Garden-path Model Andlexicalist Constraints-based Model of Sentence Processing

What are the main claims of the Garden-Path Model and the Lexicalist Constraints-Based Model of sentence processing? How are these claims supported by empirical evidence?

Introduction

Both the Garden-path and Lexicalist Constraints Models have clear empirical evidence to support their claims, however the basis in which they are founded on differs. The Garden-Path model relies on the idea of a syntactic structure, shaping sentence processing, reverting to methods such as minimal attachment and late closure to guide us past temporary ambiguity. With the Lexicalist-Constraint models, all syntactic and non-syntactic cues are given equal weight, with aspects such as lexical frequency and context playing as large a role as syntax.

The Garden-Path Model

The garden-path model of sentence processing is an example of modular-based learning – that different modules/areas in the brain are responsible for certain processes (Fodor, 1983). This idea of processing is commonly thought to compete with an interactive model (discussed later on).

Within the garden-path model:

“The sentence processor initially employs only information about the syntactic structure of the sentence to adopt a single analysis in (temporarily) ambiguous sentences.” (Rueschemeyer & Gaskell, 2007)

In other words, syntactic cues such as case, word order and agreement are utilized to achieve comprehension. Sentences are broken down/parsed into separate entities and often understood through correction. Other referents such as semantic and contextual cues are said to have influence in the second stage of the model, when ambiguity often occurs. The garden-path model supports the idea of sentence processing being incremental, deciphering words as they appear, one by one. Speakers ensure as soon as new information or units of language appear, processing immediately begins to integrate such information into the current sentence. “Garden-path sentences are sentences which contain a temporary ambiguity” (Poesio, 2004) and when this ambiguity occurs in sentence processing, the model relies on two principles: minimal attachment and late closure. Both principles account for all languages and are therefore universal (Rueschemeyer & Gaskell, 2007).

Minimal Attachment

Minimal attachment suggests that when processing sentences with ambiguity, incoming material should be attached into the phrase marker using the fewest nodes possible.

Example: (from Frazier & Rayner, 1982)

A) The city council argued the mayor’s position forcefully.

B) The city council argued the mayor’s position was incorrect.

When looking at these two sentences, sentence A would be treated in terms of minimal attachment.

Syntactic tree diagrams illustrated in

(Carroll, 2007):

Within this example, ambiguity is seen with the noun phrase the mayor’s position. If we were to use the concept of minimal attachment, this phrase would be placed within the constituent containing the verb argued and therefore reduce the number of nodes in the overall sentence (sentence A). However with sentence B, the noun phrase would form its own constituent and thus increase the number of nodes in the overall sentence. If a reader were to make the mayor’s position the head of its own constituent in sentence A, the sentence would be misinterpreted, with the adverbial forcefully being some sort of modifier to the noun phrase the mayor’s position. The principle of minimal attachment therefore ensures temporarily ambiguous (garden-path) sentences such as the example shown above can be correctly processed.

Minimal attachment does not account for all temporarily ambiguous sentences. For instance “if two analyses of an ambiguous structure have an equal number of tree structure nodes, the late closure principle applies.” (Rueschemeyer & Gaskell, 2007).

Late Closure

The idea of late closure states that

“If grammatically permissible, attach new items into the clause or phrase currently being processed (i.e. the phrase or cause postulated most recently).” (Frazier, 1987)

Example: (from Carroll, 2007)

Jessie put the book Kathy was reading in the library…

Within this example, the prepositional phrase in the library would make grammatical and semantic sense when appearing immediately after put or reading. However according to the concept of late closure, the sentence processor would rather it modified the verb reading.

Example: (from Rueschemeyer & Gaskell, 2007)

The steak with the sauce that was tasty didn’t win a prize.

Within this example, the relative clause that was tasty would rather attach to the sauce, since it is the most recent item (temporarily ambiguous item) being processed in sentence comprehension.

The Active Filler Strategy

Closely related to the garden-path model is the Active Filler Strategy (AFS). This strategy “accounts for the way in which people process unbounded dependencies, as found in relative clauses” (Rueschemeyer & Gaskell, 2007) and wh- questions. Take these two sentences:

Example: (from Rueschemeyer & Gaskell, 2007)

A) The housekeeper from Germany urged his guests to consider X

B) Who did the housekeeper from Germany urge the guests to consider?

When looking at the transformation from sentence A to B, X has been moved from modifying the verb consider and replaced with the pronoun who. However the AFS suggests that when processing sentence B, it is assumed who has replaced the direct object from sentence A – his guests. This is of course incorrect. The processors desire to fill the “gap” as early as possible has led to the interpretation – The housekeeper from Germany urged who? From this reanalysis would need to occur to ensure a correct interpretation of this sentence is made – who were the guests urged to consider?

The AFS bares relation with the garden-path model of sentence processing, more specifically with minimal attachment.  “Both strategies attempt to find some grammatical analysis of a sentence as soon as possible” (Weiner, 2003), demonstrating the garden-paths common problem of premature comprehension in temporarily ambiguous sentences.

Reanalysis

Within the garden-path model, when the initial use of syntactic cues fails to dissolve any ambiguity, other semantic cues are used for reanalysis.

“It is usually assumed that reanalysis occurs when the initially adopted analysis is inconsistent with later syntactic information, i.e. syntactic information makes the initial analysis ungrammatical.”  (Rueschemeyer & Gaskell, 2007)

Furthermore if, after employing both minimal attachment and late closure to an ambiguous sentence, other methods were used for analysis, reading and comprehension time would be increased.

Example: (from Ferreira & Clifton, 1986)

A) The defendant examined by the lawyer turned out to be unreliable.

B) The evidence examined by the lawyer turned out to be unreliable.

When following rules of minimal attachment, the verb examined would naturally attach to the subject the defendant/the evidence. This is of course incorrect and the verb examined is intended to be the main verb of the sentence. Ferreira and Clifton state how with a competent knowledge of semantics, reanalysis occurs to correct this initial comprehension. Moreover:

“If the parser still initially constructs a Minimal Attachment analysis, even in the face of the semantic information, then the most the semantic information can do is to hasten reanalysis, resulting perhaps in an earlier and less long-lasting disruption. “ (Ferreira & Clifton, 1986)

Recognising that the evidence cannot be called an agent of the verb examined would lead to the correction of any previous minimal attachment, demonstrating the garden-path’s core principle of syntactic cues first, followed by semantic and contextual cues in the second stage of sentence processing.

Evidence for sematic cues following syntactic cues can be seen when looking at ERP (event related potential) studies.  (Van Gompel, 2013) refers to the work of (Friederici, 2002; Friederici & Weissenborn, 2007) in terms of “a model of adult processing, based on ERP findings”. When looking at initial syntactic analysis (first stage of the garden-path model), ELAN (Early Left Anterior Negativity) was said to have occurred at “150-300 ms” after exposure to the sentence. When looking at semantic and other so called second stage cues, the N400 (responsible for semantic processing) was said to have occurred “300-500 ms” after exposure to the sentence. Finally, P600 readings (responsible for reanalysis) were recorded at “500-1000 ms” after exposure to the sentence. Although rather small, this gradual increase in ERP results demonstrates garden-path claims of Syntax first, semantics second.

Evidence for the Garden-Path model

In terms of empirical evidence, late closure has been shown to have effect, largely as a result of Frazier and Rayner’s 1982 study:

Making and correcting errors during sentence comprehension: Eye movements in the analysis of structurally ambiguous sentences.

The basis of this study involved video recordings of participants’ eye movement when reading temporarily ambiguous sentences. Results subsequently found that in sentences adhering to garden-path strategies such as minimal attachment and late closure, reading times were shorter (Frazier and Rayner, 1982). This was in comparison to sentences that did not comply with the parsing strategies mentioned and thus the use of such strategies led to misinterpretation.

Example: (from Frazier & Rayner, 1982):

Since Jay always jogs a mile seems like a short distance to him

With this example, ambiguity lies with the phrase a mile. In accordance to the garden-path model, readers would interpret this sentence in a certain way (i.e. that a mile modifies the verb jogs) but only to soon realise this interpretation was incorrect. Rather, a mile marks the beginning of a new phrase/node. When looking at Frazier and Rayners results, eye recordings showed longer time spent reading the latter stages of this sentence rather than the beginning. This would suggest participants had naturally attached the phrase a mile to the current constituent (jogs). Not only does this show a garden-path way of processing a sentence, but demonstrates participants use of late closure to find ways of disambiguating the sentence. When looking at the alternative sentence below, reading time would be much faster than of the previous version.

Since Jay always jogs a mile this seems like a short distance to him

This is a result of the inclusion of the pronoun this. As Frazier and Rayner discuss, the decision to use late closure methods would not result in ambiguity. The pronoun this would mark the beginning of a new constituent and therefore show success in the garden-path model.

Lexicalist Constraint-Based Model

As an alternative theory of sentence processing, the Lexicalist Constraints-Based model supports the idea of an interactive account. The basis of this theory suggests that rather than just syntactic cues (garden-path), all information/cues available to the reader are utilised from the very beginning. Language comprehension is  “accomplished through the satisfaction of multiple probabilistic constraints” (MacDonald, 1997), i.e. syntactic and non-syntactic cues that are used in parallel. Constraint-based models are though (like garden-path models), incremental in that words are processed step-by-step. The use of certain information to process a sentence is said to depend on the level of association between aspects of the sentence and level of activation of certain cues. This can cause complications, as when for instance two separate contextual cues are activated on the same level when reading a sentence, conflicting solutions can cause processing difficulty (Rueschemeyer & Gaskell, 2007).

Probably the most famous example of lexical ambiguity is the example shown below from (Bever, 1970):

A) The horse raced past the barn

B) The horse raced past the barn fell

Here ambiguity lies with the verb raced – in A taking a past tense form and in B taking a past participle form (i.e. the horse was actually raced past the barn). As (MacDonald, Pearlmutter, & Seidenberg, 1994) state, when one sentence takes the verb form of the other, analysis will be incorrect. Unlike with syntactic problems, ambiguity here lies with lexical meaning, rather than structure itself.

Lexical Frequency

A key concept to the Lexicalist constraints-based idea is that “most ambiguous words have alternative meanings that differ considerably in frequency” (MacDonald, Pearlmutter, & Seidenberg, 1994).

A well-known study into lexical frequency and ambiguous sentences can be seen in (Rayner & Duffy, 1986)’s work. Similar to Frazier and Rayner’s 1982 study of garden-path sentences, eye fixation times were recording when participants’ read ambiguous words with two equally frequent meanings (Rayner & Duffy, 1986) – “experiment 2”. The study found that participants spent longer reading ambiguous words with equally frequent meanings compared to unambiguous words with equally frequent meanings. Moreover results saw less time spent reading two ambiguous words where one meaning was more frequent. These clear findings suggested, “that word frequency and the presence of two highly likely meanings may affect lexical access” (Rayner & Duffy, 1986). When words have equal frequency in meaning, processing cues responsible for determining correct semantic meaning struggled to compete, increasing overall comprehension time of ambiguous lexis.

Further studies such as Duffy et al. (1998) went on to investigate the relationship between context, ambiguous words and eye fixation times. When altering the context a certain word appeared in, one meaning was naturally favoured over the other, showing how “contextual information can result in activation of only a single meaning of an ambiguous word” (MacDonald, Pearlmutter, & Seidenberg, 1994).

Verb Bias

Another key aspect of the constraint-based model is the idea of early integration of verb bias. In sentences containing ambiguous verb-phrases, (Snedeker & Trueswell, 2004) found that verbs posses a certain bias towards specific features. For example verbs that tended towards association with noun-modifier prepositional phrases:

Choose the frog with the feather

Since the verb choose was said to show bias towards the prepositional phrase, possible ambiguity would naturally be averted (i.e. thinking the frog is in possession of the feather).  Another example from (Snedeker & Trueswell, 2004) can be seen (below), in this case with verbs that tended towards association with instrumental prepositional phrases:

Tickle the frog with the feather

Again, ambiguity could occur if interpreting this sentence as a frog in possession of the feather, but processing abilities are said to usually refrain from this interpretation in favor of the correct meaning.

This idea of verb bias would support a constrain-based theory of sentence processing. The natural inclination towards verb association would appear at the same stage, or even before any syntactic analysis.

Conclusion

Although both principles bare in-depth analysis into the way we process sentences, it could be said the modular/garden-path model approach to sentence processing bares more complete findings. Although the Lexicalist Constraint-based model allows for complete sentence analysis from the onset, its slight haziness in competing cues and at what specific point they are actually utilised can be problematic. As Frazier states (below), the garden-path model accounts for possible ambiguity and possible ways of resolutions:

“The simplicity of these principles (minimal attachment/late closure/filler strategies) is the most impressive evidence possible for the current modular approach to language and mind.” (Frazier, 1987)

A real concrete definition of sentence processing is yet to be found, but there is scope for both models having real applications in the way we tackle sentence comprehension.

References

• Bever, T G. (1970). The cognitive basis for linguistic structure. New York: Wiley.
• Carroll, D. (2007). Psychology of language (pp. 134-137). Cengage Learning.
• Duffy. S. A., Morris, R. K., & Rayner, K. (1988). Lexical ambiguity and fixation times in reading. Journal of Memory and Language, 27, (pp. 429- 446)
• Ferreira & Clifton (1986). The independence of syntactic processing. Journal of Memory and Language 25, (pp. 348-368)
• Fodor, J. (1983). Modularity of Mind: An Essay on Faculty Psychology. Cambridge, Massachusetts: MIT Press
• Frazier, L. (1987). Sentence processing: a tutorial review. In Attention and performance XII: The Psychology of Reading (pp. 561–586). Erlbaum.
• Frazier & Rayner (1982), cited in Carroll, D. (2007). Psychology of language (pp. 134-137). Cengage Learning.
• Frazier, L., & Rayner, K. (1982). Making and correcting errors during sentence comprehension: Eye movements in the analysis of structurally ambiguous sentences. Cognitive Psychology, 14(2), (pp. 178-210.)
• http://dx.doi.org/10.1016/0010-0285(82)90008-1

• MacDonald, M. (1997). Lexical representations and sentence processing (pp. 126-133). Psychology Press.
• McDonald, M., Pearlmutter, N., & Seidenberg, M. (1994). Lexical nature of syntactic ambiguity resolution. American Psychological Association.
• Poesio, M. (2004). UNNDERSPECIFICATION AND INCREMENTALITY Week 2 Handout: Incrementality in Parsing, 1.
• Rueschemeyer, S., & Gaskell, M. (2007). The Oxford handbook of psycholinguistics (pp. 289-307). Oxford University Press.
• Snedeker & Trueswell (2004). The developing constraints on parsing decisions: The role of lexical biases and referential scenes in child and adult sentence processing. Cognitive Psychology 49, (pp. 238-299)
• Van Gompel, R. (2013). Sentence processing (pp. 209-212). London: Psychology Press.
• Weiner, I. (2003). Handbook of psychology, Experimental Psychology (pp. 530-533). John Wiley & Sons.