vignettes/E2_YESAI_2024.Rmd
E2_YESAI_2024.RmdWith a very simple model of English grammar and a little vocabulary, we can randomly generate ‘normal-sounding’ sentences.
Here is a vector of elegant nouns.
nouns = c("Mercedes", "Tiffany", "Lamborghini", "Maserati", "Tosca", "La Traviata",
"Diamonds", "Crystal", "Lenox")We’ll continue to set up some vectors of verbs (past tense), prepositions, adjectives, prepositions, and indicatives.
pastverbs = c("radiated", "glowed", "luxuriated", "ecstatized", "overwhelmed",
"hyperventilated", "raptured", "raved", "stupendousized")
preps = c("on", "over", "beyond", "above", "beneath", "within", "across")
adjectives = c("magnificent", "majestic", "sophisticated", "exquisite",
"glorious", "regal")
preps = c("on", "over", "beyond", "above", "beneath", "within", "across")
indic = c("the", "a", "any", "some")Our model for a sentence is
(*) indicative + noun + verb(tensed) + preposition + indicative + adjective + nounWe produce a sentence by sampling purely randomly from the relevant vocabulary vectors. We use sprintf to substitute our choices into a template of seven strings.
get1 = function (x) sample(x, size=1)
up1 = function(x) paste0(toupper(substring(x,1,1)), substring(x,2))
sentence = function() {
cat(sprintf("%s %s %s %s %s %s %s.", up1(get1(indic)), get1(nouns), get1(pastverbs),
get1(preps), get1(indic), get1(adjectives), get1(nouns)))
}
set.seed(12345) # reproducible
sentence()## A Lamborghini raved over the sophisticated Crystal.
sentence()## A La Traviata hyperventilated across a majestic Mercedes.
sentence()## Some Crystal raptured within a magnificent Maserati.
sentence()## Some Lamborghini stupendousized within the exquisite Diamonds.The potential repetitiousness of purely random sampling can be avoided. In the following sentence generator, we prevent reuse of the initial noun or indicative.
sentence_nr = function() {
i1 = get1(indic)
n1 = get1(nouns)
pv1 = get1(pastverbs)
p1 = get1(preps)
i2 = get1(setdiff(indic,i1)) # don't repeat indicative
ad = get1(adjectives)
n2 = get1(setdiff(nouns,n1)) # don't repeat noun
sprintf("%s %s %s %s %s %s %s.", up1(i1), n1, pv1, p1, i2, ad, n2)
}
sentence_nr()## [1] "Some Tiffany ecstatized beyond the magnificent Tosca."
sentence_nr()## [1] "Some La Traviata stupendousized beneath any sophisticated Mercedes."We can also compose longer sentences, using the conjunction “because”:
linked = function() cat(sprintf("%s,\n because %s", sub(".$", "", sentence_nr()), sentence_nr()))
linked()## Any Mercedes overwhelmed over a sophisticated Maserati,
## because The Lamborghini radiated over some magnificent Lenox.
linked()## Any Maserati glowed above some glorious Diamonds,
## because Any Diamonds raptured across a glorious La Traviata.Exercise 1. What does the up1 function do? Apply it to the string “foo”. What is the result?
Exercise 2. There is something wrong with the last two productions. What is it and how would you go about correcting it?
Exercise 3. Propose an alternate schematic to the one noted above as (*). Implement it in a function altsentence and illustrate its use with our vocaularies. For example
noun + "is not" + nouncould be implemented as
altsentence = function() {
sprintf("%s is not %s", get1(nouns), get1(nouns))
}
altsentence()## [1] "Diamonds is not Crystal"
altsentence()## [1] "Lamborghini is not Tosca"Come up with something more interesting than this!
I asked perplexity.ai ($20/mo) to produce 10 lexically tagged sentences. By “lexical tagging” we refer to a process of associating with each word or phrase a category defining the word’s role in sentence meaning. Examples of tags include
- DT: Determiner
- NN: Noun, singular or mass
- NNP: Proper noun, singular
- VB: Verb, base form
- VBZ: Verb, 3rd person singular present
- JJ: Adjective
- IN: Preposition or subordinating conjunction
- CD: Cardinal numberThe full set of “part of speech” tags can be found at the U Penn linguistics site.
Here’s what came back from perplexity.ai:
[The/DT] [Eiffel/NNP] [Tower/NNP] [is/VBZ] [a/DT] [wrought-iron/JJ] [lattice/NN] [tower/NN] [on/IN] [the/DT] [Champ/NNP] [de/NNP] [Mars/NNP] [in/IN] [Paris/NNP] [,/,] [France/NNP] [./.]
[Photosynthesis/NN] [is/VBZ] [the/DT] [process/NN] [by/IN] [which/WDT] [plants/NNS] [use/VBP] [sunlight/NN] [,/,] [water/NN] [and/CC] [carbon/NN] [dioxide/NN] [to/TO] [produce/VB] [oxygen/NN] [and/CC] [energy/NN] [in/IN] [the/DT] [form/NN] [of/IN] [sugar/NN] [./.]
[The/DT] [Great/NNP] [Wall/NNP] [of/IN] [China/NNP] [is/VBZ] [a/DT] [series/NN] [of/IN] [fortifications/NNS] [and/CC] [walls/NNS] [built/VBN] [across/IN] [the/DT] [historical/JJ] [northern/JJ] [borders/NNS] [of/IN] [ancient/JJ] [Chinese/JJ] [states/NNS] [./.]
[William/NNP] [Shakespeare/NNP] [was/VBD] [an/DT] [English/JJ] [playwright/NN] [,/,] [poet/NN] [,/,] [and/CC] [actor/NN] [widely/RB] [regarded/VBN] [as/IN] [the/DT] [greatest/JJS] [writer/NN] [in/IN] [the/DT] [English/JJ] [language/NN] [./.]
[The/DT] [human/JJ] [genome/NN] [contains/VBZ] [approximately/RB] [3/CD] [billion/CD] [base/NN] [pairs/NNS] [of/IN] [DNA/NN] [./.]
[Albert/NNP] [Einstein/NNP] [developed/VBD] [the/DT] [theory/NN] [of/IN] [relativity/NN] [,/,] [one/CD] [of/IN] [the/DT] [two/CD] [pillars/NNS] [of/IN] [modern/JJ] [physics/NN] [./.]
[The/DT] [Amazon/NNP] [rainforest/NN] [,/,] [covering/VBG] [much/JJ] [of/IN] [northwestern/JJ] [Brazil/NNP] [and/CC] [extending/VBG] [into/IN] [Colombia/NNP] [,/,] [Peru/NNP] [and/CC] [other/JJ] [South/NNP] [American/NNP] [countries/NNS] [,/,] [is/VBZ] [the/DT] [world’s/NN$] [largest/JJS] [tropical/JJ] [rainforest/NN] [./.]
[The/DT] [periodic/JJ] [table/NN] [is/VBZ] [a/DT] [tabular/JJ] [arrangement/NN] [of/IN] [the/DT] [chemical/JJ] [elements/NNS] [,/,] [organized/VBN] [on/IN] [the/DT] [basis/NN] [of/IN] [their/PRP$] [atomic/JJ] [numbers/NNS] [,/,] [electron/NN] [configurations/NNS] [,/,] [and/CC] [recurring/VBG] [chemical/JJ] [properties/NNS] [./.]
[The/DT] [United/NNP] [Nations/NNPS] [(/] [UN/NNP] [)/] [is/VBZ] [an/DT] [intergovernmental/JJ] [organization/NN] [that/WDT] [aims/VBZ] [to/TO] [maintain/VB] [international/JJ] [peace/NN] [and/CC] [security/NN] [,/,] [develop/VB] [friendly/JJ] [relations/NNS] [among/IN] [nations/NNS] [,/,] [achieve/VB] [international/JJ] [cooperation/NN] [,/,] [and/CC] [be/VB] [a/DT] [center/NN] [for/IN] [harmonizing/VBG] [the/DT] [actions/NNS] [of/IN] [nations/NNS] [./.]
[The/DT] [Mona/NNP] [Lisa/NNP] [is/VBZ] [a/DT] [half-length/JJ] [portrait/NN] [painting/NN] [by/IN] [Italian/JJ] [artist/NN] [Leonardo/NNP] [da/NNP] [Vinci/NNP] [,/,] [considered/VBN] [an/DT] [archetypal/JJ] [masterpiece/NN] [of/IN] [the/DT] [Italian/NNP] [Renaissance/NNP] [./.]
The sentnce data do not possess a regular structure amenable to tabulation. There is a hierarchical format: arbitrary numbers of language particles and annotations are ordered within each sentence, and the sentences form a group of ten.
A popular format for irregularly structured data is “JSON”. We asked perplexity.ai to produce the tagged sentences in JSON format. R can ingest JSON for programmatic analysis using the jsonlite package.
library(jsonlite)
myj = system.file("json/tensent.json", package="YESCDS")
sdat = fromJSON(myj, flatten=FALSE)
frames = sdat[[2]]
frames[[1]]## word tag
## 1 The DT
## 2 Eiffel NNP
## 3 Tower NNP
## 4 is VBZ
## 5 a DT
## 6 wrought-iron JJ
## 7 lattice NN
## 8 tower NN
## 9 on IN
## 10 the DT
## 11 Champ NNP
## 12 de NNP
## 13 Mars NNP
## 14 in IN
## 15 Paris NNP
## 16 , ,
## 17 France NNP
## 18 . .We see that for each sentence we have a data.frame with one row for each English word and associated part of speech.
The overall frequency distribution of parts of speech in this little collection is:
##
## ( ) NN$ NNPS PRP$ VBP JJS RB TO VBD WDT CD VBG VBN VB CC
## 1 1 1 1 1 1 2 2 2 2 2 4 4 4 5 9
## VBZ . , NNS IN JJ DT NNP NN
## 9 10 15 16 25 25 28 29 39Here’s how we might use R to determine how many words into a sentence we need to go to find a verb:
## [1] 4 2 6 3 4 3 5 4 7 4Exercise 1. What is the position of the verb in the fourth sentence? What is the verb, and what type is it? Consult the link to U Penn linguistics given above.
Exercise 2. Which sentence has the verb in the “latest” position, and why is the position number so high? Should the counting procedure be adjusted?
In this section we will assume that you are working on a machine that is able to use software and data organized at huggingface.co, specifically a version of BERT.
We will ‘source’ a github gist that will work if python is present, and if it and the reticulate package have been configured to be able to find python packages torch and transformers.
This code chunk will produce a table of scores and completions for the prompt The capital of Spain is [MASK]..
devtools::source_gist("https://gist.github.com/vjcitn/0e1b0289eb1a219b98d60b5b863f8d62")This turns out to be difficult to use on the ‘free’ galaxy instance. The software can be demonstrated at https://vjcitn.shinyapps.io/AIapp. Of interest:
The code in the gist can be modified to change the model, and the prompt can be changed to seek scored responses to different prompts.