Day 3

Since Sherman (1975) and Bell (1978), typologists have been concerned with the problems inherent in sampling languages for typological research. Several biases have been raised since the outset, most prominently geographical and genealogical biases, but also bibliographical, socio-cultural biases and structural biases (Bell 1978, Rijkoff & Bakker 1998). Greenbergian typology has dealt with these statistical biases by treating them as confounds to be addressed by balancing language samples, in particular focusing on geographical and genealogical balance. Distributional Typology (Bickel 2015) argues that these biases should be tested for directly, and advocates the dense sampling of both genealogical units and particular geographic areas in order to identify tendencies of language change.

In this talk we identify a bias that has not been, to the best of our knowledge, discussed before, which we call TEMPORAL BIAS. Temporal bias has several aspects. The first (‘Chronological Bias’), stems from the fact that previous stages of human languages might differ from present-day languages, both qualitatively and quantitatively. For example, earlier human languages might have been substantively different, at least in terms of their sound systems (Blasi et al., 2019). The distribution of linguistic properties in the world’s languages may also have been different. For example, the frequency of objectinitial basic word order might have been greater in the past. As such, time-before-present may be a covert variable that shapes the distribution of linguistic properties in any sample of languages.

A second aspect of temporal bias (‘Phylogenetic Depth Bias’), relates to the fact that the languages we have data for are recursively-nested intermediate nodes in phylogenetic units; even reconstructed proto-languages, up to the level of stocks (Nichols 1992) plausibly have ancestors. Furthermore, phylogenetic units are diverse in terms of their size and diversification. Temporal bias means that typological-historical research faces problems of data sparsity: if one wants to take into account both Chronological Bias and Phylogenetic Depth Bias, one has increasingly fewer data points for each previous time period. In order to compare languages spoken at different time depths and at different nodes in a family tree, one would need to eliminate nodes from the more diversified families or conversely inflate the genealogical complexity of the less diversified families, again leading to data sparsity.

One way to overcome the temporal bias is to investigate specific language families, regardless of their age, by comparing properties of proto-languages directly with the currently available data of their daughter languages. In this way, we can compare diachronic trends across phylogenetic units. We implement this on the basis of three large-scale databases: PHOIBLE (Moran & McCloy 2019), a largescale database of present-day segment inventories; SEGBO, a database of phonological segment borrowing in the world’s languages; and BDPROTO, a database of ancient and reconstructed phonological systems. We examine whether individual phylogenetic units show different rates of change in consonants versus vowels, finding that each language family shows different preferences (some show a faster rate of change in consonants or vowels). Bickel (2013) argues that situations in which there is a cross-family preference for change while individual families show particular preferences point to (i) universal pressures leading to preferred directions of change, and (ii) the relative strength of these pressures. The findings from our study provide preliminary evidence for both pressures that lead to an expansion of consonant inventories and pressures that lead to an expansion of vowel inventories. As such, our findings suggest differential rates of change in consonant and vowel inventories in different languages families -- a finding reported by Greenhill et al., (2018), who find different rates of lexical change per language family.

Background and aims

Zipf’s (1935) law of abbreviation states that more frequent words tend to be shorter than less frequent words. The correlation between length and frequency is regarded as a language universal (Bentz & Ferrer-i-Cancho 2016). However, Piantadosi et al. (2011) have recently argued that informativity, or average surprisal, based on conditional probability of a word given left context, is a better predictor of word length than simple frequency. They used large corpus data from a sample of eleven Indo-European languages. According to Gibson et al. (2019), Zipf worked before information theory, so he could not know the length of a signal in an efficient code depends on its predictability from context, not its frequency. The aim of the present study is to investigate the relationships between formal length, frequency and predictability (a) using a more diverse sample of languages, (b) considering different corpus sizes, (c) distinguishing between different word classes, and (d) testing the predictability based on right context.

Data and method

For this investigation, we use large web-based corpora of typologically diverse languages (Arabic, Czech, English, Finnish, German, Hindi, Hungarian, Indonesian, Russian, Spanish) from Leipzig Corpora Collection (Goldhahn et al. 2012), as well as the web ngram data used by Piantadosi et al. (2011), for the sake of comparability. We use corpus samples of different sizes. The average surprisal is computed based on 1 and 2 words on the left and on the right. Word length is measured as the number of utf-8 characters. In order to measure the effect of frequency and predictability given the left and right context, we use partial correlations. The parts of speech were annotated with the help of the Universal Dependencies software (Straka & Strakovà 2017).

Preliminary results

Our pilot study reveals that the partial correlations vary greatly across the languages. While the (Indo)-European languages mostly show that frequency is less strongly correlated with predictability, confirming the results reported by Piantadosi et al. (2011), some other languages (Finnish and Hungarian, in particular), display a stronger effect of frequency. Second, corpus size matters significantly. The Zipfian effect is stronger in smaller corpora. This has to do not only with the less reliable estimates of contextual predictability due to data sparseness, but also with the fact that a larger corpus contains more low-frequency words for which discrete frequencies are not fine-grained enough. In the same vein, the Zipfian effect is stronger for high-frequency words in the sample than for low-frequency words, and more for function words (and partially verbs) than for nouns. Finally, some of the languages (Hindi and, surprisingly, Spanish) display a stronger correlation between length and predictability based on the right context than one between length and predictability based on the left context. All these results raise questions about a uniform explanation of word lengths in terms of predictability and optimal code.

Morphological paradigms differ widely across languages: some feature relatively few contrasts, and others, dozens. This variation is surprising when considering the difficulty in acquiring languages with large morphological systems, a challenge that, under usage-based theories, children must face without pre-existing knowledge (Tomasello, 2000). Recent work on morphological complexity has resolved this paradox by arguing that certain features of even very large paradigms make them easy to learn and use. Specifically, Ackerman and Malouf (2013) propose an information-theoretic measure, i-complexity, which captures the extent to which forms in one part of a paradigm predict other parts of the paradigm. They contrast this measure with another commonlyused measure, e-complexity (e.g., Bickel & Nichols, 2005), which captures the number of distinctions made by the language and the different ways to mark each grammatical function. They show that languages which differ widely in their e-complexity exhibit similarly low i-complexity; this suggests that having predictive relationships between inflections (i.e. low i-complexity) reduces the learnability challenge for learners even when the morphological paradigm makes many contrasts (i.e. has high e-complexity). While this seems like a promising solution, the idea that predictive relationships among forms in the paradigm are critical for learning has only limited evidence (Seyfarth et al., 2014). Here, we test these measures of morphological complexity, to evaluate whether i-complexity in fact influences the learnability of morphological paradigms as predicted by Ackerman & Malouf.

In this study we tested whether i-complexity predictsthe learnability of inflectional paradigmsfor recurrent neural networks (RNNs) and human participants (adults). We created two languages, matched on e-complexity but differing in i-complexity. Only in the low i- complexity paradigm, the singular form of a word always predicts its form in dual. Furthermore, we compared the effect of i-complexity on learning with that of ecomplexity, by manipulating the e-complexity of inflectional paradigms. Results from simulations with RNNs and behavioural experiments with human learners both suggest that e-complexity has a more robust effect on learning of inflectional paradigms. By contrast, the effect of i-complexity was present but weaker in neural networks and absent in human learners.

These results cast doubt on Ackerman and Malouf’s hypothesis that i-complexity rather than e-complexity drives the learnability of morphological paradigms. Our results suggest that measures such as e-complexity are more predictive of learning morphological systems than the proposed i-complexity.

For over a century, phonological theory has relied on articulatory facts in determining wellformedness of sequences of consonants. With time, frequency has been taken into account as a factor possibly explaining the structure of consonant clusters. According to the first view, clusters can be evaluated in relation to universals and phonological principles such as sonority. The proposed sonority hierarchies (e.g. Clements 1990, Vennemann 1988) have been unable to account for data coming from phonotactically rich languages such as Slavic. Similarly, such languages have ben shown to violate the universal preference for the CV structure (Greenberg 1978) by displaying numerous long sequences of consonants. Another approach states that preferred structures tend to be more frequent. According to this view, structures emerge from frequency (Bybee 2003). The goal of this paper is to test whether rankings of clusters established on the basis of universals, phonological principles and usage coincide. In order to answer the question, we analyse 4 languages of varying phonotactic complexity: moderate Germanic languages and complex Slavic language (Maddieson 2013).

The ranking of clusters based on phonological properties is drawn from Orzechowska & Wiese (2015) and Orzechowska (2019). Word-initial clusters are analysed in terms of four dimensions; complexity, place and manner of articulation and voicing. The dimensions are represented by 26 phonological parameters, which allow to observe patterns which contribute to a preferred feature configuration in clusters. For instance, English clusters tend to start with a voiceless segment (70%) and end with a voiced one (93%), while in Polish voice agreement is preferred (69%). Such percentages are further transformed into scores, which eventually lead to numerical rankings of clusters. The usage-based ranking of clusters is based on type and token frequencies extracted from dictionaries and corpora (e.g. Wawrzyńczyk et al. 2007, Leipziger Wortschatz Portal). The analysis is based on 54 English, 56 German, 423 Polish and 327 Russian clusters.

The main part of the analysis involves correlating the variables and the rankings. Several types of analyses are performed (correlation, regression). The results suggest that preferred phonological structures and their usage frequencies do not need to coincide. Phonology and frequency constitute two domains which can be used as separate evidence for the emergence of structure. The analyses also demonstrate that different types of preferences govern the structure of clusters cross-linguistically. For instance, place and voice features are particularly relevant for Polish, while English favours preferences preserving the sonority-based profile.

In this paper I examine the affordances that structural resonance offers for interaction in naturally occurring dialogue. Structural resonance is pervasive in dialogue, to the extent that interlocutors selectively reuse some of the words, meanings, and constructions used by a prior speaker, often in the same order (Du Bois, 2014; Moscoso del Prado Martín & Du Bois, 2015). Consistent ordering of resonating elements across utterances creates parallelism, inviting heightened perception of affinities, including both similarities and differences (Jakobson, 1966). The phenomenon has been interpreted in terms of structural priming (Branigan & Pickering, 2016; Pickering & Garrod, 2006). A large-scale metanalysis of the experimental priming literature confirms a robust effect of structural priming, which is amplified further by lexical repetition (Mahowald, James, Futrell & Gibson, 2016). In corpus data, some researchers have found priming effects for individual constructions (Gries, 2005), but others have questioned whether the effects are illusory, being largely reducible to lexical priming (Patrick G.T. Healey, Howes & Purver, 2010; Patrick G.T. Healey, Purver & Howes, 2014).

Turning to the question of function, structural priming has been described as a cognitive facilitation that "makes conversation easy" (Garrod & Pickering, 2004). But this begs the question: Easy for what? The structural priming approach tends to neglect the question of the specific interactional projects that interlocutors seek to accomplish their goal-driven use of lexical and structural parallelism.

To address this question, we make use of a recently released computational tool for analyzing lexical and structural resonance, called Rezonator (Du Bois, 2019). Rezonator provides a front-end markup tool that allows users to representing the complexity of multilevel resonance in dialogue.

In this paper we present in-depth qualitative analysis of naturally occurring dialogue in the Santa Barbara Corpus of Spoken American English (Du Bois et al. 2000-2005), arguing that (1) Interlocutors build multi-level resonance by combining the effects of resonance at all linguistic levels, including lexicon, syntax, and semantics, positing a surface-oriented representation that incorporates the interlocutors' collaborative contributions to building a single unified alignment structure for resonance and coherence in dialogue. (2) Having established these joint structures, interlocutors put them to use for a broad array of interactional functions, including collaborative problem-solving; stance alignment; repair; questioning; and so on.

Interactional approaches to meaning have shown that the meaning of expressions in conversation is not determined in advance, but situated, negotiated, and locally elaborated through verbal and gestural means (Linell 2009; Depperman 2005, 2011; Deppermann & De Stefani 2019). One recurring pattern whereby local elaboration is accomplished is described by Couper-Kuhlen and Thompson (2005) as concessive repair, that is, the retrospective revise of claims deemed by participants as overstated or in need of qualification. Previous research has shown that affirmative claims are typically retracted by verbal negation, or by a counter-assertion (Couper-Kuhlen and Thompson 2005; Deppermann 2014). In this talk, we explore the multimodal realization of a comparable practice – a downgrading self-repeat – in spoken Israeli Hebrew based on a collection of instances coming from conversational media talk.

In our data, we observe that speakers may retract a formulation just produced through a second production of the same formulation (a same-turn self-repeat), however with an observably different prosodic and embodied design. One embodied resource that can occur in such sequences is the Quoting Gesture (QG), as demonstrated in the following excerpt. This excerpt features an Israeli actress describing how she was once forced to stay in New York due to bad weather conditions:

1 Liraz hayta li diRA m:aksima-

was.3sgf to.me apartment lovely

I had a lovely apartment

2 be Gramersi she be nyuYORK?

in Gramercy that in New.York

in Gramercy in New York

3 ve-

and

4 (1.5) ve nitKAti sham

and got.stuck.1sg there

and I got stuck there

5 *(0.4)nitKAti. (0.3)*

I got stuck

*...quoting gesture,,,,*

smiling

6 nitKAti xodesh va xetsi be nyuYORK?

got.stuck.1sg month and half in New York

I got stuck for a month and a half in New York

7 ve hizMANti et tom she yavo e: ;

and invited.1sg ACC Tom that will.come.3sgm uh

and I invited Tom to come uh

8 e lihyot iti e:

uh to.be with.me uh

be with me uh

9 ba tsiluMIM?

in.the filming

during filming

After describing favourably her living conditions in New York (lines 1-2), the speaker continues to report about her inability to leave the city by saying ‘I got stuck there’ (line 4). Following a short silence, she repeats the verbal phrase ‘I got stuck’ in lower pitch and decreased loudness, while co-extensively producing the QG (line 5).

The gesture shows clear articulation into its preparation phase, stroke (a double flexing of the fingers), and retraction, resulting thus in a relatively long duration of the gesture as a whole (1.9 seconds). By using the QG, the speaker retroactively attributes her initial production of nitkati ‘I got stuck’ to another voice, thereby distancing herself and backing down on this formulation. The conceding in this case underscores the incompatibility between the standard negative interpretation of the verb, proposed by its first non-modified production, and the report on the positive living experience in New York (lines 1-2). The articulated and elongated performance of the gesture brings into focus the action of repair and provides therefore a noticeable display of accountability for what may have otherwise been perceived as an inapposite choice of words. The speaker does not propose a revised statement, but leaves it open for inference. She then employs the same verb nitkati for the third time (line 6), however now conveying the qualified meaning imparted by the QG in the previous production.

By attributing an initial formulation to another voice, the QG signals a break in the default assumption that the roles of principal, author and animator are conflated in a single speaker (Goffman 1981; Kotthoff 2002), so that the speaker’s responsibility for that formulation is reduced. Foregrounding the presence of another voice invites the recipient to reconsider the meaning of the formulation, serving as yet another overt practice through which participants can self-appropriate and negotiate semantics.

Narrowing (e.g., finger > 'index finger') and broadening (e.g., metaphors) processes are implicated in online meaning adaptations and in linguistic change. Both depend on a prototype category structure for lexical meanings, which (i) comprise multiple denotations, and (ii) are bounded by norigid boundaries. Narrowing selects a subset of the lexeme's senses, and broadening incorporates concepts outside the lexical meaning. Standard lexemes belong to sparse lexical systems, where competitors differ on multiple parameters. Hence, there's a conceptual no-man's-land between them (e.g., between nice and attractive), which interpretations routinely encroach upon (e.g., 'nice, close to attractive').

Based on corpus studies (SBC ,LSAC, BNC) and questionnaires I argue that numerals belong to a dense lexical domain, which blocks prototype category structure. The competitors in this dense system (e.g., 6, 7 for 5) only minimally differ from each other, and don't admit no-man's-lands. Each numeral is therefore distinct and literally faithful (interpreted as 'exactly N'). Indeed, numerals are remarkably monosomous synchronically (84.5% of bare numerals interpreted as 'exactly N', Ariel, 2002), and extremely stable diachronically (Pagel and Meade, 2017).

I here tested several predictions: (1) Narrowing numerals is quite impossible, since they have single denotations. Indeed, participants accepted Hebrew narrower dey 'pretty much' (Ziv, 2016) + standard lexemes, but rejected dey N(umeral) combinations. (2) English narrower+numeral tokens are either unattested (kinda), or, when rarely found (sorta) receive broadening, rather than narrowing. (3) Broadening is a routine procedure for standard lexemes, hence, only 3.1% of nice, blue, green, and red were overtly broadened. But broadening numerals (e.g., about) is interpretatively marked, and hence, quite frequently overtly marked: 24.7% for two, seven, ten, seventeen and thirty. (4) It is the sparse system, not a precise versus imprecise core meaning that determines the category structure of lexemes: Since, unlike numerals, the literally precise straight, parallel and middle participate in sparse systems, only 1.1% were overtly broadened. Crucially, this is so, despite the fact that only 10.8% of straight and middle tokens received 'exactly' interpretations. Minimal pairs too attest to the role of sparse versus dense systems: the denser zero and 50% were interpreted as 'exactly' much more often (60.8%) than their sparse-system synonyms nothing and half (12.2%). And when interpretatively broadened, the numerals were overtly modified much more than the quantity expressions (53.6% versus 16.2%).

These differential distributions naturally follow from my proposal that numerals do not have prototype category structure, because they belong to a dense lexical system.

The current study compares two types of accounts for the interpretation of or constructions. One, a set of 'Inclusivist' theories [1-4], assumes an 'inclusive disjunction' semantics for or, and derives an 'exclusive disjunction' reading via scalar implicatures. The other, 'Alternativist' theory [5-6], assumes a mere procedural core meaning, whereby the explicit disjuncts must manifest an 'alternativity' relation, namely mutual exclusivity on some level.

A crucial difference between the two types of theories leads to different predictions for cases where there’s a discrepancy between the speaker-intended 'what is said' ('w.i.s') (say, 'one of X and Y' for X or Y) and an actual state of affairs (say, 'both X and Y'). Inclusivist theories assume that 'w.i.s' fully specifies the truth conditions of the proposition. They therefore predict that the proposition will be judged false in such cases. ‘Alternativists’, on the other hand, predict variability in truth judgments. The reason is that Truth-compatible inferences (TCIs) (e.g. that ‘both alternatives’ is compatible with ‘one alternative’) may optionally bridge this gap, should the interlocutor decide that the gap is not discourse-relevant [7].

We tested these predictions in an experiment where the difference between 'one alternative' and 'both alternatives' was not discourse-relevant (what was relevant was 'at least one alternative’, e.g., Oren is entitled to some discount should he come with at least one woman). We created two tasks: an interpreta­tion task, which taps ‘w.i.s’, where participants match a state of affairs to a sentence, and a verification task which elicits truth judgments in view of a 'both' state of affairs. A control condition without or was included.

The results show that participants’ (N=64) choice of ‘one alternative’ in the critical condition (with or) differed significantly across task types (β = 1.09, χ² = 34.17, p < .001). In interpreta­tion, participants predomi­nantly chose ‘One’ (83%). In verification, only 55% of the participants judged 'Both=F' (testifying to a ‘One’ interpretation). Response patterns differed between critical and control trials.

Only the ‘Alternativists’ predictions are borne out, since only they account for the gap between interpreta­tion and verification results. The variability of responses in verification, as well as the discrepancy between interpretation and verification provide further support for the reality of TCIs, which can only play a role in truth judgments.

Recent literature on Information Structure questions the necessity of the category of focus, arguing that relevant interpretations arise as a by-product of diverse poorly understood language-specific categories (Matić and Wedgwood 2013). This paper focuses on the apparent focus marker lɛ́ (also having variants tɛ́ , nɛ́ , floating high tone) in Mano (Guinea and Liberia). We delineate its distribution and argue that the invariant function is a request of an attention-shift and referent-identification, which trigger the contextual effects of counter-expectation (including narrow focus) or discourse-shift (including topicalization).

Considering contexts like (1), lɛ́ could be assumed to mark narrow focus. Since it is also found on the S-arguments in thetic statements and presentative constructions it could be regarded as a detopicaliser (Güldemann 2010).

(1)

e ̰̋ pàpá bɛ ̰̋ɛ ̰̋ ɓī bɛ ̰̋ɛ ̰̋ sà̰ a ̰̀ ŋwɔ

INJ papa also 2SG.EMPH also mocking issue:CS

lɛ́ ɓà kɛ ɓɛ !

?FOC 2SG.SBJV>3SG do:IPFV DEM

Áà gèè ŋwɔ́ gáà lɛ́

3SG.JNT say:JNT issue strong ?FOC

ŋ gèē wɛ o!

1SG.SBJV>3SG say:IPFV DEM INJ

‘Eh! papa, you, you are JOKING (lit.: it is a joke that you do) – he says: I am talking SERIOUSLY (lit.: it is truth that I am saying)!’

However, the marker is also used in Left Dislocation-like constructions on the detached topical constituent. Here, de-topicalising analysis is troublesome.

(2)

ŋwɔ́ tɛ́ yā kàkò à pɛ́ ̰̋

issue ?FOC DEM 1PL.SBJV 1SG pray

kó pɛ́ ̰̋ wálà mɔ

1PL.CONJ>3SG pray God on

‘(Your son comes to church, your grandson comes to church, your son’s wife comes to church) That issue (that was just explained) – let’s pray for it, let’s pray for it to God.’

The usage in (2) is homologous to the most common relativization strategy in Mano (3), where referent identification is followed by a relative clause, marked by the backgrounding device ɔ (deriving from the demonstrative yā in 2) which is then followed by the main clause where the informational contribution of the utterance is concentrated.

(3)

[gɔ ̰̀ vɔ yààkā lɛ́ ō sɛ́lɛ́ ɓɛ dɔ ɔ ]

man PL three ?FOC 3PL.PST soil DEM stop BCKGR

óò gèē ō lɛ ɛ Bōòdā gbē nì.

3PL.IPFV>3SG say:IPFV 3PL for P.N. son PL

‘The three meni who founded this village, they call themi sons of Booda.’ (nat.)

It would be awkward to postulate a unified information role to the head NPs in (1) and in (2)-(3). Instead, the cooptation (Mauri and Sansò 2011) of the attention-request is a likely source for this variety of use. Indeed, what is common to the contexts of narrow/contrastive focus in (1) and LD-like topicalization in (2) and relativization in (3) is a request to attend to counter-expected and/or discourse-shifting information. Thus, the analysis of the overall function of lɛ́ in terms of attention management instead of information management allows us both to address its various functions and to demonstrate the source of the focus-like effects in (1).

This research problematizes the assumption that predicating ‘Truth’ over an assertion simply produces an evaluation of that assertion, as true or false, with respect to some state of affairs. I argue against this assumption claiming that if this were the case, predicating Truth should have been redundant since speakers are anyway assumed to contribute only truthful assertions (Grice, 1989). I will claim that the prime motivation to predicate Truth in interaction is to express a (dis)agreement stance (Strawson, 1949, following Ramsey, 1926). To support this claim, I will present corpus data where I analyze the use of the predicate true in discourse.

The analysis carried for the predicate true is based on two spoken American English corpora (LSAC, 5 million words, SBC, half a million words). A total of 368 true tokens were collected and analyzed. Findings reveal that the intersubjective indexing of (dis)agreement is the prominent function for ~75% of all interactional true tokens. Interestingly, the corpus data also reveals two distinct discourse profiles for ‘true’ vs. ‘agree’ uses of the predicate true. In general, ‘true’ uses appear in a more objective context, whereas ‘agree’ uses appear in a more subjective context.

To support the relation between Truth and Agreement, I will also present: (1) an inquiry of the common etymological source for ‘truth’ and ‘agreement’ markers (Gadish and Katzir [n.d.]); (2) the use of Truth markers, e.g., true, indeed, in concessive constructions to emphasize agreement (König, 1988). These together will provide converging evidence to the inherent relation between Truth and Agreement.

To conclude, I argue that predicting Truth in interaction is prominently a social act used to index an intersubjective stance of (dis)agreement. This claim challenges the conventional assumptions regarding the predication of Truth by performing an empirical investigation using multiple methods, thus, offering a fresh perspective on what it takes to evaluate some assertion as true.

Researchers going back to Ross (1967) and Chomsky (1973) have hypothesized that the source of the unacceptability of long-distance syntactic dependencies such as in (1a) may be a syntactic constraint (like Chomsky’s proposed Subjacency constraint):

(1) a. extraction from subject: ?? Which sports-car did the color of __ delight the baseball player?

b. extraction from object: Which sports-car did the baseball player love the color of __ ?

One of the strongest predictions of the syntactic account is that the extraction acceptability is independent of the construction in which the extraction occurs (Schutze et al., 2015): it is the syntax that blocks the long-distance dependency, not the meaning (the construction). In recent work, Abeillé et al (2020) show that this approach cannot be correct, because extraction acceptability crucially depends on the construction. Extraction from relative clauses, for example, are permitted in similar configurations as (1):

(2) a. extraction from subject: The dealer sold the sports-car of which [the color __] delighted the baseball player

b. extraction from object: The dealer sold the sports-car of which the baseball player loved [the color __]

Here I will describe the approach taken by Abeillé et al. — the Focus-Background Conflict Constraint — and show how it accounts for many effects that have been hard to explain with the syntactic approach. Following Erteschik-Shir (1973), Kuno (1987), Goldberg (2006), Sag (2010) and Chaves & Putnam (2020), I speculate that the unacceptability of an island structure is almost never due to the syntax; instead the unacceptability of an island structure is almost always explained in terms of discourse, frequency or memory (cf. Sprouse et al. 2012; Sprouse et al., 2016; for an opposing view).