Home

UttutPipe

ref: https://hackmd.io/DWNGXRTJS9eQQ35lzJutwQ?both

Forward

Input RAW DATA

• label = [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 0, 0, 0, 2, 2, 3, 3, 3, 0, 0, 0]
• utt = "><><我想要喝200.5元的@@珍奶10杯!!!"

Step_1: normalize

• label = [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 0, 0, 0, 2, 2, 3, 3, 3, 0, 0, 0]
• utt = "><><我想要喝200.5元的@@珍奶10杯!!!"
• forward_edits = [(0, 4, ''), (8, 13, '3float1'), (15, 17, ''), (19, 21, '2int')]
• reverse_edits = [(0, 0, '><><'), (4, 13, '200.5'), (17, 23, '10'), (24, 24, '!!!')]

Step_2: tokenize

• label = [0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 2, 2, 3, 3, 3, 3, 3, 3, 3]
• utt = "我想要喝_3float1_元的珍奶_2int_杯"
• spans = [(0, 1), (1, 3), (3, 4), (4, 13), (13, 14), (14, 15), (15, 17), (17, 23), (23, 24)]
• reverse: 好像不用

Step_3: add sos and eos

• label = [0, 0, 0, 1, 1, 0, 2, 3, 3]
• utt = ['我', '想要'. '喝', '3float1', '元', '的', '珍奶', '2int', 杯']
• forward_edits = [(0, 0, ['sos']), (9, 9, ['eos'])]
• reverse_edits = [(0, 1, []), (10, 11, [])]

Step_4: padding

• label = [0, 0, 0, 0, 1, 1, 0, 2, 3, 3, 0]
• utt = ['sos', '我', '想要'. '喝', '3float1', '元', '的', '珍奶', '2int', 杯', 'eos']
• forward_edits = [(11, 11 , ['pad', 'pad'])]
• reverse_edits = [(11, 13, [])]

Step_5: token2int

• label = [0, 0, 0, 0, 1, 1, 0, 2, 3, 3, 0]
• utt = ['sos', '我', '想要'. '喝', '3float1', '元', '的', '珍奶', '2int', 杯', 'eos']
• forward_edits = [(0, 1, [0]), (1, 2, [1]), (2, 3, [2]), (3, 4, [3]), (4, 5, [4]), ...]
• reverse_edits = [(0, 1, ['sos']), (1, 2, ['我']), (2, 3, ['想要']), ...]

Output

• label = [0, 0, 0, 0, 1, 1, 0, 2, 3, 3, 0, 0, 0]
• utt = ['sos', '我', '想要', '喝', '3float1', '元', '的', '珍奶', '2int', '杯', 'eos', 'pad', 'pad']

Prediction

• prediction = [2, 1, 0, 3, 0, 1, 2, 0, 1, 0, 3, 1, 2]

=========================================

Backward

Input

• prediction = [2, 1, 0, 3, 0, 1, 2, 0, 1, 0, 3, 1, 2]
• transformed_utt = ['sos', '我', '想要', '喝', '3float1', '元', '的', '珍奶', '2int', '杯', 'eos', 'pad', 'pad']

Step_4: Unpadding

• forward_edits = [(11, 11 , ['pad', 'pad'])]
• reverse_edits = [(11, 13, [])]
• propagated_prediction = [2, 1, 0, 3, 0, 1, 2, 0, 1, 0, 3]

Step_3: remove sos and eos

• forward_edits = [(0, 0, ['sos']), (9, 9, ['eos'])]
• reverse_edits = [(0, 1, []), (10, 11, [])]
• propagated_prediction = [1, 0, 3, 0, 1, 2, 0, 1, 0]

Step_2: Untokenize

• spans = [(0, 1), (1, 3), (3, 4), (4, 13), (13, 14), (14, 15), (15, 17), (17, 23), (23, 24)]
• propagated_prediction = [1, 0, 0, 3, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 0, 0, 1, 1, 1, 1, 1, 1, 0]

Step_1: denormalize

• forward_edits = [(0, 4, ''), (8, 13, '3float1'), (15, 17, ''), (19, 21, '2int')]
• reverse_edits = [(0, 0, '><><'), (4, 13, '200.5'), (17, 23, '10'), (24, 24, '!!!')]
• propagated_prediction = [0, 0, 0, 0, 1, 0, 0, 3, 0, 0, 0, 0, 0, 1, 2, 0, 0, 1, 1, 0, 0, 0, 0]
• default label 0
• reduce func

OUTPUT

• raw_utt = "><><我想要喝200.5元的@@珍奶10杯!!!"
• [0, 0, 0, 0, 1, 0, 0, 3, 0, 0, 0, 0, 0, 1, 2, 0, 0, 1, 1, 0, 0, 0, 0]

Aggregate

• h_output = [我 -> 1, 喝 -> 3, 元 -> 1, 的 -> 2, 10 -> 1]