使用 Seq2Seq 實現中英文翻譯

1. 介紹

1.1 Deep NLP

天然語言處理(Natural Language Processing,NLP)是計算機科學、人工智能和語言學領域交叉的分支學科,主要讓計算機處理或理解天然語言,如機器翻譯,問答系統等。可是因其在表示、學習、使用語言的複雜性,一般認爲 NLP 是困難的。近幾年,隨着深度學習(Deep Learning, DL)興起,人們不斷嘗試將 DL 應用在 NLP 上,被稱爲 Deep NLP,並取得了不少突破。其中就有 Seq2Seq 模型。 html

1.2 來由

Seq2Seq Model是序列到序列( Sequence to Sequence )模型的簡稱,也被稱爲一種編碼器-解碼器(Encoder-Decoder)模型,分別基於2014發佈的兩篇論文:python

  • Sequence to Sequence Learning with Neural Networks by Sutskever et al.,
  • Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation by Cho et al.,

做者 Sutskever 分析了 Deep Neural Networks (DNNs) 因限制輸入和輸出序列的長度,沒法處理未知長度和不定長的序列;而且不少重要的問題都使用未知長度的序列表示的。從而論證在處理未知長度的序列問題上有必要提出新解決方式。因而,創新性的提出了 Seq2Seq 模型。下面讓咱們一塊兒看看這個模型究竟是什麼。 git

2. Seq2Seq Model 之不斷探索

爲何說是創新性提出呢? 由於做者 Sutskever 通過了三次建模論證,最終才肯定下來 Seq2Seq 模型。並且模型的設計很是巧妙。讓咱們先回顧一下做者的探索經歷。
語言模型(Language Model, LM)是使用條件機率經過給定的詞去計算下一個詞。這是 Seq2Seq 模型的預測基礎。因爲序列之間是有上下文聯繫的,相似句子的承上啓下做用,加上語言模型的特色(條件機率),做者首先選用了 RNN-LM(Recurrent Neural Network Language Model, 循環神經網絡語言模型)。
github

rnn.png

上圖,是一個簡單的 RNN 單元。RNN 循環往復地把前一步的計算結果做爲條件,放進當前的輸入中。
適合在任意長度的序列中對上下文依賴性進行建模。可是有個問題,那就是咱們須要提早把輸入和輸出序列對齊,並且目前尚不清楚如何將 RNN 應用在不一樣長度有複雜非單一關係的序列中。爲了解決對齊問題,做者提出了一個理論上可行的辦法:使用兩個 RNN。 一個 RNN 把輸入映射爲一個固定長度的向量,另外一個 RNN 從這個向量中預測輸出序列。
double RNN.png

爲何說是理論可行的呢?做者 Sutskever 的博士論文 TRAINING RECURRENT NEURAL NETWORKS (訓練循環神經網絡)提出訓練 RNN 是很困難的。由於因爲 RNN 自身的網絡結構,其當前時刻的輸出須要考慮前面全部時刻的輸入,那麼在使用反向傳播訓練時,一旦輸入的序列很長,就極易出現梯度消失(Gradients Vanish)問題。爲了解決 RNN 難訓練問題,做者使用 LSTM(Long Short-Term Memory,長短時間記憶)網絡。
lstm0.png

上圖,是一個 LSTM 單元內部結構。LSTM 提出就是爲了解決 RNN 梯度消失問題,其創新性的加入了遺忘門,讓 LSTM 能夠選擇遺忘前面輸入無關序列,不用考慮所有輸入序列。通過3次嘗試,最終加入 LSTM 後,一個簡單的 Seq2Seq 模型就創建了。
seq2seq1.png

上圖,一個簡單的 Seq2Seq 模型包括3個部分,Encoder-LSTM,Decoder-LSTM,Context。輸入序列是ABC,Encoder-LSTM 將處理輸入序列並在最後一個神經元返回整個輸入序列的隱藏狀態(hidden state),也被稱爲上下文(Context,C)。而後 Decoder-LSTM 根據隱藏狀態,一步一步的預測目標序列的下一個字符。最終輸出序列wxyz。值得一提的是做者 Sutskever 根據其特定的任務具體設計特定的 Seq2Seq 模型。並對輸入序列做逆序處理,使模型能處理長句子,也提升了準確率。
seq2seq1.png

上圖,是做者 Sutskever 設計的真實模型,並引覺得傲一下三點。第一使用了兩個 LSTM ,一個用於編碼,一個用於解碼。這也是做者探索並論證的結果。第二使用了深層的 LSTM (4層),相比於淺層的網絡,每加一層模型困難程度就下降10% 。第三對輸入序列使用了逆序操做,提升了 LSTM 處理長序列能力。

3. 中英文翻譯

到了咱們動手的時刻了,理解了上面 Seq2Seq 模型,讓咱們搭建一個簡單的中英文翻譯模型。 網絡

3.1 數據集

咱們使用 manythings 網站的一箇中英文數據集,現已經上傳到 Mo 平臺了,點擊查看。該數據集格式爲英文+tab+中文。
app

image.png

3.2 處理數據

from keras.models import Model
from keras.layers import Input, LSTM, Dense
import numpy as np

batch_size = 64  # Batch size for training.
epochs = 100  # Number of epochs to train for.
latent_dim = 256  # Latent dimensionality of the encoding space.
num_samples = 10000  # Number of samples to train on.
# Path to the data txt file on disk.
data_path = 'cmn.txt'

# Vectorize the data.
input_texts = []
target_texts = []
input_characters = set()
target_characters = set()
with open(data_path, 'r', encoding='utf-8') as f:
    lines = f.read().split('\n')
for line in lines[: min(num_samples, len(lines) - 1)]:
    input_text, target_text = line.split('\t')
    # We use "tab" as the "start sequence" character
    # for the targets, and "\n" as "end sequence" character.
    target_text = '\t' + target_text + '\n'
    input_texts.append(input_text)
    target_texts.append(target_text)
    for char in input_text:
        if char not in input_characters:
            input_characters.add(char)
    for char in target_text:
        if char not in target_characters:
            target_characters.add(char)

input_characters = sorted(list(input_characters))
target_characters = sorted(list(target_characters))
num_encoder_tokens = len(input_characters)
num_decoder_tokens = len(target_characters)
max_encoder_seq_length = max([len(txt) for txt in input_texts])
max_decoder_seq_length = max([len(txt) for txt in target_texts])

print('Number of samples:', len(input_texts))
print('Number of unique input tokens:', num_encoder_tokens)
print('Number of unique output tokens:', num_decoder_tokens)
print('Max sequence length for inputs:', max_encoder_seq_length)
print('Max sequence length for outputs:', max_decoder_seq_length)
複製代碼

3.3 Encoder-LSTM

# mapping token to index, easily to vectors
input_token_index = dict([(char, i) for i, char in enumerate(input_characters)])
target_token_index = dict([(char, i) for i, char in enumerate(target_characters)])

# np.zeros(shape, dtype, order)
# shape is an tuple, in here 3D
encoder_input_data = np.zeros(
    (len(input_texts), max_encoder_seq_length, num_encoder_tokens),
    dtype='float32')
decoder_input_data = np.zeros(
    (len(input_texts), max_decoder_seq_length, num_decoder_tokens),
    dtype='float32')
decoder_target_data = np.zeros(
    (len(input_texts), max_decoder_seq_length, num_decoder_tokens),
    dtype='float32')

# input_texts contain all english sentences
# output_texts contain all chinese sentences
# zip('ABC','xyz') ==> Ax By Cz, looks like that
# the aim is: vectorilize text, 3D
for i, (input_text, target_text) in enumerate(zip(input_texts, target_texts)):
    for t, char in enumerate(input_text):
        # 3D vector only z-index has char its value equals 1.0
        encoder_input_data[i, t, input_token_index[char]] = 1.
    for t, char in enumerate(target_text):
        # decoder_target_data is ahead of decoder_input_data by one timestep
        decoder_input_data[i, t, target_token_index[char]] = 1.
        if t > 0:
            # decoder_target_data will be ahead by one timestep
            # and will not include the start character.
            # igone t=0 and start t=1, means 
            decoder_target_data[i, t - 1, target_token_index[char]] = 1.
複製代碼

3.4 Context(hidden state)

# Define an input sequence and process it.
# input prodocts keras tensor, to fit keras model!
# 1x73 vector 
# encoder_inputs is a 1x73 tensor!
encoder_inputs = Input(shape=(None, num_encoder_tokens))

# units=256, return the last state in addition to the output
encoder_lstm = LSTM((latent_dim), return_state=True)

# LSTM(tensor) return output, state-history, state-current
encoder_outputs, state_h, state_c = encoder_lstm(encoder_inputs)

# We discard `encoder_outputs` and only keep the states.
encoder_states = [state_h, state_c]
複製代碼

3.5 Decoder-LSTM

# Set up the decoder, using `encoder_states` as initial state.
decoder_inputs = Input(shape=(None, num_decoder_tokens))

# We set up our decoder to return full output sequences,
# and to return internal states as well. We don't use the
# return states in the training model, but we will use them in inference.
decoder_lstm = LSTM((latent_dim), return_sequences=True, return_state=True)

# obtain output
decoder_outputs, _, _ = decoder_lstm(decoder_inputs,initial_state=encoder_states)

# dense 2580x1 units full connented layer
decoder_dense = Dense(num_decoder_tokens, activation='softmax')

# why let decoder_outputs go through dense ?
decoder_outputs = decoder_dense(decoder_outputs)

# Define the model that will turn, groups layers into an object 
# with training and inference features
# `encoder_input_data` & `decoder_input_data` into `decoder_target_data`
# model(input, output)
model = Model([encoder_inputs, decoder_inputs], decoder_outputs)

# Run training
# compile -> configure model for training
model.compile(optimizer='rmsprop', loss='categorical_crossentropy')
# model optimizsm
model.fit([encoder_input_data, decoder_input_data], 
          decoder_target_data,
          batch_size=batch_size,
          epochs=epochs,
          validation_split=0.2)
# Save model
model.save('seq2seq.h5')
複製代碼

3.6 解碼序列

# Define sampling models
encoder_model = Model(encoder_inputs, encoder_states)
decoder_state_input_h = Input(shape=(latent_dim,))
decoder_state_input_c = Input(shape=(latent_dim,))
decoder_states_inputs = [decoder_state_input_h, decoder_state_input_c]
decoder_outputs, state_h, state_c = decoder_lstm(decoder_inputs, initial_state=decoder_states_inputs)
decoder_states = [state_h, state_c]
decoder_outputs = decoder_dense(decoder_outputs)
decoder_model = Model(
    [decoder_inputs] + decoder_states_inputs,
    [decoder_outputs] + decoder_states)

# Reverse-lookup token index to decode sequences back to
# something readable.
reverse_input_char_index = dict(
    (i, char) for char, i in input_token_index.items())
reverse_target_char_index = dict(
    (i, char) for char, i in target_token_index.items())

def decode_sequence(input_seq):
    # Encode the input as state vectors.
    states_value = encoder_model.predict(input_seq)

    # Generate empty target sequence of length 1.
    target_seq = np.zeros((1, 1, num_decoder_tokens))
    # Populate the first character of target sequence with the start character.
    target_seq[0, 0, target_token_index['\t']] = 1.
    # this target_seq you can treat as initial state

    # Sampling loop for a batch of sequences
    # (to simplify, here we assume a batch of size 1).
    stop_condition = False
    decoded_sentence = ''
    while not stop_condition:
        output_tokens, h, c = decoder_model.predict([target_seq] + states_value)

        # Sample a token
        # argmax: Returns the indices of the maximum values along an axis
        # just like find the most possible char
        sampled_token_index = np.argmax(output_tokens[0, -1, :])
        # find char using index
        sampled_char = reverse_target_char_index[sampled_token_index]
        # and append sentence
        decoded_sentence += sampled_char

        # Exit condition: either hit max length
        # or find stop character.
        if (sampled_char == '\n' or len(decoded_sentence) > max_decoder_seq_length):
            stop_condition = True

        # Update the target sequence (of length 1).
        # append then ?
        # creating another new target_seq
        # and this time assume sampled_token_index to 1.0
        target_seq = np.zeros((1, 1, num_decoder_tokens))
        target_seq[0, 0, sampled_token_index] = 1.

        # Update states
        # update states, frome the front parts
        states_value = [h, c]

    return decoded_sentence
複製代碼

3.7 預測

for seq_index in range(100,200):
    # Take one sequence (part of the training set)
    # for trying out decoding.
    input_seq = encoder_input_data[seq_index: seq_index + 1]
    decoded_sentence = decode_sequence(input_seq)
    print('Input sentence:', input_texts[seq_index])
    print('Decoded sentence:', decoded_sentence)
複製代碼

該項目已公開在 Mo 平臺上,Seq2Seq之中英文翻譯,建議使用GPU訓練。
介紹 Mo 平臺一個很是貼心實用的功能: API Doc,(在開發界面的右側欄,第二個)。
oop

推廣1.png


在 Mo 平臺寫代碼能夠很方便的實現多窗口顯示,只要拖動窗口的標題欄就可實現分欄。
推廣2.png

4. 總結與展望

提出經典的 Seq2Seq 模型是一件了不得的事情,該模型在機器翻譯和語音識別等領域中解決了不少重要問題和 NLP 沒法解決的難題。也是深度學習應用於 NLP 一件里程碑的事件。後續,又基於該模型提出了不少改進和優化,如 Attention 機制等。相信在不遠的將來,會有嶄新的重大發現,讓咱們拭目以待。
項目源碼地址(歡迎電腦端打開進行fork):momodel.cn/explore/5d3…
post

5. 引用

論文:Sequence to Sequence Learning with Neural Networks
博客:Understanding LSTM Networks
代碼:A ten-minute introduction to sequence-to-sequence learning in Keras學習

相關文章
相關標籤/搜索