fastrtext︱R語言使用facebook的fasttext快速文字分類演算法
FastText是Facebook開發的一款快速文字分類器,提供簡單而高效的文字分類和表徵學習的方法,不過這個專案其實是有兩部分組成的。理論介紹可見部落格:NLP︱高階詞向量表達(二)——FastText(簡述、學習筆記)
本輪新更新的fastrtext,同樣繼承了兩個功能:訓練詞向量 + 文字分類模型訓練
來源:
相關文件地址:
相關部落格:
.
一、安裝
1.安裝
# From Cran
install.packages("fastrtext")
# From Github
# install.packages("devtools")
devtools::install_github("pommedeterresautee/fastrtext")
.
2.主函式介紹
The following arguments are mandatory:
-input training file path
-output output file path
The following arguments are optional:
-verbose verbosity level [2]
The following arguments for the dictionary are optional:
-minCount minimal number of word occurences [5 ]
-minCountLabel minimal number of label occurences [0]
-wordNgrams max length of word ngram [1]
-bucket number of buckets [2000000]
-minn min length of char ngram [3]
-maxn max length of char ngram [6]
-t sampling threshold [0 .0001]
-label labels prefix [__label__]
The following arguments for training are optional:
-lr learning rate [0.05]
-lrUpdateRate change the rate of updates for the learning rate [100]
-dim size of word vectors [100]
-ws size of the context window [5]
-epoch number of epochs [5]
-neg number of negatives sampled [5]
-loss loss function {ns, hs, softmax} [ns]
-thread number of threads [12]
-pretrainedVectors pretrained word vectors for supervised learning []
-saveOutput whether output params should be saved [0]
The following arguments for quantization are optional:
-cutoff number of words and ngrams to retain [0]
-retrain finetune embeddings if a cutoff is applied [0]
-qnorm quantizing the norm separately [0]
-qout quantizing the classifier [0]
-dsub size of each sub-vector [2]
也就是execute()時候,可以輸入的函式是啥。
-dim,向量長度,預設100維;
-wordNgrams,詞型別,一般可以選擇2,二元組
-verbose,輸出資訊的詳細程度,0-2,不同層次的詳細程度(0代表啥也不顯示)。
-lr:學習速率[0.1]
-lrUpdateRate:更改學習率的更新速率[100]
-dim :字向量大小[100]
-ws:上下文視窗的大小[5]
-epoch:迴圈數[5]
-neg:抽樣數量[5]
-loss:損失函式 {ns,hs,softmax} [ns]
-thread:執行緒數[12]
-pretrainedVectors:用於監督學習的預培訓字向量
-saveOutput:輸出引數是否應該儲存[0]
.
二、官方案例一 —— 文字分類模型訓練
2.1 載入資料並訓練
library(fastrtext)
data("train_sentences")
data("test_sentences")
# prepare data
tmp_file_model <- tempfile()
train_labels <- paste0("__label__", train_sentences[,"class.text"])
train_texts <- tolower(train_sentences[,"text"])
train_to_write <- paste(train_labels, train_texts)
train_tmp_file_txt <- tempfile()
writeLines(text = train_to_write, con = train_tmp_file_txt)
test_labels <- paste0("__label__", test_sentences[,"class.text"])
test_texts <- tolower(test_sentences[,"text"])
test_to_write <- paste(test_labels, test_texts)
# learn model
execute(commands = c("supervised", "-input", train_tmp_file_txt, "-output", tmp_file_model, "-dim", 20, "-lr", 1, "-epoch", 20, "-wordNgrams", 2, "-verbose", 1))
其中可以看到與之前熟知的機器學習相關模型不同,其模型執行是通過execute來得到,並儲存。
其中:
來看看輸入資料長啥樣子:
資料是char格式的,之前__label__XXX 是該文字的標籤,然後空格接上文字內容。
執行結果:
##
Read 0M words
## Number of words: 5060
## Number of labels: 15
##
Progress: 100.0% words/sec/thread: 1457520 lr: 0.000000 loss: 0.300770 eta: 0h0m
.
2.2 驗證集+執行模型
# load model
model <- load_model(tmp_file_model)
# prediction are returned as a list with words and probabilities
predictions <- predict(model, sentences = test_to_write)
load_model模型檔案位置,test_to_write是驗證文字,長這樣(其實跟訓練集長一樣):
顯示:
print(head(predictions, 5))
## [[1]]
## __label__OWNX
## 0.9980469
##
## [[2]]
## __label__MISC
## 0.9863281
##
## [[3]]
## __label__MISC
## 0.9921875
##
## [[4]]
## __label__OWNX
## 0.9082031
##
## [[5]]
## __label__AIMX
## 0.984375
.
2.3 模型驗證
計算準確率
# Compute accuracy
mean(sapply(predictions, names) == test_labels)
計算海明距離
# because there is only one category by observation, hamming loss will be the same
get_hamming_loss(as.list(test_labels), predictions)
## [1] 0.8316667
.
2.4 一些小函式
檢視監督模型的label有哪些,get_labels函式。
如果已經訓練好模型,放了一段時間,又不知道里面有哪些標籤,可以這麼找一下。
model <- load_model(model_test_path)
print(head(get_labels(model), 5))
#> [1] "__label__MISC" "__label__OWNX" "__label__AIMX" "__label__CONT"
#> [5] "__label__BASE"
檢視模型的引數都用了啥get_parameters:
model <- load_model(model_test_path)
print(head(get_parameters(model), 5))
#> $learning_rate
#> [1] 0.05
#>
#> $learning_rate_update
#> [1] 100
#>
#> $dim
#> [1] 20
#>
#> $context_window_size
#> [1] 5
#>
#> $epoch
#> [1] 20
#>
.
三、官方案例二 —— 計算詞向量
3.1 載入資料 + 訓練
library(fastrtext)
data("train_sentences")
data("test_sentences")
texts <- tolower(train_sentences[,"text"])
tmp_file_txt <- tempfile()
tmp_file_model <- tempfile()
writeLines(text = texts, con = tmp_file_txt)
execute(commands = c("skipgram", "-input", tmp_file_txt, "-output", tmp_file_model, "-verbose", 1))
commands 裡面的引數是:”skipgram”,也就是計算詞向量,跟word2vec一致。
輸入的文字內容,不用帶標籤資訊:
.
3.2 詞向量
model <- load_model(tmp_file_model)
載入詞向量的檔案,載入的是bin檔案
# test word extraction
dict <- get_dictionary(model)
print(head(dict, 5))
## [1] "the" "</s>" "of" "to" "and"
dict 就是詞向量的字典,
# print vector
print(get_word_vectors(model, c("time", "timing")))
顯示一下,詞向量的維度。
.
3.3 計算詞向量距離——get_word_distance
# test word distance
get_word_distance(model, "time", "timing")
## [,1]
## [1,] 0.02767485
.
3.4 找出最近鄰詞——get_nn
get_nn引數只有三個,最後數字代表選擇前多少個近義詞。
library(fastrtext)
model_test_path <- system.file("extdata", "model_unsupervised_test.bin", package = "fastrtext")
model <- load_model(model_test_path)
get_nn(model, "time", 10)
#> times size indicate access success allowing feelings
#> 0.6120564 0.5041215 0.4941387 0.4777856 0.4719051 0.4696053 0.4652924
#> dictator amino accuracies
#> 0.4595046 0.4582702 0.4535145
.
3.5 詞的類比——get_analogies
library(fastrtext)
model_test_path <- system.file("extdata", "model_unsupervised_test.bin", package = "fastrtext")
model <- load_model(model_test_path)
get_analogies(model, "experience", "experiences", "result")
#> results
#> 0.726607
類比關係式:
get_analogies(model, w1, w2, w3, k = 1)
w1 - w2 + w3
也即是:
experience - experiences + result