Lec10: LLMの利用 (Fall 2024)

Lecture10: LLMの利用

• “text”package
• Github: “text” package
• The text-package: An R-package for Analyzing and Visualizing Human Language Using Natural Language Processing and Deep Learning

パッケージのインストール

install.packages("text")
install.packages("reticulate")

Install a conda environment with text required python packages

library(text)
library(reticulate)

textrpp_install()
textrpp_initialize(save_profile = TRUE)

conda_list()
use_condaenv("r-reticulate")
use_condaenv(conda_list()$python[1])

textEmbed関数の実行

Subword Tokenization

embeddings <- textEmbed("word embeddings")

Completed layers output for texts (variable: 1/1, duration: 6.213706 secs).

Completed layers aggregation for word_type_embeddings. 

Completed layers aggregation (variable 1/1, duration: 0.073271 secs).

Completed layers aggregation (variable 1/1, duration: 0.087532 secs).

embeddings$tokens

$texts
$texts[[1]]

NANA

分析テキストURL

Category A: Retro Japan
  article1, article7, article8
Category B: Research/Survey
article2, article9, article12
Category C: Politics
  article3, article10, article11
Category D: Sports (Football)
  article4, article5, article6

記事URLの情報取得

article_urls <- readLines("Lec09_ArticleURL_info")

Term-Document Matrixの作成

contents <- lapply(article_urls, getArticleContent)
tmMtx <- getDocumentTermMTX(contents, term="lemma")

Processed document 10 of 12

dim(tmMtx)

[1] 1119   12

head(tmMtx)

article1抜粋データの単語埋め込み情報

単語埋め込み情報計算と結果保存

• 参考コード: the Text Tutorial

n <- length(contents)  # Use all elements in contents
article_tibble <- tibble(!!!set_names(contents, paste0("article", seq_along(contents))))
article_tibble

article.embeddings <- textEmbed(
  article_tibble,
  model = "bert-base-uncased", aggregation_from_layers_to_tokens = "concatenate", aggregation_from_tokens_to_texts = "mean", keep_token_embeddings = FALSE
)

saveRDS(article.embeddings, "article.embeddings.rds")

単語埋め込みデータの読み込み

article.embeddings <- readRDS("article.embeddings.rds")

データ整形（単語埋め込み情報の行結合）

article.embeddingsLst <- c()
for (current_article in article.embeddings$texts){
  colnames(current_article) <- paste0("Dim", 1:ncol(current_article))
  article.embeddingsLst <- rbind(article.embeddingsLst, current_article)
}

print(dim(article.embeddingsLst))

[1]  12 768

article.embeddingsLst[1:5, 1:5]

文書間の類似度（euclidean）

textSimilarityMTX.euclid<-textSimilarityMatrix(article.embeddingsLst, method = "euclidean")

View表示

View(round(textSimilarityMTX.euclid,2))

文書間の類似度（cosine）

textSimilarityMTX.cosine<-textSimilarityMatrix(article.embeddingsLst, method = "cosine")

結果表示

as.dist(round(textSimilarityMTX.cosine,2))

      1    2    3    4    5    6    7    8    9   10   11
2  0.84                                                  
3  0.83 0.81                                             
4  0.76 0.70 0.78                                        
5  0.71 0.73 0.72 0.76                                   
6  0.76 0.72 0.78 0.89 0.90                              
7  0.96 0.87 0.83 0.75 0.72 0.75                         
8  0.97 0.84 0.85 0.79 0.72 0.78 0.96                    
9  0.82 0.90 0.80 0.68 0.71 0.70 0.83 0.81               
10 0.82 0.83 0.86 0.74 0.72 0.75 0.82 0.82 0.85          
11 0.82 0.85 0.88 0.74 0.73 0.75 0.82 0.82 0.85 0.96     
12 0.78 0.92 0.76 0.67 0.69 0.68 0.80 0.78 0.85 0.78 0.79

View表示

View(round(textSimilarityMTX.cosine,2))

比較：出現頻度による文書間のcosine類似度

library(proxy)
library(cleanNLP)

cnlp_init_udpipe()
tmMtx <- getDocumentTermMTX(contents, term="lemma", punct=TRUE)

Processed document 10 of 12

結果表示

round(simil(t(tmMtx), method="cosine"),2)

      1    2    3    4    5    6    7    8    9   10   11
2  0.61                                                  
3  0.61 0.55                                             
4  0.60 0.55 0.51                                        
5  0.66 0.65 0.55 0.60                                   
6  0.72 0.58 0.55 0.66 0.73                              
7  0.84 0.67 0.62 0.56 0.69 0.70                         
8  0.84 0.65 0.69 0.56 0.66 0.69 0.84                    
9  0.65 0.71 0.62 0.54 0.60 0.62 0.67 0.70               
10 0.72 0.67 0.62 0.60 0.63 0.66 0.70 0.72 0.73          
11 0.72 0.67 0.63 0.59 0.69 0.68 0.72 0.71 0.69 0.81     
12 0.73 0.63 0.59 0.58 0.64 0.71 0.69 0.72 0.69 0.74 0.71

比較：出現頻度による文書間のcosine距離

dist.cos<-proxy::dist(t(tmMtx), method="cosine", diag=T)
round(dist.cos,2)

      1    2    3    4    5    6    7    8    9   10   11   12
1  0.00                                                       
2  0.39 0.00                                                  
3  0.39 0.45 0.00                                             
4  0.40 0.45 0.49 0.00                                        
5  0.34 0.35 0.45 0.40 0.00                                   
6  0.28 0.42 0.45 0.34 0.27 0.00                              
7  0.16 0.33 0.38 0.44 0.31 0.30 0.00                         
8  0.16 0.35 0.31 0.44 0.34 0.31 0.16 0.00                    
9  0.35 0.29 0.38 0.46 0.40 0.38 0.33 0.30 0.00               
10 0.28 0.33 0.38 0.40 0.37 0.34 0.30 0.28 0.27 0.00          
11 0.28 0.33 0.37 0.41 0.31 0.32 0.28 0.29 0.31 0.19 0.00     
12 0.27 0.37 0.41 0.42 0.36 0.29 0.31 0.28 0.31 0.26 0.29 0.00

階層的クラスター分析

• 参考サイト1: 階層的クラスター分析
• 参考サイト2: 距離行列

階層的クラスター分析

• Distance measure: cosine
• Agglomeration method: ward

出現単語頻度データ

hc.freq.cosine <- hclust(proxy::dist(t(tmMtx), method="cosine"), method = "ward.D2")
plot(hc.freq.cosine)

単語埋め込みデータ

distMtx <- as.dist(textDistanceMatrix(article.embeddingsLst, center = TRUE, method="cosine"))
hc.embeddings.cosine <- hclust(distMtx, method = "ward.D2")
plot(hc.embeddings.cosine)