使用更多2个输入功能训练神经网络

Question

I already referred to the keras guide on using multiple inputs . However, I am still stumped as I am new to RNNs and CNNs. I am working with keras to train a neural network classifier. In my csv file, I have 3 features.

• Sentence
• Probability
• Target

Each sentence is a sentence with exactly 5 words and there are 1860 such sentences. The probability is a float value in the range of [0,1] and the target is the field that needs to be predicted (0 or 1).

I first randomly initiate the sentences with embeddings as shown below.

import string
import nltk
from nltk.tokenize import word_tokenize
from nltk.corpus import stopwords
import numpy as np
import gensim
import pandas as pd
import os
from tensorflow.python.keras.preprocessing.text import Tokenizer
from tensorflow.python.keras.preprocessing.sequence import pad_sequences
from gensim.models import Word2Vec, KeyedVectors
from keras.models import Sequential
from keras.layers import Dense, Embedding, LSTM, GRU
from keras.layers.embeddings import Embedding
from keras.initializers import Constant
from sklearn.metrics import precision_recall_fscore_support
from sklearn.model_selection import train_test_split
from termcolor import colored
from keras.utils import to_categorical
import tensorflow as tf

import warnings
warnings.filterwarnings("ignore")

nltk.download('stopwords')
# one hot encode

seed = 42
np.random.seed(seed)
tf.set_random_seed(seed)


df = pd.DataFrame()
df = pd.read_csv('../../data/sentence_with_stv.csv')
sentence_lines = list()
lines = df['sentence'].values.tolist()
stv = df['stv'].values.tolist()

for line in lines:
    tokens = word_tokenize(line)
    tokens = [w.lower() for w in tokens]
    table = str.maketrans('','',string.punctuation)
    stripped = [w.translate(table) for w in tokens]
    words = [word for word in stripped if word.isalpha()]
    stop_words = set(stopwords.words('english'))
    words = [w for w in words if not w in stop_words]
    sentence_lines.append(words)

print('Number of lines', len(sentence_lines)))
EMBEDDING_DIM = 200

#Vectorize the text samples into a S2 integer tensor
tokenizer_obj = Tokenizer()
tokenizer_obj.fit_on_texts(sentence_lines)
sequences = tokenizer_obj.texts_to_sequences(sentence_lines)

print(colored(sequences,'green'))

This gives me an output such as,

Number of lines: 1860
[[2, 77, 20, 17, 81], 
 [12, 21, 17, 82], 
 [2, 83, 20, 17, 82], 
 [2, 20, 17, 43], 
 [12, 21, 17, 81], 
 ...

Now, I need to append the probability to each of of these lines such that the new sequence should resemble the following.

[[2, 77, 20, 17, 81, 0.456736827], 
 [12, 21, 17, 82, 0.765142873], 
 [2, 83, 20, 17, 82, 0.335627635], 
 [2, 20, 17, 43, 0.5453652], 
 [12, 21, 17, 81, 0.446739202],
 ...

I tried taking each of the sequence's row and appending the probability as,

sequence[x] = np.append(sequence[x], probability[x], axis=1)

where, the probability is an array of the same size, 1860, consisting of only the probability values. After doing this for all the rows, I print a single row to check if the value got appended. But, I get an output as shown below.

[2.     77.     20.     17.     81.     0.456736827]

Any suggestions in this regard will be much appreciated.

Answer 1

You should use word indices as an input to a neural net. Each index corresponds to different words and the indexing does not have the semantics of numbers. (10 is twice as much as 5, nothing like this holds for categorical variables.) If append a float to the indices in numpy, the indices get converted into floats.

The correct solution is using an embedding layer for the word inputs. Embedding layer assigns a trainable vector to every item in your dictionary. Embeddings are typically followed by an RNN or a CNN to get o single vector, or you can just concatenate the embeddings.

Anyway, you cannot easily combine the inputs in numpy, you need to do it TensorFlow. First, embed the words and once you have a tensor of continuous values, only then you can append you numerical features.