Part of communicating successfully involves recognising the intentions underlying what speakers choose to say (Grice, 1975, Levinson, 2000). While this recognition seems straightforward in that listeners map the incoming signal onto words and meaning, language is not always produced in such a direct or transparent manner that a speaker’s utterance and intended meaning go hand in hand. Oftentimes, language is used indirectly and utterances do not transparently map onto the intended meaning. In these cases, successful communication requires listeners to go beyond the transparent mapping and infer additional meaning in order to determine how the utterance is relevant to the discourse. One way such inference is achieved is by reasoning about a speaker’s communicative goals. When successful, indirect communication presents an efficient way for speakers to communicate an array of context-sensitive meanings in finite time. However, the availability of such inferences presents listeners with a challenge; when should they engage in the search for additional meaning and draw an inference?

In the present work we examine the availability of relevance inferences arising from trivial utterances, utterances that are neither blatantly underinformative (failing to provide a sufficient Quantity of information per Grice, 1975) nor explicitly redundant (“overinformative” per Rubio-Fernandez, 2019). We use trivial utterances as a way of testing the interpretation of a type of utterance that appears in natural communication. Pragmatic inferencing, typically, is studied couched in terms of informativity based on world knowledge, entailment relations, and referential communication. Trivial utterances offer the opportunity to examine a broader variety of utterances that prompt inferencing beyond the informational value of a description (how much information is contained in an utterance). Informational value emphasizes the amount of information and can be understood in a variety of ways: information as reduction of uncertainty about the world (Levinson, 2000, Shannon, 1948) or as an utterance’s discriminatory value e.g. modification (the banana vs the yellow banana vs the orange banana; see Sedivy, 2003, Rubio-Fernandez, 2019), or the specificity of a label (the animal, the bird, the crow; Brennan & Clark, 1996; Engelhardt et al., 2006; see also Davies & Arnold, 2019). In contrast, consider a speaker who utters the following as commentary on a situation, without any preceding question inquiring about this information:(1)

The library walls are blue

Upon hearing such an utterance, an addressee is faced with a choice to (a) interpret the utterance as transparently informing them about the state of the walls (i.e. what was said: “The library walls are blue”) or (b) reason about the speaker’s goals in producing such an utterance to determine how it is relevant; perhaps the speaker noticed a change (i.e. what was meant: the library used to be a different colour) or the speaker was surprised at the choice (i.e. what was meant: the library walls are an unusual colour). We refer to the additional meaning arising in (b) as a relevance inference, in the sense that a listener reasons about why a speaker has made the discourse contribution that they have.

Relevance inferences are prompted by an addressee reasoning about a speaker’s choice to produce an utterance rather than say nothing. By creating an overt signal, the utterance carries an assumption or expectation of relevance. We assume that utterances produced are relevant to the current situation and that determining how they are relevant involves reasoning about the speaker’s intention (Sperber & Wilson, 1995/2004; Wearing, 2015).

Consider the example from Grice (1975, p. 51):(2)

A is standing by an obviously immobile car and is approached by B;

A: I am out of petrol.

B: There is a garage round the corner.

In this exchange, B’s contribution is only a relevant contribution if they believe that the garage is open and has petrol. These additional elements of meaning are not explicitly stated and must be inferred through reasoning about why B chose to utter such an utterance. Without the assumption of relevance and without the inference of the additional meaning that the garage is open and has petrol, B’s utterance is not a cooperative response in this dialogue.

To the extent that comprehenders draw relevance inferences, it is an open question as to the availability of these inferences during language comprehension. Here we present four studies using offline and online measures to investigate the computation of these inferences. We first examine linguistic and extralinguistic cues that may signal to comprehenders that an inference can or ought to be computed. We then probe comprehension effort both during the processing of the sentence itself (sentence reading times) and at a later time point when the participant is asked whether they endorse the inference meaning (response reaction times). Our results show that, as with other types of pragmatic inferencing, whether relevance inferences arise depends on extralinguistic cues such as speaker-specific characteristics (e.g. their knowledge and their speech style). We show these inferences arise most easily when a speaker is understood to be knowledgeable about the situation they are talking about and when the speaker is understood to be someone whose communicative choices reflect a preference to filter what they say.

In the processing data, we see a more nuanced picture: Inferring additional meaning appears to require a small increase in processing effort when processing the sentence itself, however, this finding requires further investigation. We see a clearer cost after the sentence has been processed when participants are prompted to endorse or refute the additional meaning. These findings parallel other work on pragmatic inferences that shows that the inference of additional meaning is readily available to comprehenders but not necessarily cost-free. The experiments presented here provide an avenue for broadening our understanding of pragmatic inferencing beyond traditionally studied inferences.

Language use is often characterised as cooperative joint action between interlocutors (Garrod and Pickering, 2009, Grice, 1975, Searle, 1992). Successful communication involves a speaker selecting the appropriate signals to convey their intended meaning and an addressee recovering that intended meaning (Geurts, 2010, Grice, 1975). When a speaker utters “Can you open the window?”, the addressee is faced with a choice – interpret the utterance transparently and respond to the inquiry about their ability to open the window or recognise that the speaker is intending to be understood as making a request and therefore open the window. Conventionally, questions in English such as “Can you X?” are understood as requests, rather than as inquiries into ability. However, under certain circumstances the question may be interpreted as an inquiry about ability; for example, if the addressee has recently broken their arm and the speaker is genuinely asking if they are able to perform the action or not. Crucially, the success of this communicative interaction hinges on correctly retrieving the speaker’s intended goal. It is often assumed that the goal of a speaker is to convey information or prompt their addressee to action. However, there are instances of phatic communication where language is used to establish and maintain social relationships rather than convey information or ideas (Crystal, 1991, Laver, 1974, Malinowski, 1923, Žegarac and Clark, 1999). Consider the example (3):

A speaker who utters (3) could be communicating phatically, without a goal or agenda beyond avoiding an awkward silence and offering an observation as small talk. However, it is also possible that this utterance is intended non-phatically and the speaker might be specifically conveying the information that it is hot in the room. In that case, (3) would be a redundant utterance if the speaker and addressee are co-located, since presumably they would both be experiencing the same temperature in the room. Thus, an addressee may infer a further goal on the speaker’s part to get a window opened or a fan switched on. This ambiguity presents somewhat of a challenge to the addressee to identify the goal of the speaker in producing such an utterance, and the eventual interpretation may involve computing an inference.

The availability of indirect meaning raises the question of when comprehenders decide to go beyond transparent meaning and make an inference. One salient cue is an assessment that the utterance fails to meet their expectations for what constitutes an appropriate conversational contribution. There are a number of theories about what these specific expectations are (e.g. Geurts, 2010, Grice, 1975, Levinson, 2000, Bohn et al., 2022). The central expectation is that our conversational partner is communicating rationally—namely, that they care about the goals of the exchange and formulate their contributions in a manner that allows them to communicate their message (Grice, 1975). Specifically, a rational speaker’s utterance should be relevant, informative, produced simply, and convey all necessary information for the conversational exchange (e.g. Bohn et al., 2022, Brown and Dell, 1987, Grice, 1975, Horn, 2004, Levinson, 2000, Sperber and Wilson, 1995/2004).

Pragmatic inferencing is often studied in terms of rational reasoning regarding the quantity of information stated or implied by an utterance; comprehenders expect an appropriate amount of information to be conveyed (see Quantity maxim Grice, 1975; Q- & R-principles Horn, 2004; and Q heuristic Levinson, 2000). If the message is deemed to have an insufficient quantity of information, a listener a listener may search for additional enrichments that would ensure that the utterance meets the expectations of cooperative discourse. Consider scalar implicatures, a type of quantity, or informativity-based inference, which can arise when a speaker has been seemingly underinformative by choosing not to use a stronger, more informative, expression. For example, a listener who hears some students passed or the tea is warm may draw an inference. The expressions some and warm are both members of lexical scales ordered on informativity (<some, all > and < warm, hot > ) and their use is compatible with stronger meanings (some and possibly all, warm and possibly hot). When a speaker uses a less informative expression from the scale, listeners try to reconcile their expectations regarding speaker informativity with the seeming underinformativeness of the expression by inferring that the stronger (more informative) expression does not hold, thereby deriving a scalar implicature (Gazdar, 1979; Geurts, 2010, Grice, 1975, Horn, 1972, 1989; Levinson, 2000).

A common feature of scalar inferences is that there are specific lexical triggers whose underinformativeness prompts the derivation of those inferences.1 In the case of scalar implicatures, it is often a quantifier such as “some” (see van Tiel et al., 2016). Models of pragmatic inferencing suggest that, providing the speaker is cooperative, the same scalar inferences would likely arise regardless of who specifically produces the utterance. In other words, the specific goals of a specific speaker are not what drives the emergence of scalar inferences. Rather, these emerge due to the inherent informativity of the lexical items used. Scalar implicature is a reliable but not specifically creative type of inference, in that it is often context independent. However, our linguistic experience is rich and many of the inferences that arise in conversations are not limited to the use of scalar terms that invoke reasoning about speaker underinformativeness.

Utterances need not be only lacking in information for comprehenders to compute an inference. In a series of studies, Kravtchenko and Demberg (2022) investigated pragmatic inferencing from overly informative, or redundant, propositions where the quantity of information provided was in excess of what would be expected. Through exploiting script knowledge of common events (e.g. going to a restaurant), they found that when a narrative explicitly mentioned an action that is easily inferred from the context (i.e. eating in a restaurant), comprehenders were found to interpret this action as atypical for the protagonist in the narrative (i.e. that the character does not typically eat when going to a restaurant). Upon encountering informationally redundant utterances, comprehenders worked to infer the speaker’s communicative goals, leading to inferences that the action was mentioned because it was atypical for this protagonist and hence newsworthy (a so-called atypicality inference). That is, the speaker’s choice to produce the utterance at all, given the obvious nature of the content, is often interpreted by listeners as a reflection of that feature’s atypicality and a signal for an addressee to compute an inference (e.g. Bergey and Yurovsky, 2022, Horowitz and Frank, 2016, Kreiss and Degen, 2020, Ryzhova et al., 2023). Thus, the decision to say anything at all may be a strong enough cue for a comprehender to expect a suitably interesting or informative contribution; when the utterance fails to meet these criteria, additional meaning may be computed through reasoning about the speaker’s goals.

Along with conveying a certain amount of information, speakers are also expected to convey this information in an appropriate way. Consider Levinson’s (2000) I-principle, what is not said is the obvious, and M−principle, what is said in an abnormal way is not normal, which apply to the use of (4) and (5).(4)

Bill caused the car to stop.

(5)

Bill stopped the car.

Under the M−principle, (4) implies that the car was stopped in a non-stereotypical way. Choosing to express a stereotypical event in this manner is not the conventional formulation. The marked nature of the utterance can lead addressees to try to reason why the utterance was produced in an unconventional manner when a more conventional alternative exists as in (5). Perhaps the speaker is trying to convey that Bill did not stop the car in the conventional manner e.g. rather than using the brakes, he used a tree.

Under the I-principle, we do not expect speakers to state the obvious or tell us information that is easily inferable. The choice to utter (5) is also marked since the event is stereotypical and the utterance can imply that, although the car was stopped stereotypically, there was a particular reason for doing so (as it is not usually noteworthy to mention that a car was stopped unless the object of interest is the cause, e.g. Bill stopped the car because the police flagged him down).

A core assumption within the literature on inferences that arise from apparent violations of manner or informativity is that the utterances are somehow relevant to the discourse. Without this assumption, it is unclear why a listener would ever go to the trouble of computing an inference. For scalar implicatures, reconciling the violation of informativity involves reasoning about why a speaker chose not to utter the stronger alternative and therefore how the utterance produced is relevant to the current discourse. For redundant utterances or those with unconventional formulations, reconciling the seeming violation of conversational norms involves reasoning about why the speaker produced so much extra information or why they used extra words to communicate it (yielding atypicality inferences). In each of these instances, a comprehender is reasoning about how the utterance is relevant as stated; why did the speaker choose to say what they said?

Successfully communicating requires: (a) the production of a signal, (b) the intention of a speaker to inform an addressee of something, and (c) an addressee recognising the speaker’s intention to inform. Thus, by choosing to produce an utterance (c) is realised and producing any utterance comes with a presumption of its own relevance (Sperber & Wilson, 1995/2004; Wearing, 2015). How an utterance is relevant is not always immediately clear.

Consider the following utterances:(6)

The library has books.

(7)

The library opens at 9:30am.

(8)

The library walls are blue.

Each of these utterances conveys some information about the library, but they vary in their informativity or usefulness. It is possible, for each of these instances, to imagine situations in which they are relevant contributions to a conversation.

It is unlikely that (6) would be considered an informative utterance2 since it states easily inferable information about a library and could be classed as overinformative or informationally redundant. One way that a listener might reconcile a speaker choosing to produce (6) might be to draw an atypicality inference: The speaker is cooperative and tries to provide newsworthy information; mentioning the presence of books at a library isn't normally newsworthy (books are an inherent property of libraries), but perhaps this library temporarily lacked books and the update about the new-found presence of books is worthy of note (e.g. Kravtchenko & Demberg, 2022).

The utterance in (7) is informative in that it provides useful information about the library; it presents information that is often sought and is actionable (i.e. opening times to allow for visiting the library) and reduces a listener’s uncertainty about the world.

In (8) however, the informativity and relevance of the utterance is less immediately obvious. While it does reduce uncertainty about the world by presenting information about the library, this information is not particularly useful in that there is no clear action that a listener can take after hearing this. Similarly, the information doesn't necessarily induce an atypicality inference; blue walls aren't an inherent property of libraries, thus the utterance of (8) fails to be specifically overinformative and redundant.

Utterances such as (8) present listeners with a choice; they could accept this as a piece of information about the world, a transparent mapping between utterance and meaning. However, the fact that a speaker chose to produce (8) at all may invite a listener to search for possible enrichments: why is such a trivial utterance relevant, what are the speaker’s goals, and potentially what additional meaning can be inferred that would make this a relevant contribution.

In a case like (9) below, additional meaning can arise if the addressee believes the speaker is: (i) sufficiently cooperative to adhere to the general goal of conveying sufficient information for the current exchange and (ii) knowledgeable of the situation over time. If these assumptions hold, one way of reconciling the production of an utterance about a fairly trivial situation is to infer (9a), whereby what is being communicated by the utterance is not just the colour of the walls but the (rarer) event of them having been painted a new colour. In other words, (9) permits the possible inference of (9b).(9)The library walls are blue.a.

→the situation has changed.

b.

→ the walls used to be different.

It is important to note that we are not claiming that the something has changed inference is a particular category of inference. Rather, it is an illustration of a potential reading that could arise from the speaker’s choice to produce an utterance when the alternative was to remain silent. Note that an addressee may evaluate the utterance differently depending on their assessment of the content, the speaker, and the communicative context: In some cases, they may not retrieve any additional meaning and rather accept the information that the library walls are blue as simply the speaker’s presentation of a fact about the world; alternatively, they may retrieve another inference, for example that the speaker thinks painting library walls blue is unusual; lastly, they may take the utterance as a social gesture of small talk, without much transfer of contentful information.

Previous research shows that conversational partners engage in a wide range of reasoning about each other, guiding both speakers’ production decisions (e.g. Brown-Schmidt and Konopka, 2011, Lockridge and Brennan, 2002, Nadig and Sedivy, 2002, Jara-Ettinger and Rubio-Fernandez, 2021, Yoon et al., 2012) and the meanings derived by listeners (e.g. Cai et al., 2017, Davies et al., 2022, Grodner and Sedivy, 2011, Loy et al., 2019, Regel et al., 2010, Yoon et al., 2021). The decision to draw an inference can be influenced by a variety of speaker-specific factors such as speaker reliability (e.g. Diesendruck et al., 2010; Grodner & Sedivy, 2011; Sobel et al., 2012) or speaker fluency (e.g. Arnold et al., 2004, Arnold et al., 2007, 2007; Loy et al., 2019, Orena and White, 2015, Yoon et al., 2021).

One crucial factor for inferencing is speaker knowledge. In a standard model of pragmatic inferencing, information about the speaker’s knowledge state is used in the derivation process (Geurts, 2010). That is, in order for a listener’s inference of a speaker’s intended meaning to be licensed, the listener needs to assume that the speaker has knowledge (or an opinion) about the proposition of the utterance. Indeed, empirical evidence supports this view as inferencing occurs more reliably when speakers are deemed knowledgeable (Bergen and Grodner, 2012, Moty and Rhodes, 2021). Bergen and Grodner (2012) found that participants were less likely to compute a scalar implicature when the speaker had “skimmed” the document they were mentioning compared to a speaker who had “read meticulously”. Moty and Rhodes (2021) showed that the inferences drawn from generic utterances can be influenced by speaker knowledgeability; inferences about group characteristics only arose when the speaker was knowledgeable. More recently, work from Zhang and Wu (2023) demonstrates that both native and non-native speakers are sensitive to speaker knowledge states and are more likely to draw inferences from knowledgeable speakers.

In the present work we ask how linguistic and extralinguistic speaker-specific characteristics contribute to the interpretation of trivial utterances and the potential inferences that arise. As noted, one challenge faced by comprehenders is recognising when to draw an inference since a key characteristic of pragmatic inferences is they are not strictly part of what is said. Although a speaker is never on record as having asserted the additional enriched content, an addressee may nonetheless be inclined to identify and infer such meaning because doing so may be necessary for maintaining the sense and relevance of a conversational contribution (Sperber and Wilson, 1995, Carston, 2004). However, the felicity of such an inference may be dependent on Speaker Knowledge or Speaker Style.

In all experiments reported here, we manipulate Speaker Knowledge by varying whether the speaker is described as being familiar or unfamiliar with the location they are talking about. In Experiment 1, we investigate whether the decision for a speaker to produce an utterance is sufficient for a comprehender to infer a goal or reason behind the utterance. In other words, do participants accept trivial utterances as simple statements of fact about the world or are they inclined to draw an inference about why the speaker chose to produce that particular utterance? We further investigate whether the rate at which participants draw such inferences is influenced by the presence/absence of a linguistic emphasis cue (“Hey, guess what”). The results show that participants draw inferences even in the absence of such a cue, suggesting that the availability of that inference does not depend on an explicit signal to pay attention to a speaker’s contribution. Instead, there is a more general expectation that a speaker will have a reason for producing a particular utterance when they could have said nothing (see assumption of relevance, Sperber and Wilson, 1995, Wearing, 2015).

In Experiment 2, we additionally vary the Speaker Style, including or excluding a description that the speaker “isn’t normally very chatty.” Such a speaker presumably has a higher threshold for what is worth mentioning, and so if they produce a seemingly trivial utterance, comprehenders may be more likely to ascribe greater meaning to the choice to produce that utterance. We demonstrate that relevance inferences are computed more often when trivial utterances are produced by a knowledgeable speaker who is characterised as quiet.

A common question regarding pragmatic inferences relates to whether or not the process of inferencing is costly. A number of studies demonstrate that it can be costly to compute pragmatic inferences. For example, in eye-tracking studies participants show a delay in fixating on the implicature-consistent target (Huang & Snedeker, 2009); in sentence verification tasks, responses consistent with an implicature response take longer than non-implicature responses (Bott & Noveck, 2004); and ERP data demonstrates a processing cost for implicature computation (Noveck & Posada, 2003). This finding that implicatures are costly does not always hold (e.g. Breheny et al., 2013, Grodner et al., 2010). Indeed, inferences are drawn with differing levels of ease, dependent on a number of constraints (e.g. type of trigger: van Tiel & Schaeken, 2017; knowledge of speaker: Bergen & Grodner, 2012; Kampa and Papafragou, 2020, Papafragou et al., 2018; contextual support: Breheny et al., 2006; Degen & Tanenhaus, 2015; Huang & Snedeker, 2018; Singh, 2019, van Tiel et al., 2019). The standard account for why implicatures may be costly is that a comprehender is posited to interpret the literal meaning first (e.g. Grice, 1975, Geurts, 2010), at which point they notice the violation of their expectations in the speaker's utterance and look for a way to enrich the meaning that would allow the utterance to constitute a cooperative contribution to the conversation. This process of identifying the seeming violation, rejecting the literal meaning, reasoning about the speaker’s intentions, and computing the inference is posited to be costly. In Relevance Theoretic terms, comparing the positive cognitive effects of the additional meaning against the effort of doing the inference calculation may itself be costly (Sperber & Wilson, 1995).

The process for computing relevance inferences may be much the same, whereby listeners who encounter a trivial utterance may seek to reconcile the mismatch between their expectations for a cooperative utterance and the trivial utterance encountered. However, it is not known whether their computation is costly or not. In Experiments 3 and 4, we test whether or not relevance inferences are costly to compute and at what point in processing these inferences arise. If there is a cost to computing relevance inferences, this cost may be visible early in processing, i.e. upon encountering a trivial utterance, or at a later stage when prompted to endorse the inferred meaning.