Morphologically rich languages often lack the annotated linguistic resourcesrequired to develop accurate natural language processing tools. We proposemodels suitable for training morphological taggers with rich tagsets forlow-resource languages without using direct supervision. Our approach extendsexisting approaches of projecting part-of-speech tags across languages, usingbitext to infer constraints on the possible tags for a given word type ortoken. We propose a tagging model using Wsabie, a discriminativeembedding-based model with rank-based learning. In our evaluation on 11languages, on average this model performs on par with a baselineweakly-supervised HMM, while being more scalable. Multilingual experiments showthat the method performs best when projecting between related language pairs.Despite the inherently lossy projection, we show that the morphological tagspredicted by our models improve the downstream performance of a parser by +0.6LAS on average.
展开▼
