Natural Language Processing

Language models have become remarkably fluent. Fluency, however, is not the same as understanding. Our natural language research examines how meaning is represented computationally, how it connects to the world, and how language can carry reasoning rather than merely resemble it.

Research topics

  • Language understanding
  • Computational semantics
  • Multilingual modeling
  • Grounded language learning
  • Information retrieval
  • Language reasoning
  • Speech–language systems
Fig. 04Schematic — not experimental data
thesignalisreadasatraceofitssourcequery: “trace” → weighted context
Fig. 04Schematic. Arcs between tokens illustrate the pairwise dependencies an attention mechanism can represent; line weight indicates relative strength.

Why it matters

Why it matters

Language is how people state goals, share knowledge, and justify decisions. Systems that work with language therefore sit close to human intent — which makes their failure modes consequential. A model that misreads an instruction, invents a citation, or reasons fluently to a false conclusion can do so convincingly.

There is also a question of reach. Most language technology performs best in a handful of high-resource languages. Building models that represent meaning in ways that transfer across languages is both a scientific problem and a matter of who benefits from the technology.

Key research questions

Key research questions

  1. Q1

    What evidence would distinguish a model that represents meaning from one that models surface co-occurrence?

  2. Q2

    How can retrieval and generation be combined so that claims remain traceable to sources?

  3. Q3

    Do multilingual models learn shared semantic representations, and how can low-resource languages benefit from them?

  4. Q4

    When a model produces a step-by-step explanation, how faithfully does it reflect the computation that produced the answer?

  5. Q5

    How should language be grounded in perception and action to support reference to the physical world?

Primary subfields

Primary subfields

Computational semantics
Formal and learned accounts of meaning, entailment, and compositionality.
Multilingual modeling
Shared representations across languages and transfer to under-resourced ones.
Retrieval and grounding
Connecting language to documents, databases, and perceptual input.
Reasoning in language
Multi-step inference, argumentation, and the faithfulness of explanations.
Speech and language
Joint modeling of spoken and written language.

Representative methodologies

Representative methodologies

  • M1

    Targeted evaluation sets

    Minimal pairs and controlled variations that test one linguistic phenomenon at a time.

  • M2

    Attention and attribution analysis

    Tracing which inputs a model relies on for a given output.

  • M3

    Retrieval-augmented systems

    Coupling models to external sources and measuring attribution quality.

  • M4

    Cross-lingual transfer

    Training on some languages and evaluating on others.

Relationships to other domains

Relationships to other domains

Working on questions in Natural Language Processing?

We welcome conversations with researchers and organizations working on related questions — whether to compare notes, discuss a collaboration, or share work in progress.