Theoretically founded algorithms for the automatic production of analogy tests in NLP

点击翻译按钮获取中文摘要

During the second year, work on the tasks (a) to (c) have been pursued in parallel.(a) Two approaches were tried to cast vector representations of strings: The first one directly used Parikh vector representations and the second one a one hidden-layer neural network. Recall and precision were measured on various data.(b) Several directions were explored. (b.1) A series of experiments to approximate real-valued vectors to integer-valued vectors was run. Several analogy test sets in several languages were used. The new version of the programs with acceleration, implemented during the first fiscal year, was used. No parallelogram representing analogies between vectors can be discovered in none of the settings. This result has been published in a an international conference. (b.2) Work on casting words from word analogy test sets into their definitions, i.e., sentences, was done. The definitions with the analogical structure induced by the word analogies were used to fine-tune a sentence embedding space with contrastive learning. Such fine-tuned spaces delivered better performance in semantic similarity tasks. (b.3) Programs have been written to automatically extract series of analogies from a subspace around a given word. Preliminary experiments were run on classical examples. The obtained analogies are almost always formal, although they originate from an embedding space built using the distributional hypothesis.(c) Parallelisation of programs is considered as finished in the first fiscal year.