emb,t Sample Clauses

emb,t. = We(i). In the experiments, I use the same matrix for questions, evi- t dence, and answers. KB entities are considered to be individual tokens, while predicates and type names are tokenized into constituent words, i.e., by split- ting on underscore and dot characters for Freebase predicates. A layer-wise architecture of EviNets is displayed on Figure 3.9. The evidence matching module ((a) on Figures 3.8 and 3.9) estimates the relevance of each statement in the memory, and computes their weights using the softmax function. The evidence aggregation module (b) uses multiple ways to compute the aggre- gated statistics of evidences, mentioning each answer candidate. Finally, the answer scoring module (c) uses a fully connected network to predict a score for each answer candidate. EviNets selects the answer with the highest score as the final response. Evidence Matching Module Evidence matching is responsible for estimating the relevance of each of the pieces of evidence to the question, i.e., we = softmax(match(q, e)). The function match(q, e) can be implemented using any of the recently proposed semantic similarity estimation architectures16. One of the sim- plest approaches is to average question and each evidence token embed- 16e.g., see the ACL Wiki on Question Answering (State of the art). Lq Σ= · e t dings and score the similarity using the dot product: qemb = 1 Σ emb emb .