predict_responses.Rmd

---
title: "Twitter Response Generator"
output: html_document
---

```{r setup, include=FALSE}
# Required R package installation:
# These will install packages if they are not already installed
# Set the correct default repository
r = getOption("repos")
r["CRAN"] = "http://cran.rstudio.com"
options(repos = r)


if (!require("knitr")) {
  install.packages("knitr")
  library(knitr)
}

if (!require("kableExtra")) {
  install.packages("kableExtra")
  library(kableExtra)
}

if (!require("httr")) {
  install.packages("httr")
  library(httr)
}

knitr::opts_chunk$set(echo = TRUE)

source("text_helpers.R")
```

```{r, echo=FALSE}
response_sampler_url <- "http://localhost:8080/batchsampleresponses"
embedding_type <- "sbert"
embedding_dims <- if (embedding_type=="use_large") 512 else 384
```

### Predict responses to a prompt tweet:

```{r}
# Prompt author (e.g., WHO, CDCgov, CDCDirector, etc...)
prompt_authors <- list("WHO", "CDCgov", "CDCDirector")#, "ECDC_EU")
# Prompt message
prompt_messages <- list("4 steps to stay fit to beat #COVID19: eat a healthy diet #notobacco be physically active stop harmful use of alcohol",
                        "Thank you for doing your part beat #COVID19! We are excited to share these 4 steps anybody can take to stay fit against #COVID19: eat a healthy diet, reduce tobacco use, be physically active, and drink alcohol in moderation.")

# Number of responses to predict
response_sample_size <- 30
output_latex <- FALSE

# Generation hyperparams
num_beams <- 3
temperature <- 1.5
random_seed <- 42
```


```{r, echo=FALSE}
sample_model_responses <- function(prompt_author, prompt_message, response_sample_size, num_beams, 
                                   temperature, random_seed) {
  
  body <- list(sample_size = response_sample_size, 
               num_beams = num_beams,
               temperature = temperature,
               prompts = list(list(author=prompt_author, message=clean_text(prompt_message))))
  
  if (isFALSE(is.null(random_seed)) && isFALSE(is.na(random_seed))) {
    body["random_state"] <- random_seed
  }
  
  res <- POST(url=response_sampler_url, encode="json", body=body)
  
  res.list <- content(res)
  model_responses  <- lapply(1:length(res.list), function(i) {
    return (list(responses=data.frame(quoted_status.screen_name=body$prompts[[i]]$author,
                                      quoted_status.full_text=body$prompts[[i]]$message,
                                      full_text=unlist(res.list[[i]][[1]]),
                                      sentiment=unlist(res.list[[i]][[3]])),
                 responses.vectors=matrix(sapply(do.call("rbind", res.list[[i]][[2]]), function(x) x), 
                                          nrow=response_sample_size, ncol=embedding_dims)))
  })
  return(model_responses)
}
```

```{r, echo=FALSE}
if (output_latex){
  output <- list()
} else {
  output <- htmltools::tagList()
}

output_idx <- 1
for (prompt_message in prompt_messages) {
  for (prompt_author in prompt_authors) {
    results <- sample_model_responses(prompt_author, prompt_message, response_sample_size, num_beams, 
                                      temperature, random_seed)
    avg_sentiment <- mean(results[[1]]$responses$sentiment)
    sd_sentiment <- sd(results[[1]]$responses$sentiment)
    caption <- paste0("Responses to ", prompt_author, ": ", prompt_message, " (mean: ", round(avg_sentiment, 3), 
                      "; sd: ", round(sd_sentiment, 3), ")")
    if (output_latex) {
      results_latex <- kable(results[[1]]$responses[,c("full_text", "sentiment")], "latex",  
                             col.names=c("Generated Responses:", "Sentiment:"), caption=caption, booktabs=TRUE)
      output[[output_idx]] <- results_latex
    } else {
      results_table <- kable(results[[1]]$responses[,c("full_text", "sentiment")], 
                             col.names=c("Generated Responses:", "Sentiment:"), caption=caption) %>% 
        kable_styling()
      output[[output_idx]] <- htmltools::HTML(results_table)
    }
    output_idx <- output_idx + 1
  }
}
output
```