Q2Stress: A database for multiple cues to stress assignment in Italian

Spinelli, Giacomo; Sulpizio, Simone; Burani, Cristina

doi:10.3758/s13428-016-0845-7

Q2Stress: A database for multiple cues to stress assignment in Italian

Published: 30 December 2016

Volume 49, pages 2113–2126, (2017)
Cite this article

Download PDF

Behavior Research Methods Aims and scope Submit manuscript

Q2Stress: A database for multiple cues to stress assignment in Italian

Download PDF

Giacomo Spinelli¹,
Simone Sulpizio^2,3 &
Cristina Burani^4,5

1533 Accesses
7 Citations
1 Altmetric
Explore all metrics

Abstract

In languages where the position of lexical stress within a word is not predictable from print, readers rely on distributional information extracted from the lexicon in order to assign stress. Lexical databases are thus especially important for researchers willing to address stress assignment in those languages. Here we present Q2Stress, a new database aimed to fill the lack of such a resource for Italian. Q2Stress includes multiple cues readers may use in assigning stress, such as type and token frequency of stress patterns as well as their distribution with respect to number of syllables, grammatical category, word beginnings, word endings, and consonant-vowel structures. Furthermore, for the first time, data for both adults and children are available. Q2Stress may help researchers to answer empirical as well as theoretical questions about stress assignment and stress-related issues, and more in general, to explore the orthography-to-phonology relation in reading. Q2Stress is designed as a user-friendly resource, as it comes with scripts allowing researchers to explore and select their own stimuli according to several criteria as well as summary tables for overall data analysis.

Making Sense of Nonce Word Stress in English

Canadian English Word Stress: A Corpora-Based Study of National Identity in a Multilingual Community

Sensitivity to derivational morphology as cues to lexical stress among English as second language learners

Article 12 December 2023

Mapping orthography onto phonology is the basic skill underlying literacy. A full understanding of how this is done is the main challenge for researchers interested in the interface between orthography and phonology. Most empirical investigation has focused on segmental material and addressed the relationship between graphemes and phonemes; in recent years, however, researchers have started to properly consider information specified at the suprasegmental level. The assignment of stress, in particular, has proved to be especially challenging for models of polysyllabic word recognition and reading (e.g., Pagliuca & Monaghan, 2010; Perry, Ziegler, & Zorzi, 2010; Rastle & Coltheart, 2000).

In languages where stress position is not fixed, such as English, Spanish, or Italian, stress may not be predictable from print in a straightforward manner, i.e., by means of rules alone. Rather, readers may use multiple sources of information in order to assign stress, including lexical look-up, distributional information, and grammatical category. For instance, consider the English word JUStice.^{Footnote 1} Where does information concerning its stress pattern come from? In the first place, readers may recall that JUStice is stressed on the first syllable – rather than the second syllable, as in poLICE – by retrieving its entry in the lexicon (Perry et al., 2010; Rastle & Coltheart, 2000). Second, they may rely on distributional information. In English, first syllable stress is the dominant pattern and may act as a default (Brown, Lupker, & Colombo, 1994; Cutler & Carter, 1987; Monsell, Doyle, & Haggard, 1989; Rastle & Coltheart, 2000). Besides, first syllable stress is also the typical pattern of English nouns – verbs, on the contrary, show a bias towards second syllable stress, as in adVISE (Arciuli & Cupples, 2006; Kelly & Bock, 1988; Sereno, 1986). Finally, most English words ending in -ice bear stress on the first syllable (Arciuli & Cupples, 2006).

In the last 30 years, several studies have shed light on the role of different cues to stress assignment in free-stress languages such as English or Italian and how they interact within the reading system. However, we would not have made much progress without the aid of corpus analysis. With the growing interest in statistical learning approaches, lexical databases have turned out to be a fundamental support for research on stress (as well as many other linguistic aspects): As a source of distributional information, databases have played a crucial role in driving hypotheses about stress assignment and inspiring behavioral and modeling studies. The availability of a lexical database is a prerequisite for those who are interested in research on stress. Quite disappointingly, however, this support is not available for some languages. For English, researchers can count on several resources – most notably CELEX (Baayen, Piepenbrock, & van Rijn, 1993) and the analyses made by Howard and Smith (2002) and Arciuli and Cupples (2006, 2007). Despite the remarkable number of studies concerned with lexical stress (see Sulpizio, Burani, & Colombo, 2015, for a review), no useful database for the investigation of stress assignment in Italian is currently available to the scientific community. Information concerning the distribution of stress patterns is also absent from (Italian) child corpora. With the present paper, we aim to fill this gap.

We begin by briefly reviewing the findings from the studies that addressed stress assignment in Italian and the resources available so far. Then, we present Q2stress, a lexical resource that provides information for multiple variables related to stress and constitutes a powerful tool for researchers interested in investigating stress issues in word recognition and reading. The following section reports the main results concerning the distribution of stress patterns in both adult- and child-directed Italian corpora. Finally, we discuss some of the implications these data have for future research.

Stress assignment in Italian

Italian is especially appealing for research on stress. Like Spanish and unlike English, it is highly transparent at the segmental level, i.e., there is an almost one-to-one correspondence between graphemes and phonemes. Like English and unlike Spanish, however, Italian is highly opaque at the suprasegmental level, as in most cases stress patterns cannot be derived from orthographic rules or stress diacritics.^{Footnote 2} Thus, on the one hand, stress in Italian can be studied rather independently from segmental phonology. On the other hand, the fact that in most cases stress is unpredictable allows researchers to inspect carefully what happens when no straightforward way to assign stress is available. Inquiring cues to stress is, in fact, much harder for languages where orthographic rules apply, as rules (e.g., diacritics that mark the stressed syllable) may cover the vast majority of word forms (for Greek see, e.g., Protopapas, Gerakaki, & Alexandri, 2006; 2007). In sum, Italian presents a valuable opportunity to inquire how stress-processing occurs in the reading system.

Research addressing reading aloud and word recognition in Italian has highlighted several factors involved in stress assignment. Similar to English, word knowledge is a primary source of information for assigning stress to known words (Colombo, 1992). Syllabic structure also plays some role, as words with a closed penultimate syllable (i.e., ending with a consonant) bear penultimate stress, although some exceptions apply (e.g., MANdorla ‘almond’, FINferli ‘chanterelles’, POlizza ‘policy’). This behavior will be referred to as Italian’s phonological rule.

Most studies investigating stress assignment in reading focused on distributional information, which can be represented at different levels of specificity. At the most general level, there is a dominant stress pattern in Italian: most words bear penultimate stress (i.e., stress on the penultimate syllable). In reading aloud, beginning readers show a bias towards the dominant penultimate stress (Colombo, Deguchi, & Boureux, 2014; Sulpizio & Colombo, 2013); with the increase of reading skills, however, the bias disappears, since skilled readers rely on a more specific source of information, namely stress neighborhood (Burani & Arduino, 2004; Burani, Paizi, & Sulpizio, 2014; Sulpizio, Arduino, Paizi, & Burani, 2013). Stress neighborhood is the proportion of words sharing the same orthographic word ending and stress pattern on the total number of words with that ending. For instance, the stress neighborhood for the word ending -ola is mainly associated with antepenultimate stress, since most words ending in -ola bear antepenultimate stress. Because of its mainly antepenultimate stress neighborhood and despite its non-dominant stress, an antepenultimate stress word ending in -ola, such as PENtola ‘pot,’ is read faster and more accurately than a penultimate stress word with the same ending, such as piSTOla ‘gun’ (Burani & Arduino, 2004; Burani et al., 2014; Colombo, 1992). Stress neighborhood affects also nonword reading: A nonword ending in -ola is more likely to be assigned antepenultimate than penultimate stress (Sulpizio et al., 2013).

Taken together, findings from reading aloud indicate that, with a limited lexicon and weak orthography-to-phonology connections, younger readers assign stress mostly relying on general distributional information (i.e., stress dominance). As reading skills increase, general information is replaced by specific distributional information (i.e., stress neighborhood), which becomes the main source of information for stress assignment. In contrast, research on visual word recognition offers a very different picture. Studies conducted on adult readers suggest that stress neighborhood may play a weak role in lexical decision tasks (Burani & Arduino, 2004; Colombo & Sulpizio, 2015), whereas an advantage for (dominant) penultimate stress words was reported (Colombo, 1992; Colombo & Sulpizio, 2015). In sum, stress distributional information is specified at different levels, and both different stages of development and different tasks may recruit such information at different levels.

Cues to stress

While the literature focused on stress dominance and stress neighborhood, little is known concerning other factors that may be involved in stress assignment, such as grammatical category and word beginnings. Some studies conducted in languages other than Italian suggest that both grammatical category and word beginnings may have a role in driving stress assignment. Using both reading aloud and lexical decision tasks, Arciuli and Cupples (2006) and Jouravlev and Lupker (2014) found that grammatical category interacts with stress neighborhood in English and Russian, respectively (for English see also Kelly, 1988; Kelly & Bock, 1988; Smith et al. 1982). Arciuli and collaborators also found that not only word endings but also word beginnings affect English stress assignment in both adults (Arciuli & Cupples, 2007) and children (Arciuli, Monaghan, & Seva, 2010). According to a recent corpus analysis (Monaghan, Arciuli, & Seva, 2016), word beginnings, though less predictive than word endings, are good predictors of stress patterns in several languages, including Italian. Importantly, the role of both beginnings and endings in stress assignment seems to be independent from the morphological properties of those units (Arciuli & Cupples, 2006, 2007). That said, it is known that affixes either repel or attract stress (e.g., Jarmulowicz, Taran, & Hay, 2008; Ktori, Tree, Mousikou, Coltheart, & Rastle, 2016; Rastle & Coltheart, 2000), suggesting an important role for morphology in stress assignment.

Other variables that received little or no attention are orthographic consonant-vowel (CV) structure (i.e., the sequence of consonants and vowels composing a word; e.g., the orthographic CV-structure of macchia is CVCCCVV) and number of syllables. Whether the association between stress patterns and CV-structures in the lexicon may affect stress assignment is an open issue. The same is true for the possible role played by the number of syllables. One may ask, e.g., whether words differ in the distribution of stress patterns according to the number of their syllables, and how this information interacts with other sources for stress. In particular, a still unanswered question concerns which words should be considered when computing stress neighborhood. While so far stress neighborhood has been computed by counting all words sharing an ending regardless of the number of their syllables, it is still unclear whether words with a different number of syllables may give a different contribution to stress neighborhood.

To sum up, research focused on stress dominance and stress neighborhood, considering them the main factors affecting stress assignment in Italian. However, we still do not know whether and to what extent other properties of words also work as cues to stress assignment. Good candidates include grammatical category, word beginnings and endings of different length, CV-structures, and number of syllables.

Resources for investigating Italian stress

The lack of research on the role of the properties discussed above in stress assignment may be due to the inadequacy of current databases. Since Colombo’s (1992) seminal study, investigations addressing stress assignment in Italian have relied on three major frequency dictionaries: The dictionary by Bortolini, Tagliavini, and Zampolli (1971), the Barcelona corpus (Istituto di Linguistica Computazionale 1989), and CoLFIS, a more recent corpus and frequency count of more than 3 million written Italian word occurrences (Bertinetto et al., 2005; Laudanna, Thornton, Brown, Burani, & Marconi, 1995). Since no one of the above databases included stress information, researchers annotated stress information manually. Thus, those resources were unsuited to give a picture of how stress patterns distribute in Italian. To date, the only study that has provided such information is BDVDB: A database for the Italian basic dictionary (Thornton et al. 1997). According to Thornton and collaborators (1997), most Italian words with more than two syllables have penultimate stress (85.3%), some have antepenultimate stress (12.7%), and very few have ultimate stress (1.8%).^{Footnote 3} The same is true when the number of syllables and grammatical category are taken into account: Penultimate stress is always the dominant pattern, with a minority of antepenultimate stress forms. However, some doubts can be cast upon such data. BDVDB is based on just the 5,000 most frequent (plus the 2,000 most “available”) Italian words, collected by De Mauro (1991) as a list of lemmas. As a result, information is provided for neither less frequent words nor word forms in general. Note that in Italian each lemma may correspond to several word forms, and that this is especially true for verbs, which show different inflectional markers for each of the twelve simple tenses, each of the three conjugations, and each person. Most importantly, verbs show alternating stress patterns and number of syllables within inflectional paradigms. Thus, counts made on lemmas may differ dramatically from those made on word forms. Besides, nouns are possibly given too much weight in BDVDB, since they cover the vast majority of words in De Mauro’s (1991) list (4,580 on 7,000).

Another major limitation in current research is the lack of a child database providing stress information. Indeed, the distribution of stress patterns in the written materials children are exposed to may have different stress properties with respect to adult materials. So far, however, statistics for studies conducted on children were drawn from BDVDB, and the Lessico elementare (Elementary Lexicon, Marconi, Ott, Pesenti, Ratti, & Tavella, 1993) served as the main source for experimental materials. Being built on both child-directed texts and texts written by children, the latter database is arguably the most appropriate resource for children. However, it does not provide any phonological information, and thus, researchers had to annotate stress manually. As such, longitudinal studies on the development of stress assignment have been fairly impossible to date.

The present work aims to make up for these and other limitations. In the following, we describe Q2Stress (Cue-To-Stress), a lexical database providing information for multiple variables of interest for research on stress assignment in Italian. As the name suggests, Q2Stress relies on the claim that several sources of distributional information may act as cues to stress assignment, including not only stress dominance and stress neighborhood, but also variables for which no empirical evidence is currently available for Italian, such as grammatical category, word beginnings and endings of different length, CV-structure, and number of syllables. Q2Stress will thus be a powerful tool to shed light on issues such as: (1) which are the independent roles of the different variables in assigning stress, (2) how these variables may interact, and (3) which cross-linguistic differences and similarities are characteristic of stress assignment in Italian. In addition, Q2Stress includes data coming from both child-directed and adult corpora, thus allowing researchers to examine to what extent differences in performance between children and adults are due to either the properties of their lexicons or the development of cognitive skills. Finally, being derived from large corpora of word forms (rather than lemmas), Q2Stress is not affected by the drawbacks of Thornton and collaborators (1997).

Methodology

Q2Stress has been built up on the basis of both an adult and a child-directed corpus. The basis for the first (adult) part of Q2Stress is phonItalia (Goslin, Galluzzi, & Romani, 2014), an open access lexical database providing orthographic and phonological information for 120,000 word forms extracted from CoLFIS (Bertinetto et al., 2005; Laudanna et al., 1995). The choice of phonItalia instead of CoLFIS was motivated by the fact that the former, but not the latter, contains information about the stress pattern of word forms. Q2Stress is based on phonItalia version 1.1 (http://www.phonitalia.org) and derives all of its fields from phonItalia’s main lexicon (“phonItalia 1.10 – word forms.txt”).

The second part of Q2Stress derives from Lessico elementare [Elementary lexicon] (Marconi et al., 1993), a frequency count of about 1,000,000 occurrences based on materials equally divided between those written by adults for children (age range: 6–10 years) and those written by children themselves. Since our focus is on how reading, rather than writing, is affected by distributional knowledge about stress, we performed our analyses only on the child-directed half (about 25,000 word forms). For each word form, the corpus included information concerning frequency, lemma, and grammatical category. Grammatical category markers were converted to phonItalia’s codification in the first place. Then, stress information, as well as phonological representations, number of syllables, syllabic boundaries, number of letters and phonemes, and orthographic and phonological CV-structure, were obtained from phonItalia by comparing the two databases. A remarkable proportion of words included in the child-directed corpus (over 20%) did not occur in the larger adult corpus, which suggests that the two databases differ considerably. As we shall see, the child-directed corpus shows a relatively higher number of verbs than the adult corpus. An analysis of the child-exclusive forms reveals that about half are verbs, especially in enclitic form (e.g., acchiapparci “to catch us”). The remaining forms are typically found in texts for children, such as diminutive and augmentative nouns and adjectives (e.g., regalini “little gifts”), and interjections. Information for those forms was obtained using the phonological transcription module from the Italian Festival text-to-speech system (Cosi, Gretter, & Tesser, 2001), which is the same software used by Goslin and collaborators (2014). The output of the software was manually checked. Cases for which stress position is uncertain in Italian (e.g., RUbrica vs. ruBRIca “address book”) were assigned the first stress pattern listed in the De Mauro Italian Dictionary (De Mauro, 2000).

Note that the purpose of the phonological transcription process was to obtain the stress pattern and number of syllables for each word form. Phonological representations (e.g., /’kɔsto/ “cost”), though available, were not used. Although Italian is a transparent language, phonological representations are sometimes less ambiguous with respect to stress position than orthographic representations are. For example, in such Italian varieties as Tuscan and Roman, open-mid vowels (e.g., /ɔ/) are more predictive of stress position than close-mid vowels (e.g., /o/) are, but in print, they correspond to the same grapheme (i.e., <o>). As Q2Stress aims to model the Italian reader’s knowledge about stress when he/she is engaged in reading aloud, only orthographic units shall be used, thus maintaining the ambiguity of Italian spelling with respect to stress position.

Monosyllabic words were excluded from the analysis of both databases, as they are of no interest for stress assignment (at least in single-word reading). Although stress assignment to disyllabic stimuli is straightforward in Italian (i.e., they have penultimate stress unless a stress diacritic marks the last syllable), disyllables were included since they may contribute to the overall distribution of stress patterns, e.g., by enhancing a bias towards the dominant pattern. Finally, all words with three or more syllables were included.

Analyses were made on both word types and tokens.^{Footnote 4} Among types, homographs and words occurring in idioms were counted just once. Thus, type analyses were made only on non-identical word forms, which amount to 94,051 in phonItalia and 22,081 in Lessico elementare. Tokens, on the other hand, amount to 2,179,735 in phonItalia and 272,754 in Lessico elementare.

In both the adult and the child-directed parts of the database, we considered the following variables of interest for stress assignment: (1) number of syllables; (2) grammatical category; (3) word beginnings (from one to four letters long); (4) word endings (from one to four letters long); (5) word endings’ CV-structure (three-letter endings only); and (6) whole-word CV-structure. Unfortunately, because phonItalia does not include morphological information, it was not possible to investigate the role of affixes on stress assignment at this time.

A few remarks should be made concerning the variables (3), (4), and (5). With regard to word beginnings, since no criteria have been developed until now, we chose to consider word-initial units of increasing length, from one up to four letters. Corpus analyses show that longer beginnings are better predictors of stress position than shorter ones (Monaghan et al., 2016); however, future research will shed light on which size best accommodates results from empirical investigations.

Being the focus of several investigations, word endings are more clearly defined. In Italian, stress neighborhood refers to word endings starting from the nucleus of the penultimate syllable. According to this definition, word endings may vary in length. However, since three-letter endings (with a VCV structure, e.g., ab-ete “fir”) are the most frequent in Italian, almost the totality of the literature is concerned with them. Endings may be simply defined in terms of their length, though (e.g., Monaghan et al., 2016). Thus, it might be the case that endings as short as one or two letters provide a sufficient account of how stress is assigned in Italian. Alternatively, readers might use units as long as three or even four letters. As an approach to answering this empirical question, and mirroring the data for word beginnings, we provided data for endings ranging from one to four letters, regardless of the position of the nucleus of the penultimate syllable. We decided, however, to restrict information concerning word endings’ CV-structure to three-letter endings only. This was done in the interest of simplicity, as inspection of word endings’ CV-structures revealed that three-letter endings are by far the most informative in terms of stress pattern distribution. Accordingly, we will spend more time commenting on three-letter endings relative to other endings in the following sections.

The database

Q2Stress is freely available at http://www.istc.cnr.it/grouppage/databases. Being thought as a device for researchers, Q2Stress comes with two tools subserving two main research purposes: a set of functions for exploration and selection of stimuli (see the “scripts” folder), and a set of Summary Tables for overall data analysis (see the “summary tables” folder).

The functions are written in R (https://www.r-project.org/) and can be easily used also by R neophytes. The functions allow the user to specify a word beginning (e.g., to-) or a word ending (e.g., -ola) and a stress pattern (e.g., penultimate stress), and obtain all words starting with that beginning and bearing the specified stress pattern (e.g., torrone “nougat,” tovaglia “tablecloth,” etc.). More detailed queries taking into account other parameters (e.g., the number of syllables, grammatical category) can also be done. Note that the output of each query is saved in a txt file. Instructions to use the functions can be found in the Supplementary materials.

The summaries of the overall type and token counts for word beginnings, word endings, and CV-structures can be found in the tables, which report data for adults and children separately. Summary tables report the percentage and the number of words with each type of stress pattern (ultimate, penultimate, antepenultimate, or preantepenultimate) in which each unit (beginning, ending, or CV-structure) appears. For example, they show that in the adult type count the word ending -ero appears in 24 penultimate stress words and 220 antepenultimate stress words, corresponding to 9.8% and 90.2% of all words ending in -ero, respectively. More detailed information about summary tables can be found in the Supplementary Materials. In what follows, we report some of the main findings resulting from these tables.

Descriptive analyses for the adult frequency count

We report descriptive analyses for phonItalia 1.1. The variables we focus on distribute as follows. Three-syllable and four-syllable words are the most frequent in terms of types (31.3% and 35.5%, respectively; disyllables: 11.4%), but are outweighed by disyllables in terms of tokens (two syllables: 44.5%; three syllables: 31.6%; four syllables: 16.7%). Verbs, nouns, and adjectives are the most frequent grammatical categories when either types (35.8%, 28.5%, and 20.1%, respectively) or tokens (25.8%, 33.7%, and 14.7%, respectively) are considered. Finally, VCV word endings cover more than half of the types (56.9%) and half of the tokens (49.6%) of the database, followed by CCV (30.8% of types, 36.2% of tokens) and CVV word endings (6.4% of types, 9.6% of tokens). Data from word beginnings and endings, as well as whole-word CV-structures, are more sparse.

Overall distribution

As shown in Table 1, our data confirm the picture provided by Thornton and collaborators (1997) on lemmas, with the frequency order of Italian stress patterns being: (1) penultimate, (2) antepenultimate, and (3) ultimate.^{Footnote 5}

Table 1 Overall distribution of stress patterns in Italian – data from the adult frequency count

Full size table

This holds true when both types and tokens are taken into account, with a slightly stronger penultimate-stress dominance in token than in type percentages (87% vs. 76.7%): This difference is driven by disyllables, which account for 44.9% of tokens and bear penultimate stress most of the time. When disyllables are excluded, the difference between types (penultimate: 75.9%; antepenultimate: 20%; ultimate: 3.8%) and tokens (penultimate: 79.9%; antepenultimate: 17.3%; ultimate: 2.6%) is smaller. Note that type measures better fit results from studies addressing stress assignment: In a reading aloud task Burani and Arduino (2004, Experiment 2) manipulated stress neighbors in terms of types while controlling them in terms of tokens and found a larger stress neighborhood effect for word endings with higher types (see also Sulpizio et al., 2013). Accordingly, in what follows, only type counts will be discussed. The following tables and figures, however, will report measures for both types and tokens.

Number of syllables and grammatical category

Penultimate stress is the most frequent pattern even when one considers different numbers of syllables and grammatical categories. However, the higher the number of syllables in a word, the more likely the word is to have an antepenultimate stress pattern (Fig. 1).

As far as grammatical category is concerned, we found (Fig. 2) that, although penultimate stress is the dominant pattern for each grammatical category in Italian, verbs have a lower proportion of penultimate stress words as compared to nouns (70.3% vs. 84.9%). Adjectives stand half-way between nouns and verbs (76.2%), while adverbs are the most biased towards penultimate stress (95.3%). In sum, Italian shows some variability in the distribution of stress patterns within grammatical categories; however, while languages like English show a preference for the dominant (first-syllable) stress pattern in disyllabic nouns and a preference for the nondominant (second-syllable) stress pattern in disyllabic verbs (Arciuli & Cupples, 2006; 2007), the dominant pattern is always preferred in all grammatical categories in Italian. Besides, all categories except adverbs are in line with the overall tendency of a decrease of penultimate stress patterns with the increase of number of syllables.

Word beginnings

We extracted word beginnings of one up to four letters. Almost all one-letter and two-letter beginnings mirror the overall stress pattern distribution (i.e., they are associated with the penultimate stress pattern). Exceptions are few (less than 5% of all beginnings), cover few word forms, and are not strongly associated with any other stress pattern. Longer beginnings vary to a larger extent than shorter ones, as they include beginnings showing no bias towards penultimate stress (e.g., acr-, apos- appear more often in antepenultimate stress words) all the way up to beginnings appearing always in penultimate stress words (e.g., paz-).

Besides percentages, it is important to consider also the total number of words each beginning appears in. In fact, while some beginnings are very frequent (e.g., the first five one-letter beginnings appear in more than half of the word forms), others are not. Since numerosity (i.e., the raw number of words having the same cue) was found to play a role for word endings (see below), this may hold true also for beginnings. Accordingly, high-frequency beginnings may be better cues to stress than low-frequency ones. However, overall, beginnings seem to be less informative of stress position than endings, confirming the analyses reported by Monaghan and collaborators (2016).

Word endings

The most common one-letter and two-letter endings (e.g., -a, -i, -re, -to) are strongly associated with penultimate stress, and thus appear to offer few cues to stress position. Deviations from this trend concern endings that are pretty rare in Italian, such as those appearing mostly in French and English loanwords (e.g., -l, -t, -er, -in).

For three-letter endings, we focused on the words that bear no stress diacritic. These included VCV, CCV, CVV, CVC, and VVV endings. The distribution of stress patterns for such ending CV- structures is reported in Fig. 3. VCV and CCV are the most frequent endings, accounting for 56.9% and 30.8% of all word forms, respectively. Although penultimate stress is the dominant pattern for both endings, CCV endings show a higher proportion of penultimate stress words (94.1%) than VCV ones (71.4%). In fact, CCV endings include a large proportion of words that bear penultimate stress because of the phonological rule (i.e., they have a closed penultimate syllable, e.g., conCERto “concert”); they also include a small proportion of words with an open penultimate syllable (i.e., not ending in a consonant) with a stop + liquid onset in the last syllable, which can bear any stress pattern (cf. puLEdro “colt” and VERtebra “vertebra”). Statistics for CVV and VVV endings are similar to CCV endings. In contrast, CVC endings are peculiar, as they show a tendency towards ultimate stress. The reason for this is that CVC is not a regular Italian ending, and most words with such ending are either English or French loanwords.

Because of their low frequency, we will not go into details of other CV-structures. Note, however, that penultimate stress is the dominant pattern for most of them as well (for details, see types.txt and tokens.txt in the CV-ending folder). As it might be expected, we found an overwhelming number (325) of different VCV endings, which account for over 90% of all word forms for which stress cannot be derived by rule. Like beginnings, VCV endings vary in both the total number (numerosity) of words they appear in and the proportion of penultimate and antepenultimate stress patterns those words have. Both measures were found to have a role in stress assignment in Italian (Burani & Arduino, 2004; Burani et al., 2014; Sulpizio et al., 2013): In nonword reading, for example, word endings with many stress friends (i.e., words sharing the ending and stress pattern) received more responses consistent with stress neighborhood than word endings with the same percentages of stress friends but lower numerosity. Sulpizio and collaborators (2013) argue that stress neighborhood effects are to be found only for endings with both a large proportion and number of stress friends. In contrast, both small-sized and balanced neighborhoods are ambivalent, since they provide no clear cue for stress (Colombo et al., 2014). Q2Stress lists word endings of both classes. Most endings, such as -ofo and -abo, have low to medium frequency, and are thus closer to ambivalence. The same is true for endings, such as -isi and -ene, that have good-sized neighborhoods but show an almost equal proportion of penultimate and antepenultimate stress friends. The most part of word forms, however, is covered by endings that do have large neighborhoods and are clearly biased towards either penultimate (e.g., -ore, -ari) or antepenultimate stress (e.g., -olo, -ico), with the former outweighing the latter.

Four-letter endings confirm the tendency of many endings associated with few word forms and few endings associated with many word forms, either mostly stressed on the penultimate or mostly stressed on the antepenultimate syllable. Interestingly, there are remarkable differences between four-letter endings that share the last three letters: For example, -tina is strongly associated with penultimate stress (97.5%), whereas for -mina penultimate and antepenultimate patterns are pretty balanced (about 50%). Although this is not always the case, such occurrences suggest that four-letter endings may be better predictors of stress position than three-letter endings are.

When one takes number of syllables and grammatical category into account the picture becomes less clear-cut. As noted above, there is a tendency for words with increasing number of syllables to bear antepenultimate stress more often. This tendency is clearly visible in the most common one-letter endings. However, the situation changes as longer endings are considered. While some two-letter endings show a steep increase in the proportion of antepenultimate stress words as the word gets longer (e.g., -ci, -he), others show no sign of this tendency (e.g., -ni, -va). Similarly, one can distinguish three-letter and four-letter endings for which stress is and is not attracted backwards as the word length increases. Overall, the distribution of stress patterns for most endings associated with a large number of word forms changes minimally with increasing number of syllables, but a few endings (e.g., -ile, -done) have different behaviors. For instance, -ile has, overall, an antepenultimate stress neighborhood; however, three-syllable words ending in -ile show pretty balanced proportions of penultimate and antepenultimate stress words. Readers might be sensitive to such item-specific information. Indeed, Sulpizio and collaborators (2013, Experiment 2) failed to report a stress neighborhood effect for three-syllable nonwords ending in –ile. ^{Footnote 6}

For the sake of simplicity, we restrict discussion of grammatical category to three-letter endings with a VCV structure. Several such endings show remarkable differences in how stress patterns distribute within grammatical categories. Verbs, in particular, behave differently from other categories. For instance, consider -era, a word ending with a prevalently penultimate stress neighborhood. While nouns and adjectives ending in -era (e.g., panTEra ‘panther’, alTEra ‘proud’) show a preference for penultimate stress, the opposite is true for verbs (e.g., conSIdera ‘he/she/it considers’), which bear antepenultimate stress most of the time. Thus, while grammatical categories per se have little impact on the distribution of stress patterns at the general level, they seem to have a role at the item-specific level, that is when word endings are considered.

CV-structures

Almost all CV-structures are biased towards penultimate stress. Variability among CV-structures is mainly accounted for by the structure of the ending, with CCV endings showing a higher proportion of penultimate stress words than VCV and other endings. However, if the analysis is restricted to CV-structures with the same ending (e.g., VCV), some noteworthy differences remain (e.g., between VCVCV and CVCCCVCV, that bear penultimate stress 62.5% and 83% of the times, respectively).

Descriptive analyses for the child-directed frequency count

The distributional properties of the child-directed part of Q2Stress mirror those of the adult part to a large extent. Three-syllable and four-syllable words are more frequent than disyllables in terms of types (two syllables: 15.8%; three syllables: 38.1%; four syllables: 33.8%), but are outweighed by disyllables in terms of tokens (two syllables: 50.6%; three syllables: 32.9%; four syllables: 12.8%). Similarly to the adult database, verbs, nouns, and adjectives are the most frequent grammatical categories in terms of both types and tokens. However, verbs have dramatically more types than any other category (59.9% of all word types, that is twice as many as nouns). Despite being numerous, verb forms tend to occur with a low frequency, being second to nouns in terms of tokens, with similar proportions to the adult database (29.8% vs. 37.8%, respectively). Counts for ending CV-structures also mirror the adult part pretty much: VCV endings cover more than half of the types (56.1%) and almost half of the tokens (48.5%), followed by CCV (32.3% of types, 35% of tokens) and CVV endings (4.9% of types, 9.6% of tokens).