I’ve started by analysing a quick summary of the data.
# Define file paths
blogs_path <- "/home/rstudio/en_US.blogs.txt"
news_path <- "/home/rstudio/en_US.news.txt"
twitter_path <- "/home/rstudio/en_US.twitter.txt"
# Read files
blogs <- readLines(blogs_path, warn = FALSE, encoding = "UTF-8")
news <- readLines(news_path, warn = FALSE, encoding = "UTF-8")
twitter <- readLines(twitter_path, warn = FALSE, encoding = "UTF-8")
# Function to calculate word count per dataset
word_count <- function(text_data) {
sum(sapply(strsplit(text_data, "\\s+"), length))
}
# Compute summary statistics
data_summary <- data.frame(
Dataset = c("Blogs", "News", "Twitter"),
Lines = c(length(blogs), length(news), length(twitter)),
Words = c(word_count(blogs), word_count(news), word_count(twitter)),
Max_Line_Length = c(max(nchar(blogs)), max(nchar(news)), max(nchar(twitter))) # Longest line in each dataset
)
# Print summary
print(data_summary)
## Dataset Lines Words Max_Line_Length
## 1 Blogs 899288 37334131 40833
## 2 News 1010242 34372530 11384
## 3 Twitter 2360148 30373543 140
To prepare text data for an n-gram model, we need to remove:
Stopwords: Common words like “the,” “is,” and “and” that don’t add meaning.
Punctuation: To ensure clean tokenization.
Numbers: Unless they contribute meaning.
Whitespace: Extra spaces or line breaks.
Special Characters & Non-ASCII Text: Emojis, symbols, and foreign characters.
Working with large text files is computationally expensive. Key strategies:
Sampling: Only use a subset of data (readLines(n = 50000)).
Parallel Processing: Use parallel::mclapply() to speed up n-gram extraction.
Sparse Matrices: Use slam or Matrix package instead of dense matrices.
Efficient Data Storage: Save intermediate results as .rds files.
library(ggplot2)
library(data.table)
library(stringi)
library(tokenizers)
# Load sample (limit to 50k for memory)
set.seed(123)
file_path <- "/home/rstudio/en_US.twitter.txt"
lines <- readLines(file_path, warn = FALSE, n = 50000)
# Clean text efficiently
cleaned <- tolower(lines)
cleaned <- stri_replace_all_regex(cleaned, "[^a-z\\s]", " ")
cleaned <- stri_replace_all_regex(cleaned, "\\s+", " ") # remove extra spaces
cleaned <- trimws(cleaned)
Historgrams
Utilizing the ‘twitter’ text file I have created a historgram to examine frequency of common words.
# Load necessary libraries
library(ggplot2)
library(dplyr)
##
## Attaching package: 'dplyr'
## The following objects are masked from 'package:data.table':
##
## between, first, last
## The following objects are masked from 'package:stats':
##
## filter, lag
## The following objects are masked from 'package:base':
##
## intersect, setdiff, setequal, union
library(tm)
## Loading required package: NLP
##
## Attaching package: 'NLP'
## The following object is masked from 'package:ggplot2':
##
## annotate
library(Matrix) # For sparse matrices
# Read dataset (Example: Twitter)
file_path <- "/home/rstudio/en_US.twitter.txt"
twitter <- readLines(file_path, warn = FALSE)
# Sample a smaller subset (e.g., 10,000 tweets) to reduce memory consumption
twitter_sample <- sample(twitter, 10000)
# Convert to Corpus
twitter_corpus <- Corpus(VectorSource(twitter_sample))
# Clean text
twitter_corpus <- tm_map(twitter_corpus, content_transformer(tolower)) # Convert to lowercase
## Warning in tm_map.SimpleCorpus(twitter_corpus, content_transformer(tolower)):
## transformation drops documents
twitter_corpus <- tm_map(twitter_corpus, removePunctuation) # Remove punctuation
## Warning in tm_map.SimpleCorpus(twitter_corpus, removePunctuation):
## transformation drops documents
twitter_corpus <- tm_map(twitter_corpus, removeNumbers) # Remove numbers
## Warning in tm_map.SimpleCorpus(twitter_corpus, removeNumbers): transformation
## drops documents
twitter_corpus <- tm_map(twitter_corpus, removeWords, stopwords("en")) # Remove stopwords
## Warning in tm_map.SimpleCorpus(twitter_corpus, removeWords, stopwords("en")):
## transformation drops documents
# Remove empty documents after transformation
twitter_corpus <- twitter_corpus[sapply(twitter_corpus, function(x) nchar(x) > 0)]
# Create a Document-Term Matrix with sparse representation
dtm <- TermDocumentMatrix(twitter_corpus, control = list(weighting = weightTfIdf))
## Warning in TermDocumentMatrix.SimpleCorpus(twitter_corpus, control =
## list(weighting = weightTfIdf)): custom functions are ignored
## Warning in weighting(x): empty document(s): 397 457 853 1220 1509 2001 2051 2150
## 2442 2737 2767 2841 3111 3834 4340 4501 4525 4882 4939 4953 5011 5117 5412 5834
## 5911 5939 7383 7694 7791 7926 8232 8262 8554 8624 8917 9402 9687 9691 9848 9942
# Convert to sparse matrix
m <- as.matrix(dtm)
word_freq <- sort(rowSums(m), decreasing = TRUE)
# Convert to Data Frame for plotting
word_freq_df <- data.frame(word = names(word_freq), freq = word_freq)
# Plot histogram of top 30 word frequencies
ggplot(word_freq_df[1:30, ], aes(x = reorder(word, -freq), y = freq)) +
geom_bar(stat = "identity", fill = "steelblue") +
theme_minimal() +
labs(title = "Top 30 Most Frequent Words in Twitter Dataset", x = "Word", y = "Frequency") +
theme(axis.text.x = element_text(angle = 45, hjust = 1))
Bi & Trigram analysis
To exapand on this analysis, also looking at the ‘Twitter’ data set, I have examined Bi & Trigram frequencies.
# Sample 10,000 lines for n-gram extraction
set.seed(456)
sampled <- sample(cleaned, size = min(10000, length(cleaned)))
# Tokenize bigrams and trigrams using `tokenizers`
bigrams <- unlist(tokenize_ngrams(sampled, n = 2))
trigrams <- unlist(tokenize_ngrams(sampled, n = 3))
# Frequency tables using data.table
bigram_dt <- data.table(table(bigrams))[order(-N)][1:20]
trigram_dt <- data.table(table(trigrams))[order(-N)][1:20]
# Plot bigrams
ggplot(bigram_dt, aes(x = reorder(bigrams, N), y = N)) +
geom_bar(stat = "identity", fill = "tomato") +
coord_flip() +
theme_minimal() +
labs(title = "Top 20 Bigrams", x = "Bigram", y = "Frequency")
# Plot trigrams
ggplot(trigram_dt, aes(x = reorder(trigrams, N), y = N)) +
geom_bar(stat = "identity", fill = "seagreen") +
coord_flip() +
theme_minimal() +
labs(title = "Top 20 Trigrams", x = "Trigram", y = "Frequency")
Predicting unseen words is a challenge. We can use smoothing techniques:
Laplace Smoothing: Adds a small probability to unseen n-grams.
Kneser-Ney Smoothing: Adjusts based on lower-order n-grams.
Simple Laplace Smoothing Implementation
Model Selection For word prediction, common models include:
Markov Chains: Simple probabilistic transition between words.
Backoff Models: If a higher-order n-gram is missing, fall back to lower n-grams.
Interpolation: Combines different n-gram probabilities.
The Shiny App will:
Take user input
Predict next words using an n-gram model
Display results interactively
I look forward to presenting the next stage of the assignment where I will dive deeper into modelling.