Example: Fitting a model to English novels

Corpus

We will be using texts of the first lines of novels by Jane Austen, George Eliot, and Elizabeth Gaskell. Excerpts were obtained from the full texts of novels available on Project Gutenberg: http://gutenberg.org.

data(novels_excerpts)

	title	author	text
1	A Dark Night’s Work	Gaskell, Elizabeth Cleghorn	In the county town of a certain shire there lived (about forty years ago) one Mr. Wilkins, a conveyancing attorney of considerable standing. The certain shire was but a small county, and the principal town in it contained only about four thousand inhabitants; so in saying that Mr. Wilkins was the principal lawyer in Hamley, I say very little, unless I add that he transacted all the legal business of the gentry for twenty miles round. His grandfather had established the connection; his father had consolidated and strengthened it, and, indeed, by his wise and upright conduct, as well as by his professional skill, had obtained for himself the position of confidential friend to many of the surrounding families of distinction.
4	Brother Jacob	Eliot, George	Among the many fatalities attending the bloom of young desire, that of blindly taking to the confectionery line has not, perhaps, been sufficiently considered. How is the son of a British yeoman, who has been fed principally on salt pork and yeast dumplings, to know that there is satiety for the human stomach even in a paradise of glass jars full of sugared almonds and pink lozenges, and that the tedium of life can reach a pitch where plum-buns at discretion cease to offer the slightest excitement? Or how, at the tender age when a confectioner seems to him a very prince whom all the world must envy–who breakfasts on macaroons, dines on meringues, sups on twelfth-cake, and fills up the intermediate hours with sugar-candy or peppermint–how is he to foresee the day of sad wisdom, when he will discern that the confectioner’s calling is not socially influential, or favourable to a soaring ambition?
8	Emma	Austen, Jane	Emma Woodhouse, handsome, clever, and rich, with a comfortable home and happy disposition, seemed to unite some of the best blessings of existence; and had lived nearly twenty-one years in the world with very little to distress or vex her. She was the youngest of the two daughters of a most affectionate, indulgent father; and had, in consequence of her sister’s marriage, been mistress of his house from a very early period. Her mother had died too long ago for her to have more than an indistinct remembrance of her caresses; and her place had been supplied by an excellent woman as governess, who had fallen little short of a mother in affection. Sixteen years had Miss Taylor been in Mr. Woodhouse’s family, less as a governess than a friend, very fond of both daughters, but particularly of Emma. Between them it was more the intimacy of sisters.

The corpus should have at least one variable by which the texts can be grouped — the most common examples are a “speaker” or “author” attribute. Here, we will use novels_excerpts$author.

unique(novels_excerpts$author)
#> [1] "Gaskell, Elizabeth Cleghorn" "Eliot, George"              
#> [3] "Austen, Jane"

Using `stylest_select_vocab`

This function uses n-fold cross-validation to identify the set of terms that maximizes the model’s rate of predicting the speakers of out-of-sample texts. For those unfamiliar with cross-validation, the technical details follow:

The terms of the raw vocabulary are ordered by frequency.
A subset of the raw vocabulary above a frequency percentile is selected; e.g. terms above the 50th percentile are those which occur more frequently than the median term in the raw vocabulary.
The corpus is divided into n folds.
One of these folds is held out and the model is fit using the remaining n-1 folds. The model is then used to predict the speakers of texts in the held-out fold. (This step is repeated n times.)
The mean prediction rate for models using this vocabulary (percentile) is calculated.

(Vocabulary selection is optional; the model can be fit using all the terms in the support of the corpus.)

Setting the seed before this step, to ensure reproducible runs, is recommended:

set.seed(1234)

Below are examples of stylest_select_vocab using the defaults and with custom parameters:

vocab_with_defaults <- stylest_select_vocab(novels_excerpts$text, novels_excerpts$author)

Tokenization selections can optionally be passed as the filter argument; see the corpus package for more information about text_filter.

filter <- corpus::text_filter(drop_punct = TRUE, drop_number = TRUE)

vocab_custom <- stylest_select_vocab(novels_excerpts$text, novels_excerpts$author, 
                                     filter = filter, smooth = 1, nfold = 10, 
                                     cutoff_pcts = c(50, 75, 99))

Let’s look inside the vocab_with_defaults object.

# Percentile with best prediction rate
vocab_with_defaults$cutoff_pct_best
#> [1] 90

# Rate of INCORRECTLY predicted speakers of held-out texts
vocab_with_defaults$miss_pct
#>           [,1]      [,2]      [,3]     [,4]     [,5]     [,6]
#> [1,]  66.66667  66.66667  66.66667 66.66667 66.66667 66.66667
#> [2,]  60.00000  60.00000  60.00000 60.00000 20.00000 40.00000
#> [3,] 100.00000 100.00000 100.00000 25.00000 25.00000 75.00000
#> [4,] 100.00000 100.00000 100.00000 60.00000 60.00000 60.00000
#> [5,]  50.00000  50.00000  50.00000 50.00000 50.00000 50.00000

# Data on the setup:

# Percentiles tested
vocab_with_defaults$cutoff_pcts
#> [1] 50 60 70 80 90 99

# Number of folds
vocab_with_defaults$nfold
#> [1] 5

Fitting a model

Using a percentile to select terms

With the best percentile identified as 90 percent, we can select the terms above that percentile to use in the model. Be sure to use the same text_filter here as in the previous step.

terms_90 <- stylest_terms(novels_excerpts$text, novels_excerpts$author, 90, filter = filter)

Fitting the model: basic

Below, we fit the model using the terms above the 90th percentile, using the same text_filter as before, and leaving the smoothing value for term frequencies as the default 0.5.


mod <- stylest_fit(novels_excerpts$text, novels_excerpts$author, terms = terms_90, filter = filter)

The model contains detailed information about token usage by each of the authors (see mod$rate); exploring this is left as an exercise.

Fitting the model: adding custom term weights

A new feature is the option to specify custom term weights, in the form of a dataframe. The intended use case is the mean cosine distance from the embedding representation of the word to all other words in the vocabulary, but the weights can be anything desired by the user.

An example term_weights is shown below. When this argument is passed to stylest_fit(), the weights for terms in the model vocabulary will be extracted. Any term not included in term_weights will be assigned a default weight of 1.


term_weights <- data.frame("word" = c("the", "and", "Floccinaucinihilipilification"),
                           "mean_distance" = c(0.1,0.2,0.001))

term_weights
#>                            word mean_distance
#> 1                           the         0.100
#> 2                           and         0.200
#> 3 Floccinaucinihilipilification         0.001

Below is an example of fitting the model with term_weights:

mod <- stylest_fit(novels_excerpts$text, novels_excerpts$author, 
                   terms = terms_90, filter = filter,
                   term_weights = term_weights,
                   weight_varname = "mean_distance")

The weights are stored in mod$weights.

Using the model

Calculating speaker log odds


odds <- stylest_odds(mod, novels_excerpts$text, novels_excerpts$author)

We can examine the mean log odds that Jane Austen wrote Pride and Prejudice (in-sample).

# Pride and Prejudice
novels_excerpts$text[14]
#> [1] "It is a truth universally acknowledged, that a single man in possession of a good fortune, must be in want of a wife. However little known the feelings or views of such a man may be on his first entering a neighbourhood, this truth is so well fixed in the minds of the surrounding families, that he is considered the rightful property of some one or other of their daughters. \"My dear Mr. Bennet,\" said his lady to him one day, \"have you heard that Netherfield Park is let at last?\" Mr. Bennet replied that he had not. \"But it is,\" returned she; \"for Mrs. Long has just been here, and she told me all about it.\" Mr. Bennet made no answer. \"Do you not want to know who has taken it?\" cried his wife impatiently. \"_You_ want to tell me, and I have no objection to hearing it.\" This was invitation enough."

odds$log_odds_avg[14]
#> [1] -0.001925658

odds$log_odds_se[14]
#> [1] 0.001015172

Predicting the speaker of a new text

In this example, the model is used to predict the speaker of a new text, in this case Northanger Abbey by Jane Austen.

Note that a prior may be specified, and may be useful for handling texts containing out-of-sample terms. Here, we do not specify a prior, so a uniform prior is used.


na_text <- "No one who had ever seen Catherine Morland in her infancy would have supposed 
            her born to be an heroine. Her situation in life, the character of her father 
            and mother, her own person and disposition, were all equally against her. Her 
            father was a clergyman, without being neglected, or poor, and a very respectable 
            man, though his name was Richard—and he had never been handsome. He had a 
            considerable independence besides two good livings—and he was not in the least 
            addicted to locking up his daughters."

pred <- stylest_predict(mod, na_text)

Viewing the result, and recovering the log probabilities calculated for each speaker, is simple:

pred$predicted
#> [1] Eliot, George
#> Levels: Austen, Jane Eliot, George Gaskell, Elizabeth Cleghorn

pred$log_probs
#> 1 x 3 Matrix of class "dgeMatrix"
#>      Austen, Jane Eliot, George Gaskell, Elizabeth Cleghorn
#> [1,]    -1.111991      -1.02821                   -1.160115

Influential terms

stylest_term_influence identifies terms’ contributions to speakers’ distinctiveness in a fitted model.


influential_terms <- stylest_term_influence(mod, novels_excerpts$text, novels_excerpts$author)

The mean and maximum influence can be accessed with $infl_avg and $infl_max, respectively.

The terms with the highest mean influence can be obtained:

	term	infl_avg	infl_max
1	the	0.3151681	0.5978128
6	a	0.0488151	0.0864017
40	an	0.0484018	0.1089041
3	and	0.0125561	0.0196711
2	of	0.0000000	0.0000000
4	in	0.0000000	0.0000000

And the least influential terms:

	term	infl_avg	infl_max
141	time	0	0
142	two	0	0
143	under	0	0
144	whom	0	0
145	without	0	0
146	world	0	0

Citation

If you use this software, please cite:

Huang, L., Perry, P., & Spirling, A. (2020). A General Model of Author “Style” with Application to the UK House of Commons, 1935–2018. Political Analysis, 28(3), 412-434. https://doi.org/10.1017/pan.2019.49

@article{huang2020general,
  title={A General Model of Author “Style” with Application to the UK House of Commons, 1935--2018},
  author={Huang, Leslie and Perry, Patrick O and Spirling, Arthur},
  journal={Political Analysis},
  volume={28},
  number={3},
  pages={412--434},
  year={2020},
  publisher={Cambridge University Press}
}

Issues

Please submit any bugs, error reports, etc. on GitHub at: https://github.com/leslie-huang/stylest/issues.

stylest vignette

Leslie Huang

2021-03-04

About `stylest`

Installation

Load the package

Example: Fitting a model to English novels

Corpus

Using `stylest_select_vocab`

Fitting a model

Using a percentile to select terms

Fitting the model: basic

Fitting the model: adding custom term weights

Using the model

Calculating speaker log odds

Predicting the speaker of a new text

Influential terms

Citation

Issues

stylest vignette

Leslie Huang

2021-03-04

About stylest

Installation

Load the package

Example: Fitting a model to English novels

Corpus

Using stylest_select_vocab

Fitting a model

Using a percentile to select terms

Fitting the model: basic

Fitting the model: adding custom term weights

Using the model

Calculating speaker log odds

Predicting the speaker of a new text

Influential terms

Citation

Issues

About `stylest`

Using `stylest_select_vocab`