big_and_many_ldas.Rmd

---
title: "Searching the Super Topic Model - Update"
author: "Andreas Blaette, Christoph Leonhardt"
date: "`r Sys.Date()`"
bibliography: "`r path.expand('~/Lab/github/cookbook/references.bib')`"
editor_options: 
  chunk_output_type: console
---

## System requirements

- The code in this recipe has been developed on macOS and is expected to run on
Linux systems. Whether it works on Windows has not yet been tested.

- We use the parallelized approach of [Mallet](https://mimno.github.io/Mallet/index)
for efficiently calculating a set of LDA topic models.

```{r, eval = TRUE}
(cores_to_use <- parallel::detectCores() - 1L)
```

- Computing topic models may require a lot of RAM. Matrix operations when
calculating metrics for topics are particularly excessive. Allocate a lot of 
memory to the Java Virtual Machine (JVM) before loading the rJava package.

```{r, eval = FALSE}
options(java.parameters = "-Xmx28g")
```

- Even a lot of RAM may not be enough for keeping several big LDA models in
memory at the same time. So we store LDA models on disc, which is also a good 
idea for later usage if the process breaks for any reason. Note that the size 
of the model can be several hundred megabytes, or 1 GB and more - depending on 
the size of the data. Make sure you have sufficient disk space.

```{r, eval = TRUE}
outdir <- "~/Data"
```


## Loading packages and R configuration 

- The data we use is a CWB indexed corpus. We use the polmineR package to 
access the data

```{r, eval = TRUE}
library(polmineR) # get dev version, minimum v0.8.6.9008
```

- To access the parallelized topic model of Mallet, we have written the biglda
R package. It includes an installation mechanism to get the Mallet Java code.

```{r, eval = TRUE}
library(biglda) # get dev version from PolMine/biglda, minimum v0.0.0.9006
if (!mallet_is_installed()) mallet_install()
```

- For robust file operations, we like to use functionality from the
[fs](https://CRAN.R-project.org/package=fs) package.

```{r, eval = TRUE}
library(fs)
```

- Finally, we use the visualisation of topic model metrics of the
[ldatuning](https://CRAN.R-project.org/package=ldatuning) package. Note that the
biglda package includes the implementation of more efficient calculations for
the metrics used.

```{r}
library(ldatuning)
```

- Concerning configuration, note that the computation of metrics benefits 
substantially from the installation of a BLAS optimized for your system. See [this](https://cran.r-project.org/bin/macosx/RMacOSX-FAQ.html) instruction for
using the optimized BLAS on macOS.


## Defining the input documents 

### Get speeches as character vectors

- We use speeches covered by theGermaParl corpus (beta version) as an example.
To be more precise, we use a subset covering the period between 1987 and 2021.

- Here we choose to split up the corpus into speeches with a gap of 50. We end
up with more shorter speeches than with the default gap of 500 which
concatenates more utterances from the same speaker.

- We apply a minimum length: The substantive argument would be that short
documents are probably not topical, for example because they are mainly from
presidential speakers and thus procedural. The technical argument would be that
they do contain little information and thus the model's inference could be
random.

- Often, min_length of 100 seems alright if removing more procedural utterances
is the goal. Empirically: min_length of 100L removes about half the speeches
which is more or less expected when every other speech is the introduction of a
new speaker by the president. To keep some of the more substantial short
contributions here, we set this value to 50 for now, given the shorter gap.
Interestingly, this still filters out about a half of the speeches, perhaps
indicating a break-off point here which splits longer substantive speeches and
shorter ones.


```{r make_speeches, echo = FALSE, eval = FALSE}
speeches <- corpus("GERMAPARL_BETA") %>%
  as.speeches(
    s_attribute_name = "speaker",
    subset = {date >= as.Date("1987-01-01")},
    gap = 50, # gap is 500 by default
    mc = cores_to_use
  ) %>% 
  get_token_stream(p_attribute = "word", min_length = 50L, collapse = " ")
```

### Define stopwords

- There is a debate if, how and when to remove stopwords before fitting topic
models [e.g. @schofield-etal-2017-pulling]. Here, we remove some stopwords,
mainly to speed up the estimation and to somewhat minimize potential memory
limitations associated with larges matrices.

- The stopword list we use is based on the German stopword list in the tm R
package [@tm].

```{r, eval = FALSE}
## Get German Stopwords 
stopwords <- c(
  tm::stopwords("de"),
  paste(
    toupper(substr(tm::stopwords("de"), 1, 1)),
    substr(tm::stopwords("de"), 2, nchar(tm::stopwords("de"))),
    sep = ""
  )
)

punctuation_to_drop <- grep(
  "^[[:punct:]]+$",
  terms("GERMAPARL_BETA", p_attribute = "word"),
  value = TRUE,
  perl = TRUE
)

terms_to_drop <- c(
  stopwords,
  punctuation_to_drop,
  c("dass", "Dass", "Damen", "Herren", "Beifall")
)
```


### Creating the Instance List

- Everything is ready and set now to transfer the data to Java

```{r make instance_list}
instance_list <- as.instance_list(speeches, stopwords = terms_to_drop)
rm(speeches); gc()
```


### Fitting Topic Models

```{r fit_models, eval = FALSE}
for (k in c(200, 250, 300, 350, 450, 500)){
  message("... fitting model for k: ", k)
  fname <- sprintf("%s_%s_%s.bin", "GERMAPARL_BETA", Sys.Date(), k)
  outfile <- path.expand(fs::path(outdir, fname))
  
  lda <- BigTopicModel(n_topics = k, alpha_sum = 5.1, beta = 0.1)
  lda$addInstances(instance_list)
  lda$setNumThreads(cores_to_use)
  lda$setNumIterations(2000L)
  lda$setTopicDisplay(100L, 10L) 
  lda$estimate()
  lda$write(rJava::.jnew("java/io/File", outfile))
  rm(lda)
}
```


## Evaluating topic models

```{r calc_metrics, eval = FALSE}
metrics <- lapply(
  Sys.glob(fs::path(outdir, "*.bin")),
  function(fname){
    rJava::J("java/lang/Runtime")$getRuntime()$gc()
    
    message("... reading: ", fname)
    mallet_model <- mallet_load_topicmodel(fname)
    
    message("cao")
    cao2009 <- FastCao2009(mallet_model)
    
    message("deveaud")
    deveaud2014 <- FastDeveaud2014(mallet_model, cl = cores_to_use)
    
    message("arun")
    arun2010 <- FastArun2010(mallet_model)
    
    data.frame(
      Cao2009 = cao2009,
      Deveaud2014 = deveaud2014,
      Arun2010 = arun2010
    )
  }
)
model_metrics <- do.call(rbind, metrics)
model_metrics$topics <- as.integer(gsub("^.*_(\\d+)\\.bin$", "\\1", Sys.glob(fs::path(outdir, "*.bin"))))
model_metrics <- model_metrics[, c("topics", "Cao2009", "Deveaud2014", "Arun2010")]
write.table(x = model_metrics, file = fs::path(outdir, "model_metrics.csv"))
```

```{r, eval = TRUE}
if (!exists("model_metrics")) model_metrics <- read.table(file = fs::path(outdir, "model_metrics.csv"))
ldatuning::FindTopicsNumber_plot(model_metrics)
```


## Annex: Data transformation

```{r, eval = FALSE}
for (fname in Sys.glob(fs::path(outdir, "*.bin"))){
  message("... reading: ", fname)
  mallet_model <- mallet_load_topicmodel(fname)
  
  message("... processing model")
  lda <- as_LDA(mallet_model)
  
  ldafile <- sprintf("%s.rds", tools::file_path_sans_ext(fname))
  message("... writing: ", ldafile)
  saveRDS(lda, file = ldafile)
}
```