Can Symbol Grounding Improve Low-Level NLP? Word Segmentation as a Case Study