table the non-ambiguous K entries of the DELAF. For each ambiguity level we indicate the ambiguity classes involved. The number of different inflected forms in the DELAF concerns only K forms. We calculated their total frequency in the corpus and the number of occurrences matched by the grammar.

33The impact of K forms in the total PROD of CETEMPúblico (17,691,627 ; see §2.1) is only 1.934 %, but as one can see, it involves 145,998 occurrences.

a) Amb=1 (non-ambiguous K forms)

34Theoretically, the non ambiguous K forms (Amb = 1) should had been all be matched by the grammar. In fact success rate is high (88.279 %) but 1,031 occurrences of K forms were not matched by the grammar12.

35Looking through the concordances of non-ambiguous K not detected by the grammar, we found many forms that can also function as adjectives, in predicative construction with copulatives ser and estar and even in attributive position next to a noun. These forms were not marked as adjectives in the DELAF :

Segundo a jovem, Emmet Roy foi acometido de dores está fundeado na Baía de Cascais

36In many cases, theses adjectives are attested in adnominal attribute position :

90 dias num barco fundeado a 500 metros de Monróvia

37Some non-ambiguous K appear in the <haver> <V : K> compound tense, not described here :

muitas ambulâncias haviam afluído ao local

38There are some K forms that are ambiguous with proper nouns. This is the case of Chiado (a commercial area in downtown Lisbon ; 158 occurrences) and Granado (a proper (family) name ; 28 occurrences) :

queremos recuperar o prestígio do velho Chiado
a resposta (oral) que dei ao jornalista António Granado

39Of course, ambiguity between K forms and proper names were not considered.

b) Amb > 1 (ambiguous K forms)

40The number of matches of ambiguous K forms with Amb>1 is very low (an average of 14 %). This comes naturally from the fact that the more a word has other categories, the less likely is it to appear as a K in a compound tense. Other disambiguation grammars should be built in order to solve some of the predicative and attributive uses of adjectives belonging to the A/V : k ambiguity type.

41Even for the matched occurrences, success rate is not satisfactory, because the grammar eliminates correct analysis, i.e., it produces significant silence. For instance, there are some N of the N/V : k ambiguity type that can have ter as a support verb. Ambiguity class A_INTERJ_N_V is one of these cases. It contains three forms : cuidado, obrigado, sentido. Looking to the concordances where these forms appear, it was possible to determine the silence produced by the grammar :

Table 11 : Ambiguous <V : K> in the DELAF : Ambiguity class A_INTERJ_N_V

42The success rate varies considerably : while cuidado was incorrectly tagged most of the times, with obrigado it was just the opposite ; in the case of sentido, about half of the tags were correct.

43However, this may not be always the case. With class A_N_V_V13, in spite of the fact that there is potential N analysis, the grammar shows a 100 % success rate.

44Considering that any disambiguation grammar should avoid producing silence, it is possible to increase precision. This can be done by reducing the scope of the grammar using the information of ambiguity class. In most of the cases where correct analysis were eliminated, this was due to the fact the word was not a <V : K>, but a <N>. If, instead of just using the tag <V : K>, we use the same grammar but with tags stating the ambiguity class, leaving out only the classes were an analysis as <N> is potentially possible, it should be possible to obtain higher success rates. This was in fact the case : for the A_V class alone, success rate is very high (around 99 %)14.

45For the remaining ambiguity classes, the use of ambiguity information to identify problematic ambiguous word forms can help devising different strategies to constraint the general grammar shown above and to improve its performance.

IV. Exotic words

1. A filter for ambiguous exotic words

46There are many ambiguous words in the DELAF that present some analyses that are very unlikely to occur in texts. For example, the words uma and umas, among other analysis (as DET), can be tagged as inflected forms of an extremely rare verb umar (an intransitive verb said about wood that gets wet). The word uma belongs to class DET_DET_V (Amb = 3), while umas is a DET_V (Amb = 2). These words appear 82,302 and 664 times in the corpus, which results in a very high impact in text global ambiguity : PROD = 248,234. If we remove the V analysis from the DELAF, using a filter, we obtain substantial reduction of PROD to 165,268.

47The general purpose of building lexical filters is to reduce the noise produced during lexical analysis of texts, but maintaining a low and controlled level of silence, in order to improve the following steps in the text’s processing. It is only a facultative tool, to be used with care, surely. But it is also a useful tool, especially when it is meant to be a part of applications were high efficiency is paramount to the objective of 100 % correctness (at the current state of the art, this goal is still a mirage).

48In a previous paper (Baptista & Faísca 2001), we have used frequency information associated to the ambiguous words of CetemPúblico (Part 1) to identify the 560 most frequent ambiguous exotic word forms15 and to produce a filter that eliminates those exotic analyses. The 17,047 % ambiguity reduction thus obtained16 may be viewed as an important contribution to the disambiguation task in the general processing of texts.

49This approach can be seen as complementary with the approach that consist of structuring the lexicon by levels (Garrigues 1992). Furthermore, it presents the advantage that only some of the inflected forms of a lexical entry (the ambiguous forms) are affected.

50It is possible to use ambiguity information to guide the discovery of exotic words and to help constructing different types of filters. Two examples are given below.

2. X-áveis

51Many second person-plural past-imperfective verb forms are ambiguous with A-vel plural forms :

52This is another case of a A/V ambiguity type. The use of the second-plural person is becoming obsolete in European Portuguese. The general idea is that we would not produce any significant silence if we would remove these <V : I2p> analysis from the DELAF, especially in this kind of journalistic texts such as the CetemPúblico.

53There are 1,611 ambiguous X-áveis word forms. Only 4 of them do not involve A/V ambiguity type. The remaining 1,607 are distributed as follows (Table 12) :

Table 12 : X-áveis word forms with ambiguity type A/V

54In (almost17) 100 % of the A_V cases, the <V : I2p> analysis is incorrect :

55For the A_N_V words also, the <V : I2p> analysis is incorrect. However, is not possible to fully disambiguate between the N or the A analyses, except for the case of adjective ultrapassáveis (“possible to be by-passed”), for which we cannot conceive a natural N construction.

56The impact of ambiguity in the corpus resulting from the elimination of the verb analysis would represent a PROD of 5,721, which corresponds to a reduction of PROD of 3,417 (37.393 % of total PROD before reduction).

3. X-ais

57There are 662 ambiguous X-ais word forms in the DELAF. Only 262 of these words appear in the corpus, totalling 75,388 occurrences. A single word (mais, ADV_N, “plus / more”) is responsible for 43,499 (57,7 %) of these occurrences. The remaining words belong to 16 different ambiguity classes, but only 11 are in fact present in the corpus, representing 31,889 occurrences. From these, there are 299 word forms presenting the A/V ambiguity type, of which 137 appear in the corpus, totalling 14, 120 occurrences (that is, 44.279 % of all ambiguous X-ais word forms, excluding mais). All verb forms correspond to <V : P2p> for verbs ending in -ar or <V : S2p : Y2p> for verbs ending in -ir.

Table 12 : X-ais word forms with ambiguity type A/V

58Most of these V analyses correspond, in fact, to exotic words. They represent a PROD of 14,120 (33,708 % of the initial PROD, before filtering).

59The fact that we can pin-point ambiguous words via their ambiguity level (Amb), ambiguity class (AC) or ambiguity type (AT) makes it possible to ascertain which words are more or less likely to occur in texts. Such intuitions can be confirmed on large corpora in order to decide which would be more reasonable to include in filters. Ambiguity grammars may include this new type of tags.

Conclusion

60In this paper, we mapped the ambiguity of words of the Portuguese DELAF and evaluated its impact on a large corpus. Linguistically oriented disambiguation techniques can profit from the ambiguity information retrieved in this manner. It is possible to use this kind of information to trace precisely the situations where a disambiguation grammar is most efficient or, otherwise, shows poor success rate. It can also help to maintain dictionaries, making easier to find lacunae or eventual mistakes. Finally, it may guide us to determine general rules to produce lexical filters.