String acceptors and transducers are critical technologies for natural language and speech systems. They flexibly capture many kinds of stateful, left-to-right substitution; simple transducers can be composed into more complex ones; and they are trainable. Tree acceptors and transducers provide even more transformational power. Still, strings and trees are both weak at representing linguistic structure involving semantics and reference ("who did what to who"). Viewing semantic structures as directed acyclic graphs, we take a look at probabilistic acceptors and transducers for them, demonstrate some linguistic transformations, and point toward a foundation for semantics-based statistical machine translation.
Truth-conditional semantics for natural language provides for clear concepts of denotation, compositionality, and entailment. While it has formed the basis for a deep and detailed description of the structure of meaning in natural language, it falls dramatically short when it comes to robustness and coverage. Distributional semantics, on the other hand, enables efficient and comfortable acquisition of wide-coverage lexical semantic information from raw text corpora, but it does not easily lend itself to the modeling of structured information about meaning. Nor does it offer a straightforward story about the relationship between language and the world.
The truth-conditional and distributional paradigms are clearly complementary in their strengths and weaknesses. It is not at all obvious, however, how the views on semantics and semantic processing provided by the two respective frameworks might combine into a comprehensive and consistent picture.
In this talk, I will inspect recent approaches that aim to interleave the two paradigms, and I will discuss how much progress we have made towards a unified framework for computational semantics.