Typical techniques for sequence classification are designed forwell-segmented sequences which have been edited to remove noisy or irrelevantparts. Therefore, such methods cannot be easily applied on noisy sequencesexpected in real-world applications. In this paper, we present the TemporalAttention-Gated Model (TAGM) which integrates ideas from attention models andgated recurrent networks to better deal with noisy or unsegmented sequences.Specifically, we extend the concept of attention model to measure the relevanceof each observation (time step) of a sequence. We then use a novel gatedrecurrent network to learn the hidden representation for the final prediction.An important advantage of our approach is interpretability since the temporalattention weights provide a meaningful value for the salience of each time stepin the sequence. We demonstrate the merits of our TAGM approach, both forprediction accuracy and interpretability, on three different tasks: spokendigit recognition, text-based sentiment analysis and visual event recognition.
展开▼
