Discourse Context Predictability Effects in Hindi Word Order

Sidharth Ranjan, Marten van Schijndel, Sumeet Agarwal, Rajakrishnan Rajkumar

Introduction

In languages with flexible word order like Hindi, the choice of how to arrange constituents is rarely random. Linguists have long posited that various cognitive pressures guide these choices, such as the desire to minimize memory load by keeping related words close together (Dependency Locality) or the preference for placing known information before new information (Information Status). However, the role of discourse-level predictability—how the preceding linguistic context influences the structure of the current sentence—remains less understood. This research investigates whether syntactic priming and discourse context are significant predictors of Hindi word order and how these factors interact with established cognitive constraints.

Methodology

The researchers framed word order choice as a classification problem, aiming to distinguish naturally occurring sentences from grammatically plausible alternatives.

Data Generation and Features

Using approximately 2,000 reference sentences from the Hindi-Urdu Treebank (HUTB), the team generated thousands of variants for each sentence by permuting the order of words preceding the verb. They then trained a logistic regression model to identify the original corpus sentence among its variants using several features:

  • Dependency Locality: Measuring the distance between syntactically related words.
  • Surprisal Metrics: Standard trigram and PCFG surprisal to measure local predictability.
  • Adaptive LSTM Surprisal: A novel measure designed to capture discourse predictability.

The Adaptive Approach

To simulate discourse-level priming, the researchers utilized a neural language model (LSTM) trained on the 1-million-word EMILLE Hindi Corpus. The "adaptive" element involved fine-tuning the LSTM on the sentence immediately preceding the target sentence. This process models how a human listener's brain might be "primed" by recent context, making subsequent similar structures less surprising and easier to process.

Key Results

The inclusion of adaptive discourse predictability significantly enhanced the model's performance.

  • Classification Accuracy: The final model achieved a high accuracy of 94.06% in identifying the original corpus sentence against its generated variants.
  • Statistical Significance: Adding adaptive features provided a statistically significant improvement over non-adaptive baselines ($\chi^2 = 66.81, p < 0.001$).
  • Human Validation: To ensure the corpus sentences were truly the most natural choices, the researchers conducted a forced-choice study with native Hindi speakers. The results showed 89.92% agreement with the corpus, validating the study's foundational data.

Insights and Discussion

Predictability vs. Locality

A striking finding of the study was that human-chosen sentences often exhibited longer dependency lengths than their artificial variants. This suggests that in Hindi, the drive to maximize discourse predictability can override the pressure to minimize dependency length. Speakers may choose a "harder" structure in terms of memory load if it makes the sentence more predictable within the broader conversation.

The Inverse Frequency Interaction

The effects of discourse predictability were most pronounced in rare, non-canonical structures, such as Indirect Object (IO) fronting. For these rare structures, the adaptive model showed a 0.6% increase in accuracy over the baseline. This suggests an "inverse frequency interaction": when a structure is uncommon, its appearance provides a stronger priming effect for subsequent instances compared to common, expected structures.

Short-term Memory Constraints

The researchers also tested whether adapting to a larger window of context (the last five sentences) improved the model. Interestingly, it did not. This indicates that syntactic priming in this context is likely driven by short-term "residual activation" in the brain rather than a longer-term learning process during a single discourse.

Conclusion

This study demonstrates that discourse context is a vital component in understanding the word order choices of Hindi speakers. By using adaptive LSTMs to model priming, the researchers showed that predictability is a powerful cognitive pressure that often outweighs dependency locality, particularly for rare syntactic structures. While the study focused on written news text from the HUTB, future work may explore how these preferences shift in spontaneous speech, where real-time cognitive demands are even higher.

Summary updated

BibTeX
@inproceedings{ranjan2022discourse,
  author = {Sidharth Ranjan and Marten van Schijndel and Sumeet Agarwal and Rajakrishnan Rajkumar},
  title = {{Discourse Context Predictability Effects in Hindi Word Order}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2022},
  doi = {10.18653/v1/2022.emnlp-main.710},
}
Figure