Functional Specialization and Non-Linear Processing in the Human Language Network

This project investigates how different regions within the human language network respond to various measures of linguistic complexity during naturalistic story comprehension. We extend beyond previous network-level analyses by Shain et al. (2020) to examine regional specialization patterns and non-linear complexity processing mechanisms.

Overview

This repository contains code and analysis for a computational psycholinguistics study that challenges assumptions about uniform complexity processing in the language network. Our research reveals functional specialization across brain regions and uncovers a surprising semantic processing paradox.

The provided code is focused on the analysis itself. If you would like to see how the GPT-2 surprisal values or the GloVe embeddings were derived, please refer to the accompanying Jupyter notebooks, where the data generation process is fully documented.

Key Research Questions

1. Regional Specialization: Do different areas within the language network exhibit specialized response patterns to different types of prediction difficulty (syntactic vs. lexical vs. contextual)?

2. Non-Linear Processing: Does linguistic complexity processing exhibit capacity constraints and non-linear patterns rather than uniform linear scaling?

3. Semantic Complexity Paradox: How does story-specific semantic complexity influence language network activation during naturalistic comprehension?

Major Findings

• Functional Specialization: Frontal regions (LangMFG, LangIFG) showed strongest sensitivity to syntactic complexity, temporal regions (LangAntTemp, LangPostTemp) to lexical predictability, while angular gyrus (LangAngG) responded most to GPT-2 contextual complexity.

• Non-Linear Complexity Processing: Systematic deviations from linearity across all regions and measures, contradicting linear assumptions of previous approaches.

• Semantic Processing Paradox: Stories with highest semantic diversity produced language network deactivation, while coherent narratives elicited maximal activation, suggesting adaptive processing mechanisms.

Installation

Prerequisites

• Python 3.11+
• Conda or Miniconda
• Git

Setup Instructions

1. Clone the repository:

   git clone https://github.com/LizMelamed/lang-network-specialization.git
   cd functional-specialization-language-network

2. Create the conda environment:

   conda env create -f environment.yml

3. Activate the environment:

   conda activate cogsci-proj

4. Download the Data Folder:

   python code/data_download.py

Usage

Quick Start - Run Complete Analysis Pipeline

Open-ended analysis (primary contributions):

# Regional specialization analysis
python code/open_task_surprisal.py

# Semantic complexity analysis  
python code/open_task_semantic_analysis.py

Structured replication tasks:

# Open Jupyter notebook for structured tasks
jupyter notebook notebooks/Closed_Task1.ipynb
jupyter notebook notebooks/Closed_Task2.ipynb
jupyter notebook notebooks/Closed_Task3.ipynb

Interactive exploration:

jupyter notebook notebooks/Open_Task_Semantic_Analysis.ipynb
jupyter notebook notebooks/Open_Task_Surprisal.ipynb

Analysis Components

1. Complexity Binning Analysis: Creates percentile-based surprisal bins to identify non-linear response patterns
2. Regional Specialization: Analyzes differential responses across six language network regions
3. Multi-Model Comparison: Compares PCFG, 5-gram, and GPT-2 surprisal measures
4. Semantic Characterization: Uses GloVe embeddings to quantify story semantic properties
5. HRF Convolution: Properly accounts for hemodynamic response delays

Repository Structure

├── code/                           # Analysis scripts
│   ├── open_task_surprisal.py             # Regional specialization analysis
│   ├── open_task_semantic_analysis.py     # Semantic complexity analysis
│   └── utils/                             # Helper functions
├── data/                           # Input datasets
│   ├── LANG_MD_ts_X.csv                   # Combined predictors (PCFG, 5-gram surprisal)
│   ├── LANG_MD_y_train.csv                # Combined BOLD responses (LANG + MD networks)  
│   ├── pred_use.csv                       # Preprocessed predictor variables
│   ├── resp_use.csv                       # Preprocessed brain response data
│   ├── merged.csv                         # Combined analysis-ready dataset
│   ├── gpt2_surprisal_results.csv         # GPT-2 surprisal computations
│   └── merged_df_with_vectors.parquet     # Generated dataset (appears after running data_download.py)
├── notebooks/                      # Jupyter notebooks  
│   ├── Closed_Task1.ipynb                 # Structured Task 1: FastText vs GloVe decoding
│   ├── Closed_Task2.ipynb                 # Structured Task 2: Static vs contextual embeddings
│   ├── Closed_Task3.ipynb                 # Structured Task 3: Neural encoding models
│   ├── Open_Task_Semantic_Analysis.ipynb  # Semantic complexity analysis
│   └── Open_Task_Surprisal.ipynb          # Regional specialization analysis
├── results/                        # Auto-generated outputs
│   ├── all_fROIs_GPT_2.png               # GPT-2 regional responses
│   ├── all_fROIs_PCFG.png                # PCFG regional responses  
│   ├── all_fROIs_5_gram.png              # 5-gram regional responses
│   ├── bold_by_bins_*.png                # Complexity binning visualizations
│   ├── semantic_*.png                    # Semantic analysis plots
│   ├── story_progression_*.png           # Story-specific response patterns
│   └── surprisal_overview.png            # Summary visualizations
├── environment.yml                 # Conda environment specification
└── README.md                      # This file

Data Description

Primary Datasets

Natural Stories fMRI Corpus (Shain et al., 2020):

• Participants: 78 native English speakers
• Stories: 8 naturalistic narratives (990-1099 words each)
• Brain Networks: Language (LANG) and Multiple Demand (MD) networks
• Regions: 6 language network regions (LangAngG, LangAntTemp, LangIFG, LangIFGorb, LangMFG, LangPostTemp)
• Predictors: PCFG surprisal, 5-gram surprisal, word timing information
• Responses: BOLD signals temporally aligned to word-level predictors

Pereira et al. (2018) Structured Tasks:

• Participants: 16 fluent English speakers
• Stimuli: Individual words (180), sentence passages (384), naturalistic sentences (243)
• Purpose: Baseline validation for neural decoding approaches

Key Variables

Predictor Variables:

• totsurp: PCFG surprisal values (syntactic complexity)
• fwprob5surp: 5-gram surprisal values (lexical predictability)
• gpt2_surprisal: GPT-2 surprisal values (contextual complexity)
• word: Word tokens with precise timing
• docid, subject, time, fROI: Experimental identifiers

Response Variables:

• BOLD: Brain activation values (HRF-convolved)
• network: Brain network labels (LANG, MD)
• fROI: Functional region of interest within networks

Semantic Variables:

• semantic_coherence: Story-level semantic consistency (GloVe-based)
• semantic_diversity: Story-level semantic complexity (1 - coherence)
• vocabulary_size: Unique words per story

Methodology

Analysis Pipeline

1. Data Preprocessing:

   • Temporal alignment of word-level predictors with BOLD responses
   • Hemodynamic Response Function (HRF) convolution
   • Z-score normalization across all surprisal measures

2. Complexity Binning:

   • Divide each surprisal measure into 10 equal-count percentile bins
   • Ensure balanced sample sizes (~9,500 observations per bin)
   • Capture full range from predictable to highly complex constructions

3. Regional Specialization Analysis:

   • Linear correlations between surprisal measures and BOLD responses
   • Region-specific response profiles across complexity bins
   • Cross-measure comparisons (PCFG vs 5-gram vs GPT-2)

4. Non-linearity Testing:

   • Systematic deviations from linear fits
   • Identification of peaked responses at intermediate complexity
   • Statistical validation of capacity-limited processing

5. Semantic Analysis:

   • GloVe embedding-based story characterization
   • Semantic coherence and diversity metrics
   • Story-specific brain response patterns

Novel Methodological Contributions

• Multi-scale surprisal analysis: First comparison of classical (PCFG, n-gram) and modern (GPT-2) measures in naturalistic fMRI
• Regional complexity binning: Beyond network-level analysis to understand within-network specialization
• Semantic-neural mapping: Story-level semantic properties linked to regional brain responses
• Non-linear complexity modeling: Systematic testing of capacity constraints in language processing

Key Results

Regional Functional Specialization

• LangMFG: Primary executive control hub with strongest correlations across all surprisal measures (PCFG: r = 0.903)
• LangIFG: Core syntactic processor with preferential sensitivity to structural complexity
• LangPostTemp: Integrates semantic complexity across all measures
• LangAntTemp: Specializes in lexical processing (5-gram: r = 0.703)
• LangAngG: Strongest response to GPT-2 contextual complexity (r = 0.736)
• LangIFGorb: Narrative complexity gate with extreme story-specific variability

The Semantic Processing Paradox

Stories with highest semantic diversity ("Tulips": 0.390 diversity) produced language network deactivation, while coherent narratives ("Elvis": 0.768 coherence) elicited maximal activation. This suggests:

• Inverted-U relationship between semantic complexity and network engagement
• Adaptive processing mechanisms preventing resource depletion
• Potential shifting to MD network under excessive semantic load

Non-Linear Complexity Processing

• Systematic deviations from linearity across all regions and measures
• Peaked responses at intermediate complexity levels rather than monotonic increases
• Most consistent patterns observed for GPT-2 surprisal, suggesting modern neural language models better capture human comprehension mechanisms

Notes and Considerations

Technical Requirements

• Memory: GPT-2 analysis requires substantial RAM (8GB+ recommended)
• Reproducibility: All random operations use fixed seeds
• Platform: Tested on Python 3.11 with provided conda environment
• Data Size: Large fMRI files not included in repository due to size constraints

Important Methodological Notes

• No synthetic data generation: All analyses use real fMRI and linguistic data
• HRF convolution: Proper temporal alignment accounts for 4-6 second hemodynamic delays
• Balanced binning: Percentile-based approach ensures equal sample sizes across complexity levels
• Multiple comparison correction: Statistical results account for multiple regions and measures

Performance Considerations

• Code optimized for clarity and reproducibility rather than maximum speed
• For large-scale analysis, consider parallelization options in utils modules
• GPT-2 processing can be memory-intensive; reduce batch size if needed

Citation and References

Primary Paper

Melamed, E., & Bamberger, N. (2025). Functional Specialization and Non-Linear Processing in the Human Language Network. Course Project Report, Language, Computation and Cognition (00960222), Technion.

Primary Data Sources

Shain, C., Blank, I. A., van Schijndel, M., Schuler, W., & Fedorenko, E. (2020). fMRI reveals language-specific predictive coding during naturalistic sentence comprehension. Neuropsychologia, 138, 107307. OSF Repository

Pereira, F., Lou, B., Pritchett, B., Ritter, S., Gershman, S. J., Kanwisher, N., & Fedorenko, E. (2018). Toward a universal decoder of linguistic meaning from brain activation. Nature Communications, 9(1), 963.

Key Software Dependencies

• Neural Language Models: Transformers library (Hugging Face) for GPT-2 analysis
• Brain Analysis: Nilearn, SciPy, PyGAM for fMRI data processing
• Embeddings: GloVe for semantic analysis
• Standard Stack: NumPy, Pandas, Matplotlib, Seaborn, Statsmodels

Theoretical Framework

• Predictive Processing: Bastos et al. (2012), Kuperberg & Jaeger (2016)
• Language Network Organization: Fedorenko et al. (2013)
• Neural Language Models: Caucheteux & King (2022), Goldstein et al. (2022)

---

Contact: For questions about this analysis, to report issues, or to request access to processed data outputs, please open a GitHub issue or contact the project maintainers at elizavetam@campus.technion.ac.il or noabamberger@campus.technion.ac.il.

GitHub Repository: [[(https://github.com/LizMelamed/lang-network-specialization.git)]]

Academic Affiliation: Data and Decision Sciences, Technion - Israel Institute of Technology