English. We present a study on the degree of homonymy between the lexicon of a morphological analyser for Latin and an Onomasticon. To understand the impact of homonymy, we discuss an experiment on four Latin texts of different era and genre.
Ambiguity affects linguistic analysis at various levels. In particular, homonymy plays a substantial role in the analysis of single words. Indeed, when considered out of context, one same word can be assigned different Parts of Speech (PoS), morphological features, lemmas and meanings. Contextual disambiguation is the task of Natural Language Processing (NLP) tools like PoS-taggers, morphological analysers, lemmatisers and word-sense disambiguators.
The problem of ambiguity is particularly
remarkable for NLP when Named Entity
Recognition (NER) is concerned. In order to
automatically classify the textual occurrences of
(multi)words into categories such as names of
persons, locations and organisations, NER faces
that specific kind of ambiguity consisting in the
homonymy between proper names and other
words in the lexicon
The recent extension of the lexical basis of the
morphological analyser and lemmatiser for Latin
Lemlat with an Onomasticon (i.e. a list of proper
names) makes it possible to evaluate the degree
of homonymy of proper names in Latin and,
thus, to understand the extent of the
disambiguation task
Since the large majority of lemmas in O are nouns (19,599 out of 26,250), we will focus on them here, first by comparing their distribution in L and O. Table 1 shows the number of nouns and their percentage (on the total of nouns) in L and O by inflectional category.
5,009 (22.26%) 6,651 (33.94%) 7,466 (33.17%) 7,235 (36.92%) 8,677 (38.54%) 4,464 (22.77%) 980 (4.35%) 101 (0.45%) 278 (1.23%) 22,511 58 (0.29%)
6 (0.03%) 1,185 (6.05%) 19,599
While third declension nouns are more frequent in L than in O, the opposite holds for first declension and (to a lesser extent) second declension nouns. The main difference between L and O concerns uninflected nouns, which are much more in O than in L because of the large number of loans recorded in O.
Also gender-based distribution of nouns by inflectional category shows substantial differences between L and O. Among the most relevant is that O includes more first declension masculine nouns than L (1,626 vs. 562). Instead, the number of second declension neuter nouns is larger in L than in O (4,005 vs. 1,523), because O tends to include more proper names of persons than of places, the latter being often assigned the neuter gender. As for third declension, feminine nouns are more than masculine in L (5,112 vs. 2,590), while the opposite holds in O (2,847 masculine vs. 1,185 feminine). 3
To categorise nominal homonymy in L and O, we defined three kinds of homonymy: (a) Full Homonymy (FH): words with the same lemma, PoS, inflectional category and gender in L and O; (b) Partial Homonymy (PH): words with the same lemma in L and O, but with different PoS, inflectional category or gender (the last for nouns only); (c) Mixed Homonymy (MH): words with the same lemma in L and O and with more than one PoS, inflectional category or gender, thus resulting partly into FH and partly into PH.
An example of FH in our data is the word spes, which means “hope” in L and “the Goddess 2,043 1,038
912
Most of the PH instances for first declension lemmas are due to different gender. An example is the first declension noun caligula, which is feminine in L (“a small military boot”) while it is masculine in O (a cognomen). Second declension shows several cases of PoS change, like in the case of severus, which is an adjective in L (“serious”) and a noun in O (a proper name). Instead, a large number of verb-noun changes holds for third declension. This mostly occurs for imparisyllable nouns ending in –o, like cato, which is a first conjugation verb in L (“to see”) and a noun in O (a proper name). of Hope” in O. PH is represented, for instance, by the word augustus, which is an adjective in L (“majestic”) and a noun in O (a cognomen given to Octavius Caesar as emperor). The word spina is a case of MH, being a first declension feminine noun in L (“thorn”) and both a first declension feminine noun (an old town in Aemilia) and a third declension masculine noun with genitive in –anis (a river God) in O, the former thus showing FH and the latter PH.
Table 2 presents the rates of homonymy in L and O by each kind per inflectional category. The total number of homonyms is provided as well (column “H”). This corresponds to the number of nouns of an inflectional category that are graphically identical in L and O. For instance, the first row of table 2 shows that there are 556 lemmas recorded as first declension nouns (in L or O) that are identical to a lemma occurring in the other section of the lexical basis of Lemlat. 383 lemmas out of these show FH, i.e. they share not only the same graphical form but also the same PoS, inflectional category and gender in L and O (column “FH”). Instead, 163 lemmas occur as graphically identical in L and O but do not have in common at least one among PoS, inflectional category or gender (column PH”). Finally, 10 lemmas show MH.
556 752 584 85
6 60
FH 383 307 334 9 5 0
PH 163 389 226 73
1 60
MH 10 56 24 3 0 0 93
PH does not raise any tricky issue for NLP, the task of PoS/morphological taggers being just that of disambiguating contextually PoS and morphological features. Conversely, FH (including the FH-like part of MH) represents a challenging question for NLP. Indeed, if upper/lowercase distinction is not available in input data, only context-based semantic properties can disambiguate between candidate lemmas affected by FH. For instance, in the clause “spes est expectatio boni” (“hope is expectation of good”, Cicero, Tusculanae, 4, 37, 80) there is nothing but semantics to help us to understand that the word spes is an occurrence of the noun from L instead of the proper name from O. In order to evaluate the extent of homonymy in real texts and to understand how much big the impact of FH is, we performed the experiment discussed in the next section. 4
We run Lemlat on four Latin texts of similar size and different genre and era.1 Table 3 shows the number of distinct words out of the total (column “Types”) analysed by the original version of Lemlat (column “Lemlat”) and by the one enhanced with the Onomasticon (column “LemlatON”).
(1) (2) (3) (4) 3,092 5,057 3,542 4,589 2,888 4,717 3,357 4,292 3,039 5,005 3,487 4,537 +151 +288 +130 +245
Beside the words analysed by LemlatOn only (column “Improv.”), there is a certain degree of overlapping between Lemlat and LemlatOn. The words falling in this ‘grey zone’ are those that are analysed both by Lemlat and by LemlatOn, as they are lemmatised both under a lemma from L and under one from O. Among these words, those affected by homonymy are to be found. 1 (1) Caesar, De Bello Gallico, 1 (Classical Lat., prose); (2) Virgil, Aeneid, 1 & 2 (Classical Lat., poetry); (3) Tertullian, Apologeticum (Late Lat., prose); (4) Claudian, De Raptu Proserpinae (Late Lat., poetry). All the texts were downloaded from the Perseus Digital Library (www.perseus.tufts.edu). (1) (2) 618 1,207
686 1,062 405 799 486 706 303 546 330 469 88 186 120 177 14 67 36 60
Column “L/O” in table 4 reports the number of words for each text that are analysed both by Lemlat and by LemlatON. The other columns show the homonymy rates by the kinds described in Section 3. For instance, in the text of Caesar (1) there are 618 words analysed by both the versions of Lemlat (L/O). 405 out of them share the same lemma in at least one analysis (H). This is further detailed: 303 out of 405 show FH, 88 PH and 14 MH. An example of a word analysed by both the versions of the tool that does not share the same lemma in all analyses is acie, which is lemmatised under acies (“dagger”) by Lemlat (fifth declension feminine noun) and also under the proper name acius by LemlatON (second declension masculine noun). The word constantia is an example of H: it is lemmatised as a form of both lemmas consto (“to agree”; first conjugation verb) and constantia (“steadiness”; second declension feminine noun) by Lemlat, and also as a form of both proper names constantius (second declension masculine noun) and constantia (second declension feminine noun) by LemlatON. The word constantia is also an example of FH, as the analyses provided by the two versions of Lemlat that share the same lemma have in common even the same inflectional category and gender. PH is shown by the word crassi, which is assigned the same lemma (crassus) both by Lemlat and by LemlatON, but while it is a first class adjective in the former (“solid”), it is a second declension masculine noun in the latter (a proper name). An example of MH is the word amico, which is lemmatised under the lemma amicus (“friend”) both by Lemlat and LemlatON. The lemma amicus is both an adjective and a second declension masculine noun in Lemlat, but only the latter analysis is shared with LemlatON, because the lemma amicus in the Onomasticon is recorded only as a noun and not also as an adjective. Thus, when the word amico is assigned PoS noun it shows FH, while when it is assigned PoS adjective it shows PH.
The proportions between the kinds of homonymy remain quite similar for all the texts. Words affected by H tend to be more than half of L/O; among them the large majority is affected by FH. By comparing columns “FH” and “MH” in table 4 with column “Types” in table 3, one can see that slightly more than 10% of the words of all the texts is affected by FH. This is the percentage rate of words whose lemmatisation cannot be disambiguated by a PoS tagger, because semantic features only are here at work to choose between candidate lemmas. For instance, if a PoS tagger assigns to one occurrence of the word constantia PoS noun and gender feminine, it cannot disambiguate between the two (fully morphologically identical) lemmas constantia provided by LemlatON.
If we focus on textual occurrences (tokens) instead of distinct words (types), the rates of FH (+MH) range between 8.44% (Caesar) and 13.19% (Tertullian), as shown by table 5. This result represents the extent of the impact of FH on the texts that we used in the case-study.
(1) (2) (3) (4)
8,171 10,045 7,317 6,991
FH+MH 690 (8.44%) 1,325 (13.19%) 668 (9.13%) 797 (11.4%)
Most of the words showing FH can be easily disambiguated (at least, manually) according to peculiarities of single texts. For instance, the word amicitiam (from Caesar’s text) belongs to lemma amicitia both in L (“friendship”) and in O (“the Goddess of Friendship”), thus showing FH. However, it is more likely that the former is the one occurring in Caesar than the latter. Conversely, in the same text the word galli (lemma gallus) is more likely a proper name from O (“Gauls”) than a noun from L (“cock”). 5
We presented a study about the degree of homonymy between the lexical basis of a morphological analyser for Latin and an Onomasticon recently added in the tool. We have shown the impact of nominal homonymy on a number of Latin texts of different era and genre.
Since the analysis of many homonymous words can be disambiguated according to the features of single texts (and authors), in the near future we foresee to enhance such words in Lemlat with information about their distribution in a number of manually tagged reference texts.