sunflower is a package designed to assist clinicians and researchers in the fields of Speech Therapy and Neuropsychology of Language. Its primary goal is to facilitate the management of multiple response data and compute formal similarity indices to assess the quality of oral and written productions in patients with aphasia and related disorders, such as apraxia of speech, in Spanish. Additionally, the package allows for the classification of these productions according to classical typologies in the field, prior to computing formal and semantic similarity measures. For the computation of the latter, sunflower partially relies on natural language processing models such as word2vec. The outputs provided by this package facilitate statistical analyses in R, a widely-used tool in the field for data wrangling, visualization, and analysis.

sunflower can be installed as an R package with:

install.packages("devtools")
devtools::install_github("ismaelgutier/sunflower")

The sunflower package works using the pipe operator (%>%) from the tidyverse package, allowing it to work seamlessly with functions from other packages in the tidyverse, such as dplyr for data wrangling, readr for data reading, and ggplot2 for data visualization. This can significantly enhance our workflow.

Once installed, we only need to load the sunflower package. However, as previously mentioned, the tidyverse package can also be valuable for other complementary tasks.

require("sunflower")
require("tidyverse")

df_to_formal_metrics = sunflower::IGC_long_phon %>% 
  dplyr::select(-c(modality, task_modality,task_type, test, task)) # keep some relevant columns


formal_metrics_computed = df_to_formal_metrics %>% 
    get_formal_similarity(item_col = "item", 
                          response_col = "response",
                          attempt_col = "Attempt",
                          group_cols = c("ID", "item_ID"))
#> The function get_formal_similarity() took 1.75 seconds to be executed

formal_metrics_computed %>% head(8) %>% knitr::kable()

Note: Move the dataframe to the right to see all the columns and metrics.

positions_accuracy = formal_metrics_computed %>% 
  positional_accuracy(match_col = "itemL_adj_strict_match_pos", last_ID_col = "targetL")

positions_accuracy %>% head(8) %>% knitr::kable()

Note. A plot depicting the positions’ accuracy of 14,418 datapoints.

errors_classified = df_to_classify %>% 
  check_lexicality(item_col = "item", response_col = "Response", criterion = "database") %>%
  get_formal_similarity(item_col = "item", response_col = "Response", 
                        attempt_col = "Attempt", group_cols = c("ID", "item_ID")) %>%
  get_semantic_similarity(item_col = "item", response_col = "Response", model = m_w2v) %>%
  classify_errors(response_col = "Response", item_col = "item",
                  access_col = "accessed", RA_col = "RA", also_classify_RAs = T)
#> The function check_lexicality() took 2.60 seconds to be executed
#> The function get_formal_similarity() took 3.40 seconds to be executed
#> The function get_semantic_similarity() took 3.86 seconds to be executed
#> The function classify_errors() took 3.89 seconds to be executed

errors_classified %>% head(8) %>% knitr::kable()

Notes. Move the dataframe to the right to see all the columns and errors.

sunflower allows for the classification of production errors once some indexes related to responses to a stimulus have been obtained and contextualized based on whether they come from repeated attempts or single productions. This process involves three steps. First, a lexicality check of the response is performed using the lexicality_check() function, which involves determining whether the response is a real word. To do this, the package searches for the response in a database such as BuscaPalabras (BPal) and compares its frequency with the target word to determine if it is a real word based on whether it has a higher frequency or not when the parameter criterion = "database" is set. Alternatively, the response can be checked against a dictionary (sunflower searches for responses among entries from the Real Academia Española, RAE) when the parameter criterion = "dictionary" is used.

Next, similarity measures between the targets and the responses are obtained using various algorithms within the get_formal_similarity() function. Finally, the cosine similarity between the two productions is computed if possible using the get_semantic_similarity() function, based on an NLP model. In our case, the parameter model = m_w2v refers to a binary file containing a Spanish Billion Words embeddings corpus created using the word2vec algorithm. This file is included in the zip file (for more information, see the markdown in the vignettes) located within the dependency-bundle zip, which can be found in our supplementary OSF repository mirror.

Thanks to Cristian Cardellino for making his work on the Spanish Billion Word Corpus and Embeddings publicly available.

Any suggestions, comments, or questions about the package’s functionality are warmly welcomed. If you’d like to contribute to the project, please feel free to get in touch. 🌻