TL;DR
This paper investigates whether the syntactic structure of a sentence, specifically its dependency relationships, influences where our eyes look back to during reading. By analyzing eye-tracking data alongside dependency parses of text, the researchers found that regressions are more likely to land on words that are syntactically related to the current word, suggesting that eye movements play a role in online syntactic analysis.
What Problem Does This Paper Address?
When we read, our eyes don't just move smoothly from left to right. They also make occasional backward jumps, known as regressions. These regressions have been hypothesized to serve two main purposes: either as a mechanism to re-analyze text when comprehension falters or as an integral part of the normal reading process, aiding in the construction of linguistic representations. This paper delves into the latter hypothesis, specifically exploring whether these regressions are guided by the underlying syntactic structure of the sentence. The core question is: do our eyes "look back" to words that are structurally connected to the word we're currently processing, even if there's no obvious comprehension difficulty?
Methods
The researchers employed a mixed-effects regression analysis to model the probability of eye regressions. They used eye-tracking data from participants reading narrative texts and compared two models:
- Null Model: This model included basic predictors like word order, word frequency, and surprisal (a measure of how predictable a word is in its context). These are standard controls in reading research, accounting for general processing difficulty.
- Full Model: This model built upon the null model by adding predictors derived from dependency parses of the sentences. Specifically, it included measures of the number of "head" and "dependent" relationships a word has with preceding words in the sentence.
A dependency parse represents a sentence as a network of directed relationships between words, where one word (the head) governs another (the dependent). This contrasts with constituency parsing, which groups words into hierarchical phrases. The paper argues that dependency structures offer a "shallow" representation of syntax.
Datasets
The study utilized an existing eye-tracking dataset originally collected for a study on mental simulation during literary reading. This dataset comprised:
- Participants: 102 native Dutch speakers.
- Stimuli: Three published short stories in Dutch, with lengths ranging from 2143 to 2988 words.
- Eye-tracking Data: Collected using an EyeLink1000+ system at a 500 Hz sampling rate.
- Dependency Parses: Generated using the ALPINO toolbox for Dutch natural language processing.
No new datasets were introduced.
Key Insights and Ideas
- Regressions are not random: The study provides evidence that backward eye movements are not purely reactive to comprehension errors but are also influenced by the sentence's syntactic architecture.
- Syntactic connections guide gaze: The most striking finding is that regressions are significantly more likely to land on words that are syntactically linked (as head or dependent) to the current word. This suggests that readers are actively using syntactic information to guide their eye movements.
- Local dependency parsing: The effect of dependency relations on regressions appears to be strongest for short distances (within the preceding four words). This implies that eye movements are primarily involved in local, rather than long-range, dependency parsing.
- Dependency structure as a predictor: The inclusion of dependency-based measures significantly improved the model's ability to predict regressions, highlighting the importance of syntactic structure beyond general text difficulty.
Key Results
- The full model, incorporating dependency measures, showed a significantly better fit to the data than the null model (χ² = 738.87, p < .0001).
- Both
N_head(number of words where the current word is a dependent) andN_dependents(number of words where the current word is a head) were significant predictors of regressions.N_headhad a positive effect (β = 0.242, p < .0001), andN_dependentsalso had a positive effect (β = 0.046, p < .0005). - Approximately 40% of regressions landed on words that were either heads or dependents of the word they originated from ("matches").
- The association between dependency relations and regressions was significant for distances of -1, -2, -3, and -4 words, but not for longer distances.
Limitations and Future Work
- Parser limitations: The study acknowledges that dependency parses are generated independently for each sentence, and no relations are assigned between sentences. This limits the analysis to intra-sentence dependencies.
- "Matches" vs. "Misses": While 40% of regressions were "matches" (landing on a related word), a significant portion (around 60%) were "misses" (landing on unrelated words). The paper suggests that the parser assigns a head to every word, whereas regressions are relatively rare events, explaining this discrepancy.
- Future directions: The authors suggest that these findings support the idea that humans engage in online syntactic analysis, at least at a local level, and that eye movements reflect this process. Future work could explore how these findings generalize to different types of text or linguistic structures. The study also opens avenues for investigating the role of regressions in more complex syntactic phenomena.
Summary updated