TL;DR
This paper introduces "topical surprisal," a novel computational measure of how predictable a word is based on broader discourse context, not just the immediate sentence. By combining this with traditional lexical surprisal and fMRI data from naturalistic listening, they demonstrate that different brain regions are involved in processing local vs. broad contextual information during language comprehension.
What Problem Does This Paper Address?
Understanding how humans comprehend language involves figuring out how we predict upcoming words. A key concept in this area is "surprisal," which quantifies how unexpected a word is given the preceding context. Traditionally, surprisal has been calculated using only the immediate linguistic context (e.g., the previous few words or grammatical structure). However, human comprehension also relies on broader context, like the overall topic of a conversation or story. This paper addresses the gap of how to computationally model and neurobiologically investigate this broader contextual influence on language processing. They want to know if processing local (sentence-level) and broad (topic-level) contextual information recruits distinct neural mechanisms.
Methods
The core of this paper is the introduction of topical surprisal, a new computational predictor of word predictability. This is calculated by:
- Topic Modeling: Using Latent Dirichlet Allocation (LDA) to identify underlying topics in a large corpus (COCA).
- Contextual Topic Probability: For a given word in a text, defining its "context" as the preceding 30 seconds of audio. This context is then treated as a document to estimate the probability of each topic being relevant.
- Topical Surprisal Calculation: The surprisal of a word is then computed as the weighted average of its probability given each topic, where the weights are the estimated topic probabilities for that context.
This new measure is then used alongside traditional lexical surprisal (calculated using n-gram and LSTM language models based on sentence-level context) as predictors of brain activity. The researchers used fMRI data collected while participants listened to a story. They employed a General Linear Model (GLM) to regress the fMRI BOLD signal against these surprisal predictors, allowing them to identify brain regions whose activity correlates with different types of contextual prediction.
Datasets
- Training Data for Language Models:
- N-gram surprisal: Trained on the Corpus of Contemporary American English (COCA), a large, genre-balanced corpus of over one billion words.
- LSTM surprisal: Trained on 90 million words of English Wikipedia.
- Topic Model (LDA): Trained on 219,380 documents from COCA to estimate topic distributions.
- fMRI Data: Collected from 56 participants listening to an audiobook of "The Little Prince" (English translation). This is a naturalistic listening paradigm, which is highlighted as a strength.
Key Insights and Ideas
- Beyond the Sentence: The paper convincingly shows that language comprehension isn't just about predicting the next word based on the immediate sentence. Broader topical context plays a significant role.
- Distinct Neural Networks: The most striking finding is that local (lexical/syntactic) and broad (topical) contextual prediction engage different brain regions. This suggests a functional specialization within the language network.
- Right Hemisphere Involvement: Topical surprisal, representing broader context, showed significant activation in the right hemisphere, particularly in the Precuneus and Middle Temporal Gyrus. This is a key differentiator from the typically left-lateralized networks associated with local processing.
- LSTM vs. N-gram: The study also provides evidence that LSTM models, unlike n-gram models, capture aspects of syntactic representation that are reflected in distinct neural activity patterns (specifically, implicating the left Superior Temporal Gyrus for LSTM surprisal).
- Computational Linguistics for Neuroscience: The work underscores the power of using computational linguistics tools (like topic modeling and advanced language models) to operationalize and test hypotheses about neurocognitive mechanisms in language processing.
Key Results
- Analysis 1 (N-gram vs. Topical Surprisal): Found distinct, non-overlapping brain regions for local (n-gram) and broad (topical) context. N-gram surprisal activated bilateral Anterior Temporal Lobe and left Inferior Frontal Gyrus. Topical surprisal activated the right Precuneus and right Middle Temporal Gyrus.
- Analysis 2 (LSTM vs. Topical Surprisal): Similar to Analysis 1, local (LSTM) and broad (topical) context showed distinct neural correlates. LSTM surprisal activated bilateral Anterior Temporal Lobe and left Superior Temporal Gyrus. Topical surprisal again showed right-hemisphere activation in the Precuneus and Middle Temporal Gyrus.
- Behavioral Results: Participants showed high comprehension accuracy (90% on average), confirming they were attentively listening to the story.
- Statistical Significance: All whole-brain effects survived stringent FWE voxel correction (p < 0.05).
Limitations and Future Work
- Sample Representativeness: The study acknowledges that the sample was drawn from a university population and may not be representative of the general population.
- Safety Protocols: Participants were excluded if they had metal implants, fresh tattoos, or certain medical conditions, which is standard but limits the participant pool.
- Future Work: The authors suggest exploring the differences between LSTM and n-gram models further, particularly in narrowing down the nature of syntactic dependencies captured (e.g., sequential vs. structural, short vs. long-distance). They also propose further investigation into the connection between temporal receptive windows and the implicated brain regions.
Summary updated