Learning to Discover, Ground and Use Words with Segmental Neural Language Models
This addresses the problem of learning language structure and meaning from raw data for AI systems that need to process multimodal information.
The researchers developed a segmental neural language model that simultaneously discovers word-like units from unsegmented character sequences, learns sentence structure, and grounds words in visual context. Experiments showed the unconditional model outperformed character LSTM models in predictive distributions, matched nonparametric Bayesian models in word discovery, and visual conditioning improved performance on both tasks.
We propose a segmental neural language model that combines the generalization power of neural networks with the ability to discover word-like units that are latent in unsegmented character sequences. In contrast to previous segmentation models that treat word segmentation as an isolated task, our model unifies word discovery, learning how words fit together to form sentences, and, by conditioning the model on visual context, how words' meanings ground in representations of non-linguistic modalities. Experiments show that the unconditional model learns predictive distributions better than character LSTM models, discovers words competitively with nonparametric Bayesian word segmentation models, and that modeling language conditional on visual context improves performance on both.