R text mining how to segment document into phrases not terms - r

Question

When do text mining using R, after reprocessing text data, we need create a document-term matrix for further exploring. But in similar with Chinese, English also have some certain phases, such as "semantic distance", "machine learning", if you segment them into word, it have totally different meanings, I want to know how to segment document into phases but not word(term).

Answer 1

You can do this in R using the quanteda package, which can detect multi-word expressions as statistical collocates, which would be the multi-word expressions that you are probably referring to in English. To remove the collocations containing stop words, you would first tokenise the text, then remove the stop words leaving a "pad" in place to prevent false adjacencies in the results (two words that were not actually adjacent before the removal of stop words between them).
require(quanteda)
pres_tokens <-
tokens(data_corpus_inaugural) %>%
tokens_remove("\\p{P}", padding = TRUE, valuetype = "regex") %>%
tokens_remove(stopwords("english"), padding = TRUE)
pres_collocations <- textstat_collocations(pres_tokens, size = 2)
head(pres_collocations)
# collocation count count_nested length lambda z
# 1 united states 157 0 2 7.893307 41.19459
# 2 let us 97 0 2 6.291128 36.15520
# 3 fellow citizens 78 0 2 7.963336 32.93813
# 4 american people 40 0 2 4.426552 23.45052
# 5 years ago 26 0 2 7.896626 23.26935
# 6 federal government 32 0 2 5.312702 21.80328
# convert the corpus collocations into single tokens, for top 1,500 collocations
pres_compounded_tokens <- tokens_compound(pres_tokens, pres_collocations[1:1500])
tokens_select(pres_compounded_tokens[2], "*_*")
# tokens from 1 document.
# 1793-Washington :
# [1] "called_upon" "shall_endeavor" "high_sense" "official_act"
Using this "compounded" token set, we can now turn this into a document-feature matrix where the features consist of a mixture of original terms (those not found in a collocation) and the collocations. As can be seen below, "united" occurs alone and as part of the collocation "united_states".
pres_dfm <- dfm(pres_compounded_tokens)
head(pres_dfm[1:5, grep("united|states", featnames(pres_dfm))])
# Document-feature matrix of: 5 documents, 10 features (86% sparse).
# 5 x 10 sparse Matrix of class "dfm"
# features
# docs united states statesmen statesmanship reunited unitedly devastates statesman confederated_states united_action
# 1789-Washington 4 2 0 0 0 0 0 0 0 0
# 1793-Washington 1 0 0 0 0 0 0 0 0 0
# 1797-Adams 3 9 0 0 0 0 0 0 0 0
# 1801-Jefferson 0 0 0 0 0 0 0 0 0 0
# 1805-Jefferson 1 4 0 0 0 0 0 0 0 0
If you want a more brute-force approach, it's possible simply to create a document-by-bigram matrix this way:
# just form all bigrams
head(dfm(data_inaugural_corpus, ngrams = 2))
## Document-feature matrix of: 57 documents, 63,866 features.
## (showing first 6 documents and first 6 features)
## features
## docs fellow-citizens_of of_the the_senate senate_and and_of the_house
## 1789-Washington 1 20 1 1 2 2
## 1797-Adams 0 29 0 0 2 0
## 1793-Washington 0 4 0 0 1 0
## 1801-Jefferson 0 28 0 0 3 0
## 1805-Jefferson 0 17 0 0 1 0
## 1809-Madison 0 20 0 0 2 0