Combining Machine Learning with Formal Linguistic Semantics
Early computational linguistics in the 1970's and 80's encoded formal methods in the higher-level languages Lisp and Prolog. These systems were unacceptably slow on the computers available at the time. For example one system processed a single long sentence in 12 hours. Meanwhile statistical methods with Markov models had proven successful in speech recognition. A pendulum swing towards statistical methods ensued, and the past twenty years of computational linguistics has emphasized machine learning in opposition to formal linguistic rules. Meanwhile, computers are far more powerful and linguistic science has advanced, especially in formal semantics, creating an opportunity to combine linguistic and machine learning to advance towards intelligent natural language understanding software.
Following the Thread of Dialogue with Linguistic Semantics.
Linguistic semantics forms the basis of more intelligent and natural interaction with devices such as TVs and phones than is possible with machine-learning only methods. Formal linguistics in combination with statistical methods enables devices to understand more of the user’s utterances and to respond appropriately. The profound improvements with linguistic semantics are especially noticeable when users employ anaphoric expressions such as pronouns, demonstratives and definite descriptions.