logos (λόγος)

logos is a family of experiments designed to explore the acquisition of semantic representations of natural-language sentences by neural networks. It uses @clay-lab's transductions library to train Seq2Seq models on datasets and analyze the results.

logos uses Featural Context-Free Grammars from nltk to produce training data consisting of input sentences, a transformation token sem, and target outputs of predicate logic. transductions models may then be trained on these datasets.

The experiments directory contains the trained models and logs for several different experiments run with logos datasets:

• Alice-*: The Alice-* family of experiments explore the ability of Seq2Seq networks to generalize knowldge of anaphors (reflexive pronouns) to novel antecedents. The training data consists of transitive sentences of the form PERSON-1 VERBS {PERSON-2, him/herself}, where PERSON-1 and PERSON-2 may be distinct, and intransitive sentences of the form PERSON VERBS. In each experiment, certain reflexive combinations are withheld from the training data and we test the networks' abilities to generalize to these new antecedents.