1 Loading packages

library(tidyverse)
library(tidytext)
library(quanteda)
library(stringr)
library(jiebaR)
library(readtext)

2 Objectives

In this tutorial, I would like to talk about the relationship between a construction and words. Words may co-occur to form collocation patterns. When words co-occur with a particular morphosyntactic pattern, they would form collostruction patterns.

Here I would like to introduce a widely-applied method for research on the meanings of constructional schemas—Collostructional Aanalysis (Stefanowitsch & Gries, 2003). This is the major framework in corpus linguistics for the study of the relationship between words and constructions.

The idea behind collostructional analysis is simple: the meaning of a morphosyntactic construction can be determined very often by its co-occurring words.

In particular, words that are strongly associated (i.e., co-occurring) with the construction are referred to as collexemes of the construction.

Collostructional Analysis is an umbrella term, which covers several sub-analyses for constructional semantics:

• Collexeme Analysis (cf. Stefanowitsch & Gries (2003))
• Co-varying Collexeme Analysis (cf. Stefanowitsch & Gries (2005))
• (Multiple) Distinctive Collexeme Analysis (cf. Stefan T. Gries & Stefanowitsch (2004))

This tutorial will first focus on collexeme analysis, whose principles can be extended to the other analyses.

Also, I will demonstrate how we can conduct a collexeme analysis by using the R script written by Stefan Gries (Collostructional Analysis).

3 Corpus

As a demonstration, we use the Apple News Corpus (demo_data/applenews10000.tar.gz).

Our target is the Chinese construction X + 起來. The goal is to discover which words most frequently fill the X slot and what this reveals about the semantics of the construction.

That is, we want to identify the collexemes of X + 起來 and examine how coherent they are in meaning.

We begin by loading the corpus and creating a corpus object.

## Load Data
apple_df <-
  readtext("demo_data/applenews10000.tar.gz", encoding = "UTF-8") %>% ## loading data
  filter(!str_detect(text, "^\\s*$")) %>% ## removing empty docs
  mutate(doc_id = row_number()) ## create index

• Define how we would like to preprocess the data:

## Text normalization function
## Define a function
normalize_document <- function(texts) {
  texts %>%
    str_replace_all("\\p{C}", " ") %>%  ## remove control chars
    str_replace_all("[\\s|\\n]+", " ") %>%  ## replace whitespaces/linebreaks with one whitespace
    trimws
}

• An example:

## Example usage of `normalize_document()`
## before cleaning
apple_df$text[1]

## [1] "《蘋果體育》即日起進行虛擬賭盤擂台，每名受邀參賽者進行勝負預測，每周結算在周二公布，累積勝率前3高參賽者可繼續參賽，單周勝率最高者，將加封「蘋果波神」頭銜。註:賭盤賠率如有變動，以台灣運彩為主。\n資料來源：NBA官網http://www.nba.com\n\n金塊(客) 103：92 76人騎士(主) 88：82 快艇活塞(客) 92：75 公牛勇士(客) 108：82 灰熊熱火(客) 103：82 灰狼籃網(客) 90：82 公鹿溜馬(客) 111：100 馬刺國王(客) 112：102 爵士小牛(客) 108：106 拓荒者\n\n"

## after cleaning
normalize_document(apple_df$text[1])

## [1] "《蘋果體育》即日起進行虛擬賭盤擂台，每名受邀參賽者進行勝負預測，每周結算在周二公布，累積勝率前3高參賽者可繼續參賽，單周勝率最高者，將加封「蘋果波神」頭銜。註:賭盤賠率如有變動，以台灣運彩為主。 資料來源：NBA官網http://www.nba.com 金塊(客) 103：92 76人騎士(主) 88：82 快艇活塞(客) 92：75 公牛勇士(客) 108：82 灰熊熱火(客) 103：82 灰狼籃網(客) 90：82 公鹿溜馬(客) 111：100 馬刺國王(客) 112：102 爵士小牛(客) 108：106 拓荒者"

• Preprocess all texts in the corpus

## Apply cleaning to all docs
apple_df$text <- normalize_document(apple_df$text)

4 Word Segmentation

Next we segment the texts into words using jiebaR. The steps are:

• Initialize a jiebar segmentation model with worker().
• Define a function word_seg_text() to tokenize the text and, if needed, add POS tags.
• Apply this function to all texts and create a new column containing the segmented strings.

This new column will serve as the basis for later construction extraction.

## Initialize jiebar
segmenter <- worker(user = "demo_data/dict-ch-user.txt", ## User-defined Dictionary
                    bylines = FALSE, ## ignore lines
                    symbol = FALSE)  ## ignore symbols

## Define function
word_seg_text <- function(text, jiebar) {
  segment(as.character(text), jiebar) %>% ## word segmentation
    str_c(collapse = " ")                 ## concatenate word tokens into long strings
}

## Apply the function
apple_df <- apple_df %>%
  mutate(text_tag = map_chr(text, ~ word_seg_text(.x, segmenter)))

word_seg_text(apple_df$text[1], segmenter)

## [1] "蘋果 體育 即日起 進行 虛擬 賭盤 擂台 每名 受邀 參賽者 進行 勝負 預測 每周 結算 在 周二 公布 累積 勝率 前 3 高 參賽者 可 繼續 參賽 單周 勝率 最高者 將 加封 蘋果 波神 頭銜 註 賭盤 賠率 如有 變動 以 台灣 運彩 為主 資料 來源 NBA 官網 http www nba com 金塊 客 103 92 76 人 騎士 主 88 82 快艇 活塞 客 92 75 公牛 勇士 客 108 82 灰熊 熱火 客 103 82 灰狼 籃網 客 90 82 公鹿 溜 馬 客 111 100 馬刺 國王 客 112 102 爵士 小牛 客 108 106 拓荒者"

5 Extract Constructions

With segmented texts, we can now extract our target construction X + 起來 using a regular expression and unnest_tokens().

• Define the constructional pattern:

## Define regex
pattern_qilai <- "[^\\s]+\\s起來\\b"

• Extract the patterns from the corpus

## Extract patterns
apple_qilai <- apple_df %>%
  select(-text) %>% ## dont need original texts
  unnest_tokens(
    output = construction, ## name for new unit
    input = text_tag, ## name of old unit
    token = function(x) ## unnesting function
      str_match_all(x, pattern = pattern_qilai)
  )

• Check the results

## Print
apple_qilai

6 Exploratory/Descriptive Analysis

Usually before the statistical/quantitative analysis, we examine our data in an exploratory way.

apple_qilai %>%
  group_by(construction) %>%
  summarize(freq = n(),
            docfreq = n_distinct(doc_id)) -> apple_qilai_dist

apple_qilai_dist %>% 
  arrange(-freq) %>%
  top_n(50, freq)

7 Distributional Information Needed for Collostruction Analysis

To perform the collexeme analysis, which is essentially a statistical analysis of the association between words and a specific construction, we need to collect necessary distributional information of the words (X) and the construction (X + 起來).

In particular, to use Stefan Gries’ R script of Collostructional Analysis, we need three types of frequency information:

1. Joint frequencies of each word with the construction.
2. Word frequencies in the entire corpus.
3. Total corpus size (number of tokens).

## create word freq list
apple_word_freq <- apple_df %>%
  select(-text) %>% ## dont need original raw texts
  unnest_tokens( ## tokenization
    word,  ## new unit
    text_tag,  ## old unit
    token = function(x)  ## tokenization function
      str_split(x, "\\s+")
  ) %>%
  filter(nzchar(word)) %>% ## remove empty strings
  count(word, sort = T)

apple_word_freq %>%
  head(100)

## Joint frequency table
apple_qilai_freq <- apple_qilai %>%
  count(construction, sort = T) %>%  ## get joint frequencies
  tidyr::separate(col = "construction", ## restructure data frame
                  into = c("word", "construction"),
                  sep = "\\s") %>%
  ## identify the freq of X in X_起來
  mutate(word_freq = apple_word_freq$n[match(word, apple_word_freq$word)])

apple_qilai_freq

## prepare a tsv
## for coll analysis
apple_qilai_freq %>%
  select(word, word_freq, n) %>%
  write_tsv("qilai.tsv")

## corpus information
cat("Corpus Size: ", sum(apple_word_freq$n), "\n")

## Corpus Size:  2615074

cat("Construction Size: ", sum(apple_qilai_freq$n), "\n")

## Construction Size:  545

## save info in a text
sink("qilai_info.txt") ## start flushing outputs to the file not the terminal
cat("Corpus Size: ", sum(apple_word_freq$n), "\n")
cat("Construction Size: ", sum(apple_qilai_freq$n), "\n")
sink() ## end flushing

8 Running Collostructional Analysis

We are now ready to run the analysis by sourcing coll.analysis.r, available on Stefan Gries’ website.

######################################
##      Preloading Gries' Functions ##
######################################
source("https://www.stgries.info/teaching/groningen/coll.analysis.r")


## Run the interactive analysis
coll.analysis()

The script runs interactively, asking you to provide the required information.

For our example, the answers would be:

• Analysis to perform: 1
• Name of construction: QL
• Corpus size: 2615074
• Fisher-Yates: no
• Tab-delimited input data: qilai.tsv

If everything runs correctly, you will get a tab-delimited CSV output file in your working directory with the results.

9 Interpretations

The output is a tab-delimited file with statistics for each candidate collexeme.

A sample result file is available in the root directory 2025-10-XX XX-XX-XXXX.csv.

• We load the results and check the top collexemes:

## convert into CSV
collo_table<-read_tsv("2025-10-03 08-35-47.915105.csv")

## auto-print
collo_table %>%
  filter(RELATION =="attraction") %>%
  arrange(desc(LLR)) %>%
  head(50) %>%
  select(WORD, LLR, everything())

The output includes several useful COLLSTRENGTH measures, specifying the association between the collexemes and the slot of the construction (V + 起來).

• LLR/PMI
• Log Odds Ratio
• Delta P
• Kullback-Leibler Divergence (KLD)

With these statistics, we can identify the strongest collexemes according to different measures.

Here we focus on the top 10 collexemes by four metrics:

1. QL: joint frequency of word and construction
2. DELTAPC2W: delta P (construction → word)
3. DELTAPW2C: delta P (word → construction)
4. LLR: log-likelihood ratio

## from wide to long
collo_table %>%
  filter(RELATION == "attraction") %>%
  filter(QL >=5) %>%
  select(WORD, QL, 
         DELTAPC2W, 
         DELTAPW2C,
         LLR) %>%
  pivot_longer(cols=c("QL", 
                      "DELTAPC2W", 
                      "DELTAPW2C",
                      "LLR"),
               names_to = "METRIC",
               values_to = "STRENGTH") %>%
  mutate(METRIC = factor(METRIC, 
                         levels = c("QL", 
                                   "DELTAPC2W",
                                   "DELTAPW2C",
                                   "LLR"))) %>%
  group_by(METRIC) %>%
  top_n(10, STRENGTH) %>%
  ungroup %>%
  arrange(METRIC, desc(STRENGTH)) -> coll_table_long

## plot
coll_table_long %>%
    mutate(WORD = reorder_within(WORD, within= METRIC, by=STRENGTH)) %>%
    ggplot(aes(WORD, STRENGTH, fill=METRIC)) +
    geom_col(show.legend = F) +
    scale_fill_brewer(palette = "Dark2") +
    coord_flip() +
    facet_wrap(~METRIC,scales = "free") +
    tidytext::scale_x_reordered() + 
    labs(x = "Collexemes",
         y = "Strength",
         title = "Collexemes Ranked by Different Metrics")

The bar plots above show the top 10 collexemes based on four different metrics: QL, DELTAPC2W, DELTAPW2C, and LLR.

10 Distinctive Collexeme Analysis

We can also use Stefan Gries’ script coll.analysis.r to perform a Distinctive Collexeme Analysis.

To illustrate this method, let’s use data from the following study:

Huang, P. W., & Chen, A. C. H. (2022). Degree adverbs in spoken Mandarin: A behavioral profile corpus-based approach to language alternatives. Concentric: Studies in Linguistics, 48(2), 285–322.

In this work, the focus is on differences among four Mandarin degree adverbs: 很, 蠻/滿, 超, and 太.

As before, we run the interactive function coll.analysis().

For this example, the answers to the prompts would be:

• analysis to perform: 2 (Distinctive Collexeme Analysis)
  
• number of distinctive categories: 2 (or more if there are 3+ alternatives)
  
• tab-delimited input data: <demo_data/input-distcollexeme.tsv>

Once the script runs, it generates the results as a tab-delimited CSV file in the root directory.

An example output file is included in this project.

### Distinctive Collexeme Analysis: Results Exploration
dist_results <- read_tsv("2025-10-02 04-10-56.334203.csv")

dist_results

And our next step is to explore the semantic patterns from the output:

dist_results %>%
  group_by(LARGESTPREF) %>%
  top_n(10, SUMABSDEV) %>%
  select(COLLOCATE, SUMABSDEV, LARGESTPREF) %>%
  ungroup %>%
  arrange(LARGESTPREF) %>%
  mutate(COLLOCATE = reorder_within(COLLOCATE, within= LARGESTPREF, by=SUMABSDEV)) %>%
  ggplot(aes(COLLOCATE, SUMABSDEV, fill=SUMABSDEV)) +
    geom_col(show.legend = F) +
    scale_fill_distiller(palette="YlOrRd", direction=1)+
    coord_flip() +
    facet_wrap(~LARGESTPREF,scales = "free") +
    tidytext::scale_x_reordered() + 
    labs(x = "COLLOCATE",
         y = "STRENGTH(SUMABSDEV)",
         title = "DISTINCTIVE COLLEXEMES")