The instruction understanding in the eld of music aim for intention judgment and slot lling. In this paper, we rst employ BiLSTM model to recognize named entities for slot lling, then adopt heuristic rules for intention judgment. The instruction understanding in the eld of music includes two tasks, namely intention judgment and slot lling. We rst employ Bi-LSTM model to recognize named entities for slot lling, then adopt heuristic rules for intention judgment.
2.1
Bi-LSTM Model for slot lling
The slot lling task aims to recognize named entities in the led of music. The named entity recognition task can be seen as a sequence labeling problem, that is, given a sequence X =< x1, ..., xn > from texts, the goal is to label X with tag sequence Y =< y1, ..., yn >. We adopt the BIEO tag scheme, where B indicates \xi is the beginning of a clinical named entity", I indicates \xi is inside a clinical named entity", E indicates \xi is the end of a clinical named entity", and O indicates \xi is outside a clinical named entity". Take X =<请, 播, 放, 周, 杰, 伦, 的, 稻, 香> as an example, its corresponding Y =<O, O, O, B-artist, I-artist, E-artist, O, B-song, E-song>.
We employ a Bi-LSTM Model model for this task. Note that we exploit
character embedding rather than word embedding, and has no additional features
used. As shown in Fig. 1, the raw natural language input sentence is processed
into sequence of characters X = [x]1T . The character sequence is fed into the
embedding layer, which produces dense vector representation of characters. The
softmax
forwardLSTM
backwardLSTM
charcaterembedding
character vectors are then fed into a bidirectional LSTM layer [
c~t ot = (Woet + Uoht 1 + bo) ht = ot tanh(ct) (1) (2) (3) (4) (5) (6) where h, i, f , o 2 Rdh are dh-dimensional hidden state (also called output vector), input gate, forget gate and output gate, respectively; Wi, Wf , Wc, Wo 2 R4dh de , Ui, Uf , Uc, Uo 2 R4dh dh and bi, bf , bc, bo 2 R4dh are the parameters of the LSTM; is the sigmoid function, and denotes element-wise production.
In the bidirectional LSTM, for any given sequence, the network computes both a left, h!t, and a right, ht, representations of the sequence context at every input, xt. The nal representation is created by concatenating them as ht = [h!t; ht]. Finally, ht is fed into a fully connected layer with a softmax function to predict the tag yt. 2.2
Heuristic Rules for Intention Judgment
The intention judgment task can be seen as a classi cation task. We simply adopt heuristic rules to deal with the task. If a sentence contains named entities recognized by Bi-LSTM model, it is music related. If not, it is music unrelated. Speci cally, if the sentence contains key words such as \唱", \放", \歌" , it will be labeled as \Random" . If not, it will be labeled as \No music" 3
Experimental Settings
We rst clean the utterances and only process the third sentence in it. When
training, we use all the labeled data as training data. The Bi-LSTM model is
trained by Adam algorithm [
As for the Bi-LSTM model, each Chinese character is projected into the continuous space of 256-dimensional Euclidean space to reduce the dimensionality. The size of LSTM hidden layer is 256, and the batch size of the model is 32.