Perspective is an API that uses machine learning models to score the perceived impact a comment might have on a conversation. Website.

peRspective provides access to the API using the R programming language.

For an excellent documentation of the Perspective API see here.

This is a work-in-progress project and I welcome feedback and pull requests!

Setup

Get an API key

Follow these steps as outlined by the Perspective API to get an API key.

NOTE: Perspective API made some changes recently and you now have to apply for an API key via a Google Form.

Quota and character length Limits

Be sure to check your quota limit! You can learn more about Perspective API quota limit by visiting your google cloud project’s Perspective API page.

Models

For detailed overview of the used models see here.

Languages in production

Currently, Perspective API has production TOXICITY and SEVERE_TOXICITY attributes in the following languages:

  • English (en)
  • Spanish (es)
  • French (fr)
  • German (de)
  • Portuguese (pt)
  • Italian (it)
  • Russian (ru)

Toxicity attributes

Attribute name Type Description Available Languages
TOXICITY prod. A rude, disrespectful, or unreasonable comment that is likely to make people leave a discussion. English (en), Spanish (es), French (fr), German (de), Portuguese (pt), Italian (it), Russian (ru)
TOXICITY_EXPERIMENTAL exp. A rude, disrespectful, or unreasonable comment that is likely to make people leave a discussion. Arabic (ar)
SEVERE_TOXICITY prod. A very hateful, aggressive, disrespectful comment or otherwise very likely to make a user leave a discussion or give up on sharing their perspective. This attribute is much less sensitive to more mild forms of toxicity, such as comments that include positive uses of curse words. en, fr, es, de, it, pt, ru
SEVERE_TOXICITY_EXPERIMENTAL exp. A very hateful, aggressive, disrespectful comment or otherwise very likely to make a user leave a discussion or give up on sharing their perspective. This attribute is much less sensitive to more mild forms of toxicity, such as comments that include positive uses of curse words. ar
IDENTITY_ATTACK prod. Negative or hateful comments targeting someone because of their identity. de, it, pt, ru, en
IDENTITY_ATTACK_EXPERIMENTAL exp. Negative or hateful comments targeting someone because of their identity. fr, es, ar
INSULT prod. Insulting, inflammatory, or negative comment towards a person or a group of people. de, it, pt, ru, en
INSULT_EXPERIMENTAL exp. Insulting, inflammatory, or negative comment towards a person or a group of people. fr, es, ar
PROFANITY prod. Swear words, curse words, or other obscene or profane language. de, it, pt, ru, en
PROFANITY_EXPERIMENTAL exp. Swear words, curse words, or other obscene or profane language. fr, es, ar
THREAT prod. Describes an intention to inflict pain, injury, or violence against an individual or group. de, it, pt, ru, en
THREAT_EXPERIMENTAL exp. Describes an intention to inflict pain, injury, or violence against an individual or group. fr, es, ar
SEXUALLY_EXPLICIT exp. Contains references to sexual acts, body parts, or other lewd content. en
FLIRTATION exp. Pickup lines, complimenting appearance, subtle sexual innuendos, etc. en

New York Times attributes

These attributes are experimental because they are trained on a single source of comments—New York Times (NYT) data tagged by their moderation team—and therefore may not work well for every use case.

Attribute name Type Description Language
ATTACK_ON_AUTHOR exp. Attack on the author of an article or post. en
ATTACK_ON_COMMENTER exp. Attack on fellow commenter. en
INCOHERENT exp. Difficult to understand, nonsensical. en
INFLAMMATORY exp. Intending to provoke or inflame. en
LIKELY_TO_REJECT exp. Overall measure of the likelihood for the comment to be rejected according to the NYT’s moderation. en
OBSCENE exp. Obscene or vulgar language such as cursing. en
SPAM exp. Irrelevant and unsolicited commercial content. en
UNSUBSTANTIAL exp. Trivial or short comments. en

A character vector that includes allpeRspective supported models can be obtained like this:

c(
  peRspective::prsp_models,
  peRspective::prsp_exp_models
)
#>  [1] "TOXICITY"                     "SEVERE_TOXICITY"             
#>  [3] "IDENTITY_ATTACK"              "INSULT"                      
#>  [5] "PROFANITY"                    "SEXUALLY_EXPLICIT"           
#>  [7] "THREAT"                       "FLIRTATION"                  
#>  [9] "ATTACK_ON_AUTHOR"             "ATTACK_ON_COMMENTER"         
#> [11] "INCOHERENT"                   "INFLAMMATORY"                
#> [13] "LIKELY_TO_REJECT"             "OBSCENE"                     
#> [15] "SPAM"                         "UNSUBSTANTIAL"               
#> [17] "TOXICITY_EXPERIMENTAL"        "SEVERE_TOXICITY_EXPERIMENTAL"
#> [19] "IDENTITY_ATTACK_EXPERIMENTAL" "INSULT_EXPERIMENTAL"         
#> [21] "PROFANITY_EXPERIMENTAL"       "THREAT_EXPERIMENTAL"

Usage

First, install package from GitHub:

devtools::install_github("favstats/peRspective")

Load package:

Also the tidyverse for examples.

library(tidyverse)

Define your key variable.

peRspective functions will read the API key from environment variable perspective_api_key. In order to add your key to your environment file, you can use the function edit_r_environ() from the usethis package.

usethis::edit_r_environ()

This will open your .Renviron file in your text editor. Now, you can add the following line to it:

perspective_api_key="YOUR_API_KEY"

Save the file and restart R for the changes to take effect.

Alternatively, you can provide an explicit definition of your API key with each function call using the key argument.

prsp_score

Now you can use prsp_score to score your comments with various models provided by the Perspective API.

my_text <- "You wrote this? Wow. This is dumb and childish, please go f**** yourself."

text_scores <- prsp_score(
           text = my_text, 
           languages = "en",
           score_model = peRspective::prsp_models
           )

text_scores %>% 
  tidyr::gather() %>% 
  dplyr::mutate(key = forcats::fct_reorder(key, value)) %>% 
  ggplot2::ggplot(ggplot2::aes(key, value)) +
  ggplot2::geom_col() +
  ggplot2::coord_flip() +
  ggplot2::ylim(0, 1) +
  ggplot2::geom_hline(yintercept = 0.5, linetype = "dashed") +
  ggplot2::labs(x = "Model", y = "Probability", title = "Perspective API Results")

A Trump Tweet:

trump_tweet <- "The Fake News Media has NEVER been more Dishonest or Corrupt than it is right now. There has never been a time like this in American History. Very exciting but also, very sad! Fake News is the absolute Enemy of the People and our Country itself!"

text_scores <- prsp_score(
           trump_tweet, 
           score_sentences = F,
           score_model = peRspective::prsp_models
           )

text_scores %>% 
  tidyr::gather() %>% 
  dplyr::mutate(key = forcats::fct_reorder(key, value)) %>% 
  ggplot2::ggplot(ggplot2::aes(key, value)) +
  ggplot2::geom_col() +
  ggplot2::coord_flip() +
  ggplot2::ylim(0, 1) +
  ggplot2::geom_hline(yintercept = 0.5, linetype = "dashed") +
  ggplot2::labs(x = "Model", y = "Probability", title = "Perspective API Results")

Instead of scoring just entire comments you can also score individual sentences with score_sentences = T. In this case the Perspective API will automatically split your text into reasonable sentences and score them in addition to an overall score.

trump_tweet <- "The Fake News Media has NEVER been more Dishonest or Corrupt than it is right now. There has never been a time like this in American History. Very exciting but also, very sad! Fake News is the absolute Enemy of the People and our Country itself!"

text_scores <- prsp_score(
           trump_tweet, 
           score_sentences = T,
           score_model = peRspective::prsp_models
           )

text_scores %>% 
  tidyr::unnest(sentence_scores) %>% 
  dplyr::select(type, score, sentences) %>% 
  tidyr::gather(value, key, -sentences, -score) %>% 
  dplyr::mutate(key = forcats::fct_reorder(key, score)) %>% 
  ggplot2::ggplot(ggplot2::aes(key, score)) +
  ggplot2::geom_col() +
  ggplot2::coord_flip() +
  ggplot2::facet_wrap(~sentences, ncol = 2) +
  ggplot2::geom_hline(yintercept = 0.5, linetype = "dashed") +
  ggplot2::labs(x = "Model", y = "Probability", title = "Perspective API Results")

You can also use Spanish (es) for TOXICITY, SEVERE_TOXICITY and _EXPERIMENTAL models.

spanish_text <- "gastan en cosas que de nada sirven-nunca tratan de saber la verdad del funcionalismo de nuestro sistema solar y origen del cosmos-falso por Kepler. LAS UNIVERSIDADES DEL MUNDO NO SABEN ANALIZAR VERDAD O MENTIRA-LO QUE DICE KEPLER"


text_scores <- prsp_score(
           text = spanish_text, 
           languages = "es",
           score_model = c("TOXICITY", "SEVERE_TOXICITY", "INSULT_EXPERIMENTAL")
           )

text_scores %>% 
  tidyr::gather() %>% 
  dplyr::mutate(key = forcats::fct_reorder(key, value)) %>% 
  ggplot2::ggplot(ggplot2::aes(key, value)) +
  ggplot2::geom_col() +
  ggplot2::coord_flip() +
  ggplot2::geom_hline(yintercept = 0.5, linetype = "dashed")  +
  ggplot2::labs(x = "Model", y = "Probability", title = "Perspective API Results")

NOTE: Your provided text will be stored by the Perspective API for future research. This option is the default. If the supplied texts are private or any of the authors of the texts are below 13 years old, doNotStore should be set to TRUE.

prsp_stream

So far we have only seen how to get individual comments or sentences scored. But what if you would like to run the function for an entire dataset with a text column? This is where prsp_stream comes in. At its core prsp_stream is a loop implemented within purrr::map to iterate over your text column. To use it let’s first generate a mock tibble.

text_sample <- tibble(
       ctext = c("You wrote this? Wow. This is dumb and childish, please go f**** yourself.",
                 "I don't know what to say about this but it's not good. The commenter is just an idiot",
                 "This goes even further!",
                 "What the hell is going on?",
                 "Please. I don't get it. Explain it again",
                 "Annoying and irrelevant! I'd rather watch the paint drying on the wall!"),
       textid = c("#efdcxct", "#ehfcsct", 
                  "#ekacxwt",  "#ewatxad", 
                  "#ekacswt",  "#ewftxwd")
       )

prsp_stream requires a text and text_id column. It wraps prsp_score and takes all its arguments. Let’s run the most basic version:

text_sample %>%
  prsp_stream(text = ctext,
              text_id = textid,
              score_model = c("TOXICITY", "SEVERE_TOXICITY"))
#> Binding rows...
#> # A tibble: 6 x 3
#>   text_id  TOXICITY SEVERE_TOXICITY
#>   <chr>       <dbl>           <dbl>
#> 1 #efdcxct   0.959           0.846 
#> 2 #ehfcsct   0.932           0.534 
#> 3 #ekacxwt   0.0652          0.0248
#> 4 #ewatxad   0.605           0.319 
#> 5 #ekacswt   0.0734          0.0257
#> 6 #ewftxwd   0.350           0.137

You receive a tibble with your desired scorings including the text_id to match your score with your original dataframe.

Now, the problem is that sometimes the call might fail at some point. It is therefore suggested to set safe_output = TRUE. This will put the function into a purrr::safely environment to ensure that your function will keep running even if you encounter errors.

Let’s try it out with a new dataset that contains text that the Perspective API can’t score

text_sample <- tibble(
       ctext = c("You wrote this? Wow. This is dumb and childish, please go f**** yourself.",
                 "I don't know what to say about this but it's not good. The commenter is just an idiot",
                 ## empty string
                 "",
                 "This goes even further!",
                 "What the hell is going on?",
                 "Please. I don't get it. Explain it again",
                 ## Gibberish
                 "kdlfkmgkdfmgkfmg",
                 "Annoying and irrelevant! I'd rather watch the paint drying on the wall!",
                 ## Gibberish
                 "Hippi Hoppo"),
       textid = c("#efdcxct", "#ehfcsct", 
                  "#ekacxwt",  "#ewatxad", 
                  "#ekacswt",  "#ewftxwd", 
                  "#eeadswt",  "#enfhxed",
                  "#efdmjd")
       )

And run the function with safe_output = TRUE.

text_sample %>%
  prsp_stream(text = ctext,
              text_id = textid,
              score_model = c("TOXICITY", "SEVERE_TOXICITY", "INSULT"),
              safe_output = T)
#> # A tibble: 9 x 5
#>   text_id  error                                TOXICITY SEVERE_TOXICITY  INSULT
#>   <chr>    <chr>                                   <dbl>           <dbl>   <dbl>
#> 1 #efdcxct "No Error"                             0.959           0.846   0.958 
#> 2 #ehfcsct "No Error"                             0.932           0.534   0.960 
#> 3 #ekacxwt "Error in .f(...): HTTP 400\nINVALI~  NA              NA      NA     
#> 4 #ewatxad "No Error"                             0.0652          0.0248  0.0309
#> 5 #ekacswt "No Error"                             0.605           0.319   0.322 
#> 6 #ewftxwd "No Error"                             0.0734          0.0257  0.0396
#> 7 #eeadswt "No Error"                             0.143           0.0871  0.131 
#> 8 #enfhxed "No Error"                             0.350           0.137   0.440 
#> 9 #efdmjd  "Error in .f(...): HTTP 400\nINVALI~  NA              NA      NA

safe_output = T will also provide us with the error messages that occured so that we can check what went wrong!

Finally, there is one last argument: verbose = TRUE. Enable this argument and thanks to crayon you will receive beautiful console output that guides you along the way, showing you errors and text scores as you go.

text_sample %>%
  prsp_stream(text = ctext,
              text_id = textid,
              score_model = c("TOXICITY", "SEVERE_TOXICITY"),
              verbose = T,
              safe_output = T)

Or the (not as pretty) output in Markdown

#> 11.11% [2021-07-13 11:52:36]: 1 out of 9 (11.11%)
#> text_id: #efdcxct
#>  0.96 TOXICITY
#>  0.85 SEVERE_TOXICITY
#> 
#> 22.22% [2021-07-13 11:52:37]: 2 out of 9 (22.22%)
#> text_id: #ehfcsct
#>  0.93 TOXICITY
#>  0.53 SEVERE_TOXICITY
#> 
#> 33.33% [2021-07-13 11:52:38]: 3 out of 9 (33.33%)
#> text_id: #ekacxwt
#> ERROR
#> Error in .f(...): HTTP 400
#> INVALID_ARGUMENT: Comment must be non-empty.
#> NO SCORES
#> 
#> 44.44% [2021-07-13 11:52:39]: 4 out of 9 (44.44%)
#> text_id: #ewatxad
#>  0.07 TOXICITY
#>  0.02 SEVERE_TOXICITY
#> 
#> 55.56% [2021-07-13 11:52:40]: 5 out of 9 (55.56%)
#> text_id: #ekacswt
#>  0.60 TOXICITY
#>  0.32 SEVERE_TOXICITY
#> 
#> 66.67% [2021-07-13 11:52:41]: 6 out of 9 (66.67%)
#> text_id: #ewftxwd
#>  0.07 TOXICITY
#>  0.03 SEVERE_TOXICITY
#> 
#> 77.78% [2021-07-13 11:52:42]: 7 out of 9 (77.78%)
#> text_id: #eeadswt
#>  0.14 TOXICITY
#>  0.09 SEVERE_TOXICITY
#> 
#> 88.89% [2021-07-13 11:52:43]: 8 out of 9 (88.89%)
#> text_id: #enfhxed
#>  0.35 TOXICITY
#>  0.14 SEVERE_TOXICITY
#> 
#> 100.00% [2021-07-13 11:52:45]: 9 out of 9 (100.00%)
#> text_id: #efdmjd
#> ERROR
#> Error in .f(...): HTTP 400
#> INVALID_ARGUMENT: Attribute SEVERE_TOXICITY does not support request languages: ja-Latn
#> NO SCORES
#> # A tibble: 9 x 4
#>   text_id  error                                        TOXICITY SEVERE_TOXICITY
#>   <chr>    <chr>                                           <dbl>           <dbl>
#> 1 #efdcxct "No Error"                                     0.959           0.846 
#> 2 #ehfcsct "No Error"                                     0.932           0.534 
#> 3 #ekacxwt "Error in .f(...): HTTP 400\nINVALID_ARGUME~  NA              NA     
#> 4 #ewatxad "No Error"                                     0.0652          0.0248
#> 5 #ekacswt "No Error"                                     0.605           0.319 
#> 6 #ewftxwd "No Error"                                     0.0734          0.0257
#> 7 #eeadswt "No Error"                                     0.143           0.0871
#> 8 #enfhxed "No Error"                                     0.350           0.137 
#> 9 #efdmjd  "Error in .f(...): HTTP 400\nINVALID_ARGUME~  NA              NA

Citation

Thank you for using peRspective! Please consider citing:

Votta, Fabio. (2019). peRspective: A wrapper for the Perspective API. Source: https://github.com/favstats.

Icons made by Freepik from www.flaticon.com is licensed by CC 3.0 BY