Advertisement

The Division of Labor Between Structure Building and Feature Checking During Sentence Comprehension

  • Markus BaderEmail author
Chapter
Part of the Studies in Theoretical Psycholinguistics book series (SITP, volume 48)

Abstract

The two-stage architecture of the Garden-Path Theory with its separation of first- and second-pass parsing has been replaced by simpler architectures in certain probability-based models of the human parser, including the Surprisal Theory. Based on evidence from subject-object ambiguities in German, this paper argues that the two-stage architecture still provides a better account of the garden-path strength observed for object-before-subject sentences in German. In the first part of the paper, corpus findings concerning the relationship between animacy and word order are discussed. Although animacy information is an important predictor of word-order in German, the Surprisal Theory does not predict differences in garden-path strength related to this information because animacy constrains word order only in combination with the verb’s argument structure. Because garden-path strength in verb-final clauses, as they are found in German, is a function of the verb’s expectedness according to the Surprisal Theory, verb specific information itself cannot affect garden-path strength in this theory. In the second part of the paper, a specific implementation of a two-stage model of garden-path recovery, the Linking-and-Checking model, is discussed. This model accounts for the dependence of garden-path strength in object-before-subject sentences on animacy as well as for findings concerning the use of subject-verb agreement for garden-path recovery.

1 Introduction

According to the classical two-stage architecture of the human parsing mechanism pioneered by Lyn Frazier in collaboration with various colleagues, an obligatory initial stage of analysis is followed by an optional second stage of reanalysis. The purpose of the initial analysis stage—the stage of first-pass parsing—is to assign the incoming word string a phrase structure representation that captures all relevant syntactic properties of the string. The phrase structure tree computed during first-pass parsing is the input to the subsequent process of semantic interpretation. Analysis is the parser’s main business—and in most cases, it is all that the parser has to do. Due to the existence of local ambiguities, however, analysis is not guaranteed to succeed. When the parser selects an analysis for a locally ambiguous string that is contradicted by later input, the second parsing stage—the stage of second-pass parsing—is called for. The purpose of this stage is to reanalyze the syntactic structure computed on first pass parsing in order to arrive at a correct analysis. This stage is accordingly also known as the reanalysis stage. Depending on how difficult it is to revise the initial syntactic structure, a more or less severe garden-path effect results.

The two-stage architecture of the Human Sentence Processing Mechanism (HSPM) was introduced with the Garden-Path Theory proposed in Frazier and Rayner (1982), following earlier work on the Sausage Machine (Frazier & Fodor, 1978). The Garden-Path Theory was developed based on experimental results for English, but was soon applied to other languages as well. Frazier (1987) extended the theory to Dutch. This in turn inspired research on German, which is syntactically quite similar to Dutch. Sentences which are locally or globally ambiguous between a subject-before-object (SO) analysis and an object-before-subject (OS) analysis have attracted most attention in these languages. Frazier (1987) found a preference for the SO structure in an experiment investigating Dutch sentences, and for German a similar preference has been found for a large range of sentence types exhibiting this kind of ambiguity (see references below).

An authentic German sentence illustrating the ambiguity between subject and object is shown in (1).

Because proper names are not case-marked in German, this is a globally ambiguous sentence in which either the first NP is the subject and the second NP the object, or the other way round. Only by taking world knowledge into account can we know that this sentence must be understood as an OS sentence. As this is not compatible with the initially preferred SO structure, the sentence leads to some sense of surprise on first reading.

The two-stage architecture of the human parser is not without alternatives, but in the following I will argue that this architecture still provides a viable model of the human parser. To do so, I will draw on research on how the parser assigns syntactic functions when processing ambiguous German sentences. Section 2 introduces the relevant data and shows why they do not follow from a recent alternative to the two-stage architecture, the Surprisal Theory of Levy (2008). A two-stage model developed to account for these data, the Linking-and-Checking Model proposed by Bader and Bayer (2006), is presented in Sect. 3. The paper closes with a short summary in Sect. 4.

2 Reanalysis and Garden-Path Strength

When a sentence containing a local syntactic ambiguity is disambiguated toward a non-preferred structural alternative, a garden-path effect arises. It is for this situation that a second stage of parsing—the reanalysis stage—has been included within the Garden-Path Theory. The main question with regard to the reanalysis stage is why in cases of local syntactic ambiguity disambiguation toward the non-preferred analysis sometimes causes severe processing disruptions whereas in other cases reanalysis proceeds without much ado (see Fodor & Ferreira, 1998, for an overview of findings and theories).

In German, case syncretism is a major source of syntactic ambiguity. Only masculine singular NPs distinguish morphologically between nominative and accusative case, whereas feminine and neuter singular NPs and plural NPs do not. Furthermore, NPs without determiners—including proper names and bare NPs—are often three-way ambiguous between nominative, accusative and dative case. Subject-object ambiguities are the most prominent type of syntactic ambiguity caused by case syncretism. A globally ambiguous example of this kind of ambiguity was provided in (1). A locally ambiguous example is given in (2).

Open image in new window

The embedded clauses in (2a) and (2b) are locally ambiguous up to the clause-final auxiliary. Since the two NPs Maria and einige Kinder (‘some children’) are morphologically ambiguous between nominative and accusative case, the embedded clauses in (2) are compatible with a subject-before-object (SO) and an object-before-subject (OS) analysis until the clause-final auxiliary disambiguates the sentences via the obligatory agreement between subject and verb. In (2a), the final auxiliary is marked for singular, and therefore, the first NP must be the subject. Sentence (2b), in contrast, ends with a plural auxiliary, making the second NP the subject.

Inspired by Lyn Frazier’s seminal work on subject-object ambiguities in Dutch (Frazier, 1987), subject-object ambiguities in German have been investigated intensively (for comprehensive overviews, see Bader & Bayer, 2006; Bornkessel & Schlesewsky, 2006). A major finding of these investigations has been that in the overwhelming majority of cases the HSPM prefers to assign a SO structure if there is a choice between a SO and an OS analysis. In addition, when a sentence contains an NP which is ambiguous between accusative and dative object, the NP will preferentially be analyzed as an accusative object. These findings can be summarized by the Case Assignment Generalization in (3).
  1. (3)

    The Case Assignment Generalization

    When encountering a case-ambiguous NP, the parser assigns the highest case on the case hierarchy nominative ≫ accusative≫ dative that is morphologically and syntactically compatible with the NP.

     

It is a controversial question how the generalization in (3) derives from properties of the HSPM. Bader and Bayer (2006) have shown how the Case Assignment Generalization can be derived from two of the parsing principles of the Garden-Path Theory—Minimal Attachment (Frazier & Rayner, 1982) and the Minimal Chain Principle (De Vincenzi, 1991), a generalization of the Active Filler Strategy (Frazier, 1987). The main alternative to deriving the Case Assignment Generalization from general parsing principles is to derive it from production frequencies, for example by assuming a parser working with a probabilistic phrase-structure grammar (Hale, 2014). Structure-driven and frequency-driven models make the same predictions in most cases and it is therefore difficult to decide between them (see Häussler & Bader, 2011, for an attempt to disentangle the two). Since the topic of this section is the reanalysis stage of the HSPM, it is not necessary for present purposes to take a stance on this issue.

Since the SO structure is preferred for a locally ambiguous sentence on first-pass parsing, disambiguation toward the OS structure usually involves processes of second-pass parsing or reanalysis in order to replace the preferred SO structure by the non-preferred OS structure. The strength of the garden-path effect caused by locally ambiguous OS sentences varies substantially, from barely detectable garden-path effects to garden-path effects which lead to a downright rejection of a sentence as ungrammatical. In the present context, the most important finding concerning garden-path strength in subject-object ambiguities concerns the syntactic means by which an object ends up in front of the subject. The grammar provides two ways to license OS order. First, a sentence that normally would appear with SO word-order can nevertheless have OS order for discourse reasons, for example if the object is old information and the subject new information. Second, a sentence can appear with OS order for lexical-conceptual reasons, either to confirm to the more general constraint according to which animate NPs preferentially precede inanimate NPs, or because it contains a verb from a set of verbs licensing OS order, including object-experiencer psych-verbs, unaccusative verbs, and passivized ditransitive verbs. Note that animacy and verb-specific properties often go hand in hand: The object of an object-experiencer verb is obligatorily animate, and the subject of such a verb is often inanimate (e.g., einfallen ‘something occurs to someone’). There are competing syntactic accounts concerning the syntactic means necessary to capture the derivation of sentences with OS order. Since nothing hinges on the particular assumptions, I follow much work in generative grammar and assume that for discourse-licensed OS sentences, the object is first base-generated behind the subject and then moved in front of it (movement derived OS structure, “Scrambling”), whereas for OS sentences licensed by lexical-conceptual reasons, the object is directly base-generated in front of the subject (base-generated OS structure) (cf. Haider, 2010).

Whether a particular sentence has a movement-derived or a base-generated OS structure can be captured by the argument structure associated with the main verb of the sentence.1 For most verbs having one or more objects, the argument structure specifies that the subject is base-generated in front of the object(s). Discourse reasons of the sort mentioned above, however, may allow the object to move from its base-generated position to a position in front of the subject. A subject-object ambiguity illustrating this point has already been given in (2). Experimental investigations of movement-derived ambiguities have shown that such ambiguities often lead to rather strong garden-path effects. For example, in the experiment reported in Bader and Meng (1999) sentences as in (2b) were judged as ungrammatical 45% of the time.

A subject-object ambiguity for which both the SO sentence and the OS sentence receive a base-generated syntactic structure is shown in (4).

Open image in new window

The embedded clause in sentence (4a) is base-generated because it represents the canonical SO order of German. In sentence (4b), the embedded clause is base-generated despite having OS order because it contains a ditransitive verb in the passive voice. The subject in this clause is the direct object in the corresponding active clause. In English, the underlying object would have to move to the subject position in a passive clause. In German, in contrast, the object can stay in situ in a passive clause. Because accusative objects follow dative objects in active clauses, the subject follows the dative object after passivization.

Sentences with base-generated OS structure have received a great deal of attention in the literature on German subject-object ambiguities (see Bader, 1996; Scheepers, 1996 and much subsequent work). A major conclusion arising from these investigations is that locally ambiguous base-generated OS sentences typically lead to only minor garden-path effects, in contrast to locally ambiguous movement-derived OS sentences like (2b) which elicit strong garden-path effects.

In accordance with what was said above about the licensing of OS order, the base-generated OS sentence in (4b) has an animate object and an inanimate subject. In the corresponding movement-derived OS sentence (2b), in contrast, the subject and the object NP are both animate NPs. Could this difference be responsible for the observed difference with regard to garden-path strength?

An obvious possibility would be that reanalysis is easy for (4b) because the ultimately correct assignment of subject and object follows from the semantics of the verb in conjunction with the animacy features of the two NPs. Because the recipient of schicken (‘send’) must be animate, it follows immediately that the first NP Maria must be the dative object and the second NP ein Buch the subject. In (2b), in contrast, both NPs are animate and the meaning of the verb einladen (‘invite’) is therefore of no help in deciding who did what to whom. On first sight, this may explain the difference with regard to garden-path strength. Things are not so easy, however. Although the two NPs in (2b) do not differ with regard to animacy, they differ with regard to number. The first NP in sentence (2b) is a singular NP and the second NP is a plural NP. The obligatory agreement between the subject and the clause-final verb therefore provides an unambiguous cue as to the correct assignment of subject and object role. But oddly, this information does not seem to be of great help to the HSPM in garden-path recovery, in contrast to the semantic information in form of the animacy constraints in the base-generated sentence (4b).

Perhaps animacy information is helpful during reanalysis because—as has been found by several corpus studies of the order of subject and object in German—the choice between SO and OS order is to a large degree a matter of animacy (Bader & Häussler, 2010; Hoberg, 1981).

Figure 1 shows how the order of subject and object varies with animacy and case in the corpus study of Bader and Häussler (2010). This corpus study analyzed several sentence sets randomly drawn from the newspaper corpus of the Institute for the German Language in Mannheim/Germany. The data shown in Fig. 1 are the results for the sentence set containing 732 complementizer-introduced embedded clauses with either SO or OS word order. Because this corpus was not syntactically analyzed, sentences were found by searching for NPs introduced by the article den (the.ACC/DAT). The sentences found in this way were annotated for various features by hand, including information concerning animacy.
Fig. 1

Order of subject and object depending on animacy of subject and object and case of object. The graphic was prepared based on the corpus data in Bader and Häussler (2010)

For sentences with animate subjects, the vast majority of sentences occur with SO order. The animacy of the object also has an effect in sentences with an animate subject, but only with regard to the case of the object and not with regard to order. Accusative objects are always more frequent than dative objects, but for animate objects the ratio of accusative to dative case is much less extreme. For sentences with an inanimate subject, word order is crucially determined by the object’s animacy. When the subject is animate and the object inanimate, sentences with OS order prevail, whereas we see again a strong predominance of sentences with SO order when both subject and object are inanimate. The relationship between animacy and order of subject and object visible in Fig. 1 can be captured by the generalization that SO order is chosen unless the object is higher on the animacy hierarchy than the subject, which amounts to a preference for OS order when the subject is inanimate and the object animate and a preference for SO order in all other cases. This generalization accounted for about 90% of all sentences in the study by Bader and Häussler (2010).

Several recent theories of syntactic ambiguity resolution make heavy use of information derived from frequency distributions as they are found in corpora of authentic texts. Do the differences in garden-path strength that are observed for movement-derived and base-generated OS sentences follow when frequency-based theories of syntactic ambiguity resolution are applied to the corpus results discussed above? For reasons of space, the Surprisal Theory of Levy (2008), which builds on earlier work by Hale (2001), is the only frequency-based theory that can be considered here. The basic intuition behind the Surprisal Theory is that a word w is easy to process when it is expected but difficult when it is unexpected, that is, when encountering w comes as a surprise. More precisely, integration difficulty is hypothesized to be proportional to the negative log probability of w in the current context, as captured by the equation in (5) (Levy, 2008:1130).
  1. (5)

    difficulty \( \propto - \log \,{\text{P}}(w_{i} |w_{1} \ldots w_{i - 1} \), CONTEXT)

     

The notion of context includes sentential context—i.e., the word string preceding w—and extrasentential context. The expectations concern the syntactic category of w, possibly annotated with additional syntactic features like case or number and subcategorization features of verbs.

Can the difference in garden-path strength observed between movement-derived OS sentences and base-generated OS sentences be explained in terms of surprisal? If so, the parser should be surprised to see a verb requiring an OS structure after having seen a sequence of two animate NPs, as in (6a), whereas it should not be surprised—or at least much less so—by seeing an OS verb in the context of an animate NP followed by an inanimate NP, as in (6b).

Open image in new window

After having processed a sentential context containing a sequence of two animate NPs, it is indeed highly surprising to encounter a verb that imposes an OS structure. After all, as shown by Fig. 1, OS sentences containing two animate NPs are quite rare. It is also surprising, however, to see an OS verb after a sequence of an animate NP followed by an inanimate NP, as in (6b). Although it is true that a typical OS sentence has an animate object followed by an inanimate subject, a sentence in which an animate NP precedes an inanimate NP can also have an SO structure, as was already shown by (4a). In the corpus study of Bader and Häussler (2010), sentences with an animate subject and an inanimate object were about 2.5 times more frequent than sentences with an inanimate subject and an animate object (see Fig. 1). In addition, the preference for SO order in the former case is much stronger than the preference for OS order in the latter case. Thus, after having seen an animate NP followed by an inanimate NP, there is a strong expectation to see a verb requiring SO order. In other words, a verb requiring OS order comes as a surprise in the context of an animate NP followed by an inanimate NP. According to the Surprisal Theory, base-generated sentences should therefore cause garden-path effects of substantial strength, but in fact they cause only mild garden-path effects.

Even stronger evidence against a surprisal account of the weak garden-path effect caused by base-generated OS sentences comes from experiments that have manipulated the number of competitors to the ultimately correct OS structure. Such a manipulation is made possible by the existence of sentences that are only ambiguous with regard to the case of the object. An example of such an ambiguity is provided in (7). Sentence (7a) differs from the base-generated OS sentence (4b) only by containing the definite NP den Studenten as object instead of the proper name Maria.

Open image in new window

Crucially, the NP den Studenten is ambiguous between accusative and dative case, but it is not compatible with nominative case. As shown by the glosses in (7), the case ambiguity of this NP is accompanied by a number ambiguity. Dative case goes hand in hand with plural number and accusative case with singular number. Due to the case ambiguity of den Studenten, the two sentences in (7) are locally ambiguous between an OS structure with a dative object, as in (7a), and an OS structure with an accusative object, as in (7b). As can be seen in Fig. 1, OS sentences with a dative object are much more frequent than OS sentences with an accusative object. When taking structural frequencies alone into account, it is therefore not surprising to see a dative OS verb after having processed the two NPs den Studenten and ein Buch, whereas seeing an accusative OS verb is surprising. These structural frequencies have to be qualified by the lexical frequencies for the article den and the following case and number-ambiguous noun. As shown in Bader and Häussler (2018), this reduces the difference between the surprisal values of dative and accusative verbs somewhat, but even then a dative verb is not surprising whereas an accusative verb is. Note furthermore that even a dative verb is more surprising in a locally ambiguous context like (7a) than in a context that is completely unambiguous, that is, in a context in which the object NP is unambiguously marked for dative case.

To sum up, how surprising it is to see an OS dative verb after having processed an animate NP followed by an inanimate NP depends strongly on whether the competitor set contains an alternative structure with SO order or not [(4b) vs. (7a)]. The experimental results of Bader and Häussler (2018) show, however, that sentences as in (4b) and sentences as in (7a) cause only mild garden-path effects, with minimal differences between the two, contrary to what would be expected under the Surprisal Theory. Thus, despite the correlation between animacy and word-order that has been found in corpus studies of German, the Surprisal Theory does not provide a solution to the puzzle of why movement-derived OS sentences cause strong garden-path effects whereas only weak garden-path effects result from base-generated OS sentences.

For language production, there is overwhelming evidence from corpus studies that the decision to produce a sentence with SO or OS order is based on animacy information. Why then does the Surprisal Theory not make the correct predictions for language comprehension, despite taking frequency information into account? The reason is that animacy information can help the parser only in combination with the verb’s lexical entry because the verb provides the semantic information that constrains the linking of syntactic functions to argumental NPs, in particular by constraints concerning animacy. During language production, both the verb and the arguments are available, and the observed corpus frequencies are the joint product of both. According to the Surprisal Theory, this verb information is not taken into account for making predictions when comprehending a verb-final clause, because the verb is predicted given the arguments and not the other way round. The Surprisal Theory thereby fails to account for the relationship between different types of OS sentences and garden-path strength.

In a two-stage architecture of the HSPM, reanalysis can operate on the arguments together with the information provided by the verb, and thus animacy constraints provided by the verb can become effective. The verb also provides the number information that would be helpful in reanalyzing a movement-derived OS sentence like the one in (2b), but the HSPM does not seem to make use of this information. Animacy features and number features thus seem to play different roles for reanalysis. This difference is accounted for in the Linking and Checking Model (LCM) that is introduced in the next section.

3 From Analysis to Reanalysis: The Linking-and-Checking Model

The Linking and Checking Model (LCM) was proposed by Bader and Bayer (2006) in order to explain the contrast in garden-path strength between movement-derived and base-generated OS sentences as well as a variety of other findings on subject-object ambiguities. The main claim of the LCM, which presupposes the classical two-stage architecture of the HSPM, is that reanalysis makes use of the normal linking and feature checking operations of the HSPM. With the diagnosis model of Fodor and Inoue (1994), the LCM shares the premise that reanalysis is a diagnostic process, and that the ease of garden-path recovery is mainly a function of the ease by which diagnosis proceeds. For the case of subject-object ambiguities, the LCM makes the further claim that diagnosis is intimately connected to the argument-linking and feature-checking operations that the HSPM has to perform for every sentence, whether ambiguous or not. In particular, the LCM hypothesizes that certain locally ambiguous object-subject sentences are automatically reanalyzed as a by-product of the normal linking and checking processes of the HSPM whereas others need a more elaborate type of reanalysis.

The LCM assumes the overall architecture of the HSPM shown in Fig. 2. Figure 2 rests on the assumption that the HSPM contains processes of structure assembly followed by processes of linking and feature handling which continuously monitor the ongoing syntactic representation assembled at the prior stage (for more discussion of the distinction between assembly and checking, cf. Mitchell, 1994). The main responsibility of the linking stage is to associate phrases within the ongoing phrase-structure representation—the current partial phrase marker (CPPM)—with thematic roles as specified by the verb of the sentence. The stage called feature handling checks whether the distribution of case and agreement features is correct. In addition, the LCM assumes that certain temporary feature violations can be repaired automatically at this stage, as detailed below. If linking and feature handling do not result in a well-formed structure, additional processes of diagnosis and repair are needed in order to correct the current structure. While additional processes of diagnosis and repair are only invoked for garden-path sentences, it is not true that all garden-path sentences require both processes.2 Some garden-path sentences are already diagnosed and repaired as a by-product of the linking and feature-handling processes that are part of the processing of every sentence, ambiguous or not.
Fig. 2

A model of the HSPM integrating linking and feature handling

The internal working of the linking and feature handling stage of the HSPM is assumed to be governed by the Linking-Based Checking Algorithm (LBCA) which is given in (8).
  1. (8)

    The Linking-Based Checking Algorithm (LBCA)

    1. (A)

      Argument Linking 3

      Link each DP within the CPPM to a position within the verb’s argument structure in the order specified in the argument structure.

       
    1. (B)
      Feature Handling
      1. a.

        Feature Checking

        Check the relevant features (Case for subject and object, number and person for subject)

         
      2. b.

        Feature Repair

        For each resulting feature mismatch, where a feature mismatch has the form “Feature value α assigned to XP instead of feature value β”, determine if the lexical material of XP would be compatible with the assignment of β.

        If so, replace α with β and—if necessary—adjust the phrase-marker accordingly.

         
       
     

The working of the Linking-Based Checking Algorithm (LBCA) is best explained by example. Consider first the base-generated OS sentence in (4b). During first-pass parsing, the initial NP Maria will be assigned nominative case and the second NP ein Buch accusative case, in accordance with the Case Assignment Generalization in (3). The argument structure of the clause-final verb geschickt wurde contains a dative marked object followed by the subject. According to the Argument Linking Step of the LBCA, the argument structure will be projected onto the CPPM with the arguments respecting their order within the argument structure. That is, the first NP within the CPPM will be associated with the first slot within the argument structure of geschickt wurde and the second NP with the second slot (cf. (9)).

Open image in new window

Given the linking pattern in (9), the feature checking part of the LBCA will detect two case violations. First, the NP Maria bears nominative case in the CPPM instead of dative case, as required by the verb; second, the NP ein Buch bears accusative case instead of nominative case. The feature-repair step of the LBCA is therefore invoked. At this step, the HSPM determines whether the feature mismatches found at the step before can be cured by local feature corrections. For sentence (4b), this is indeed possible. Due to morphological ambiguity, dative case can be assigned to Maria and nominative case to ein Buch. The only task remaining for the HSPM is to actually replace the offending case features by the correct ones. In sum, for a base-generated sentence as in (4b), the normal linking and checking procedures of the HSPM automatically deliver the information which the HSPM needs in order to repair the sentence. Processes that are specifically dedicated to reanalysis are not necessary for this purpose.

Consider next the movement-derived OS sentence (2b). As before, the HSPM will compute an SO structure for this sentence during first-pass parsing. Since the verb eingeladen haben has an argument structure with the subject preceding the object, the first NP of the CPPM will be linked to the subject role of the verb and the second NP to the object role (cf. (10)).

Open image in new window

As can be seen in (10), the resulting linking pattern does not lead to a case mismatch. However, it leads to a violation of the number agreement between subject and verb. Since at this point of processing the first NP has been analyzed as the subject, subject and verb do not agree in number: Maria is a singular NP but the verb is specified for plural. The next step to take is the feature-repair step of the LBCA. In contrast to the base-generated sentence considered before, a local feature correction is not possible for sentence (2b). Maria is unambiguously a singular NP, and the verb unambiguously a plural verb. Thus, linking and checking alone do not lead to an automatic reanalysis of sentence (2b). After linking and checking are completed, the CPPM still contains a mismatch, and additional diagnostic processes are necessary in order to determine that sentence (2b) is a sentence with an OS structure.

In summary, the LCM claims that locally ambiguous base-generated OS sentences are automatically reanalyzed as a by-product of the normal linking and checking operations of the HSPM. For locally ambiguous movement-derived OS sentences, in contrast, the HSPM’s linking and checking procedures do not automatically indicate how a successful reanalysis should proceed. For such sentences, more elaborate diagnostic processes are necessary.

These diagnostic processes can make use of a much broader range of information sources than the automatic processes of linking and checking discussed above. For example, it was pointed out above that the information provided by subject-verb agreement does not prevent strong garden-path effects in the case of movement-derived OS sentences [see example (2b)]. This does not mean however that subject-verb agreement is never helpful at all. Consider the two movement-derived OS sentences in (11). In both sentences, the initial NP Maria is three-way case ambiguous (nominative/accusative/dative) whereas the second NP is only two-way case ambiguous (nominative/accusative). Because the clause-final verb geholfen (‘helped’) assigns dative case to its object, these sentences must be analyzed as OS sentences.

Open image in new window

To conclude that (11a) is an OS sentence, the HSPM cannot rely on any direct hint because both NPs are case-ambiguous. For (11b), in contrast, the HSPM can also rely on the information provided by subject-verb agreement. Bader and Bayer (2003) investigated sentences as in (11a) and (11b) and found that both lead to relatively strong garden-path effects, as expected given that these are movement-derived OS sentences. Crucially, however, reanalysis was easier for (11b) than for (11a). For base-generated OS sentences, a similar effect of type of disambiguation was not found, providing further evidence for a distinction between reanalysis as an automatic by-product of linking and checking and reanalysis as an elaborate diagnostic process.

4 Conclusion

Lyn Frazier’s early work on subject-object ambiguities in Dutch spawned a still ongoing line of research investigating how the HSPM assigns syntactic functions. A large part of this research has been concerned with subject-object ambiguities as they are found in languages with a relatively free word order. This research has revealed a robust SO preference for sentences that are locally or globally ambiguous between a SO or an OS structure. Even if not without exceptions (e.g., Mak, Vonk, & Schriefers, 2008), this preference implies that disambiguation toward an OS structure leads to garden-path effects under many circumstances. These garden-path effects vary widely in strength, from effects that are detectable only with experimental methods to effects that are so strong that speakers attribute ungrammaticality to sentences that are in fact only locally ambiguous.

In this paper, I have argued that the variation in garden-path strength observed for subject-object ambiguities is best understood in terms of the two-stage architecture of the HSPM, which was introduced with the Garden-Path Theory of Frazier and Rayner (1982). For reasons of space, only one competing account of garden-path strength could be considered, namely the Surprisal Theory of Levy (2008). Section 2 discussed data showing that the Surprisal Theory does not provide the means necessary to capture the relationship between animacy, word order, and garden-path strength that has been observed in corpus studies and in comprehension experiments. Section 2 therefore introduced the Linking-and-Checking Model proposed by Bader and Bayer (2006). By implementing the processes of argument linking and feature checking within a two-stage architecture of the HSPM, the Linking-and-Checking Model successfully accounts for a large array of findings concerning garden-path strength caused by locally ambiguous OS sentences.

As a final note, let me point out that the above considerations do not imply that production frequencies do not play any role within the human parsing mechanism. As discussed in Sect. 2, the problem with the Surprisal Theory is not that it takes probabilities derived from frequency distributions into account in order to predict garden-path strength. The problem is a more specific one, concerning the particular kind of probabilistic information that is assumed to reflect garden-path strength. Because surprisal is the inverse of expectation, the difficulty of integrating a disambiguating word follows from the word’s expectedness in the preceding context. In the case of verb-final clauses disambiguated by the verb, this amounts to the probability that a verb of a particular kind is expected. Properties of the verb itself are not taken into account for this purpose. In the discussion above, the relevant verb properties concerned constraints on the verb’s arguments, as captured in its argument structure. Although the above discussion was framed in terms of categorical constraints, in particular with regard to animacy, the relevant constraints could also be of a probabilistic nature. That is, when trying to recover from a garden-path, the parser may well be guided in a probabilistic way by the frequency of different argument structures. Exploring this possibility, which could result in a probabilistic formulation of the Linking-and-Checking Model, must be left as a task for future research.

Footnotes

  1. 1.

    Actually, the order among arguments is a function of several factors, including verb semantics, animacy and definiteness (cf. Ellsiepen & Bader, 2018, for recent experimental evidence). Since the question of how the order of arguments is determined is orthogonal to the purposes of this paper, I stick to the simplifying assumption that the argument structure associated with the verb directly determines the order of arguments.

  2. 2.

    More generally speaking, these processes are always invoked when the HSPM detects an ungrammaticality—whether this ungrammaticality is a non-permanent one (garden-path sentences) or a permanent one (ungrammatical sentences). See Meng and Bader (2000) and Hopf et al. (2003) for the relation between garden-path sentences and ungrammatical sentences.

  3. 3.

    Given the simplifying assumption concerning word-order and argument structure, the Argument Linking Step of the LBCA is simplified accordingly. A non-simplified version can be found in Bader and Bayer (2006).

References

  1. Bader, M. (1996). Sprachverstehen: Syntax und Prosodie beim Lesen. Opladen: Westdeutscher Verlag.CrossRefGoogle Scholar
  2. Bader, M., & Bayer, J. (2003). Diagnosis and garden-path recovery—case and agreement symptoms revisited. Presented at the 9th annual conference on architecture and mechanisms for language processing, University of Glasgow, Glasgow, Scotland.Google Scholar
  3. Bader, M., & Bayer, J. (2006). Case and linking in language comprehension: Evidence from German. Dordrecht: Springer.Google Scholar
  4. Bader, M., & Häussler, J. (2010). Word order in German: A corpus study. Lingua, 120(3), 717–762.CrossRefGoogle Scholar
  5. Bader, M., & Häussler, J. (2018). Assigning syntactic functions: On the proper balance between frequency and structure. Manuscript, Goethe University Frankfurt/University Leipzig.Google Scholar
  6. Bader, M., & Meng, M. (1999). Subject-object ambiguities in German embedded clauses: an across-the-board comparison. Journal of Psycholinguistic Research, 28, 121–143.CrossRefGoogle Scholar
  7. Bornkessel, I., & Schlesewsky, M. (2006). The extended argument dependency model: A neurocognitive approach to sentence comprehension across languages. Psychological Review, 113, 787–821.CrossRefGoogle Scholar
  8. De Vincenzi, M. (1991). Syntactic parsing strategies in Italian. Dordrecht: Kluwer.CrossRefGoogle Scholar
  9. Ellsiepen, E., & Bader, M. (2018). Constraints on argument linearization in German. Glossa: A Journal of General Linguistics, 3(1), 1–36.  https://doi.org/10.5334/gjgl.258.CrossRefGoogle Scholar
  10. Fodor, J. D., & Ferreira, F. (Eds.). (1998). Reanalysis in sentence processing. Dordrecht: Kluwer.Google Scholar
  11. Fodor, J. D., & Inoue, A. (1994). The diagnosis and cure of garden paths. Journal of Psycholinguistic Research, 23, 407–434.CrossRefGoogle Scholar
  12. Frazier, L. (1987). Syntactic processing: Evidence from Dutch. Natural Language & Linguistic Theory, 5, 519–559.CrossRefGoogle Scholar
  13. Frazier, L., & Fodor, J. D. (1978). The sausage machine: A new two-stage parsing model. Cognition, 6, 291–325.CrossRefGoogle Scholar
  14. Frazier, L., & Rayner, K. (1982). Making and correcting errors during sentence comprehension: Eye movements in the analysis of structurally ambiguous sentences. Cognitive Psychology, 14, 178–210.CrossRefGoogle Scholar
  15. Haider, H. (2010). The syntax of German. Cambridge: Cambridge University Press.CrossRefGoogle Scholar
  16. Hale, J. (2001). A probabilistic Earley parser as a psycholinguistic model. Proceedings of the North American Chapter of the Association for Computational Linguistics, 159–166.Google Scholar
  17. Hale, J. T. (2014). Automaton theories of human sentence comprehension. Stanford, Calif.: CSLI Publications.Google Scholar
  18. Häussler, J., & Bader, M. (2011). Grammar- versus frequency-driven syntactic ambiguity resolution: the case of double-object constructions. In M. Lamers & P. de Swart (Eds.), Case, word order and prominence: Interacting cues in language production and comprehension (pp. 273–302). Dordrecht: Springer.Google Scholar
  19. Hoberg, U. (1981). Die Wortstellung in der geschriebenen deutschen Gegenwartssprache. München: Hueber.Google Scholar
  20. Hopf, J.-M., Bader, M., Meng, M., & Bayer, J. (2003). Is human sentence parsing serial or parallel? Evidence from event-related brain potentials. Cognitive Brain Research, 15, 165–177.CrossRefGoogle Scholar
  21. Levy, R. (2008). Expectation-based syntactic comprehension. Cognition, 106, 1126–1177.CrossRefGoogle Scholar
  22. Mak, W. M., Vonk, W., & Schriefers, H. (2008). Discourse structure and relative clause processing. Memory & Cognition, 36(1), 170–181.CrossRefGoogle Scholar
  23. Meng, M., & Bader, M. (2000). Ungrammaticality detection and garden-path strength: Evidence for serial parsing. Language and Cognitive Processes, 15, 615–666.CrossRefGoogle Scholar
  24. Mitchell, D. C. (1994). Sentence parsing. In M. A. Gernsbacher (Ed.), Handbook of psycholinguistics (pp. 375–409). San Diego etc.: Academic Press.Google Scholar
  25. Scheepers, C. (1996). Menschliche Satzverarbeitung: Syntaktische und thematische Aspekte der Wortstellung im Deutschen (Doctoral dissertation). Universität Freiburg.Google Scholar

Copyright information

© Springer Nature Switzerland AG 2019

Authors and Affiliations

  1. 1.Institute for LinguisticsGoethe University FrankfurtFrankfurtGermany

Personalised recommendations