Desire, belief, and semantic composition: variation in mood selection with desire predicates

Abstract

Mood selection properties of desire verbs provide a rich source of evidence regarding the semantics of propositional attitudes. This paper approaches the topic by providing an analysis of crosslinguistic variation in the selection patterns of the desire verbs ‘want’ and ‘hope’, focusing on Spanish and French. There is no evidence that the meanings of ‘hope’ and ‘want’ differ between these languages, and yet in Spanish esperar ‘hope’ and querer ‘want’ both take subjunctive, while in French only vouloir ‘want’ selects subjunctive and espérer ‘hope’ strongly prefers the indicative. The inclination of ‘hope’ toward the indicative is manifest also in other Romance languages. Previous theories tie mood selection tightly to the verb’s modal backgrounds and do not anticipate such variation. We explain the consistency in mood selection with ‘want’ versus the variation with ‘hope’ in terms of two key ideas: (i) moods are modal operators that encode different degrees of modal necessity, and (ii) modal backgrounds can be manipulated by the grammar. In terms of (i), we argue, building on the comparison-based theory of mood, that the indicative is a strong necessity operator, while the subjunctive encodes a weaker necessity. Regarding (ii), we propose that two backgrounds may function as one under certain well-defined circumstances. Our proposal supports the decompositional approach to attitude verbs, where mood is responsible for the quantificational force traditionally attributed to the verb.

Appendices

Appendix A: Preliminary data on mood selection in other Romance languages

In this Appendix, we give a bit more detail on our preliminary findings concerning mood selection with ‘want’ and ‘hope’ in Catalan, Italian, Portuguese, and Romanian.

• 1. Catalan allows only subjunctive with both ‘want’ and ‘hope’ across tenses.⁴⁸Voler (‘want’) is non-past oriented.

  

  Esperar (‘hope’) allows both future and present orientation with present subjunctive under a present tense matrix, and it allows past orientation (i.e., backshifting) with the imperfect under present tense. However, imperfect under imperfect does not seem to allow backshifting; a perfect form is required for that meaning.

  

• 2. Portuguese requires subjunctive with quer ‘want’, like French, Spanish, and Catalan, but shows a high degree of freedom of mood choice with esperar ‘hope’. Present tense ‘want’ requires present subjunctive, with either simultaneous or future orientation; the incompatibility with past orientation rules out the imperfect, and while Portuguese has a future subjunctive, it is not used for future orientation in this context. Past tense ‘want’ only allows imperfect subjunctive.⁴⁹

  

  ‘Hope’ allows both indicative and subjunctive, but there are two interesting facts: first, imperfect under present is ruled out for one of our consultants; this suggests that present tense esperar is non-past oriented for some speakers, like ‘want’ (note that it can also mean ‘wait for’). Second, indicative imperfect is not allowed under imperfect for one of our speakers. In contrast, the other speakers continued the pattern of both moods being allowed under ‘hope’ even into the past. Imperfect subjunctive under imperfect allows backshifted, simultaneous, and future-oriented temporal relations. The future-oriented interpretation sometimes has a counterfactual inference (and our consultants differed in terms of which be-verb, estar or ser, they preferred here); the backshifted reading is easier to access with an eventive predicate.

  

  The situation in Portuguese is complex, and the differences in judgment among our small group of speakers suggest that there is much to be discovered in a more thorough investigation of this language. Neverthess, as far as mood choice goes, the pattern is in line with our expectations.

• 3. In Italian, many speakers only accept subjunctive with volere ‘want’, but some speakers also allow a tense-matching indicative with simultaneous interpretation and a distinct meaning which we have not yet investigated ((79b) and (79f)).⁵⁰ Unlike in French and Spanish, ‘want’ allows past orientation with imperfect in both the matrix and subordinate clause, at least with an event verb as in (79e); this pattern is described as having a counterfactual interpretation similar to ‘wish’, but more study is clearly required.

  

  Our consultants for Italian showed two patterns with regard to mood choice with sperare ‘hope’, one more like Portuguese in allowing both indicative and subjunctive in many cases, and one more like Spanish and Catalan, allowing only subjunctive. ‘Hope’ allows past orientation in both the present and past.

  

  As with Portuguese, the facts for Italian are complex and in need of much futher investigation.

• 4. Romanian vrea ‘want’ selects subjunctive, as in the other languages, with present subjunctive allowed under matrix present or past, expressing simultaneous or future orientation. Past subjunctive is only possible under matrix past, with a simultaneous interpretation.⁵¹$^{,}$⁵²

  

  Spera ‘hope’ freely allows both indicative and subjunctive across all tenses. This consistent flexibility means that Romanian is likely to be an important case study for future work on mood selection with desire verbs. Mood and tense interact in ways we do not describe fully here. For example, according to our preliminary investigation, subjunctive under imperfect (82g) allows past, simultaneous, and future orientation, while the imperfect indicative under imperfect (82k) allows only past or simultaneous orientation; future is used for future orientation with the indicative.

  

Appendix B: Formal systems

In this Appendix we present our analyses of Spanish and French in a compositional way. We specify the semantics by giving meanings in an English-based metalanguage of the kind presented in Heim and Kratzer (1998). Note that the only difference between the languages is found in the meanings of the mood thematic relations (see the entries for $\theta _{indic}$ and $\theta _{subj}$ in Semantics: Lexicon below).

1.1 B.1 Version 1: Premise semantics

Syntax: Phrase structure For this demonstration fragment, we interpret lexicalizations of (83). We assume that M raises to C from somewhere within TP in order to take scope over TP.

For simplicity, we label verbs as having either an experiencer or an agent subject, but we would assume that there is a better syntactic/semantic explanation for the assignment of thematic roles to nominal arguments. Because we are focusing on selected mood, we do not consider the mood marker in the root clause.⁵³

Syntax: Lexicon Single quotes indicate that the entries represent lexical items in each relevant language.

• 1. N: ‘John’, ‘Mary’

• 2. V$_{exp}$: ‘believes’, ‘wants’, ‘hopes’

• 3. C: ‘that’

• 4. $\Theta$: $\theta _{exp}$, $\theta _{agent}$

• 5. M: $\theta _{indic}$, $\theta _{subj}$

• 6. T: present

• 7. V$_{agent}$: ‘leaves’

Semantics: Lexicon (premise semantics) Functional meanings are expressed in a predicate logic-like metalanguage. For example, $\lambda s[believing(s)]$ is the function that is defined for any situation s; it is true of s if s is a believing situation, and false of s otherwise. Variables are implicitly sorted as $x\in$ Individuals, $s\in$ Situations, $w\in$ Worlds.

• 1. $[[{`\mathsf John'}]] = john$

• 2. $[[{`\mathsf Mary'}]] = mary$

• 3. $[[{`\mathsf believes'}]] = \lambda s[believing(s)]$

• 4. $[[{`\mathsf wants'}]] = \lambda s[wanting(s)]$

• 5. $[[{`\mathsf hopes'}]] = \lambda s[hoping(s)]$

• 6. $[[{`\mathsf leaves'}]] = \lambda s[leaving(s)]$

• 7. $[[{`\mathsf that'}]] = \lambda m\in D_{\langle \langle s,t\rangle , \langle s, t\rangle \rangle }[m]$.

• 8. $[[{present}]] = \lambda p \lambda w \exists s[s<w \wedge s\circ now \wedge p(s)]$

• 9. $[[{\theta _{agent}}]] = \lambda x \lambda s[Agent(s,x)]$

• 10. $[[{\theta _{exp}}]] = \lambda x \lambda s[Experiencer(s,x)]$

• 11. $[[{\theta _{indic}}]]$ =

• Spanish: $\lambda p\lambda s[sn(p,content(s),s)]$

• French: $\lambda p\lambda s[sn(p,$ spl(content(s)), s)]

• 12. $[[{\theta _{subj}}]] =$

• Spanish: $\lambda p\lambda s[ln(p,content(s),s)]$

• French: $\lambda p\lambda s[ln(p,$ spl(content(s)), s)]

Semantics: Composition rules

• 1. For any subtree $[_{\alpha }\ \beta ]: [[{\alpha }]] = [[{\beta }]], \mathrm{unless} \,{{\alpha }} =$ Root

• 2. For any subtree $[_{\alpha }\ \beta ]$: if $\alpha$ = Root, then $[[{\alpha }]] = \exists w[[[{\beta }]](w)]$

• 3. For any subtree $[_{\alpha }\ \beta \ \gamma ]$: if $[[{\beta }]]$ is a function with $[[{\gamma }]]$ in its domain, then $[[{\alpha }]] = [[{\beta }]]([[{\gamma }]])$

• 4. For any subtree $[_{\alpha }\ \beta \ \gamma ]$: if both $[[{\beta }]]$ and $[[{\gamma }]]$ are the characteristic functions of sets of situations, then $[[{\alpha }]] = \lambda s[[[{\beta }]](s) \wedge [[{\gamma }]](s)]$

Important metalanguage functions (premise semantics)

• 1. content(s) is only defined if s is a content-bearing situation. For any content-bearing situation s, content(s) is a modal background or pair of modal backgrounds representing the cognitive or informational content borne by s. Specifically:

  • (a) If s is a believing situation, then $content(s)=dox$

  • (b) If s is a wanting situation, then $content(s)=\langle bul,dox^{+}\rangle$

    • Diversity property: $\forall s[wanting(s) \rightarrow \bigcap second(content(s))\cap p\ne \emptyset$ and $\bigcap second(content(s))\backslash p\ne \emptyset ]$

  • (c) If s is a hoping situation, then $content(s)=\langle bul^c,dox\rangle$

    • Diversity property: $\forall s[hoping(s) \rightarrow \bigcap second(content(s))\cap p\ne \emptyset$ and $\bigcap second(content(s))\backslash p\ne \emptyset ]$

    • Consistency property: $\forall s[hoping(s)\rightarrow \bigcap (bul^c(s)\cup dox(s))\ne \emptyset ]$.

• 2. sn(p, B, s) is only defined if p is a proposition, B is a single modal background, s is a content-bearing situation, and B(s) is defined. When defined, sn(p, B, s) is true iff $\bigcap B(s)\subseteq p$.

• 3. ln(p, B, s) is only defined if p is a proposition, B is a pair of modal backgrounds $\langle g, f\rangle$, s is a content-bearing situation, and f(s) and g(s) are both defined. When defined, ln(p, B, s) is true iff $\exists b[b \in BestSets(g,f,s) \wedge b\subseteq p]$.

• 4. $BestSets(g,f,s) = \{X : \forall w\forall v[(w\in X \wedge v\in X) \leftrightarrow (\{p\in g(s) : w\in p\}=\{p\in g(s) : v\in p\}\wedge w\in Best(g,f,s) \wedge v\in Best(g,f,s)\}$

• 5. $Best(g,f,s) = \{w : w\in \bigcap f(s) \wedge \lnot \exists v[v\in \bigcap f(s) \wedge v <_{g(s)} w]\}$

• 6. Unify$(\langle g, f\rangle )$ is only defined if M is a pair of modal backgrounds and $\forall s[s\in domain(f)\rightarrow (s\in domain(g) \wedge \bigcap (g(s)\cup f(s))\ne \emptyset )]$. When defined, ${\textsf {unify}}(\langle g, f\rangle ) = [\lambda s:\ s\in domain(f)\ .\ g(s)\cup f(s)]$.

• 7. spl(M) is only defined if M is a sequence of modal backgrounds. When defined, spl$(M) {=}$ unify(M) if ${\textsf {unify}}(M)$ is defined; otherwise, spl$(M) {=} M$.

  • Equivalence of ln and sn after unification: If $\textsf {unify}(\langle g,f\rangle )$ is defined, $sn(p,\textsf {unify}(\langle g,f\rangle ),s) = ln(p,\langle g,f\rangle ,s)$ (for every proposition p and situation $s\in domain(f)$).

Proof

Assume that ${\textsf {unify}}(\langle g, f\rangle )$ is defined; therefore $\bigcap (f(s)\cup g(s))\ne \emptyset$ for every s in the domain of f. Assume that s is in the domain of f.

The essence of the proof is that, when $f(s)\cup g(s)$ is consistent, $\bigcap (f(s)\cup g(s))$ is the sole member of BestSets(g, f, s), and so $\bigcap (f(s)\cup g(s))\subseteq p$ iff $\exists b[b\in BestSets(g,f,s) \wedge b\subseteq p]$.

(1) First we show that $\bigcap (f(s)\cup g(s))\in BestSets(g,f,s)$. This requires that $\forall w\forall v[(w\in \bigcap (f(s)\cup g(s)) \wedge v\in \bigcap (f(s)\cup g(s))) \leftrightarrow (\{p\in g(s) : w\in p\}=\{p\in g(s) : v\in p\}\wedge w\in Best(g,f,s) \wedge v\in Best(g,f,s))$.

• Left to right. Suppose we have a $w\in \bigcap (f(s)\cup g(s))$ and a $v\in \bigcap (f(s)\cup g(s))$. Then $w\in Best(g,f,s)$. To see this, assume $w\not \in Best(g,f,s)$; then either $w\not \in \bigcap f(s)$, or $\exists q\in g(s)$ such that for some $u\in \bigcap f(s)$, $u\in q \wedge w\not \in q$. But since $w\in \bigcap (f(s)\cup g(s))$, $w\in \bigcap f(s)$, and there is no such u or q. (Similarly for v.)

  Next, show that $\{p\in g(s) : w\in p\}=\{p\in g(s) : v\in p\}$. Suppose the contrary. Then there must be a $q\in g(s)$ such that either $w\in q \wedge v\not \in q$ or $w\not \in q \wedge v\in q$. Neither of these can be because $w,v\in \bigcap (f(s)\cup g(s))$.

• Right to left. Suppose we have w and v such that $\{p\in g(s) : w\in p\}=\{p\in g(s) : v\in p\}\wedge w\in Best(g,f,s) \wedge v\in Best(g,f,s)$ but $w\not \in \bigcap (f(s)\cup g(s))$ or $v\not \in \bigcap (f(s)\cup g(s)$. Since $w\in Best(g,f,s)$ and $g(s)\cup f(s)$ is consistent, $w\in \bigcap (f(s)\cup g(s))$, and similarly for v. To prove this, assume the contrary; then there is some $q\in f(s)\cup g(s)$ such that $w\not \in q$. If $q\in f(s)$, then $s\not \in Best(g,f,s)$, contradicting the assumption that $w\in Best(g,f,s)$; and if $q\in g(s)$, then there is a $u\in \bigcap (f(s)\cup g(s))$ such that $u<_{g(s)}) w$, contradicting the assumption. (Similarly for v.)

(2) Next we show that there is no other member of BestSets(g, f, s) than $\bigcap (f(s)\cup g(s))$. Suppose that there were some other such set X. Then there is either a world u such that $u\in X$ but $u\not \in \bigcap (f(s)\cup g(s))$ or vice versa.

Assume there exists a u such that $u\in X$ but $u\not \in \bigcap (f(s)\cup g(s))$. Then either $u\not \in \bigcap f(s)$ or $u\not \in \bigcap g(s)$. If the former, then $u\not \in Best(g,f,s)$, and so $u\not \in X$, contradicting the assumption. If the latter, then there is some $q\in g(s)$ such that $u\not \in q$. Let $u'$ be a world such that $u' \in q$ and $u' \in \bigcap (f(s)\cup g(s))$. In this case, $u' <_{g(s)} u$, and so $u\not \in Best(g,f,s)$, and so not in X, contradicting our assumption.

Alternatively, assume that there is a u such that $u\in \bigcap (f(s)\cup g(s))$ but $u\not \in X$. Unless $X=\emptyset$, there is a world $u'\in X$ such that $u'\ne u$ and $u'\in Best(g,f,s)$. If $u'\in \bigcap g(s)$, then $\{p\in g(s) : u'\in p\}=\{p\in g(s) : u\in p\}$, contrary to our assumption that $u\not \in X$. Otherwise, there is a $q\in g(s)$ such that $u'\not \in q$. But then $u <_{g(s)} u'$, contrary to our assumption that $u'\in Best(g,f,s)$.

• 8. first(X) and second(X) are only defined if X is an ordered pair. When defined, first(X) is the first member of X and second(X) is the second member of X.

1.2 B.2 Version 2: Similarity semantics

The fragment which uses similarity semantics is largely the same as the premise semantics version. The only differences are in the entries for the subjunctive theta role (Semantics: Lexicon 12, redefined below), and associated definitions.

Semantics: Lexicon (similarity semantics)

• 12. $[[{\theta _{subj}}]] =$

  • Spanish: $\lambda p\lambda s[\leadsto ^{second(content(s))(s)}_{p}\subseteq <_{first(content(s))(s)}]$

  • French: $\lambda p\lambda s[\leadsto ^{second(\mathsf{spl}(content(s))(s)}_{p}\subseteq <_{first(\mathsf{spl}(content(s))(s)}]$

Important metalanguage functions (similarity semantics) The similarity semantics relies on the definition of $\leadsto$:

• 9. For any proposition p and world v, SIM(p)(v) is the world $w\in p$ which is most similar to v; hence if $v\in p$, $v=w$.⁵⁴

• 10. For any worlds w and v, premise set B and proposition p, $w \leadsto ^{B}_{p} v$ if and only if either (i) $w\in p\cap (\bigcap B)$ and $v=SIM(\lnot p)(w)$ or (ii) $v\in \lnot p\cap (\bigcap B)$ and $w=SIM(p)(v)$.

To see how these definitions work, consider a Spanish subjunctive clause with denotation $\lambda s[\leadsto ^{second(content(s))(s)}_{p}\subseteq <_{first(content(s))(s)}]$. When applied to a wanting situation s, $\leadsto ^{second(content(s))(s)}_{p}$ forms the set of pairs $\langle w,v\rangle$ where either w is a p world in $dox^+(s)$ and v is the most similar $\lnot p$ world, or where v is a $\lnot p$ world in $dox^+(s)$ and w is the most similar p world. These are the similarity pairs which differ in the truth value of p and at least one member of which is a belief(+) world. If the set of such pairs is a subset of $<_{bul(s)}$, this is to say that each p world is more-desired than its most similar $\lnot p$ world.

In French, the definition of subjunctive contains spl. When s is a hoping situation, and so the two components of content(s) can be unified, $second(\textsf {spl}(content(s)))$ is not defined because spl${\text{(content(s)) }}$ is a singleton. Thus, when the subjunctive is used with ‘hope’ in French, the derivation fails at the point when the subjunctive-marked clause is combined with the verb.

Comments on the similarity analysis. In the version of our analysis which builds on the similarity semantics, we do not predict equivalence between French ‘hope’+indicative and Spanish ‘hope’+subjunctive. More precisely, the following two sentences would not have the same truth conditions (even assuming that Pierre and Pedro refer to the same person and Fido court and Fido corra denote the same proposition).

Because unification is successful, the semantics for the French case would be the same as under the premise semantics; it is true iff all of the most desirable worlds compatible with Peter’s beliefs are ones in which Fido runs. In contrast, the Spanish counterpart could be false even if all of the most desirable worlds compatible with Peter’s beliefs are ones in which Fido runs, for example if for one such world, the most similar world in which Fido doesn’t run is more preferable (and not a belief world). A case which brings the difference out more clearly would be (85).

The example is true in this scenario, but arguably, for some belief world w in which I pay $1, the most similar world in which I don’t pay $1 is one in which I pay $.50 and not $300. If so, ‘I hope to pay a dollar to enter the show’ will be false in Spanish according to the definitions above.

Because the notion of similarity is vague and it is not clear how the particular beliefs and desires of the experiencer factor into it, we can’t prove that this scenario or another like it is a problem. We also can’t prove it is not a problem, and thus conclude that the similarity-based approach cannot be shown to predict the equivalence between the French and Spanish counterpart ‘hope’ sentences. Given that the sentences are equivalent, as far as we know, there is reason to prefer the premise semantics approach, where the equivalence follows directly.

Heim (1992) considered a variant of the similarity approach which is like ours,⁵⁵ but in her official version both members of each similarity pair are required to be belief worlds. It is worth noting that, even with this assumption, simplification is not necessarily meaning preserving. In Spanish (without simplification), (84b) is true if each world within the set of belief worlds in which Fido runs is better than its most similar counterpart (also a belief world), and each world within the set of belief worlds where Fido does not run is worse than its most similar counterpart (also a belief world). These truth conditions require that (i) the very best belief worlds are ones in which Fido runs, and also that (ii) non-ideal belief worlds where Fido runs are better than their most similar not-running counterparts. In contrast, for French (84a) with simplification and a simple necessity semantics, only condition (i) holds. In other words, the similarity semantics with the assumption that the relevant comparisons involve pairs of belief worlds predicts that ‘hope’ sentences have weaker truth conditions in French. This point has been made by von Fintel (2018); he argues that meanings for desire sentences assigned by the similarity semantics are not just stronger, but in fact too strong. If he is correct in this, that is already a sufficient reason to prefer the premise semantics approach.