session3.Rmd

---
title: "Session 3 : Topic model"
author: "cb,jcrm,bc,oc"
date: "`r Sys.Date()`"
output: 
  html_document :
    toc: true
    toc_float: true
    toc_depth: 3
bibliography: [PLSweek2022.bib]
---

# Tools

```{r setup, include=TRUE, message=FALSE, warning=FALSE}
knitr::opts_chunk$set(echo = TRUE, include=TRUE, message=FALSE, warning=FALSE)
#from session 1 & 2
library(tidyverse)
library(udpipe)
library(flextable)
library(cowplot)
library(quanteda)
library(quanteda.textmodels)
library(quanteda.textstats)
library(quanteda.textplots)
library(syuzhet)  #analyse du sentimeent

# new for session 2
library(FactoMineR)
library(factoextra)
library(igraph)
library(ggwordcloud)
library(ggrepel)

library(Rtsne)
library(tidytext)

# new for session 3
library(cleanNLP)
library(text2vec)


theme_set(theme_minimal())

t1=Sys.time()

```

# The basic LDA

proposed by @blei_latent_2003 (nearly 20 years ago!)

## A few theory

![LDA](./Images/LDA-algorithm.png)

## Application

Prepare the data (a TfIdf approach with `cleanNLP` elegant function)

```{r 01, fig.width=12}
UD<-readRDS("./Data/UD.rds")

#library(cleanNLP) #an other technics for tf_idf
tf <- UD%>%
  filter(upos %in% c("NOUN", "VERB")) %>%
  cnlp_utils_tfidf(min_df = 0.05, max_df = 0.95, tf_weight = "raw")
```

The LDA model with `text2vec`

```{r 02, fig.width=12}
#library(text2vec)
lda_model = LDA$new(n_topics = 12, doc_topic_prior = 0.1, topic_word_prior = 0.01)

set.seed(67) #pour la reproducibilité des résultats
#On définit les paramètres du processus d'estimation : 
##n_iter = le nombre d'itérations
##convergence_tol =le seuil de convergence
doc_topic_distr = 
  lda_model$fit_transform(x = tf, 
                          n_iter = 1000, 
                          convergence_tol = 0.001, 
                          n_check_convergence = 25, 
                          progressbar = TRUE)

```

Topic description

```{r 03, fig.width=12}
lda_res<-as.data.frame(lda_model$get_top_words(n = 15, lambda = 0.30))
lda_res$rank<-as.numeric(row.names(lda_res))

lda_res<-lda_res%>% gather(variable, value, -rank)

ggplot(lda_res, aes(x=variable, y= rank, group =  value , label = value)) + 
  scale_y_reverse() + 
  geom_text(aes(color=variable,size=sqrt(26-rank)))+
  scale_color_hue()+
  guides(color="none",size="none")+
  labs(x="topics", y="par ordre de pertinence")

# un nuage de mot plutôt
```

A more interactive description (that we cannot edit here... need to be in rstudio)

```{r 04, fig.width=12}
#description des topics en fonction d'un degré de pertinence = lamba ( lambda =1 probabilité d'obtenir le terme sachant le topic)

library(LDAvis)
lda_model$plot() #mode interactif

```

## the optimal number of topics

No optimal analytics methods. Okham rule. 

but some systematic approachs by comparing index for different solutions. 

### Perplexity

http://qpleple.com/perplexity-to-evaluate-topic-models/

### package  `ldatuning`


# Some other models

Topics model are popular, and quickly new models were suggested. 

## STM : regression like

Introduce de control a third variable, as a regression approach model, very usefull to introduce a time parameter. But also let the topic being correlated, which open to hierarchical topics

## semi-supervised lda

We could help the solution by predefining key tokens. Seed_lda


# Notes

Beware to computing time! Best to sample for testing. (come back to the beginning)

```{r 20}
t2=Sys.time()
t<- t2-t1
print(t)
```


See you to later and [go to session 4](https://benaventc.github.io/NLP_lecture_PSLWeek/session4.html)


# References