overview

Chung-hong Chan 1

The validation test is called “oolong test” (for reading tea leaves). Creating oolong test for topic models and dictionary-based uses the same function: create_oolong(). The most important parameters are input_model and input_corpus. Setting each of them to NULL generates different tests.

input_model input_corpus output
Not NULL NULL oolong test for validating a topic model with word intrusion test
Not NULL Not NULL oolong test for validating a topic model with word intrusion test and topic intrusion test
NULL Not NULL oolong test for creating gold standard
NULL NULL error

Installation

You can install the development version from GitHub with:

# install.packages("devtools")
devtools::install_github("chainsawriot/oolong")

Validating Topic Models

Word intrusion test

abstracts_stm is an example topic model trained with the data abstracts using the stm package. Currently, this package supports structural topic models / correlated topic models from stm, Warp LDA models from text2vec and LDA/CTM models from topicmodels.

library(oolong)
library(stm)
#> stm v1.3.5 successfully loaded. See ?stm for help. 
#>  Papers, resources, and other materials at structuraltopicmodel.com
library(quanteda)
#> Package version: 2.0.0
#> Parallel computing: 2 of 4 threads used.
#> See https://quanteda.io for tutorials and examples.
#> 
#> Attaching package: 'quanteda'
#> The following object is masked from 'package:utils':
#> 
#>     View
library(dplyr)
#> 
#> Attaching package: 'dplyr'
#> The following objects are masked from 'package:stats':
#> 
#>     filter, lag
#> The following objects are masked from 'package:base':
#> 
#>     intersect, setdiff, setequal, union
abstracts_stm
#> A topic model with 20 topics, 2500 documents and a 3998 word dictionary.

To create an oolong test, use the function create_oolong_test.

oolong_test <- create_oolong(abstracts_stm)
oolong_test
#> An oolong test object with k = 20, 0 coded.
#> Use the method $do_word_intrusion_test() to do word intrusion test.
#> Use the method $lock() to finalize this object and see the results.

As instructed, use the method $do_word_intrusion_test() to start coding.

oolong_test$do_word_intrusion_test()

After the coding, you need to first lock the test. Then, you can look at the model precision by printing the oolong test.

oolong_test$lock()
oolong_test
#> An oolong test object with k = 20, 20 coded.
#> 95%  precision

Topic intrusion test

For example, abstracts_stm was generated with the corpus abstracts$text

library(tibble)
abstracts
#> # A tibble: 2,500 x 1
#>    text                                                                    
#>    <chr>                                                                   
#>  1 This study explores the benefits and risks featured in medical tourism …
#>  2 This article puts forth the argument that with the transfer of stock tr…
#>  3 The purpose of this study was to evaluate the effect the visual fidelit…
#>  4 Among the many health issues relevant to college students, overconsumpt…
#>  5 This address, delivered at ICA's 50th anniversary conference, calls on …
#>  6 The Internet has often been used to reach men who have sex with men (MS…
#>  7 This article argues that the literature describing the internet revolut…
#>  8 This research study examined Bud Goodall's online health narrative as a…
#>  9 Information technology and new media allow for collecting and sharing p…
#> 10 Using a national, telephone survey of 1,762 adolescents aged 12-17 year…
#> # … with 2,490 more rows

Creating the oolong test object with the corpus used for training the topic model will generate topic intrusion test cases.

oolong_test <- create_oolong(abstracts_stm, abstracts$text)
oolong_test
#> An oolong test object with k = 20, 0 coded.
#> Use the method $do_word_intrusion_test() to do word intrusion test.
#> With 25 cases of topic intrusion test. 0 coded.
#> Use the method $do_topic_intrusion_test() to do topic intrusion test.
#> Use the method $lock() to finalize this object and see the results.

Similarly, use the $do_topic_intrusion_test to code the test cases, lock the test with $lock() and then you can look at the TLO (topic log odds) value by printing the oolong test.

oolong_test$do_topic_intrusion_test()
oolong_test$lock()
oolong_test
#> An oolong test object with k = 20, 20 coded.
#> 100%  precision
#> With 25 cases of topic intrusion test. 25 coded.
#> TLO: -0.121

Suggested workflow

The test makes more sense if more than one coder is involved. A suggested workflow is to create the test, then clone the oolong object. Ask multiple coders to do the test(s) and then summarize the results.

Train a topic model.

dfm(abstracts$text, tolower = TRUE, stem = TRUE, remove = stopwords('english'), remove_punct = TRUE, remove_numbers = TRUE, remove_symbols = TRUE, remove_hyphens = TRUE) %>% dfm_trim(min_docfreq = 5, max_docfreq = 1000) %>% dfm_select(min_nchar = 3, pattern = "^[a-zA-Z]+$", valuetype = "regex") -> abstracts_dfm
docvars(abstracts_dfm, "title") <- abstracts$title
abstracts_dfm %>% convert(to = "stm", omit_empty = FALSE) -> abstracts_stm
abstracts_stm <- stm(abstracts_stm$documents, abstracts_stm$vocab, data =abstracts_stm$meta, K = 10, seed = 42)

Create a new oolong object.

oolong_test_rater1 <- create_oolong(abstracts_stm, abstracts$text)

Clone the oolong object to be used by other raters.

oolong_test_rater2 <- clone_oolong(oolong_test_rater1)

Ask different coders to code each object and then lock the object.

## Let rater 1 do the test.
oolong_test_rater1$do_word_intrusion_test()
oolong_test_rater1$do_topic_intrusion_test()
oolong_test_rater1$lock()

## Let rater 2 do the test.
oolong_test_rater2$do_word_intrusion_test()
oolong_test_rater2$do_topic_intrusion_test()
oolong_test_rater2$lock()

Get a summary of the two objects.

summarize_oolong(oolong_test_rater1, oolong_test_rater2)
#> Mean model precision: 0.3
#> Quantiles of model precision: 0.25, 0.275, 0.3, 0.325, 0.35
#> P-value of the model precision (H0: Model precision is not better than random guess): 0.0494424481559265
#> Krippendorff's alpha: 0.535714285714286
#> K Precision: 0, 0, 0, 0, 0, 1, 0, 0, 0.5, 0, 0, 1, 0, 1, 0.5, 0, 1, 0.5, 0, 0.5
#> Mean TLO: -2.08
#> Median TLO: -2.47
#> Quantiles of TLO: -7.15806610855452, -3.6173090924152, -2.465168096581, 0, 0
#> P-Value of the median TLO (H0: Median TLO is not better than random guess): 0.046

About the p-values

The test for model precision (MP) is based on an one-tailed, one-sample binomial test for each rater. In a multiple-rater situation, the p-values from all raters are combined using the Fisher’s method (a.k.a. Fisher’s omnibus test).

H0: MP is not better than 1/ n_top_terms

H1: MP is better than 1/ n_top_terms

The test for the median of TLO is based on a permutation test.

H0: Median TLO is not better than random guess.

H1: Median TLO is better than random guess.

One must notice that the two statistical tests are testing the bear minimum. A significant test only indicates the topic model can make the rater(s) perform better than random guess. It is not an indication of good topic interpretability. Also, one should use a very conservative significant level, e.g. \(\alpha < 0.001\).

About Warp LDA

There is a subtle difference between the support for stm and for text2vec.

abstracts_warplda is a Warp LDA object trained with the same dataset as the abstracts_stm

abstracts_warplda
#> <WarpLDA>
#>   Inherits from: <LDA>
#>   Public:
#>     clone: function (deep = FALSE) 
#>     components: 0 1 0 46 0 95 0 20 42 8 31 36 50 23 0 0 0 58 0 43 0 0 0  ...
#>     fit_transform: function (x, n_iter = 1000, convergence_tol = 0.001, n_check_convergence = 10, 
#>     get_top_words: function (n = 10, topic_number = 1L:private$n_topics, lambda = 1) 
#>     initialize: function (n_topics = 10L, doc_topic_prior = 50/n_topics, topic_word_prior = 1/n_topics, 
#>     plot: function (lambda.step = 0.1, reorder.topics = FALSE, doc_len = private$doc_len, 
#>     topic_word_distribution: 0 9.41796948577887e-05 0 0.00446992517733942 0 0.0086837 ...
#>     transform: function (x, n_iter = 1000, convergence_tol = 0.001, n_check_convergence = 10, 
#>   Private:
#>     calc_pseudo_loglikelihood: function (ptr = private$ptr) 
#>     check_convert_input: function (x) 
#>     components_: 0 1 0 46 0 95 0 20 42 8 31 36 50 23 0 0 0 58 0 43 0 0 0  ...
#>     doc_len: 80 68 85 88 69 118 99 50 57 88 70 67 53 62 66 92 89 79 1 ...
#>     doc_topic_distribution: function () 
#>     doc_topic_distribution_with_prior: function () 
#>     doc_topic_matrix: 0 0 0 0 0 3 111 0 0 0 0 0 90 134 0 174 0 321 0 0 109 38  ...
#>     doc_topic_prior: 0.1
#>     fit_transform_internal: function (model_ptr, n_iter, convergence_tol, n_check_convergence, 
#>     get_c_all: function () 
#>     get_c_all_local: function () 
#>     get_doc_topic_matrix: function (prt, nr) 
#>     get_topic_word_count: function () 
#>     init_model_dtm: function (x, ptr = private$ptr) 
#>     internal_matrix_formats: list
#>     is_initialized: FALSE
#>     n_iter_inference: 10
#>     n_topics: 20
#>     ptr: externalptr
#>     reset_c_local: function () 
#>     run_iter_doc: function (update_topics = TRUE, ptr = private$ptr) 
#>     run_iter_word: function (update_topics = TRUE, ptr = private$ptr) 
#>     seeds: 135203513.874082 471172603.061186
#>     set_c_all: function (x) 
#>     set_internal_matrix_formats: function (sparse = NULL, dense = NULL) 
#>     topic_word_distribution_with_prior: function () 
#>     topic_word_prior: 0.01
#>     transform_internal: function (x, n_iter = 1000, convergence_tol = 0.001, n_check_convergence = 10, 
#>     vocabulary: explor benefit risk featur medic broker websit well type ...

All the API endpoints are the same, except the one for the creation of topic intrusion test cases. You must supply also the input_dfm.

### Just word intrusion test.
oolong_test <- create_oolong(abstracts_warplda)
oolong_test
#> An oolong test object with k = 20, 0 coded.
#> Use the method $do_word_intrusion_test() to do word intrusion test.
#> Use the method $lock() to finalize this object and see the results.
abstracts_dfm
#> Document-feature matrix of: 2,500 documents, 3,998 features (98.6% sparse).
#>        features
#> docs    explor benefit risk featur medic broker websit well type persuas
#>   text1      1       2    2      2     6      3      6    1    3       1
#>   text2      0       0    1      0     0      0      0    0    1       0
#>   text3      0       1    0      0     0      0      0    0    0       0
#>   text4      1       0    0      0     0      0      0    0    0       0
#>   text5      1       0    0      0     0      0      0    0    0       0
#>   text6      0       1    1      0     0      0      0    0    0       0
#> [ reached max_ndoc ... 2,494 more documents, reached max_nfeat ... 3,988 more features ]
oolong_test <- create_oolong(abstracts_warplda, abstracts$text, input_dfm = abstracts_dfm)
#> INFO  [13:50:18.227] early stopping at 50 iteration
oolong_test
#> An oolong test object with k = 20, 0 coded.
#> Use the method $do_word_intrusion_test() to do word intrusion test.
#> With 25 cases of topic intrusion test. 0 coded.
#> Use the method $do_topic_intrusion_test() to do topic intrusion test.
#> Use the method $lock() to finalize this object and see the results.

Validating Dictionary-based Methods

Creating gold standard

trump2k is a dataset of 2,000 tweets from @realdonaldtrump.

tibble(text = trump2k)
#> # A tibble: 2,000 x 1
#>    text                                                                    
#>    <chr>                                                                   
#>  1 "In just out book, Secret Service Agent Gary Byrne doesn't believe that…
#>  2 "Hillary Clinton has announced that she is letting her husband out to c…
#>  3 "\"@TheBrodyFile: Always great to visit with @TheBrodyFile one-on-one w…
#>  4 "Explain to @brithume and @megynkelly, who know nothing, that I will be…
#>  5 "Nobody beats me on National Security. https://t.co/sCrj4Ha1I5"         
#>  6 "\"@realbill2016: @realDonaldTrump @Brainykid2010 @shl Trump leading LA…
#>  7 "\"@teapartynews: Trump Wins Tea Party Group's 'Nashville Straw Poll' -…
#>  8 "Big Republican Dinner tonight at Mar-a-Lago in Palm Beach. I will be t…
#>  9 ".@HillaryClinton loves to lie. America has had enough of the CLINTON'S…
#> 10 "\"@brianstoya: @realDonaldTrump For POTUS #2016\""                     
#> # … with 1,990 more rows

For example, you are interested in studying the sentiment of these tweets. One can use tools such as AFINN to automatically extract sentiment in these tweets. However, oolong recommends to generate gold standard by human coding first using a subset. By default, oolong selects 1% of the origin corpus as test cases. The parameter construct should be an adjective, e.g. positive, liberal, populistic, etc.

oolong_test <- create_oolong(input_corpus = trump2k, construct = "positive")
oolong_test
#> An oolong test object (gold standard generation) with 20 cases, 0 coded.
#> Use the method $do_gold_standard_test() to generate gold standard.
#> Use the method $lock() to finalize this object and see the results.

As instructed, use the method $do_gold_standard_test() to start coding.

oolong_test$do_gold_standard_test()

After the coding, you need to first lock the test and then the $turn_gold() method is available.

oolong_test$lock()
oolong_test
#> An oolong test object (gold standard generation) with 20 cases, 20 coded.
#> Use the method $turn_gold() to convert the test results into a quanteda corpus.

Example: Validating AFINN using the gold standard

A locked oolong test can be converted into a quanteda-compatible corpus for further analysis. The corpus contains two docvars, ‘answer’.

oolong_test$turn_gold()
#> Corpus consisting of 20 documents and 1 docvar.
#> text1 :
#> "Thank you Eau Claire, Wisconsin.  #VoteTrump on Tuesday, Apr..."
#> 
#> text2 :
#> ""@bobby990r_1: @realDonaldTrump would lead polls the second ..."
#> 
#> text3 :
#> ""@KdanielsK: @misstcassidy @AllAboutTheTea_ @realDonaldTrump..."
#> 
#> text4 :
#> "Thank you for a great afternoon Birmingham, Alabama! #Trump2..."
#> 
#> text5 :
#> ""@THETAINTEDT: @foxandfriends @realDonaldTrump Trump 2016 ht..."
#> 
#> text6 :
#> "People believe CNN these days almost as little as they belie..."
#> 
#> [ reached max_ndoc ... 14 more documents ]
#> Access the answer from the coding with quanteda::docvars(obj, 'answer')

In this example, we calculate the AFINN score for each tweet using quanteda. The dictionary afinn is bundle with this package.

gold_standard <- oolong_test$turn_gold()
dfm(gold_standard, remove_punct = TRUE) %>% dfm_lookup(afinn) %>% quanteda::convert(to = "data.frame") %>%
    mutate(matching_word_valence = (neg5 * -5) + (neg4 * -4) + (neg3 * -3) + (neg2 * -2) + (neg1 * -1)
           + (zero * 0) + (pos1 * 1) + (pos2 * 2) + (pos3 * 3) + (pos4 * 4) + (pos5 * 5),
           base = ntoken(gold_standard, remove_punct = TRUE), afinn_score = matching_word_valence / base) %>%
    pull(afinn_score) -> all_afinn_score
all_afinn_score
#>       text1       text2       text3       text4       text5       text6 
#>  0.33333333 -0.09090909 -0.16666667  0.45454545  0.00000000  0.00000000 
#>       text7       text8       text9      text10      text11      text12 
#>  0.16666667  0.38461538  0.00000000  0.38461538 -0.29166667  0.00000000 
#>      text13      text14      text15      text16      text17      text18 
#>  0.50000000  0.07142857  0.00000000 -0.12000000  0.28571429  0.16000000 
#>      text19      text20 
#>  0.36842105  0.38888889

Put back the vector of AFINN score into the respective docvars and study the correlation between the gold standard and AFINN.

summarize_oolong(oolong_test, target_value = all_afinn_score)
#> `geom_smooth()` using formula 'y ~ x'
#> `geom_smooth()` using formula 'y ~ x'
#> Correlation: 0.718 (p = 0)
#> Effect of content length: -0.319 (p = 0.171)

Suggested workflow

Create an oolong object, clone it for another coder. According to Song et al. (Forthcoming), you should at least draw 1% of your data.

trump <- create_oolong(input_corpus = trump2k, exact_n = 40)
trump2 <- clone_oolong(trump)

Instruct two coders to code the tweets and lock the objects.

trump$do_gold_standard_test()
trump2$do_gold_standard_test()
trump$lock()
trump2$lock()

Calculate the target value (in this case, the AFINN score) by turning one object into a corpus.

gold_standard <- trump$turn_gold()
dfm(gold_standard, remove_punct = TRUE) %>% dfm_lookup(afinn) %>% quanteda::convert(to = "data.frame") %>%
    mutate(matching_word_valence = (neg5 * -5) + (neg4 * -4) + (neg3 * -3) + (neg2 * -2) + (neg1 * -1)
           + (zero * 0) + (pos1 * 1) + (pos2 * 2) + (pos3 * 3) + (pos4 * 4) + (pos5 * 5),
           base = ntoken(gold_standard, remove_punct = TRUE), afinn_score = matching_word_valence / base) %>%
    pull(afinn_score) -> target_value

Summarize all oolong objects with the target value.

res <- summarize_oolong(trump, trump2, target_value = target_value)
#> `geom_smooth()` using formula 'y ~ x'
#> `geom_smooth()` using formula 'y ~ x'

Read the results. The diagnostic plot consists of 4 subplots. It is a good idea to read Bland & Altman (1986) on the difference between correlation and agreement.

The textual output contains the Krippendorff’s alpha of the codings by your raters. In order to claim validity of your target value, you must first establish the reliability of your gold standard. Song et al. [Forthcoming] suggest Krippendorff’s Alpha > 0.7 as an acceptable cut-off.

res
#> Krippendorff's Alpha: 0.931443661971831
#> Correlation: 0.744 (p = 0)
#> Effect of content length: -0.319 (p = 0.171)
plot(res)

References

  1. Chang, J., Gerrish, S., Wang, C., Boyd-Graber, J. L., & Blei, D. M. (2009). Reading tea leaves: How humans interpret topic models. In Advances in neural information processing systems (pp. 288-296).
  2. Song et al. (2020) In validations we trust? The impact of imperfect human annotations as a gold standard on the quality of validation of automated content analysis. Political Communication. link
  3. Bland, J. M., & Altman, D. (1986). Statistical methods for assessing agreement between two methods of clinical measurement. The lancet, 327(8476), 307-310.
  4. Chan et al. (2020) It is not listed here, because it is still under review. Do it faster! Reviewers!
  5. Nielsen, F. Å. (2011). A new ANEW: Evaluation of a word list for sentiment analysis in microblogs. arXiv preprint arXiv:1103.2903.

  1. University of Mannheim↩︎