Generative Resolution of Proportional Analogies between Sentences

In this paper, we address the task of solving proportional analogies between sentences in a generative manner. To this end, we use pre-trained word or sentence embedding models or fine-tune several pre-trained language models. Experiments on analogies between short sentences show that fine-tuning GPT-2 achieves the best performance, while combining a word embedding model with its vector-to-sequence decoder delivers competitive accuracy. However, the increased sizes of large-scale language models might not be worth the slight increase in performance observed when compared with lightweight models to simple analogy tasks. To extend this, we develop a data pipeline for collecting semantico-formal analogies between long sentences, spotting parallel associations beyond individual words. We create a complex set of analogies, each validated with evidence of relational matches on underlying concepts. Comparative analyses of three language models demonstrate that the autoregressive framework is particularly effective at learning analogy structures when sequences are long.