Encoding and Retrieval Interference in Sentence Comprehension: Evidence from Agreement

Abstract

Long-distance verb-argument dependencies generally require the integration of a fronted argument when the verb is encountered for sentence interpretation. Under a parsing model that handles long-distance dependencies through a cue-based retrieval mechanism, retrieval is hampered when retrieval cues also resonate with non-target elements (retrieval interference). However, similarity-based interference may also stem from interference arising during the encoding of elements in memory (encoding interference), an effect that is not directly accountable for by a cue-based retrieval mechanism. Although encoding and retrieval interference are clearly distinct at the theoretical level, it is difficult to disentangle the two on empirical grounds, since encoding interference may also manifest at the retrieval region. We report two self-paced reading experiments aimed at teasing apart the role of each component in gender and number subject-verb agreement in Italian and English object relative clauses. In Italian, the verb does not agree in gender with the subject, thus providing no cue for retrieval. In English, although present tense verbs agree in number with the subject, past tense verbs do not, allowing us to test the role of number as a retrieval cue within the same language. Results from both experiments converge, showing similarity-based interference at encoding, and some evidence for an effect at retrieval. After having pointed out the non-negligible role of encoding in sentence comprehension, and noting that Lewis and Vasishth's (2005) ACT-R model of sentence processing, the most fully developed cue-based retrieval approach to sentence processing does not predict encoding effects, we propose an augmentation of this model that predicts these effects. We then also propose a self-organizing sentence processing model (SOSP), which has the advantage of accounting for retrieval and encoding interference with a single mechanism.

Keywords: ACT-R; agreement; encoding; long-distance dependencies; retrieval; self-organized sentence processing; similarity-based interference; working memory.

FIGURE 1

Distribution of reading times (in ms) in the four experimental conditions for the different regions of Experiment 1 (correct trials only).

FIGURE 2

Distribution of reading times (in ms) in the four experimental conditions for the different regions of Experiment 2 (correct trials only).

FIGURE 3

As the “canoe” treelet begins to attach to the NP (Subject Node) of S, the features of this node gravitate toward <+NP, + Boat, –Plural>. Consequently, when “sailboats” arrives, it matches the Subject Node on two features (+NP and +Boat) and competes strongly for attachment as subject, pushing the Plural feature of the Subject Node toward +. By contrast, when “cabins” arrives, it matches the Subject Node on only one feature and therefore does not push the Plural feature as strongly toward +. (Note: The VP Node of S is constrained to have the same Plural value as the Subject Node. Therefore, the state of the Subject Node determines the choice of verb number).