knitr::opts_chunk$set( message = FALSE, warning = FALSE, cache = TRUE )
library(LimpiaR) library(tibble) library(dplyr) library(stringr)
We want to take a document (or later a set of documents) like:
Who was the first person on the moon?
And extract for each token (where in this case token = word) in the document its part of speech and its lemma. So that for each document we have a table like:
model <- limpiar_pos_import_model(language = "english") data <- tibble(text = c("Who was the first man on the moon?"), id = 1) data <- limpiar_pos_annotate( data, text, id, model, update_progress = 0 ) %>% select(token, lemma, pos_tag) knitr::kable(data)
In this vignette we'll be looking at how to use LimpiaR via {udpipe} to achieve this. For a primer on parts of speech see Parts of Speech Wikipedia.
pos_table <- tibble( `Part of Speech` = c("Adjective", "Adposition", "Adverb", "Auxiliary", "Coordinating Conjunction", "Determiner", "Interjection", "Noun", "Numeral", "Particle", "Pronoun", "Proper Noun", "Punctuation", "Subordinating Conjunction", "Symbol", "Verb", "Other"), `UPOS Tag` = c("ADJ", "ADP", "ADV", "AUX", "CCONJ", "DET", "INTJ", "NOUN", "NUM", "PART", "PRON", "PROPN", "PUNCT", "SCONJ", "SYM", "VERB", "X") , Definition = c( 'Usually describes or modifies a noun. It provides information about an object\'s size, color, shape, etc. Example: "big", "blue".', 'Class of words that includes prepositions and postpositions, used to express spatial or temporal relations or mark various semantic roles. Example: "in", "on", "before", "after".', 'Modifies a verb, an adjective, another adverb, or even an entire sentence. It provides information about manner, time, place, frequency, degree, etc. Example: "quickly", "very".', 'Special verb used to add functional or grammatical meaning to the clause in which it appears. It accompanies the main verb and forms different aspects, voices, or moods. Example: "is", "have", "will".', 'Joins words, phrases, or clauses of similar grammatical status and syntactic importance. Example: "and", "but", "or".', 'Introduces a noun and provides context in terms of definiteness, where it belongs in a sequence, quantity, or ownership. Example: "a", "the", "this", "those".', 'An abrupt remark, made apart from the main sentence structure, often used to express strong emotion or surprise. Example: "Oh!", "Wow!", "Ugh!".', 'Represents a person, place, thing, or idea. Nouns can be subjects, objects, or complement in a sentence. Example: "dog", "city", "happiness".', 'A word, symbol, or group of words representing a number. Example: "one", "first", "100".', 'A function word that does not fit into the other categories but is used to express grammatical relationships with other words or to specify the attitude of the speaker. Example: "up" in "stand up".', 'Replaces a noun, often used to avoid repetition. Pronouns can do most things that nouns can do. Example: "he", "they", "who".', 'Names a specific person, place, thing, or idea and is typically capitalized. Example: "Elizabeth", "London", "Microsoft".', 'Symbols that organize writing into clauses, phrases, and sentences, clarify meaning, and indicate pauses. Example: ".", ",", "!".', 'Introduces a subordinate clause and indicates the relationship between the subordinate clause and the rest of the sentence. Example: "because", "although", "when".', 'A mark or character used as a conventional representation of an object, function, or process. Example: "$", "%", "&".', 'Expresses an action, occurrence, or state of being. Verbs are central to a clause. Example: "run", "be", "have".', 'A category used for words or tokens that do not fit into the above categories. This is less common and often used in tagging to mark anomalies or unclassifiable items.' ) ) knitr::kable(pos_table, caption = "Source: universaldependencies.org/u/pos")
The main motivation for processing and extracting parts of speech is to home in on the particular aspects of language which are most informative for answering a given research question. For example, if you want to know what people think about something, you might focus on adjectives and nouns because adjectives describe things, and nouns are things. If you want to know how people are using something, you might focus more on verbs/phrasal verbs as they are 'doing words'.
library(LimpiaR) library(tibble) library(dplyr) library(stringr)
LimpiaR's PoS workflow leans on pre-existing functionality of the {udpipe} package to allow users to import a pre-trained model for parts of speech analysis whilst enabling advanced users to perform 'dependency parsing'.
What is dependency parsing?
"a technique which provides to each word in a sentence the link to another word in the sentence, which is called its syntactical head. This link between each 2 words furthermore has a certain type of relationship giving you further details about it"
see here for more info.
First, we import the model we want to use. UDPipe pre-trained models are built upon Universal Dependencies treebanks and are made available for more than 65 languages based on 101 treebanks. For demonstrative purposes we'll use language = "english"
.
model <- limpiar_pos_import_model(language = "english")
Now that we have our model imported into our session, let's get some data to tag, tokenise, lemmatise, and extract parts of speech. We need a data frame with both a text variable and an ID column which uniquely identifies each text.
( data <-tibble(text = tolower(stringr::sentences[1:100]), universal_message_id = paste0("TWITTER", 1:100)) )
So now we have some data and we also have our desired model loaded into session, the function we need next is limpiar_pos_annotate()
.
Thelimpiar_pos_annotate
function converts the text of each document into its tokens, lemmas, POS tags, and dependency relationships*. The input data frame should have one document per row, the return data frame will have one row per token. This means the return data frame will have many more rows than the input data frame!
dependency_parse default value
*the dependency_parse
argument is set to FALSE
by default as this step can be costly in time and compute if performed on large data sets. Most users will get by just fine without parsing dependencies, but if you are sure you need them set this argument to TRUE
.
limpiar_pos_annotate
can be sped up by using in_parallel = TRUE
, this argument makes the function process multiple rows of the data frame in parallel.
By default the function will print its progress every 100 rows, this way you know that it's still working and roughly how long it should take to finish running. If you don't want the progress updates to print in your console set update_progress = 0
.
# annotate texts and perform dependency parsing annotations <- limpiar_pos_annotate(data = data, text_var = text, id_var = universal_message_id, pos_model = model, dependency_parse = TRUE, in_parallel = FALSE, update_progress = 25)
Now that we have our texts tokenized, dependencies parsed, and POS annotations are complete, let's take a look at the output. After annotating we have r nrow(annotations)
rows, which is an r nrow(annotations)/nrow(data)
x increase!
annotations %>% select(-c(sentence, feats, xpos, doc_id)) %>% relocate(universal_message_id)
Output has some additional columns, we will pay closest attention to paragraph_id, sentence_id, token, lemma, pos_tag, dependency_tag.
What are these new columns?
paragraph_id
and sentence_id
tell us which of the document's paragraphs, and which sentence the token is in. Because our data set contains only single sentences, all of these values are 1.
token
is the word the rest of the columns refer to.
lemma
is the lemmatised version of the token, an example of lemmatisation is 'served' being converted to 'serve'.
pos_tag
displays PoS labels such as; Noun(NOUN), Proper Noun(PROPN), Pronoun(PRON), Verb(VERB), Adjective(ADJ) etc. For more on the exact tags and what they mean see PoS Tags or the UPOS tags table at the top of this document.
dependency_tag
displays the dependency relation for each token. Dependency relations are significantly more complicated than parts of speech. see Dependency Relations for more details.
head_token_id
, dependency_tag
and feats
will only be present if we select dependency_parse = TRUE
, they tell us the id of the token (inside the document, paragraph, sentence) which the token of that row has the dependency relation to, the dependency relation and additional features of the relationship
Finally, the column you inserted as id_var =
. This variable will help you to join the results back to your original data frame.
Now that we have adequate context, let's look at what we can do with our output.
We can combine commonly-used {dplyr} functions for data wrangling and summarisation to filter for certain parts of speech and count their occurrences.
Given the size of our data set there are limited inferences we can make from the counts, but if we just want to know which adjectives are seen most frequently:
# Count adjectives annotations %>% filter(pos_tag %in% c("ADJ")) %>% count(token, pos_tag, sort = TRUE)
Alternatively we could do the same thing for nouns, but this time count the lemma (in case of plurals).
#Count the lemma of each noun annotations %>% filter(pos_tag %in% c("NOUN")) %>% # this time selecting nouns count(lemma, sort = TRUE)
Are there any adjective -> noun pairs seen more than once? This task gets slightly more involved as we need to look at the value of each token and the value in the next row, so first we need dplyr::filter
and dplyr::lag
+ dplyr::lead
, and to group by the doc_id, paragraph_id and sentence_id.
adj_to_noun <- annotations %>% filter( pos_tag == "ADJ" & lead(pos_tag) == "NOUN" | pos_tag == "NOUN" & lag(pos_tag) == "ADJ", .by = c(doc_id, paragraph_id, sentence_id))
Then we'll need to perform some more {dplyr} magic to count the ADJ-NOUN collocations. The data is now set up so that row 1+2, 3+4, 4+5 and so on and so forth are ADJ-NOUN collocations, we can use this information to extract the pairs with a combination of floor division, exploiting how cumsum meshes with logicals in R, and a 'grouped, summarise paste with a flourish of collapse' trick.
adj_to_noun %>% mutate(collocation_id = row_number() %% 2, .before = 1) %>% mutate(collocation_id = cumsum(collocation_id == 1)) %>% summarise(collocation = paste0(token, collapse = " "), .by = collocation_id) %>% count(collocation, sort = TRUE)
The only ADJ-NOUN collocation seen more than once was 'third week'. In a larger data set we would expect to find more and higher-frequency ADJ-NOUN collocations.
However, when manipulating the data this way, the onus is on the user to know and understand what transformations should be created.
For this next example we're going to convert all pos_tag == ADJ and pos_tag == "NOUN" to their lemma, and stitch the sentences back together so that they're ready for visualisations. We'll make use of case_when and the now-familiar grouped summarise paste trick. If we group by universal_message_id we'll get back a r nrow(data)
row data frame which we can join back to our original data frame.
annotations <- annotations %>% mutate(token = case_when(pos_tag == "ADJ" ~ lemma, pos_tag == "NOUN" ~ lemma, TRUE ~ lemma)) %>% select(token, universal_message_id) %>% summarise(sentence = paste0(token, collapse = " "), .by = universal_message_id)
Once we have the information regarding the co-occurrences of nouns and adjectives within our documents, we are able to evidence what entities appear throughout our data, as well as the type of language used in conjunction with and toward them. We can later add to this, by visualizing the frequency of terms appearing within the same document as well as those that are explicitly used one after another, just like the example above, where we see 'hot sun' appear as most frequent.
The workflow for Spanish is very similar to English. We'll select language = "spanish" and then we'll take a look at the output for a Spanish document. We'll take a paragraph from El Pais and store it in a data frame named 'spanish'
"Hace muchos años, cuando llegó a La Habana, encontró el mejor acto de amor —el estudio de la escritura, que es aún mejor que el estudio de la vida— por parte de la joven Yalenis Velazco, quien dedicó un intenso ensayo a su obra."
spanish <- tibble(text = "Hace muchos años, cuando llegó a La Habana, encontró el mejor acto de amor —el estudio de la escritura, que es aún mejor que el estudio de la vida— por parte de la joven Yalenis Velazco, quien dedicó un intenso ensayo a su obra.", id = 1)
spanish_model <- limpiar_pos_import_model("spanish") # Load the model spanish <- spanish %>% limpiar_pos_annotate(text, id, spanish_model, udpate_progress = 0) # Extract Pos Tags spanish %>% select(token, lemma, pos_tag) # Select relevant columns
We see immediately that 'hace' a conjugation of the verb 'hacer' is converted to hacer, muchos -> mucho, años -> año. These changes would generally be agreeable, they'll help us to reduce the overall number of words in our data set, and our text-based visualisations should be cleaner. However, it is always a good idea to carefully consider the effects of any transformation you perform on data, rather than doing it blindly.
We can convert all verbs to lemma and stitch the document back together using the same tricks as we did for English:
spanish %>% mutate( token = ifelse(pos_tag == "VERB", lemma, token) ) %>% summarise(.by = id, text = paste0(token, collapse = " "))
Clearly the sentence is now ungrammatical, but for a topic modelling workflow or some n-gram networks, the data will be cleaner and the counts will better reflect the real frequencies for each verb.
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.