Design a word embedding specifically for query expansio...

To design an embedding for query expansion, the target distribution must consist of terms found in relevant documents. Supervision can be derived from query logs with human judgments or pseudo-relevance feedback (PRF) from the top-m documents. When training, use random non-word vectors as negatives. At serving time, the learned term scores should be interpolated with a maximum-likelihood baseline to ensure the expanded terms are both semantically relevant and statistically appropriate for the retrieval engine.