Prosodic word deletion in coordinate structures

Prosodic word deletion in partially identical coordinate structures (DUI) is a rather general phenomenon across languages. In this paper, new data is presented that contributes to a cross-linguistic understanding of the deletion process. Its phonological nature is supported by an array of prosodic requirements on the deleted unit, its counterpart, the remnant, and the way the coordinate sequence is prosodically phrased, showing that a pure morphosyntactic analysis of DUI is unattainable. On the other hand, it is also shown that syntax constrains DUI as well. These properties are argued to be general and the cross-linguistic variation found is shown to result from prosodic differences that the languages independently display. It is further argued that DUI is fundamentally different from syntactic ellipsis phenomena. The account proposed places this process at the syntax-phonology interface without any further addition to constrained models of syntax-phonology interaction.


Introduction
The topic under investigation in this paper is a deletion process that affects internal parts of words in partially identical coordinate structures.On the one hand, this phenomenon depends on the existence of coordinate structures, and thus it looks syntactic in nature, but on the other hand it targets phonological, rather than syntactic or morphological units, and is subject to phonological conditions.
Although the deletion process is found in several languages, as illustrated in (1), to our knowledge it has only been studied in depth in Dutch and in German (cf.Booij, 1985Booij, , 1988;;Wiese, 1993;Kleinhenz, 1994).
(1) landbouw en tuinbouw Dutch (Booij, 1985:147) 'agriculture and horticulture' 1.1 The target of DUI: a phonological rather than morphosyntactic constituent It is clear that the deleted unit is not a morphological constituent.As illustrated in (2)-( 4), units with the same morphological status behave asymmetrically with regard to DUI depending on their prosodic nature.In all these examples, suffixes that can be shown to form a prosodic word () independent from their base may be part of coordinate structures that undergo deletion, whereas suffixes that do not form independent prosodic words and thus belong to the same prosodic word as their morphological base do not delete (in the examples, prosodic word structure is indicated and suffixes are marked in italics).
(2) a. (storm)  (achtig)  en (regen)  (achtig)  (based on Booij, 1985) 'stormy and rainy' b. *(blauwig)  en (rodig)  'bluish and reddish' (3) a. (erkenn)  (bar)  und (begreif)  (bar)  (based on Booij, 1985) 'recognizable and comprehensible' b. *(Bestrafung)  oder (Beförderung)  'punishment of promotion' (4) a. (demorada)  (mente)  ou (rapida)  (mente)  'slowly or quickly' b. *(isolamento)  ou (revestimento)  'isolation or cover' Evidence for the prosodic structure of the coordinate members in (2-3) includes resyllabification facts, as resyllabification is blocked across prosodic words in both Dutch and German (cf.Booij, 1985Booij, , 1995;;Wiese, 1996), prevocalic schwa deletion in Dutch, which applies within but not across prosodic words (Booij, 1985(Booij, , 1995)), and glotal stop insertion in German, which applies before vowel initial units that form independent prosodic words (Booij, 1985).The prosodic distinction between the structures that allow/do not allow deletion is also straightforward in European Portuguese (EP).In (4a), each member of the coordination is formed by two prosodic words, as shown by their respective word-level stresses: the stressed vowel of the first prosodic word of each coordinate element is low (see (5a-b)), whereas if it were stressless it should have been raised by the vowel reduction processes; the stress on rapida falls on the antepenultimate syllable, which is the marked stress pattern and can only be the result of word-level stress rules (see (5c)).By contrast, each of the coordinate elements in (4b) show a single word-level stress and the vowels that would bear stress if the morphological base formed a prosodic word independent from the suffix were subject to vowel reduction (cf. (5d--e) Besides showing that the deleted unit is not a morphological constituent, the data presented in (2)-( 4) also indicates that DUI may not be the result of a syntactic process, either.Note that the deleted unit corresponds to a suffix, which is not a proper syntactic object (see Zwicky, 1977, andSelkirk, 1982, for early proposals that the internal structure of words may not be seen by syntactic operations, and Booij, 1985, for a similar observation).The examples in (6) strengthen this point: each of the units that compose the coordinate structure may not be analysed as syntactic words, as they never occur independently.Given that they form independent prosodic words, however, deletion under identity is possible.
(6) (mono)  (gâmico)  e (poli)  (gâmico)  'monogamous and poligamous' (mono)  (lingue)  e (multi)  (lingue)  'monolingual and multilingual' (homo By the same token, an analysis of basic syntactic coordination or of syntactic ellipsis, as in (7), could not account for the facts described, since each member of the coordinate structure may not be analysed as containing two syntactic words (we will return to the distinction between DUI and syntactic ellipsis in section 4).
Cases where there is a mismatch between morphological (or syntactic) and prosodic structure also argue in favour of the phonological rather than morphological (or syntactic) analysis.This is illustrated by example (8).As noticed by Villalva (1994), the suffix -ista is attached to the morphological compound and not simply to the preceding root, since a macro-economista is not 'a big economist' but rather 'a person who deals with macro-economy'.Again, a syntactic analysis is also ruled out, given that a sequence like macro--economista corresponds to a single syntactic word.Only from a prosodic point of view does the deleted unit correspond to a constituent.macro-economista ou micro-economista 'macro-economist and micro-economist' Other facts rule out any attempt to treat DUI as a syntactic operation.In languages like German, there are examples of deletion in situations where either the two coordinate members or the deleted unit and its identical counterpart have different morphosyntactic status.As pointed out by Wiese (1993), deletion is possible in German when the phonologically identical units are marked with different case features, as in ( 9).(from Kleinhenz, 1994: 16) As highlighted in Booij (1985), cases like (10a) further show that sequences like [Amerikaanse en Papoea] may not be analyzed as resulting from basic syntactic coordination, since the two words belong to different syntactic categories.
A final argument first advanced by Booij (1985: 153-154) that clearly indicates that the phenomenon may not be accounted for in morphosyntactic terms is the presence of a linking segment in coordinate structures involving morphological compounds.Firstly, a basic syntactic coordination analysis may not explain the presence of the linking vowel (see (11a, c), respectively for Dutch and EP); secondly, although the linking segment does not morphologically form a constituent with any of the roots, since it links them, it is phonologically part of the first root and thus it is only deleted when the prosodic word that includes that root is deleted, as in (11b) (11) a. pruimeboom of appelboom 'plum tree or apple tree' b. onderzoeksdoelstelling of onderzoeksbelangstelling 'research goal or research interest' c. israelo-palestiniano ou afro-palestiniano 'israelipalestinian or afropalestinian' To sum up, there is clear evidence that the process under investigation here may not be purely syntactic or morphological since (i) units with the same morphological status behave asymmetrically depending on their prosodic status, (ii) the elements that are affected by DUI may be composed of units that form a constituent from a phonological but not from a syntactic or a morphological point of view, (iii) the phonologically identical portions of coordinate structures that undergo deletion may be morphologically (and syntactically) distinct, and (iv) two phonological requirements on DUI have been identified: phonological identity between the target and its counterpart and the prosodic word status of the deleted unit.Given that the deleted unit must correspond to a prosodic word and as there are no simplex words composed of more than one prosodic word in any language (a fact that follows from the algorithms of construction of prosodic words), it is predicted that we cannot find deletion under identity between two coordinate elements that are simplex words.As far as European Portuguese, Dutch and German are concerned this prediction is borne out.

Other phonological requirements on deletion under identity
Besides phonological identity and the prosodic word status of the deleted constituent, there are additional phonological conditions for the DUI to apply.Once again, these have to do with prosodic constituency.

The nature of the remnant: also a prosodic word
The observation of EP data suggests that prosodic constituency is not relevant only for characterizing the deleted unit, but also the remnant of deletion, which always corresponds to a prosodic word as well.This is shown by the examples in ( 12), where deletion is impossible when the deleted unit is a prosodic word but the remnant is not.In the examples, we provide the prosodic structure associated to each prefixed word, according to independent evidence previously presented in Vigário (2001) The prosodic requirement on the remnant stated in ( 13) is supported by all the tests that can be used as diagnostics for the prosodic word domain in EP.The data presented in the following examples illustrates how.
(13) Prosodic requirement on the remnant The unit left as the remnant of deletion must form a prosodic word Word-level stress is a crucial property of prosodic words in any language, and it is perceived for both elements within each coordination member in all cases where deletion applies in EP.Besides the perception facts, the absence of vowel reduction in the stressed syllable also constitutes evidence for the stressed status of a given unit and hence for its prosodic word status.The examples in ( 14) show that when the stressed vowel of the remnant is not subject to vowel reduction deletion is possible, whereas when the vowel that could bear stress is subject to vowel reduction deletion is not possible.
Several other phonological facts point in the same direction, namely the blocking of vowel raising in non-final root compounds when the first root corresponds to a prosodic word (see ( 15)), the blocking of semivocalization of front vowels followed by vowel (see ( 16)), and the low realization of stressless non-high vowels followed by a sonorant coda in prosodic word final position (see ( 17 'international or intra-national' It is therefore well established that in European Portuguese the remnant must be a prosodic word similarly to the deleted unit.The Dutch and German data found in the literature suggest that this is also true for these languages.

The nature of the identical counterpart: also a prosodic word
We have seen that the coordinate member that includes the target of deletion must be formed by more than one prosodic word.It will now be shown that the coordinate member that includes the identical counterpart must contain more than one prosodic word as well.This is illustrated in (18).( 18 The explanation for this fact seems straightforward, since we have already seen that deletion requires the presence of a unit phonologically identical in the second member of coordination.Thus, in order for DUI to obtain, the deleted unit, which is a prosodic word, must be prosodized similarly to its identical counterpart.The latter must therefore form a prosodic word. 2n conclusion, the phonological identity between the deleted unit and its counterpart must be expressed in terms of segmental material and prosodic constituency.

The relevance of phasal prosodic constituency and phonological distance
Similarly to Dutch and German (Booij, 1988;Wiese, 1993), in EP the deleted unit must be adjacent to a phonological phrase ( edge (see (19a)) and the s that include each coordinate element must be adjacent within an intonational phrase (see (19a) vs. (19b)).In addition, the EP data show that the number of intervening elements between the remnant and the identical counterpart also matters: as the length of the second phonological phrase increases, the acceptability of deletion decreases.This effect is illustrated in ( 20 Thus phonological distance should be added to the conditions on DUI, further supporting the phonological nature of this process.

Non-phonological requirements on deletion under identity
In the previous section, a number of phonological requirements on DUI was established: (i) the deleted unit, the remnant and the identical counterpart must all be prosodic words, and (ii) phrasal prosodic constituency together with phonological distance constrain DUI.However, there seems to be a syntactic requirement on DUI, as well.In EP, unlike in Dutch or German, deletion is only possible when coordinate structures are involved In his account of the data in (22), Booij (1988) suggests that what is relevant is not the presence of a given syntactic structure but rather a specific prosodic configuration.He proposes a pure prosodic analysis where deletion must occur at either edge of the phonological phrase and the two phrases with identical phonological words must be adjacent.As it was shown above, this generalization holds for EP as well, but, nevertheless, it does not account for the impossibility of deletion in syntactic constructions without coordination (see the contrast between (23a) and (23b)).
In conclusion, deletion under identity is not a pure phonological process in EP, as it requires both reference to phonological and syntactic properties.

Crosslinguistic variation
It was shown in the previous sections that languages like Dutch, German and European Portuguese behave in a remarkably similar way with respect to deletion under identity (to the exception of the syntactic restriction, if at all).In this section, we will consider data that is apparently not accounted for simply on the basis of the conditions proposed so far.

Particular cases where the phonological and the syntactic conditions are met but deletion is impossible
In Dutch and German, syntactic (word+word) compounds may undergo deletion under identity.In EP, by contrast, not only syntactic compounds but also derived words with z-evaluative suffixes, which form independent prosodic words, do not allow deletion.This is shown in ( 26) and ( 27) for -members of syntactic compounds and z-evaluative suffixes, respectively.'(He) was a nice(AFFECTIVE) but stupid(AFFECTIVE) guy.' a'.Era um fulano bom mas estupidozinho.
'One can choose between a little coffee and a little tea.' b'.Pode escolher-se entre café e chazito 'One can choose between a coffee and a little tea.' These cases are apparently problematical, since they could be seen as exceptions to DUI.However, there is a condition that explains these data and therefore allows us to characterize DUI as a general and exceptionless process.In all the examples given in (26-27) the result of deletion would coincide with a string that independently exists in the language, as shown by the sequences in (a') and (b').In other words, DUI would create ambiguity.Under such conditions the process is blocked.
The same explanation accounts for cases where the deletion of the first prosodic word of the second member of coordination is not allowed in EP, as such deletion would result in a sequence ambiguous with a structure independently formed (see ( 28)).This explains as well some German cases noted in Wiese (1996) where DUI is impossible although the other conditions for deletion are met.In fact, all such cases are always reported to result in ambiguous readings.
(28) pós-sintáctico e pós-fonológico / pós-sintáctico e fonológico 'postsyntactic and postphonological' / 'postsyntactic and phonological' euro-asiático e euro-africano/ euro-asiático e africano 'euro-asian and euro-african' / euro-asian and african' Given that deletion under identity is a stylistic kind of operation, the fact that it is blocked when its result coincides with a syntactic structure independently obtained in the language may be seen as an instance of a more general principle that rules out stylistic operations that create ambiguity, namely the "up-to-ambiguity principle" (cf.Chomsky, 1965).
3.2.The prosodic nature of the remnant: are some languages more liberal than others?
In section 1, we have proposed that the target of DUI must be a prosodic word, following an initial proposal by Booij (1985) also adopted by Wiese (1993) and Kleinhenz (1994), and that the same prosodic requirement is imposed on the remnant as well (see (13) above).The prosodic word status of the remnant was put forward as a general requirement on DUI, as it is not only supported by the EP data, but is also apparent in the Dutch and German data.There are some Standard Italian and Brazilian Portuguese facts, however, that seem to cast doubt on the general nature of this requirement.We will consider each case in turn, and show that under a more careful analysis of the facts these languages are as strict as EP is as to the prosodic nature of the remnant.
According to Peperkamp (1997: chap. 3: 71-74), monosyllabic prefixes do not form independent prosodic words in Italian.Nevertheless, some of them may be involved in DUI, as (29) shows.
(29) prebellico o postbellico 'pre or postwar' (Peperkamp, 1997: 137) The arguments given in Peperkamp against the prosodic word status of monosyllabic prefixes in Italian are not entirely convincing of their non prosodic word status.Monosyllabic prefixes are said not to satisfy the minimal word requirement that prosodic words are minimally composed of a disyllabic foot.Peperkamp assumes this principle to be active in Italian because the large majority of words in this language are not monosyllabic.However, Italian is not different in this respect from (both varieties of) Portuguese, where there is clear evidence that such requirement is not active (cf.Bisol, 2000, for Brazilian, andVigário, 2001, for European Portuguese; see also Bafile, 1997, who argues against minimal word requirements in Italian).Peperkamp also notes that the vowel of the prefix pre-is mid whereas similar etymological vowels (ae) surface as low in stressed position.However, this sort of argument is not convincing in synchronic terms.So, if minimal word requirements are excluded, there is no compelling evidence against the prosodic word status of some monosyllabic prefixes in Standard Italian, like pre-.
Unlike pre-or post-, prefixes such as ri-and dis-, which are also assumed not to form independent prosodic words in Italian, do not admit DUI (see the contrast between ( 29) and ( 30)).This seems to indicate that pre-and post-but not ri-and dis-have prosodic word status in this language.
(30) *(ri(fare)) e (dis(fare)) 'redo and undo' In short, all these facts strongly suggest that the prosodic condition on the remnant holds for Italian as well.
Brazilian Portuguese (BP) facts appear to suggest that DUI in this language is not subject to the same requirements as the other languages under observation.Examples like those in (31) are reported in Moreno (1997) (cited in Schwindt, 2000) to be possible in BP.
(31) exportar e importar 'export and import' inclusive e exclusive 'incluvise and exclusive' In EP, by contrast, the exact same coordinate words do not allow deletion.This is so, because these prefixed words are not formed of two prosodic words in EP (as shown by the fact that the prefix is not perceived as stressed and the vowel of ex-surfaces as a schwa or is deleted).Schwindt (2000) argues that the BP prefixes, like the EP ones, are not prosodic words and are incorporated into the prosodic word of the morphological base.Under such an analysis, our requirement that the remnant must form an independent prosodic word would not apply to BP.To support his claim on the prosodization of prefixes, Schwindt observes their behavior with respect to a number of phonological processes: word final stressless vowel neutralization, external vowel sandhi, pretonic stressless vowel neutralization, vowel harmony and nasal assimilation.Interestingly, the only process that could support the incorporated status of one of these prefixes (in-) is the behavior of nasal consonants, which is not a pure phonological process and therefore may not be used as an argument for the prosodization of this prefix (see Vigário, 2001, for the demonstration of the latter point with EP data).By contrast, the fact that vowel harmony does not apply in exclusive [e]/*[i], but applies in estimulo [e]/[i] or estudioso [e]/[i], suggests that the prefix is not prosodically incorporated into the prosodic word of the morphological base.
Like in Italian and EP, other prefixes exist in BP whose prosodic status is clear.This is the case of re-and des-that do not form independent prosodic words, as shown in Schwindt (2000) e.g.re-does not undergo prosodic word final neutralization (see (32a)) nor external vowel sandhi (see (32b)).These facts show that it does not behave like an independent prosodic word.As predicted by our analysis, this sort of prefixes does not allow DUI (cf.(33a)): given that they do not form independent prosodic words, they may not function as remnants.Notice that Schwindt (2000) claims that deletion is not found with these suffixes because they do not contrast semantically with other prefixes.However, this is not accurate, as shown in (33b), where the structure is good without deletion.In short, the BP data also confirms the prosodic condition on the remnant.
(33) a. *desfez e refez a análise mil vezes *refez e desfez a análise mil vezes b. desfez e refez a análise mil vezes '(He) undid and redid the analysis a thousand times' We conclude that all the facts we are aware of converge towards the validity of the prosodic requirement in ( 13), here repeated under (34), in all the languages observed: (34) The constituent left as the remnant of deletion under identity must form a prosodic word

Suffixed words in Portuguese and Spanish vs. French and Italian
Several languages allow DUI with suffixed words, such as Dutch and German (see examples (2-3) above), and EP and Spanish (see ( 35)).
(35) alegremente ou tristemente EP 'happily or sadly' únicamente y exclusivamente Sp. 'only and exclusively' However, in languages like Italian and French, which also allow DUI (see the examples in (1) above), there are no suffixed words that accept deletion.This is illustrated by the contrast between ( 35) and ( 36).
(36) *normalmente e logicamente It. 'normally and logically' *lentement mais surement Fr. 'slowly but confidently' It could be thought that these languages are subject to different conditions on DUI, namely morphological instead of prosodic conditions.We propose that this is not the case, and that the difference results instead from the dissimilar prosodic structure associated to suffixed words in these languages.In fact, according to the work of Hannahs (1995) and Peperkamp (1997) on the prosodic word structure of French and Italian, respectively, in these languages there are no suffixes that form prosodic words independent of their morphological base.Thus, in these languages suffixes like -mente behave like suffixes that do not form independent prosodic words in EP, such as -mento, which does not allow deletion under identity either (see (37)).

The position of the deleted unit: German and Dutch vs. Romance languages
The preceding cases showing differences among languages are related to different prosodizations of complex words, but other sources of cross--linguistic variation on DUI structures also exist.In this section, the position of the deleted prosodic word is examined.
Whereas in European Portuguese and most other Romance languages only the final prosodic word of the first coordination member may delete under identity, in Dutch and German the first prosodic word of the second member may be deleted as well, as illustrated in (38).
(38) landbouw en tuinbouw Dutch (from Booij 1985: 147-8) 'agriculture and horticulture' regelordening en regeltoepassing 'rule ordering and rule application' This difference in behavior could result from an independent parametrical choice.However, we will argue that this results instead from differences in other properties of the languages at stake.
It could be hypothesized that the reason for the difference resides in the fact that in Romance, but not in Germanic languages, syntactic structures exist that are identical to the ones that arise from deletion under identity.In such cases deletion would be blocked simply because it generates ambiguity (see section 3.1).The contrast between Romance and Germanic languages would therefore follow from the different syntactic properties of the two groups of languages.However, ambiguity cannot explain all cases where the first member of the second coordinate is never deleted in Romance.In (39) there is no syntactic structure identical to the output of deletion, as the corresponding string without deletion is syntactically ill-formed.
(39) *um vestido verde-seco e verde-garrafa (*um vestido verde-seco e garrafa) 'A dark green dress and bottle (green)' *um casaco azul-petróleo e azul-mar (*um casaco azul-petróleo e mar) 'A coat petrol blue and sea (blue)' We propose that the observed difference between Germanic and Romance languages results from a prosodic difference between the two language groups, namely, the stress pattern of complex words formed by more than one prosodic word.These sort of complex words are shown in Vigário (2001) to be prosodized under a compound prosodic word. 6In languages like Dutch, the main stress of the compound prosodic word is assigned to its leftmost prosodic word (see ( 40)).In languages like EP, by contrast, the prominent unit of the prosodic compound is its last prosodic word (see ( 41)).
(40) Compound stress in Dutch (vakántie)  (fòto)  (compound word) (Zonneveld et al., 1999) 'holiday snapshot' (kléur)  (lòos)  (derived word) (Booij, 1995:115) 'colourless' (41) Compound stress in EP (Vigário, 2001) (pòrta)  (óculos)  (compound word) 'eyeglasses case' (alègre)  (ménte)  (derived word) 'happily' We hypothesize that the first prosodic word of the second coordinate member may only delete if it bears compound prosodic word stress (and the result of deletion is not ambiguous).As only in Germanic, but not in Romance languages, compound stress is initial, we predict that only in the languages of the former group, but not in those of the latter, deletion may target the first prosodic word of the second coordinate member.The motivation for the prominence condition proposed is likely to be the following.In languages like Dutch and German, when a stressed prosodic word is deleted, the remnant receives the level of stress of the deleted prosodic word, that is, a level of stress higher than the one that it would have if deletion had not applied (cf. (42a)).In languages like European Portuguese, by contrast, if the deleted unit were the first prosodic word of the second member of the coordination, the resulting stress pattern would be similar to the one found in structures without deletion (cf.(42b)).Note, however, that the deletion of the final prosodic word of the first coordination member, which is the only acceptable pattern in EP, yields a promotion of the prominence level of the remnant, just like in Germanic languages (cf.(42c)).Given that in all of the cases surveyed the remnant always ends up with a higher level of prominence, we have to conclude that this promotion in prominence is a property of the deletion process.In other words, we propose the following prominence condition on DUI: (43) Prominence condition: DUI changes the prominence pattern by promoting the remnant to a level of prominence higher than the one it would bear in structures without deletion.
This prominence condition may well serve the goal of enhancing the recoverability of deleted information (see Booij, 1985, for a suggestion along these lines, and Frota and Vigário, 2002, for a parallel case in constructions such as topicalization, which induce prominence effects on the remnant).

Prominence on the remnant
In Booij (1985), it is shown that the remnant (and its counterpart) is stressed and must function as a focus constituent in Dutch and German.Kleinhenz (1994) further argues that the assignment of focus to the remnant in German is a result of deletion under identity.In EP (and in Romance languages in general), by contrast, we find no focus on the remnant.
The following facts show that neither the remnant nor its counterpart are focused in EP: there is no focus pitch accent (H*L) on these units, there is no I-level prominence shift in sequences with non-final stressed units and no narrowing of the pitch range after the non-final stressed unit within its intonational phrase, unlike in all cases of phonological focus assignment (Frota, 2000;Vigário, 1998).This absence of focus on the remnant seems to be also true for other Romance languages, like Italian, where native speakers agree on the non-focused status of the remnant (Avesani, p.c.).
The difference between languages like EP or Italian and languages like Dutch and German can be accounted for on the basis of their distinct stress patterns and the prominence condition on DUI.We propose that the variation found results from the interaction between the different stress patterns of complex words in the two groups of languages (see (41-42) above) and the requirement that the remnant must bear a level of prominence higher than the one it would have in similar structures without deletion (see ( 43)).As stress in complex words is often initial in Germanic languages (e.g.Booij, 1995;Wiese, 1996;Nespor, 1999), the deletion of the final  of the first coordinate member does not result per se in the assignment of extra prominence to the remnant.This is shown in (44a), where kleur bears both the compound stress and the -level stress before and after deletion.
(44) a. ((kléur)  (lòos)  )  = ((kléur)  )  Dutch (Germanic languages) b. no stress-level modification  focus In Romance languages, by contrast, complex words are never stress-initial.This was demonstrated in Vigário (2001) for European Portuguese on the basis of various phonological arguments (e.g.stress perception, blocking of vowel deletion rules, focus assignment).Consequently, the deletion of the final  of the first coordinate member leads to the assignment of -level stress to the remnant, which gets promoted from (secondary) word-stress to -level stress, as shown in (45a).The proposal just presented has two consequences.First, it predicts that in languages like Dutch the deletion of the first  of the second coordinate member does not entail focus assignment, since it is already licensed by (46a), i.e. the remnant gets the higher level of stress that was carried by the target unit (see section 3.4).Second, the unmarked way of complying with the prominence condition on DUI is (46a), whereas (46b) should be regarded as a last resource measure and be only active when a structure needs to be saved.If these considerations are on the right track, DUI universally targets the final  of the first coordinate member and some languages, like Germanic languages, get the deletion of the first  of the second member for free, just because their stress patterns make it directly prominence compliant.unit that is not phonetically realized, and not the first one.This is shown by the example in (48) above.Finally, syntactic units that are not prosodic words, namely phonological clitics, may also be absent from the phonetic string, as shown by the following examples (example (51) is taken from Matos, 1997: 707).8 (50) Todos o aplaudiram e festejaram all 3-Sg-ACC applauded and celebrated '(they) all applauded and celebrated it' (51) Eles tinham-nas ouvido às avós e contadoaos filhos they had 3-Sg-ACC-FEM heard to-the grand-mothers and told to-the suns 'They had heard them from their grand-mothers and told them to their suns' Since the target of DUI must be a prosodic word, it follows that the deletion of phonological clitics cannot be an instance of the same phenomenon.
The two phenomena can be distinguished in a principled way if their proper objects are considered: in DUI, the deleted unit is a prosodic word, not a morphological constituent or a syntactic word; by contrast, in ellipsis the deleted unit is a syntactic word, which does not have to coincide with a prosodic word.DUI, which has the properties of a phonological process, targets phonological objects, whereas ellipsis, which has been described as a syntactic process, manipulates syntactic objects.

Conclusion
In this paper, new data was presented that contributes to a cross-linguistic understanding of deletion under identity.The phonological nature of this process is supported by an array of prosodic requirements on the deleted unit, its counterpart, the remnant, and the way the coordinate sequence is prosodically phrased, which is summarized in Table 1 (first row).A pure morphosyntactic analysis of DUI is therefore unattainable.It was also seen that the variation found across languages is not due to differences in the way DUI operates, but instead to prosodic differences that the languages independently display, namely the prosodic constituency of suffixed words in Romance languages and the stress patterns of compound prosodic words typical of Romance and Germanic languages.On the other hand, it was also shown that DUI cannot be seen as a pure phonological phenomenon either, as syntax constrains DUI as well (Table 1, second row).However, no reference to mor-phological or semantic properties is needed to account for DUI.This interplay between phonology and syntax calls for a look at syntactic ellipsis phenomena.We believe to have shown that DUI is fundamentally different from syntactic gapping (Table 1, third row).The interplay between phonology and syntax in DUI includes it in the set of phenomena already known to require both syntactic and phonological information (cf.Zec & Inkelas, 1990;Guasti & Nespor, 1999;Frota & Vigário, 1996, 2002).Like in the other cases mentioned in the literature, the syntactic information at stake seems to be available in the 'stylistic' component of the grammar (see Rochemont & Culicover, 1990;Guasti & Nespor, 1999, among others), and the phonological information involved (namely prosodic phrasing and prominence) is restricted to prosodic phonology, which is precisely an interface component between syntax and phonology (see Fig. 1).

Prosodic
In conclusion, the present account of deletion under identity places it at the syntax-phonology interface without any further addition to a constrained model of syntax-phonology interaction.

Figure 1 .
Figure 1.A view of the syntax-phonology interaction .

Table 1 .
Deletion under identity: summary