Search

Pragmatics is core to natural language, enabling speakers to communicate efficiently with structures like ellipsis and anaphora that can shorten utterances without loss of meaning. These structures require a listener to interpret an ambiguous form—like a pronoun—and infer the speaker’s intended meaning—who that pronoun refers to. Despite potential to introduce ambiguity, anaphora is ubiquitous across human language. In an effort to better understand the origins of anaphoric structure in natural language, we look to see if analogous structures can emerge between artificial neural networks trained to solve a communicative task. We show that: first, despite the potential for increased ambiguity, languages with anaphoric structures are learnable by neural models. Second, anaphoric structures emerge between models ‘naturally’ without need for additional constraints. Finally, introducing an explicit efficiency pressure on the speaker increases the prevalence of these structures. We conclude that certain pragmatic structures straightforwardly emerge between neural networks, without explicit efficiency pressures, but that the competing needs of speakers and listeners conditions the degree and nature of their emergence.

Conversational partners expect each other to communicate rationally and cooperatively and to contribute relevant and informative utterances. Occasionally, however, speakers produce trivial utterances which may violate our expectations of informativity. These utterances can prompt listeners to draw inferences about a speaker’s goals in producing such an utterance. Here we present two studies investigating how and when listeners derive informativity-based inferences. The results demonstrate that speaker knowledge plays an important role in the computation of inferences. Furthermore, the timecourse of the results suggests that these inferences do not always arise automatically and that their computation is costly.

The current project aims to investigate how the at-issue status of clauses impacts the processing of these clauses, as well as subsequent discourse. To test this, we designed two experiments, each investigating this for a different type of main clause-subordinate clause construction: Experiment 1 Sentences containing an adverbial clause that starts with 'after' Experiment 2 Relative clause constructions: subject- and object-modifying Experiment 1 is a self-paced reading experiment which will be conducted on sentences containing adverbial clauses. In these sentences, the adverbial clause is not expected to be at-issue, and the main clause is. We will also test how content that follows these sentences is processed differently based on whether it is coreferential with the at-issue content (main clause) or the not-at-issue content (adverbial clause). For example, the reading times of below bracketed clauses will be compared: (1a) [subordinate clause, low at-issue potential] After [they had dinner at a French bistro,] they went to a violin concert. (1b) [main clause, high at-issue potential] After they went to a violin concert, [they had dinner at French bistro.] In the sentence that follows these, the bracketed clause content will be coreferential with either the subordinate clause content (1a), or the main clause content (1b): (2) It was [a very small cozy restaurant.] In experiment 2, we want to find out how clauses that are ambiguous with respect to their at-issue status, object-modifying relative clauses, are processed differently compared to clauses that have a clear at-issue status, subject-modifying relative clauses (not-at-issue) and the main clauses that embed these (at-issue). Furthermore, we aim to investigate how content that follows above-mentioned relative clause constructions is processed differently based on whether its content is coreferential with at-issue content, not-at-issue content, or content that is ambiguous with respect to its at-issue status. For example, the reading times of below bracketed clauses will be compared: (3a) [subject-modifying relative clause, low at-issue potential] My dad, who [was having a snack of chips and dip,] opened a bottle of wine for my uncle. (3b) [object-modifying relative clause, medium at-issue potential] My dad opened a bottle of wine for my uncle, who [was having a snack of chips and dip.] (3c) [main clause that embeds a subject-modifying relative clause, high at-issue potential] My dad, who opened a bottle of wine for my uncle, [was having a snack of chips and dip.] In the sentence that follows these, the bracketed clause content will be coreferential with either the subject-modifying relative clause content (3a), the object-modifying relative clause content (3b), or the main clause content (3c): It was [tortilla chips with guacemole.]

Communication depends in part on the hope that people will tell us things we’re interested in with language providing a channel by which speakers can convey, among other things, newsworthy and informative messages. Despite this, research often emphasises the role of predictability, showing that comprehenders make fast and sophisticated use of knowledge about typical situations to shape their expectations about upcoming words in fine-grained and context-driven ways. This new study contrasts these two dimensions, distinguishing knowledge about situation plausibility from knowledge about what people actually choose to talk about.

This fourth registration aims to fill in gaps that were identified after carrying out experiments affiliated with the third registration part of this project: 'Processing at-issue content part 3: disentangling recency from clause type (again)'. I will first review the findings of two experiments of the third registration (experiment 5 & 6) in order to set up the goals of the current registration. These self-paced reading experiments - experiment 5 with sentences containing an adverbial clause that starts with 'after', and experiment 6 with subject- and object-modifying relative clause constructions - were designed to investigate which factor had a greater impact on the at-issue status of the individual clauses of above-mentioned sentence types: clause type, or recency. This was tested by measuring how fast a {continuation sentence} was processed that had content that was coreferential with content present in only one of the clauses in the sentence it followed. In experiment 5 this was done by comparing the reading times of {continuation sentences} in two conditions: one in which the {continuation} had content that was coreferential with content that was coreferential with content in a sentence-final adverbial clause (1a), and one in which the {continuation} had content that was coreferential with content in a sentence-early matrix clause (1b): Experiment 5 - adverbial clauses: (1a) They went to a violin concert after they had dinner at a French bistro. It was {a very small cosy restaurant.} (1b) They had dinner at a French bistro after they went to a violin concert. It was {a very small cosy restaurant.} In experiment 6 this was done by comparing the reading times of {continuation sentences} in three conditions: one in which the {continuation} had content that was coreferential with content in a sentence-early matrix clause (2a), one in which the {continuation} had content that was coreferential with content in a sentence early relative clause (2b), and one in which the {continuation} had content that was coreferential with content in a sentence-final matrix clause (2c): Experiment 6 - relative clauses: (2a) My dad opened a bottle of wine for my uncle, who was having a snack of chips and dip. It was {a dry white wine from Italy.} (2b) My dad, who opened a bottle of wine for my uncle, was having a snack of chips and dip. It was {a dry white wine from Italy.} (2c) My dad, who was having a snack of chips and dip, opened a bottle of wine for my uncle. It was {a dry white wine from Italy.} Experiment 5 yielded no significant differences in reading times for the continuations, leaving the question as to which factor - recency or clause type - is more important in assigning at-issue status to clauses, open. In experiment 6 we found that the continuation in (2c) was read faster than in (2b), but no differences were found between (2a) and any of the other conditions. This also means we cannot draw any conclusions about which factor - recency or clause type - is more important in assigning at-issue status to clauses, because in the fastest condition (2c), the {continuation} connected with a clause that was both a main clause and recent; both of these contribute to making said clause more at-issue. In fact, so far across all studies we have conducted, all instances where we found a continuation to be read significantly faster were when it was coreferential with a clause that was a recent matrix clause, or a recent relative clause (object-modifying). The current follow-up studies aim to investigate if clause type contributes to the at-issue status of clauses when these clauses are not recent. I.e., we will hold constant position (sentence-early), to try and isolate an effect of clause type. Theories predict that matrix clauses can always be at-issue, also when they are sentence-early. Our experiments so far have not been able to evidence this. Therefore, we have designed two follow-up experiments in which the at-issue status of a sentence-early matrix clause is compared to the at-issue status of other clauses in this same position that theories predict are not likely to be at-issue. In experiment 7 this will be done by comparing the reading times of {continuations} when their content is coreferential either with content in a sentence-early matrix clause (3b) or a sentence-early adverbial clause (3a): Experiment 7 - adverbial clauses: (3a) After they had dinner at a French bistro, they went to a violin concert. It was {a very small cosy restaurant.} (3b) They had dinner at a French bistro after they went to a violin concert. It was {a very small cosy restaurant.} In experiment 8 this will be done by comparing the reading times of {continuations} when their content is coreferential either with content in a sentence-early matrix clause (4b) or a main clause that is followed by another main clause (i.e.: two free-standing units/sentences). The point is to compare the same content when the intervening material is either a subordinate clause, leaving the matrix clause accessible, or a free-standing sentence, making the main clause before that harder to access. The goal for the condition with two main clauses, was to have them both answer the same Question Under Discussion of 'what was happening?', which is compatible with a coordinating relation. This makes this first main clause less accessible (not-at-issue): Experiment 8 - relative clauses: (4a) My mom opened a bottle of wine for my uncle. He was having a snack of chips and dip. It was {a dry white wine from Italy.} (4b) My mom opened a bottle of wine for my uncle, who was having a snack of chips and dip. It was {a dry white wine from Italy.} For experiment 8, the items underwent some additional changes. In the condition where there are two main clauses (4a), the wh-pronoun is replaced by a personal pronoun. To make sure that this pronoun can only refer to one of the participants in the preceding main clause, we made sure that all items contained at least one participant with an explicitly clear gender, so that the pronoun always only had one possible referent. This was done to avoid any slowed-down reading times possibly arising because of issues with ambiguous pronoun resolution.

We learn about the world through language. Communication is an inherently complicated process. On the surface it seems simple enough to take the sounds we hear and map those onto words and then onto meaning. However, listeners often go beyond the surface form of an utterance and compute additional meaning (Grice, 1975). That is, they compute pragmatic inferences. For example, if I say in conversation that there is a “yellow banana” this would give rise to the inference that there is another non-yellow banana present (Sedivy, 2003). Since the colour of a banana is typically rather mundane, as listeners we intuit why the speaker has chosen to specify the colour. Pragmatic inferences rely on the expectation that interlocutors communicate rationally and cooperatively. We have an expectation for conversational contributions to add something to the discourse (Rohde, Futrell, & Lucas, 2021). In the studies registered here we examine inferences that arise from mundane content at the propositional level. For example, if a speaker proffers a mundane utterance like “I saw that the library walls are blue” when talking about a location they are familiar with, a listener is licensed to look for a communicative reason for such a mundane contribution to the conversation (see Kravtchenko & Demberg 2015). Study 1 will examine how speaker characteristics influence the inferences drawn by listeners. Study 2 will examine how highlighting the speaker's intention to be communicative influences the inferences drawn by listeners.

This is the second pre-registration in connection with the project "Color interpretation is guided by informativity expectations, not by world knowledge about colors". We are pre-registering the analyses of an eye-tracking experiment we already conducted, but have not yet analyzed.

Our previously registered first part of this project aimed to investigate how the at-issue status of clauses impacts the processing of these clauses, as well as subsequent discourse. This was tested with two experiments: Experiment 1: sentences containing an adverbial clause that starts with 'after' Experiment 2: relative clause constructions; subject- and object-modifying In experiment 1 we found that discourse segments following a sentence containing an adverbial clause (continuations) were read faster when their content was coreferential with an element that was only present in the main clause (1a), and read slower when it was coreferential with an element that was only present in the adverbial clause (1b). The bracketed content { } represents the continuation in all examples below. {continuation} (1a) After they went to a violin concert, they had dinner at a French bistro. It is {a very small cozy restaurant.} (1b) After they had dinner at a French bistro, they went to a violin concert. It is {a very small cozy restaurant.} In experiment 2 we found that continuations were read faster when their content was coreferential with an element that was only present in the main clause (2a) or sentence-final relative clause (2b), than when it was coreferential with an element that was only present in the sentence-early relative (2c). {continuation} (2a) My dad, who opened a bottle of wine for my uncle, was having a snack of chips and dip. It was {tortilla chips with guacamole.} (2a) My dad opened a bottle of wine for my uncle, who was having a snack of chips and dip. It was {tortilla chips with guacamole.} (2c) My dad, who was having a snack of chips and dip, opened a bottle of wine for my uncle. It was {tortilla chips with guacamole.} While these results were expected, our previous design makes it impossible to disentangle the effect of recency from the effect of clause type, both of which could have an impact on the at-issue potential of discourse segments. For all conditions in which the continuation was read faster, it was coreferential with an element that was only present in the clause it immediately followed, and no difference was found when this clause was a main clause or an object-modifying relative clause. The current follow-up study will make it possible to disentangle clause type from clause position (recency) by presenting the participants with similar items, in which the continuation is (also) coreferential an element that was only present in a sentence-early main clauses (3a & 4a). Comparing the reading times for these to the reading times of continuations that are coreferential with an element only present in a sentence-final adverbial clauses (3b), as well as sentence-early relative clauses (4b) and sentence-final main clauses (4c), will help us understand better how both clause type and clause position (recency) impact the at-issue status of discourse segments. Experiment 3: {continuation} (3a) They went to a violin concert after they had dinner at a French bistro. It is {a very small cozy restaurant.} (3b) They had dinner at a French bistro after they went to a violin concert. It is {a very small cozy restaurant.} Experiment 4: {continuation} (4a) My dad opened a bottle of wine for my uncle, who was having a snack of chips and dip. It was {a dry white wine from Italy.} (4b) My dad, who opened a bottle of wine for my uncle, was having a snack of chips and dip. It was {a dry white wine from Italy.} (4c) My dad, who was having a snack of chips and dip, opened a bottle of wine for my uncle. It was {a dry white wine from Italy.}

This third registration aims to address potential confounds that were identified after carrying out experiments affiliated with the second registration part of this project: 'Processing at-issue content part 2: disentangling recency from clause type'. These self-paced reading experiments - experiment 3 with sentences containing an adverbial clause that starts with 'after', and experiment 4 with subject- and object-modifying relative clause constructions - were designed to investigate which factor had a greater impact on the at-issue status of the individual clauses of above-mentioned sentence types: clause type, or recency. This was tested by measuring how fast a {continuation sentence} was processed that had content that was coreferential with only one of the clauses in the sentence it followed. In experiment 3 this was done by comparing the reading times of {continuation sentences} with content that was coreferential with content in either a sentence-final adverbial clause (1a), or a sentence-early matrix clause (1b): Experiment 3 - adverbial clauses: (1a) They went to a violin concert after they had dinner at a French bistro. It was {a very small cosy restaurant.} (1b) They had dinner at a French bistro after they went to a violin concert. It was {a very small cosy restaurant.} In experiment 4 this was done by comparing the reading times of {continuation sentences} with content that was coreferential with content in a sentence-early matrix clause (2a), a sentence early relative clause (2b), or a sentence-final matrix clause (2c): Experiment 4 - relative clauses (2a) My dad opened a bottle of wine for my uncle, who was having a snack of chips and dip. It was {a dry white wine from Italy.} (2b) My dad, who opened a bottle of wine for my uncle, was having a snack of chips and dip. It was {a dry white wine from Italy.} (2c) My dad, who was having a snack of chips and dip, opened a bottle of wine for my uncle. It was {a dry white wine from Italy.} Neither experiments yielded any significant differences in reading times for the continuations, leaving the question as to which factor - recency or clause type - is more important in assigning at-issue status to clauses, open. While it is clear from the first study (see previous registrations for this project) that recency is an important factor, it is not yet clear whether clause type is as well. The current follow-up study will address this question once more, by carrying out follow-up studies for both experiments which will address potential confounds in experiments 3 and 4. There are three potential reasons we identified for our null results: 1. In experiments 3 & 4, the 'it is/was' pre-region was presented immediately before the {continuation}, at the line below the sentence with which 'it is/was' establishes coreference once the {continuation} is read. This was changed from experiments 1 & 2, where the 'it is/was' pre-region was presented immediately after the coreference sentence, after which the {continuation} was presented on the next line. While there is no immediate reason to believe that the position of this pre-target region with respect to on which line it is presented should have such an effect, it is the only notable difference between experiment 4 and its predecessor (experiment 2). Conditions (2b) and (2c) were both present in experiment 2, and significant differences were found between these two conditions then. Since this result was missing from experiment 4 where the pre-target region was on a different line, we want to make sure that this was not the reason for the absence of an effect. If we replicate experiment 4 and find no effect again, it might mean that the significant differences we found in experiment 2 were spurious differences. If clause type and/or recency are important factors influencing reading times of the {continuation}, experiment 6 (our replication of experiment 4, in line with the positioning of the pre-target region of experiment 2), should show that the continuation is read faster in (3c) than in (3b). Experiment 6 - relative clauses (3a) My dad opened a bottle of wine for my uncle, who was having a snack of chips and dip. It was {a dry white wine from Italy.} (3b) My dad, who opened a bottle of wine for my uncle, was having a snack of chips and dip. It was {a dry white wine from Italy.} (3c) My dad, who was having a snack of chips and dip, opened a bottle of wine for my uncle. It was {a dry white wine from Italy.} 2. In experiment 3, the sentence containing the adverbial clause was presented as a single [chunk] (4) rather than being broken up in three [chunks] (5) as was done in experiment 1(6): (4) [They went to a violin concert after they had dinner at a French bistro.] (5) [They went to a violin concert] [after] [they had dinner at a French bistro.] (6) [After] [they went to a violin concert,] [they had dinner at a French bistro.] This was done to avoid a reading where the subordinating adverbial clause was actually read as coordinating (7): (7) They went to a violin concert, after, they had dinner at a French bistro. While a single [chunk] might have facilitated the subordinating reading, it could be facilitated even more so by presenting the sentence in two individual [chunks]: one containing the matrix clause, and one containing the adverbial clause (8): (8) [They went to a violin concert] [after they had dinner at a French bistro.] By chunking the regions like this, we also avoid having one [chunk] that is much larger than all of the surrounding [chunks], making the experiment more cohesive throughout with no [chunks] that stand out to participants for their size. This leads to below conditions for experiment 5: Experiment 5 - adverbial clauses: (9a) [They went to a violin concert] [after they had dinner at a French bistro.] [It was] {a very small cosy restaurant.} (9b) [They had dinner at a French bistro] [after they went to a violin concert.] [It was] {a very small cosy restaurant.} Both of these changes mean that experiments 5 & 6 will be closer to the original design of experiments 1 & 2, where our findings were in line with our expectations and predictions. That said, if we replicate our null results, this could mean that the significant differences we previously found in experiments 1 & 2, were spurious differences. 3. Lastly, in experiment 3, the null-effect can be explained by the conditions themselves. The conditions were balanced so that in both of them the {continuation} attaches to a clause that has a characteristic that makes it likely to be at-issue, and one that makes it less likely to be at issue. In (1a), the {continuation} attaches to a sentence-late clause, but it is a subordinate clause. In (1b) the {continuation} attaches to a main clause, but it is sentence-early. If this is the (main) reason for our null-effect, we don't expect to find anything different for experiment 5 compared to experiment 3. However, we want to make sure that any other potential confounds are addressed before concluding this.

Communication depends in part on the hope that people will tell us things we’re interested in with language providing a channel by which speakers can convey, among other things, newsworthy and informative messages. Despite this, research often emphasises the role of predictability, showing that comprehenders make fast and sophisticated use of knowledge about typical situations to shape their expectations about upcoming words in fine-grained and context-driven ways. This new study contrasts these two dimensions, distinguishing knowledge about situation plausibility from knowledge about what people actually choose to talk about.

This is a pre-registration for the second study of a project looking at expectations for informativity in language processing. Comprehenders rely on real-world knowledge when guessing upcoming content. For example, taking a sip from the waterfall is more expected than taking a sip from the transmitter (Kutas & Hillyard 1980). By contrast, speakers tend to omit typical content in favour of newsworthy content: they include an optional instrument more often when it is atypical, favouring the production of stab with an ice pick over stab with a knife (Brown & Dell 1987). These findings are taken to reflect constraints on cooperative communication whereby speakers’ contributions are expected to be appropriately informative and relevant (Grice 1975). If comprehenders are rational and are sensitive to speakers’ production preferences, these should also bear upon their predictions about upcoming content: that is, they should expect the kind of content that cooperative speakers are likely to mention, rather than just the kind of content that is likely to be the case in the real world. In our first study in this project (Experiment 1), we elicited sentence completions via a Cloze task as an index of comprehenders’ expectations about upcoming material (Taylor 1953). We manipulated the salience of the speaker and showed that the more aware participants are of the speaker as an intentional communicator, the more informative they expect the speaker’s contribution to be. In this second study, we will investigate whether this expectation for informativity is further malleable depending on properties of the speaker (Grodner & Sedivy 2011). Specifically, we ask whether participants will adjust their expectations based on the speaking style of two different speakers: one speaker routinely makes uninformative utterances, and one produces utterances that are highly informative.

Language use can be characterised as transparent, stating facts about the world, or non-transparent, requiring additional meaning to be inferred. The challenge faced by addressees is recognizing when language use is transparent or not. The current study investigates two factors that may influence how readily participants interpret utterances as instances of transparent or non-transparent language use; speaker knowledgeability and utterance form. When utterances involved negation participants were more likely to recognize this as non-transparent language use and infer that the situation is usually different. Whereas speaker knowledge did not influence how utterances were understood.

This study aims to investigate whether language comprehenders are sensitive to the likelihood of color adjective use (e.g., how frequently people use the word 'yellow') for different object categories (e.g., to refer to a banana vs a shirt). Eye-tracking studies show that comprehenders rely on their conceptual knowledge about the color of objects, looking at yellow things when they hear a word for a typically yellow object (e.g., a banana when hearing 'sunflower'; Huettig & Altmann, 2004, 2011; Yee & Sedivy, 2006; Naor Raz, 2003). However, what happens when they hear a speaker mention a color? If comprehenders assume that the speaker is being informative, they should anticipate an object that has low probability of possessing that color in the real world. This yields an interesting empirical prediction: while prior works shows that comprehenders anticipate yellow objects following the word 'banana', is it possible that they disfavor bananas following the word 'yellow'? All our studies present participants with a display containing a pair of foods and a pair of clothes; the linguistic materials consist of two conditions, a Color Adjective condition and a Number Word condition. This pre-registration pertains to two new studies: we have already conducted one pilot study on Mechanical Turk (pilot, N=20) using an original set of visual materials, and a full study on Mechanical Turk (Study 0, N=52) using the same set of visual materials. The two new studies are as follows: Study 1. New web study with improved, high-resolution, uniform background images corresponding to the same food-clothing pairs as in the original study (e.g., two yellow bananas paired with two yellow shirts). This study manipulates Modification (Color Adjective vs Color Word) and food-clothing pairing (the arbitrary pairings used in Study 0 vs a second arbitrary set of pairings; e.g., two yellow bananas paired with two yellow skirts). Study 2. Eye-tracking study with a new set of high-resolution, uniform background images corresponding to the same food-clothing pairs as in Study 0. Data for Study 2 has already been collected, but no eye-tracking or response time analysis has been performed. Picture selection analysis, on the other hand, has already been performed. Given the different stages of Study 1 and Study 2, we will write two separate pre-registrations. Study 3. Corpus analysis of Google ngram data. We are not pre-registering this work because we have already conducted exploratory analyses and visualizations, and unlike Studies 1 and 2, this corpus analysis does not involve data collection by the authors.

Contains fulltext : 231577.pdf (Publisher’s version ) (Open Access) Coherence relations are often assumed to hold between clauses, but restrictive relative clauses (RCs) are usually not granted discourse segment status because they are syntactically and conceptually integrated in their matrix clauses. This paper investigates whether coherence relations can be inferred between restrictive RCs and their matrix clauses. Three experiments provide converging evidence that restrictive RCs can indeed play a role at the discourse level and should not categorically be excluded from receiving discourse segment status in discourse annotation practices. At the same time, the studies provide new insights into implicit causality verb biases, specifically about next-mention biases in concessive coherence relations, and expectations about discourse structure, upcoming referents, and upcoming coherence relations. 07 december 2020 18 p.

If speakers communicate efficiently, they should produce more linguistic material when comprehension difficulty increases. Comprehension difficulty can be impacted by the message itself (studied extensively) or by properties of the listener (studied less). Here we investigate the impact of listeners’ estimated proficiency on speakers’ productions, using referential choice as a case study. Compared to full noun phrases (fNPs, ‘the woman’), pronouns (‘she’) can convey similar content with less linguistic material: accordingly, speakers use more fNPs when the message is unpredictable (Tily & Piantadosi, 2009) and when listeners lack relevant information (Bard & Aylett, 2004). If referential choice is also impacted by listeners’ estimated proficiency, then speakers should use more fNPs when conversing with language learners. To test this, we compared participants' descriptions of the same picture book to learners and proficient speakers. Indeed, participants used more fNPs when their interlocutors were child- or adult-learners, illustrating efficient adaptation to listener type.

Are the cues that speakers produce when lying the same cues that listeners attend to when attempting to detect deceit? We used a two-person interactive game to explore the production and perception of speech and nonverbal cues to lying. In each game turn, participants viewed pairs of images, with the location of some treasure indicated to the speaker but not to the listener. The speaker described the location of the treasure, with the objective of misleading the listener about its true location; the listener attempted to locate the treasure, based on their judgement of the speaker’s veracity. In line with previous comprehension research, listeners’ responses suggest that they attend primarily to behaviours associated with increased mental difficulty, perhaps because lying, under a cognitive hypothesis, is thought to cause an increased cognitive load. Moreover, a mouse-tracking analysis suggests that these judgements are made quickly, while the speakers’ utterances are still unfolding. However, there is a surprising mismatch between listeners and speakers: When producing false statements, speakers are less likely to produce the cues that listeners associate with lying. This production pattern is in keeping with an attempted control hypothesis, whereby liars may take into account listeners’ expectations and correspondingly manipulate their behaviour to avoid detection.

When do speakers produce ambiguous expressions? How do comprehenders interpret such expressions to infer a speaker’s intended meaning? This chapter reviews a body of work on pronoun production and interpretation, considering a number of computational, linguistic, and psycholinguistic frameworks and the factors that have been posited to drive pronoun use. These factors include surface structural elements (grammatical role, syntactic parallelism), information structural cues (topichood), lexical semantics and real-world knowledge (thematic roles, causality), as well as aspects of the larger discourse (recency, rhetorical structure, coherence relations). The chapter then turns to a more general model of message production and interpretation that incorporates elements of existing pronoun models using a Bayesian approach.

Listeners' interpretations of deceit are no higher than chance, arguably because listeners are biased towards such interpretations upon encountering elements in speech that do not correlate with deceptive behaviour. Important to this project are 1) the linguistic background of the speaker, and 2) the way in which they speak. Regarding (1), research has found that L1 listeners sometimes display a lie-bias towards L2 speakers, which can be explained by processing fluency and stereotyping (e.g., Lev-Ari & Keysar, 2010; Foucart & Hartsuiker, 2021). When it comes to (2), filled pauses (e.g., uh or um) are usually reported as indices of deception (Arciulli et al., 2011). Interestingly, in an eye-tracking study, Loy et al. (2017) found that this interpretation arises early in the comprehension process. While these two cues in themselves can bias listeners' interpretations towards deceit, it is worth acknowledging that their co-presence could affect the exerted bias. For example, at the lexical level, filled pauses produced by L2 speakers do not trigger predictions as L1's do (Bosker et al., 2014). Therefore, it could be possible that L2's disfluencies do not trigger interpretations of deceit, or that such interpretation arises later compared to L1's disfluencies. In the present study, we tested 64 L1 British English listeners following Loy et al.'s (2017) treasure-hunt eye-tracking study. Listeners heard either an L1 (British English) or an L2 (L1: Spanish) speaker referred to the location of a treasure fluently or disfluent (e.g., "The treasure is behind the/thee uh apple") (see Materials component). We analysed participants' clicking behaviour and eye-movements (see Data) via Linear Mixed Models (see Analysis, with an extra section for GAMM analysis). We found a main effect of manner of delivery, whereby disfluent utterances led to more interpretations of deceit. Importantly, we did not find a main effect of speaker's linguistic background, nor an interaction between these factors. The timecourse of this interpretation, as reflected in eye movements, did not differ depending on the speaker's linguistic background: The bias exerted by the presence of a disfluency was integrated early in both scenarios, replicating previous findings. We argue that in such a constrained scenario, the stereotypes of how a liar sounds (i.e., the disfluency-as-deception bias) is a heavily anchored heuristic that can override any other biases.

Since \citet{HooperThompson1973}, many researchers have pursued the insight that V2 is licensed by assertion. H\&T categorise predicates depending on whether their complement can be asserted: e.g. communication verbs (\emph{say}) permit the assertion of their complement, in contrast to factives (\emph{be happy}). \citet{Simons2007} proposes distinguishing between embedded propositions that do or do not constitute the Main Point of Utterance (MPU) -- a sharpening of the notion of assertion: in question/response-sequences, the proposition answering the question is the MPU. Given this definition/diagnostic for assertion, factives \emph{can}, given the appropriate discourse context, embed MPU and thus should allow embedded V2 (EV2). This paper presents two experiments testing whether factives can embed MPU and whether MPU licenses EV2 in Swedish. The results support both \citegen{Simons2007} contention that factives can embed MPU, while providing new evidence that MPU does not correlate with EV2.} &nbsp

Comprehenders’ perception of the world is mediated by the mental models they construct. During discourse processing, incoming information allows comprehenders to update their model of the events being described. At the same time, comprehenders use these models to generate expectations about who or what will be mentioned next. The temporal dynamics of this interdependence between language processing and mental event representation has been difficult to disentangle. The present visual world eye-tracking experiment measures listeners’ coreference expectations during an intersentential pause between a sentence about a transfer-of-possession event and a continuation mentioning either its Source or Goal. We found a temporally dispersed but sustained preference for fixating the Goal that was significantly greater when the event was described as completed rather than incomplete (passed versus was passing). This aligns with reported offline sensitivity to event structure, as conveyed via verb aspect, and provides new evidence that our mental model of an event leads to early and, crucially, proactive expectations about subsequent mention in the upcoming discourse.

This dissertation provides a psycholinguistic investigation of the influence of discourse on language comprehension. It examines factors that allow comprehenders to follow a discourse, to form representations of the events being described, and to make predictions about how subsequent utterances will relate to prior linguistic material. Previous work has recognized the importance of prediction in sentence-internal processing: transition probabilities at the phonemic level, semantic associations in lexical access, and structural frequencies at the syntactic level. The work presented here investigates whether learnable statistical regularities also exist at the discourse level, a topic that has remained largely unexplored in the psycholinguistics literature. The dissertation presents a series of experiments testing the extent to which comprehenders use various pragmatic cues to make predictions about how a discourse will be continued. In order to quantify discourse-level information, the experiments use an inventory of coherence relations adopted from the theoretical linguistics and artificial intelligence literatures. The experimental results demonstrate that comprehenders do indeed make use of available pragmatic cues to generate expectations about upcoming coherence relations. Furthermore, the results show that the mechanisms for establishing coherence relations can inform our understanding of two well studied sentence-internal phenomena: coreference and syntactic ambiguity. The online results establish the importance of these pragmatic cues in comprehenders incremental sentence processing. The coherence-based approach taken here provides a lens through which to view previous results in the domains of both coreference and syntactic ambiguity. The fact that phenomena in both these domains appear to be sensitive to coherence-driven biases suggests that these biases may be more pervasive than has been previously acknowledged. This work indicates that future processing models of sentence and discourse processing must take into account effects that emerge from discourse coherence