Inferences about event outcomes influence text-based memory of event outcomes

2.1 Intention fulfilment, event outcomes, and relevant terms

To situate this study with respect to existing event literature, we compare ‘event outcomes’ – the focus of the present study – with the following concepts: ‘event boundary’ (e.g., Zacks et al. 2009), ‘event ending/endpoint’ (e.g., von Stutterheim et al. 2012), ‘event completion/cessation’ (Sakarias and Flecken 2019), and ‘event result’ (e.g., Santin et al. 2021; Santin, van Hout et al. 2021).

2.2 Fulfilment type versus manner/result verbs (and relevant concepts)

As the fulfilment type notion concerns the likelihood of intention fulfilment denoted by the verb, a relevant concept is manner/result verbs and manner/result complementarity in verb meaning (e.g., Rappaport Hovav and Levin 2010). Result verbs (e.g., clean) denote the result of an action but cannot denote manner information; manner verbs (e.g., wash) specify how an action is conducted but cannot denote what result is achieved. It is argued that a verb root can only denote either manner or result information, not both (e.g., Rappaport Hovav and Levin 2010). Major disagreements about manner/result complementarity come from Croft (2012) and Goldberg (2010), but their proposals disagree with manner/result complementarity to different extents.

For Croft (2012), ‘result’ can be split into ‘directed change’ (change in one particular dimension) and ‘undirected change’ (change in multiple unspecified dimensions). It is inaccurate to claim that manner verbs do not denote any change – for example, activities denoted by wash do entail some change in the affected object. The real distinction between manner/result verbs is that manner verbs do not denote ‘directed change’, while result verbs do. To illustrate, clean denotes the directed change from dirty to clean, while wash denotes undirected changes that might involve the relevant object becoming soaked and undergoing miscellaneous shape changes as it is agitated by the agent. Therefore, if result is taken to mean directed change, then manner/result complementarity holds; if result is taken to mean undirected change, manner/result complementarity would not hold. Thus, Croft (2012) is not a rebuttal per se to manner/result complementarity – he only placed a stricter definition on what counts as result.

Scholars who disagree with manner/result complementarity in the strict sense (directed change) presented counter-examples that encode both manner and directed change such as schuss, verbs of cooking (e.g., fry, stew; Goldberg 2010), and cutting and breaking verbs (e.g., slice; Chen 2007). Goldberg (2010) argued from the perspective of Frame Semantics that the only constraint on verb meaning is the Conventional Frame Constraint: “For a situation to be labeled by a verb, the situation or experience may be hypothetical or historical and need not be directly experienced, but it is necessary that the situation or experience evoke a cultural unit that is familiar and relevant to those who use the word” (Goldberg 2010: 50). Thus, when a conventional frame features both manner and directed change, it would be possible for both to be incorporated in a verb’s meaning.

Let us now relate Talmy’s fulfilment type to the frameworks above. Talmy (2000b: 261–278) did not define whether ‘intention fulfilment’ is directed or undirected change, but it more likely refers to directed change, because directed changes are arguably more likely than undirected changes to be conventionally received as intended goals. Further, intrinsic-/moot-/implied-fulfilment verbs would be semantically classified as manner verbs, and attained-fulfilment verbs as result verbs (Kou and Hohenstein 2024). Thus, the fulfilment type notion basically argues that some manner verbs point to intended results (i.e., moot-/implied-fulfilment verbs) and can even imply intention fulfilment (i.e., implied-fulfilment verbs).

Frame Semantics makes more accessible the argument that some manner verbs can imply intention fulfilment. In real life, manner and result are often associated in the mind. Human activities are primarily goal-oriented (Landau and Lakusta 2012; Zwaan and Radvansky 1998). Indeed, manner verbs often point to canonical goals (e.g., Pederson 2008; Rappaport Hovav and Levin 2010). Recall that Talmy (2000b: 262–268) argued that intrinsic-fulfilment verbs do not inherently denote specific intentions while moot-/implied-(/attained)-fulfilment verbs do. In light of Frame Semantics, this would suggest that intrinsic-fulfilment verbs do not evoke specific conventional frames linking the pertinent manners and results (e.g., kick is not associated with any specific goal), while moot-/implied-fulfilment verbs do (e.g., hunt is associated with the goal to arrest, wash is associated with the goal to clean).

While Frame Semantics provides the potential for manner verbs to be associated with result, why would some manner verbs imply no intention fulfilment? As intrinsic-fulfilment verbs are argued to independently denote no intention (Talmy 2000b: 262–264), the discussion here is mainly confined to moot- and implied-fulfilment verbs. Specifically, why would implied-fulfilment verbs imply intention fulfilment, while moot-fulfilment verbs would not? Two factors seem to contribute to this variation: (1) directness of causation between the activity and the result associated with the verb, and (2) the task difficulty of the event denoted by the verb.

First, the directness of causation (e.g., Levin and Rappaport Hovav 1994) between the causing and caused subevents in the event denoted by the verb can influence the outlook for intention fulfilment. This is reminiscent of the ‘idealised causative event’, which features direct manipulation of an entity by an agent (Lakoff and Johnson 1980: 69–71). Anticipating the verbs examined in the present study (Table 2), it is easy to see that moot-fulfilment verbs mostly denote mental efforts^[1] as they are verbs of searching (e.g., explore, investigate, hunt, seek) and verbs of obtaining (e.g., demand, request), following Levin’s (1993) verb classes. In contrast, implied-fulfilment (and intrinsic-fulfilment) verbs often denote activities involving physical manipulation. Thus, while implied-fulfilment verbs primarily involve direct causation in the events they denote, moot-fulfilment verbs often involve indirect causation. Indirect causation often features no physical contact (e.g., Ammon 1980) or transmission of physical force (Croft 2012: 198; Talmy 2000a: 409–470) between the agent and the object. The lack of physical contact and force transmission can lead to the impression that the agent has little control (e.g., Smith 1978) over how the event unfolds, thereby lowering the likelihood of intention fulfilment. Indirect causation can also involve a third party (apart from the agent and the object) who serves as an ‘intervening cause’ and is fully independent of the agent (Wolff 2003). Thus, it is not up to the agent whether intention can be fulfilled, which weakens the prospect for intention fulfilment. To illustrate, when demanding something, there is often someone other than the agent who determines whether the demanded objected can be obtained. Summarising, indirectness of causation might explain why moot-fulfilment verbs trigger no fulfilment preferences.

A further contributor to whether a verb implies intention fulfilment is task difficulty. Understandably, if a goal is more difficult to achieve, it is less likely to be achieved. The events described by different verbs naturally differ in task difficulty. For example, hunting activities are generally more difficult than washing activities. The event scenarios described with moot-fulfilment verbs tend to be relatively difficult, which might dim the outlook for intention fulfilment.

2.3 Fulfilment type versus telicity

Fulfilment types concern whether intention fulfilment (akin to change, more details below) happens in causative events, which is reminiscent of verbs’ lexical aspect, i.e., “a semantic category that concerns properties of eventualities … expressed by verbs … [such as] the presence [or absence] of some end, limit or boundary in the lexical structure of … verbs” (Filip 2012: 722). A well-known aspectual system is Vendler’s (1967) classification along three dimensions: stative/dynamic, durative/punctual, and telic/atelic. A crossing of these features yields four aspectual categories: states (stative/durative/atelic, e.g., know), activities (dynamic/durative/atelic, e.g., walk), accomplishments (dynamic/durative/telic, e.g., clean), and achievements (dynamic/punctual/telic, e.g., break). Arguably, telicity (e.g., Garey 1957) is the most relevant to fulfilment type. To enunciate the distinction between fulfilment type and telicity, we show below that intention fulfilment readings are not influenced by means that can alter (a)telicity.

It is well-known that the (a)telicity of an expression can be altered by the specificity of quantity in the direct object. Telic expressions often have direct objects that are specified for quantity (e.g., I ate two apples ; Verkuyl 1993); this specific quantity imposes an endpoint in time for the event. For example, in I ate two apples, when the two apples are fully consumed, eating stops in time. When the direct object is unspecified for quantity – I ate apples, eating could last forever when there is an endless number of apples available for consumption. If fulfilment type was the same concept as telicity, whether an event expression denotes intention fulfilment would similarly hinge on the quantity of the direct object. However, the shift from quantity-specific to quantity-unspecific direct objects would not alter the outlook for intention fulfilment (e.g., quantity-specific: I killed a person → certain that the person died; quantity-unspecific: I killed people → still certain that people died). Conversely, quantity-specific direct objects can make some atelic expressions telic. For example, while run is atelic, run two miles is telic, as two miles can ‘measure out’ (Tenny 1994) or delimit running by imposing an endpoint on it. This shift does not pertain to intention fulfilment; quantity-specific direct objects cannot turn an event with uncertain intention fulfilment into an event with certain intention fulfilment (e.g., quantity-unspecific: I investigated cases → intention fulfilment uncertain; quantity-specific: I investigated two cases → intention fulfilment still uncertain).

Such disconnection between intention fulfilment readings and the specificity of quantity of the direct object is because telicity and fulfilment type essentially concern different types of event endpoints. For telicity, an event is bounded by a point in time where an event can no longer proceed; for fulfilment type, an event boundary is a change in state. Thus, in the apple-eating scenario, when there is an indefinite number of apples that can be consumed, eating theoretically never stops. In contrast, for fulfilment type, no matter whether there is a definite or indefinite number of apples available for consumption, when apples are eaten, they are eaten, and change has occurred (to however many apples).

2.4 Section summary

A verb’s fulfilment type represents the likelihood of intention fulfilment (i.e., how likely is a resultative event completed/event result obtained/event boundary reached when the event ends). On this note, we introduced event outcomes as the focus of the present study. An event outcome can be a fulfilled intention or an unfulfilled intention. In event descriptions, verbs with different fulfilment types can trigger intention-fulfilled or intention-unfulfilled inferences for event outcomes. To illustrate, while I washed the shirt might encourage an intention-fulfilled reading (i.e., the shirt became clean), I hunted the fugitive might give a dimmer outlook for intention fulfilment (i.e., the fugitive gets arrested).

Although intention fulfilment is somewhat analogous to path in the well-studied Motion event domain, we do not find equivalent concepts to fulfilment type in motion verbs. This is because manner-of-motion verbs do not point to canonical paths (e.g., run does not denote any canonical path and can thus combine with various paths, such as run into/along/up/etc.; cf. Gerwien and von Stutterheim 2018) in the same way as many manner-of-action verbs point to canonical goals (e.g., wash inherently denotes the goal to clean). When one wishes to mention path while using manner-of-motion verbs, one obligatorily employs path satellites (e.g., run into ) or switches to a path verb (e.g., enter), which would lead to unambiguous path information. For instance, in I ran into / entered the room, it is certain that the agent ended up inside the room – there is no uncertainty regarding that path. Given this, we are not in a position to discuss likelihood of path in Motion events, as path is either non-existent (e.g., I ran) or 100 % certain (e.g., I ran into / entered the room). In contrast, manner-of-action verbs often independently denote canonical goals (except for intrinsic-fulfilment verbs) and can imply the obtainment of that goal (e.g., Kou and Hohenstein 2024; Lee and Kaiser 2021; Rappaport Hovav and Levin 1998; Talmy 2000b). This is what enables the discussion of the likelihood of intention fulfilment for manner-of-action verbs. To illustrate, in I washed the shirt, we can discuss how likely the shirt became clean; in I hunted the fugitive, we can discuss how likely the fugitive was arrested.

Such likelihood is in nature language-generated inferences. The present study concerns how such inferences might influence memory of event outcomes. To our knowledge, there is no study directly exploring this topic, so we review below research on three relevant areas, namely, the interplay between language, inferences, and memory.

3.1 The Deese-Roediger/McDermott paradigm and the pragmatic inference paradigm

Previous memory research has reported robust memory errors induced by pragmatic inferences, mainly through two experimental paradigms. The first is the Deese-Roediger/McDermott paradigm (Deese 1959; Roediger and McDermott 1995). In this paradigm, participants study lists of semantically associated words (e.g., bed, pillow, rest, awake, nap) and later recall whether a related but non-presented target word (e.g., sleep) is on the list. The target words are often falsely recalled or recognised by participants despite their absence during encoding (e.g., Deese 1959; Gallo 2010; Roediger and McDermott 1995). The presented words activate a certain domain in which the non-presented word is related to, so the non-presented word is inferred to be on the list, via relational processing (Maraver et al. 2021).

The second paradigm linking memory and inferences is the pragmatic inference paradigm (see Maraver et al. 2021 for a review). In this paradigm, participants study texts that can generate inferences (e.g., The karate champion hit the cinder block implies The karate champion broke the cinder block). Participants often misremember the inferred version of the studied material, rather than the original version (e.g., Barclay et al. 1984; Bartlett 1932; McDermott and Chan 2006).

These paradigms provide direct evidence that language-generated inferences can influence memory of previously encountered language input. While this line of research explicitly ascribes memory errors to inferences, there are other areas of memory research that do not explicitly discuss memory distortion in relation to inferences but nonetheless can be interpreted as related to inferences, as Section 3.2 illustrates.

3.2 Memory studies relevant to language-generated inferences

The following studies do not directly focus on inferences, but they are somewhat relevant as they study language-generated expectations that could be regarded as inferences. These studies can serve as indirect evidence that inferences affect memory.

First, language can influence memory of previously observed scenarios (e.g., Loftus 1975; Loftus et al. 1978). In Loftus (1975), participants viewed videos of road accidents and answered questions containing either misinformation about a non-existent barn (How fast was the white sports car going when it passed the barn while traveling along the country road?) or neutral information (How fast was the white sports car going while traveling along the country road?). Participants who received misinformation believed that they saw a barn in the video significantly more than those who received neutral information did. Further, spatial memory can be influenced by prepositions denoting spatial orientation (e.g., Feist and Gentner 2007). When ambiguous spatial images (e.g., a block balanced on the edge of a building) were described with spatial prepositions (e.g., The block is on the building), participants tended to remember the images as prototypical spatial relationships represented by the prepositions (e.g., block further away from the edge of the building). Moreover, language can influence memory of the temporal duration of events (Wang and Gennari 2019). Participants viewed videos featuring moving geometric shapes (e.g., a red rectangle moving upwards), each video labelled as either a ‘long event’ (e.g., A Chinese lantern rising up into the sky) or a ‘short event’ (e.g., A fire rocket launched into the sky). When participants recalled the length of the videos, videos described as long events were believed to last longer, compared to when the same videos were described as short events.

3.3 Language and memory of event endings

Having reviewed literature showing that memory can be influenced by inferences, we are in a better position to ask the specific question in the current study: to what extent do inferences influence memory of event outcomes? Since there seems to be no existing study on this issue, we review below previous investigations into memory of event endings.

Research into language and memory of event endings started in the motion domain with von Stutterheim et al. (2012). Languages differ as to the availability of grammatical means for expressing ongoing motion (e.g., imperfective or progressive aspect). For languages with grammatical means to express ongoingness (‘[+aspect] languages’, e.g., English, Arabic), it is not necessary to express path endings in ongoing motion events (e.g., A car is driving along a road (to a village)). For languages without grammatical means to encode ongoingness (‘[-aspect] languages’, e.g., German), path endings are much more often expressed (e.g., Ein Auto fährt zu einem Dorf ‘A car drives to a village’) (e.g., von Stutterheim et al. 2012). von Stutterheim et al. (2012) asked native speakers of seven [+aspect]/[−aspect] languages to watch videos depicting completed motion (e.g., a car reaches a village) and ongoing motion (e.g., a car drives towards a village in the distance). Participants either described or were prevented from describing the events during video display, and their eye gazes at the objects at path endings (e.g., a village house) were tracked. Later, participants viewed still images of the videos with the objects at path endpoints removed and recalled what the objects were. Speakers who typically mentioned path endpoints tended to remember endpoint objects better, compared to speakers who normally left out path endpoints in event descriptions, but significant difference was only detected when participants verbalised the events and between the most typical [-aspect] language (German) and the most representative [+aspect] language (Arabic). When participants were prohibited from verbalising the events, no language effects on memory were detected. In short, a linguistic highlight on path endpoints benefited memory of path endpoints. Although von Stutterheim et al. (2012) detected a language effect on memory of path endpoints, the study might not be very relevant to the present investigation. We examine memory of event outcomes, namely whether some intention has been fulfilled; in the context of von Stutterheim et al. (2012), this information would translate into whether the agents actually reached the path endpoints. Thus, we target a different aspect of events.

More pertinent to the present study are investigations into memory of event endings in the domain of state change. Sakarias and Flecken (2019) explored the impact of the Estonian case marking for full/partial state change on memory of event outcomes, contrasting it with a language without such differentiation (Dutch). In Estonian, objects take obligatory grammatical markers to reflect whether a relevant state change happens partially (e.g., kartuli-t, ‘potato.partitive’, a potato half peeled) or in full (e.g., kartuli, ‘potato.accusative’, a potato fully peeled). Native speakers of Estonian and Dutch viewed videos of causative events with visually salient state changes (e.g., peeling a banana) and without (e.g., polishing glasses), among which half reached the final state (‘ceased’) and the other half did not (‘ongoing’). Half of the participants described the videos shortly after viewing, while the other half were prevented from verbalising them via a distractor task. A subsequent recognition task examined participants’ memory of whether state change was fully or partially achieved. In the verbal condition, Estonian speakers outperformed Dutch speakers, but the non-verbal condition elicited the opposite pattern. Importantly, memory of Dutch speakers did not vary in verbal and non-verbal conditions, suggesting that the memory differences detected above were driven by the distinctions between verbal and non-verbal conditions for Estonian speakers. This shows that the Estonian full/partial case marking enhances memory accuracy when language is involved and weakens memory performance when language is absent, showcasing language effects on memory.

Two similar studies are Santin et al. (2020) and Santin, van Hout et al. (2021). Focussing on native speakers of Dutch, Chinese, and Spanish, the authors explored how verb semantics would influence memory of event endings. The materials and experimental paradigm in these two studies closely followed Sakarias and Flecken (2019), but Santin and colleagues further explored whether or not participants’ event descriptions entailed state change. Participants’ memory of event endings was overall enhanced by verbalisation, but verb semantics did not significantly influence memory.

As shown, research often reports that verbalising events affects event representation (e.g. Gennari et al. 2002; Sakarias and Flecken 2019). Verbalising specific event components often leads to memory enhancement for those elements. When verbalising is prohibited, language effects on memory are less frequently detected. However, as we explicitly present event expressions in all trials in the memory experiment (Section 4.2.4), the verbal/non-verbal contrast is not pertinent here. This is because our study is not designed to investigate how verbs’ fulfilment types would influence encoding of event outcomes. Rather, we focus on how verbs’ fulfilment types would trigger inferences about event outcomes, which would in turn influence previously established memory of event outcomes.

The rationale for our memory study might look clearer if we draw an analogy in motion event language and memory research in the Talmyan tradition. If this study was conducted in the motion domain, we would present event descriptions lexicalising manner and path and ask how the relevant manner information later presented impacts memory of path. An imaginary stimulus would read: I walked and then I entered the room [change-occurred outcome] / remained outside the room [no-change outcome]. At recall, participants would be cued with manner information I walked and recall whether the agent entered the room or not. This design does not make sense in the motion domain, because manner-of-motion verbs do not inherently point to canonical paths, so manner-of-motion verbs alone are not theoretically expected to generate inferences about path and modulate memory of previously encountered paths. However, this design would make sense in the current context of fulfilment type, as manner-of-action verbs can point to canonical intentions and imply intention fulfilment. Therefore, when one reads about manner in the absence of event outcome information, one can infer whether there is intention fulfilment based on the manner verb.

3.4 Section summary

The aforementioned studies indicate that memory can be distorted when language introduces inferences about the memorised content (e.g., how long an activity should last, what spatial relation should be described by a preposition). Indeed, recall is argued to be a constructive process, whereby beliefs and world knowledge can be integrated with established memory (e.g., Barclay et al. 1984; Bartlett 1932).

At this point, let us revisit Santin et al. (2020) and Santin, van Hout et al. (2021). To some extent, these studies also approached event language from the perspective of inferred event outcomes. The authors asked participants to produce descriptions for events and judged whether those descriptions entailed change in the affected objects. Why did they not detect a language effect on memory of event endings, though? This is possibly because those event descriptions were not in a position to trigger inferences for the participants who produced the event descriptions and later did the recall task, because those participants clearly observed whether change took place in the video stimuli and thus did not have to make inferences about event outcomes. Therefore, although verb semantics can trigger inferences about whether change happened, it was not in a position to do so in Santin et al. (2020) and Santin, van Hout et al. (2021). Had participants encountered those event descriptions at recall in the absence of event ending information, their recall might have been modulated by language.

4.1 Experiment 1

Experiment 1 explores whether events described with verbs with different fulfilment types could generate inferences about event outcomes. If so, there is reason to expect such inferences to affect memory of (text-based) event outcomes (to be investigated in Experiment 2). Notably, the semantic contrast between attained-fulfilment verbs and intrinsic-/moot-/implied-fulfilment verbs might encourage the impression that the former suggest intention fulfilment, while the latter suggest non-fulfilment. This is not necessarily the case, as the lack of semantic entailment of intention fulfilment does not necessarily imply non-fulfilment of intention. Thus, it is necessary to empirically explore how events described with intrinsic-/moot-/implied-fulfilment verbs are perceived in terms of event outcomes in language use.

There are studies investigating verbs from the perspective of certainty of event outcomes (e.g., Chen 2018; Esposito 2021; Lee and Kaiser 2021; Wright 2014). The research method is fairly straightforward – by asking native speakers to read event descriptions and judge the acceptability or likelihood of different outcomes. Following this approach, the present experiment is a likelihood estimation task investigating the default event outcome readings for different event descriptions so as to gauge the inferences about event outcomes.

4.2 Experiment 2

With a general understanding of how verbs with different fulfilment types generate inferences about intention fulfilment, we examined in Experiment 2 whether such inferences would impact text-based memory of previously encountered event outcomes. Depending on intention-fulfilled/intention-unfulfilled event outcomes, two predictions were formulated:

Prediction 1: For events with intention-fulfilled outcomes, memory performance would be better if post-encoding event descriptions encourage fulfilment preferences (i.e., contain intrinsic-/implied-/attained-fulfilment verbs, compared to moot-fulfilment verbs).

Prediction 2: For events with intention-unfulfilled outcomes, memory performance would be better if post-encoding event descriptions do not encourage fulfilment preferences (i.e., contain moot-fulfilment verbs, compared to intrinsic-/implied-/attained-fulfilment verbs).