Choosing anaphoric expressions: Do people take into account likelihood of reference?
Introduction
When people refer to a previously mentioned discourse entity, they can use various types of anaphoric expressions such as pronouns, proper names, or definite descriptions. An important question for models of language production is how people choose among different expressions. A general assumption is that the choice of anaphor depends on the referent’s accessibility or how activated it is in the discourse representation. When the referent is highly accessible in the discourse, reduced anaphoric expressions such as pronouns tend to be used, whereas when it is less accessible, more explicit expressions such as proper names and definite descriptions are used (Ariel, 1990, Chafe, 1976, Chafe, 1994, Givón, 1983, Grosz et al., 1995, Gundel et al., 1993).
Several factors have been identified to affect accessibility and hence the choice of anaphoric expressions. For example, a referent is more accessible when it has been mentioned more recently or more frequently in the preceding discourse, so more reduced anaphoric expressions such as pronouns tend to be used under such conditions (Ariel, 1990, Givón, 1983). Other research has shown that people use more pronouns when the referent is the grammatical subject in sentence-initial position rather than a later-mentioned object in the immediately preceding clause (Arnold, 2001, Brennan, 1995, Fletcher, 1984, Stevenson et al., 1994), consistent with theoretical accounts that claim that the referent’s accessibility is affected by structural properties of the preceding sentence such as the antecedent’s grammatical role (e.g., Brennan et al., 1987, Frederiksen, 1981, Gordon et al., 1993, Grosz et al., 1995) or surface sentence position (Gernsbacher and Hargreaves, 1988, Gordon et al., 1993).
In this article, we investigate whether the likelihood of referring to an entity, which has been argued to influence the activation of discourse entities, also constrains the form of anaphoric expressions. Many studies have shown that verb semantics influences which entity people are most likely to refer to next (Au, 1986, Brown and Fish, 1983, Caramazza et al., 1977, Garvey and Caramazza, 1974, Garvey et al., 1975, Stevenson et al., 1994). When completing sentence fragments such as (1a), people tend to start their completion by referring to John as the subject of the subsequent clause (e.g., because he/John was very clever), whereas when completing sentence fragments such as (1b), they tend to refer to Mary (e.g., because she/Mary was very clever).
- 1a
John impressed Mary because…
- 1b
John admired Mary because…
Such completion preferences have been assumed to occur because verbs such as impress have a semantic bias that attributes causality to the first mentioned noun phrase (NP1), whereas verbs such as admire have a bias that assigns causality to the second mentioned noun phrase (NP2) (often referred to as implicit causality biases). Some researchers have argued that this is because certain semantic roles are more likely to be seen as the cause of the event denoted by the verb (Au, 1986, Brown and Fish, 1983, Crinean and Garnham, 2006, Stevenson et al., 1994). For instance, impress is a stimulus–experiencer verb because NP1 has the semantic role of stimulus and NP2 the role of experiencer, whereas admire is an experiencer–stimulus verb because NP1 is the experiencer and NP2 the stimulus. When these verbs are combined with the causal connective because, people tend to associate causality with the stimulus rather than the experiencer. Other researchers have argued that completion preferences cannot always be predicted from the verb’s semantic roles and that the effects are due to more subtle properties of the event that is described by the sentence (Garvey and Caramazza, 1974, Garvey et al., 1975, Pickering and Majid, 2007). For example, modal verbs (e.g., may) (Grober, Beardsley, & Carmazza, 1978) and verb tense (Rohde, Kehler, & Elman, 2006) have been shown to affect completion preferences. In the following, we will refer to completion preferences determined by the meaning of the verb (including its modality and tense) and the connective as semantic biases.
Of interest is whether the choice of anaphor is affected by such semantic biases. Many researchers have assumed that semantic biases modulate accessibility in the discourse, so that entities congruent with the semantic bias are more accessible than those that are not. That is, following (1a), NP1 (John) is more accessible than NP2, whereas following (1b), NP2 (Mary) is assumed to be more accessible (e.g., Garvey and Caramazza, 1974, Greene and McKoon, 1995, Long and De Ley, 2000, McDonald and MacWhinney, 1995, Stevenson et al., 1994). Consistent with this, probe recognition studies have shown that names denoting bias-consistent entities are recognised more quickly than those denoting bias-inconsistent entities (e.g., Long and De Ley, 2000, McKoon et al., 1993). Similarly, people preferentially interpret ambiguous pronouns as co-referent with entities that are consistent with the semantic bias (Kehler et al., 2008, Stevenson et al., 1994), and unambiguous pronouns are comprehended faster when they refer to the bias-consistent entity (e.g., he in (1a) and she in (1b)) than the bias-inconsistent entity (e.g., Koornneef and Van Berkum, 2006, Van Berkum et al., 2007, Vonk, 1985).
Given the general assumption that choice of anaphor is affected by referents’ accessibility in the discourse, it seems plausible that anaphoric forms for bias-consistent entities are more reduced than for bias-inconsistent entities. Such an account has been proposed by Arnold, 2001, Arnold, 2008 who argued that an entity’s level of activation is affected by how likely it is to be referred to in the subsequent discourse. Arnold (2001) argued that: “in cases in which speakers are more likely to refer to entities that have played certain thematic roles, both speakers and comprehenders should find pronouns more natural than fuller forms of reference to refer to these entities” (p. 158).
Arnold’s frequency-based account is in line with functional linguists who have equated predictability with givenness (Kuno, 1972, Kuno, 1978, Prince, 1981) and accessibility (Givón, 1988, Givón, 1989). For example, Givón (1989) argued that predictable information is “more readily available (…) or (…) more vividly activated in the memory, thus more strongly attended to”, so “It does not – unlike new unpredictable information – require strong activation by massive coding” (p. 218). Thus, according to Givón, reduced referring expressions such as pronouns should occur in more predictable contexts than more explicit expressions. Importantly, he assumes that predictability is determined by various types of contextual information, including the semantics of the discourse.
Similarly, some probabilistic models of human language processing also assume a link between the predictability of a word and its form reduction (Bell et al., 2003, Gregory et al., 1999, Jurafsky et al., 2001). These models claim that the more predictable or probable a word is because of its neighbouring words, syntactic and lexical structure, semantic expectations and discourse factors, the more phonologically reduced the word is. Indeed, words tend to be shorter in duration when their neighbouring words make them statistically predictable than otherwise (Bell et al., 2003, Gregory et al., 1999, Jurafsky et al., 2001). Importantly, several researchers, including Arnold (2008) and Givón, 1988, Givón, 1989, assume that predictability affects not only acoustic reduction but also reduction of lexical form (e.g., a pronoun rather than name). For example, Fowler, Levy, and Brown (1997) argued that referring expressions for highly predictable entities are not only acoustically reduced but also lexically reduced, so more reduced referring expressions such as pronouns should be used for more predictable entities. Consistent with this, factors that have been argued to influence predictability and acoustic reduction, such as prior mention and frequency of mention (Bard and Aylett, 1999, Fowler and Housum, 1987, Fowler et al., 1997), have also been shown to influence anaphoric form (Ariel, 1990, Givón, 1983, Levy and McNeill, 1992).
According to frequency-based accounts, semantic biases should affect the choice of anaphor because what people refer to and how they refer are determined by the same constraints. Evidence that supports this claim comes from Arnold (2001). She examined the choice of anaphoric expressions by contrasting references to the goal vs. the source role in the preceding sentence. Participants had to complete auditorily presented stories such as (2). In (2a), Marguerite (the subject) is the goal, whereas Eduardo (the oblique object) is the source. In contrast, in (2b), the positions of the semantic roles are reversed: Brendan (the oblique object) is the goal and Lisa (the subject) the source.
- 2a
I hate getting sick. It always seems like everyone gets sick as soon as it’s vacation. Marguerite caught a cold from Eduardo two days before Christmas. …
- 2b
There was so much food for Thanksgiving, we didn’t even eat half of it. Everyone got to take some food home. Lisa gave the leftover pie to Brendan. …
Participants more often started their completions by referring to the goal (Marguerite in 2a, Brendan in 2b) than the source (Eduardo in 2a, Lisa in 2b). Importantly, when participants referred to the goal, the proportion of pronouns relative to repeated names was larger than when they referred to the source, suggesting that language users take the likelihood of reference into account when choosing an anaphor. In addition, participants used more pronouns than repeated names when referring to the subject (Marguerite in 2a, Lisa in 2b), whereas they used more repeated names than pronouns when referring to the oblique object (Eduardo in 2a, Brendan in 2b), consistent with other studies that have shown that the antecedent’s grammatical role or sentence position affects the choice of anaphor (Crawley and Stevenson, 1990, Fletcher, 1984, Stevenson, 2002, Stevenson et al., 1994).
However, Arnold’s (2001) results may not be fully conclusive. First, the semantic bias effect on the choice of anaphor was stronger with the oblique object than the subject. This is somewhat surprising, because the reported means suggested that the goal bias in the completions was smaller with the oblique object than the subject. If the semantic effect on the choice of anaphor is due to the likelihood of reference, the effect should have been strongest with the grammatical role that exhibited the strongest goal bias. Second, as shown in (2), the source–goal and goal–source conditions differed not only in the order of the semantic roles, but also in other respects. For example, the contexts preceding the source–goal and goal–source sentences were different between conditions, the entity that was being transferred differed (e.g., an abstract entity in 2a, a concrete entity in 2b), and the final phrase was different (two days before Christmas in 2a, the goal noun in 2b). Furthermore, the argument status of the prepositional phrase is difficult to control with these verbs: In (2b) to Brendan is an obligatory argument, whereas in (2a) from Eduardo is not. Hence, the differences between the conditions may not have been due to the goal–source manipulation.
Indeed, some researchers have suggested that the factors that determine what people are likely to refer to are different from the information they use for choosing how they refer. Stevenson et al. (1994) suggested that semantic biases determine what language users refer to, but they do not affect how they refer. Instead, how they refer is affected by the structural position of the antecedent. Kehler et al. (2008) used a Bayesian account to formalise the distinction between what and how people refer. They claimed that semantic biases, or what they call coherence-driven expectations, determine the probability of reference to a particular entity, whereas structural factors such as grammatical role affect the probability that people produce a pronoun (rather than a repeated name) given reference to a particular entity.
We suggest that the proposed dissociation between the likelihood of referring to a particular entity and how it is subsequently referred to makes sense if we assume that likelihood of reference does not affect an entity’s accessibility. Whereas the structure of the preceding sentence may determine an entity’s level of activation and hence the choice of anaphor, semantic biases may not. One possible reason for this is that people choose particular sentence structures depending on how accessible the entities are in their discourse representation (e.g., Bock and Irwin, 1980, Ward and Birner, 2004), whereas they are unlikely to choose the semantic bias of a sentence depending on the entities’ accessibility. For instance, language users could choose John impressed Mary or Mary was impressed by John depending on which discourse entity is most accessible, but if they would like to convey that John impressed Mary, it is unlikely that they produce Mary impressed John just to create a semantic bias towards Mary. Because the entities’ accessibility affects the way a sentence is structurally organised (its information packaging), the structure of the preceding sentence signals the level of activation of the entities that are mentioned in it. In contrast, the semantic bias of the preceding sentence does not, because it is unaffected by the accessibility of the entities. Hence, when people choose an anaphor, they may rely on the antecedent’s structural position, but not on semantic biases.
The decision to refer to a particular entity is presumably determined by the content of the message people would like to convey, rather than by the accessibility of the entities. Semantic biases arise because the semantics of the preceding clause influences the inferences people make, which in turn affects who or what they refer to, rather than because the preceding clause makes the bias-consistent entity more accessible than the bias-inconsistent entity. For example, following John impressed Mary because, people may infer that John is the cause of the event, regardless of whether he is the most accessible entity in the discourse. Similarly, the reason why semantic biases affect the comprehension of pronouns and probe recognition may not be because semantic biases influence accessibility, but because people use the semantics of the previous clause to predict what the next clause is about. As suggested by Koornneef and Van Berkum (2006), such predictions may not be constrained by accessibility in the discourse. For example, following John impressed Mary because, people may predict that the next clause refers to John, because they infer that he is responsible for the event, rather than because he is the most accessible discourse entity. Because semantic biases influence the language user’s predictions about what the next clause refers to (Koornneef & Van Berkum, 2006), ambiguous pronouns are preferentially interpreted as consistent with the bias or the prediction (Kehler et al., 2008, Stevenson et al., 1994), and processing a pronoun is easier when its gender disambiguates it towards the antecedent that is consistent with the prediction than inconsistent with it (e.g., Garnham et al., 1992, Koornneef and Van Berkum, 2006). Similarly, probe recognition times are faster for names that are consistent with the prediction than those that are inconsistent with it (Long and De Ley, 2000, McKoon et al., 1993). Thus, anaphor resolution processes may be affected by semantic biases because they affect language users’ predictions about what is going to be mentioned next, but such predictions may arise from a different mechanism than the mechanism that affects how accessible entities are in the discourse.
Stevenson et al.’s (1994) arguments for a dissociation between likelihood of reference and choice of anaphor were based on sentence completion experiments investigating how the meanings of verbs and connectives influence the likelihood of reference. They used pairs of verbs that assigned two semantic roles (goal–source/source–goal, agent–patient/patient–agent, stimulus–experiencer/experiencer–stimulus verbs), such that each thematic role was in the first-mentioned subject position (NP1) in one condition and in the second-mentioned object position (NP2) in the other. The verbs were followed by a full stop, and, so or because. When action verbs (goal–source/source–goal and agent–patient/patient–agent) were followed by a full stop, and or so, participants tended to refer to the semantic role that was most strongly associated with the endpoint of the event (the goal or the patient). However, when because followed the action verbs, such biases towards the endpoint were weakened, indicating that in addition to the verb meaning, the connective also contributes to semantic biases. The effect of connective was even clearer with state verbs. Following because, state verbs involving stimulus and experiencer roles had a strong stimulus bias whereas following so, there was a bias towards the experiencer.
Stevenson et al. also analysed the form of the anaphoric expressions participants used. Overall, collapsed across all verbs and connectives, the antecedent’s position or grammatical role affected the subsequent choice of expression, with more pronouns (relative to names) for NP1 than NP2 (see also Stevenson, 2002). They concluded that the choice of how to refer is affected by the structural position of the antecedent, but not by semantic biases. However, because they collapsed across verbs and connectives, it is unclear whether the effects of the antecedent’s structural position were modulated by the verb and connective bias, that is, whether semantic biases also had an effect. They did not report the means for the different anaphors broken down by verb and connective, nor did they statistically analyse the effects of semantic biases on the choice of anaphor.
More recently, Miltsakaki (2007) conducted sentence completion experiments in Greek to investigate the choice between null and overt pronouns. In one experiment, participants had to produce continuations following clauses with agent–patient verbs, where the agent was always the subject and the patient the object, by producing either a subordinate continuation (by adding connectives like because, since, when to the first clause) or a coordinated continuation (usually by adding and). Overall, participants referred somewhat more often to the patient than the agent. When producing subordinate clause continuations, participants almost always used null subject pronouns to refer to both the agent (subject) and the patient (object). When they produced coordinate continuations, they also nearly always used null pronouns for the agent, but favoured overt pronouns for the patient. Miltsakaki concluded that the semantic bias towards the patient is not sufficient to warrant reference with a null pronoun and that choice of anaphor in Greek is sensitive to syntactic factors. But this conclusion may be premature. First, when participants produced coordinated continuations, they did refer more often to the agent than the patient, so the results are in fact compatible with the view that the choice of anaphor is affected by likelihood of reference. Second, because the antecedent’s grammatical and semantic roles were confounded, it is unclear whether syntactic or semantic factors affected anaphoric choice. Finally, no significance levels were reported, so it is unclear whether the effects were statistically robust.
Finally, Rohde and Kehler (2008) tested source–goal sentences in English. Although participants referred more often to the goal (the oblique object) than the source (the subject), they used more pronouns (relative to names) to refer to the source than the goal. Rohde and Kehler concluded that the probability of referring to a particular entity and the probability of using a pronoun (rather than name) given reference to a particular entity are affected by different factors. However, because they did not include goal–source sentences, the source was always the first-mentioned subject, and the goal the second-mentioned object. It is possible that the effects of semantic bias were overridden by a stronger structural effect, with more pronouns for the first-mentioned subject than the second-mentioned object.
Therefore, to resolve the question of whether semantic biases (or likelihood of reference) affect the choice of anaphor, we conducted two experiments. In Experiment 1, we manipulated semantic biases by contrasting stimulus–experiencer (SE) and experiencer–stimulus (ES) verbs followed by the causal connective because. In Experiment 2, we used SE verbs in conjunction with either the connective because or so. As in Stevenson et al., 1994, Miltsakaki, 2007, Rohde and Kehler, 2008, we used a written sentence completion method. Written completions have been used to examine various semantic bias effects (e.g., Garvey and Caramazza, 1974, Stevenson et al., 2000, Stevenson et al., 1994) as well as other aspects of language production such as number agreement (Fayol, Largy, & Lemaire, 1994), structural priming (Pickering & Branigan, 1998) and message planning (Simner & Pickering, 2005).
For each experiment, we examined the choice of anaphor in two separate tests, using different completion methods. In pretests, we used a free completion task, as in previous studies on anaphor production (Stevenson et al., 1994, Miltsakaki, 2007), where participants completed a sentence fragment in any way they liked, and we examined whether participants produced a pronoun or repeated name. But because their completions rarely violated the bias, the analyses of choice of anaphor were based on a small number of observations for bias-inconsistent references. To maximise our chance of finding a semantic biases effect, therefore, the main experiments employed a constrained completion task, where participants had to refer to a specified entity (but they were free in how they referred to it), and we used the most strongly biased materials from the pretest.
Section snippets
Experiment 1
Experiment 1 contrasted stimulus–experiencer (SE) and experiencer–stimulus (ES) verbs. These verbs were chosen because the argument status of the semantic roles is better controlled (the stimulus and experiencer are both obligatory arguments) compared to the goal–source/source–goal verbs used by Arnold (2001) and because they are commonly used in studies investigating semantic bias effects (Au, 1986, Brown and Fish, 1983, Stevenson et al., 1994). We employed items such as (3). In (3a), scared
Main experiment
The pretest showed that following SE verbs, participants referred to NP1 (the subject) more often than NP2 (the object), whereas following ES verbs the pattern was reversed, indicating that participants referred to the stimulus more often than the experiencer in both verb conditions. However, the analyses of choice of anaphor suggested that participants used no more pronouns to refer to the stimulus than the experiencer. In order to increase the chance of finding a semantic bias effect on the
Experiment 2
Verb meaning is not the only factor that affects semantic biases. Apart from verb meaning, the semantics of connectives has been shown to influence what people refer to (Au, 1986, Grober et al., 1978, Stevenson et al., 1994, Stevenson et al., 2000). Stevenson et al. (1994) found that semantic biases for state verb sentences with stimulus and experiencer roles are due to interactions between the verb meaning and particular connectives: When these verbs were followed by because, the stimulus role
Main experiment
The pretest showed a strong connective effect on completion preferences: Following because, participants referred to NP1 (stimulus) more often than NP2 (experiencer), whereas the opposite was the case following so. If the choice of anaphor is affected by the same factors that affect completion preferences, then for NP1 references, participants should use more pronouns (relative to names) following because than so, whereas for NP2 references, they should use more pronouns following so than
General discussion
Experiment 1 investigated whether semantic biases due to the meaning of the verb affected choice of anaphor, while Experiment 2 investigated whether biases due to the connective had an effect. Neither experiment showed any evidence that semantic biases, or likelihood of reference, affected the choice of anaphor. In contrast, structural properties of the antecedent sentence did have an effect: Participants produced more pronouns (relative to names) when referring to the first-mentioned subject
Acknowledgment
We thank Maria-Nella Carminati and Amit Dubey for their comments on earlier versions of this paper.
References (75)
A verb is worth a thousand words: The causes and consequences of interpersonal events implicit in language
Journal of Memory and Language
(1986)- et al.
The psychological causality implicit in language
Cognition
(1983) Markedness and topic continuity in discourse processing
Journal of Verbal Learning and Verbal Behavior
(1984)- et al.
Talkers’ signalling of “new” and “old” words in speech and listeners’ perception and use of the distinction
Journal of Memory and Language
(1987) - et al.
Reductions of spoken words in certain discourse contexts
Journal of Memory and Language
(1997) - et al.
Factors influencing assignment of pronoun antecedents
Cognition
(1975) - et al.
Accessing sentence participants: The advantage of first mention
Journal of Memory and Language
(1988) - et al.
Building and accessing clausal representations – The advantage of 1st mention versus the advantage of clause recency
Journal of Memory and Language
(1989) - et al.
Pronouns, passives, and discourse coherence
Journal of Memory and Language
(1995) - et al.
Pronouns, names, and the centering of attention in discourse
Cognitive Science
(1993)
Parallel function strategy in pronoun assignment
Cognition
On the use of verb-based implicit causality in sentence comprehension: Evidence from self-paced reading and eye tracking
Journal of Memory and Language
Implicit causality and discourse focus: The interaction of text and reader characteristics in pronoun resolution
Journal of Memory and Language
The time course of anaphor resolution: Effects of implicit causality and gender
Journal of Memory and Language
The representation of verbs: Evidence from syntactic priming in language production
Journal of Memory and Language
Discourse constraints on syntactic processing in language production: A cross-linguistic study in English and Spanish
Journal of Memory and Language
Planning causes and consequences in discourse
Journal of Memory and Language
The time course of the influence of implicit causality information: Focusing versus integration accounts
Journal of Memory and Language
Establishing reference in language comprehension: An electrophysiological perspective
Brain Research
Accessing noun–phrase antecedents
The effect of thematic roles on pronoun use and frequency of reference continuation
Discourse Processes
Reference production: Production-internal and addressee-oriented processes
Language and Cognitive Processes
The dissociation of deaccenting, givenness and syntactic role in spontaneous speech
Effects of disfluencies, predictability, and utterance position on word form variation in English conversation
Journal of the Acoustical Society of America
Syntactic effects of information availability in sentence production
Journal of Verbal Learning and Verbal Behavior
Centering attention in discourse
Language and Cognitive Processes
Comprehension of anaphoric pronouns
Journal of Verbal Learning and Verbal Behavior
Givenness, contrastiveness, definiteness, subjects, topics, and point of view
Discourse, consciousness, and time
Comprehension and the given-new contract
Reference in single sentences and in texts
Journal of Psycholinguistic Research
The use of heuristic strategies in the interpretation of pronouns
Journal of Psycholinguistic Research
Implicit causality, implicit consequentiality and semantic roles
Language and Cognitive Processes
Subordinating conjunctions as devices for unifying sentences in memory
European Journal of Cognitive Psychology
Cognitive overload and orthographic errors – When cognitive overload enhances subject verb agreement errors – A study in French written language
Quarterly Journal of Experimental Psychology Section a-Human Experimental Psychology
Given-new ordering effects on the production of scrambled sentences in Japanese
Journal of Psycholinguistic Research
Cited by (85)
A systematic evaluation of factors affecting referring expression choice in passage completion tasks
2023, Journal of Memory and LanguageAvoiding gender ambiguous pronouns in French
2022, CognitionSemantic predictability of implicit causality can affect referential form choice
2021, CognitionCitation Excerpt :One reason that implicit causality has attracted so much attention is that people have strong expectations that the implicit cause will be mentioned in the next sentence. When previous studies found that implicit causality had no effect on pronoun production, scholars concluded that pronoun production was not influenced by next-mention biases (Fukumura & van Gompel, 2010; Kehler et al., 2008; Rohde & Kehler, 2014; Stevenson et al., 1994). Our findings potentially overturn this conclusion, suggesting that implicit causality may matter after all, and by extension, predictability as well.
The influence of the temporal characteristics of events on adults' and children's pronoun resolution
2022, Journal of Child Language