babi_rnn

Trains two recurrent neural networks based upon a story and a question. The resulting merged vector is then queried to answer a range of bAbI tasks.

The results are comparable to those for an LSTM model provided in Weston et al.: “Towards AI-Complete Question Answering: A Set of Prerequisite Toy Tasks” http://arxiv.org/abs/1502.05698

Task Number FB LSTM Baseline Keras QA
QA1 - Single Supporting Fact 50 100.0
QA2 - Two Supporting Facts 20 50.0
QA3 - Three Supporting Facts 20 20.5
QA4 - Two Arg. Relations 61 62.9
QA5 - Three Arg. Relations 70 61.9
QA6 - yes/No Questions 48 50.7
QA7 - Counting 49 78.9
QA8 - Lists/Sets 45 77.2
QA9 - Simple Negation 64 64.0
QA10 - Indefinite Knowledge 44 47.7
QA11 - Basic Coreference 72 74.9
QA12 - Conjunction 74 76.4
QA13 - Compound Coreference 94 94.4
QA14 - Time Reasoning 27 34.8
QA15 - Basic Deduction 21 32.4
QA16 - Basic Induction 23 50.6
QA17 - Positional Reasoning 51 49.1
QA18 - Size Reasoning 52 90.8
QA19 - Path Finding 8 9.0
QA20 - Agent’s Motivations 91 90.7

For the resources related to the bAbI project, refer to: https://research.facebook.com/researchers/1543934539189348

Notes:

  • With default word, sentence, and query vector sizes, the GRU model achieves:
  • 100% test accuracy on QA1 in 20 epochs (2 seconds per epoch on CPU)
  • 50% test accuracy on QA2 in 20 epochs (16 seconds per epoch on CPU) In comparison, the Facebook paper achieves 50% and 20% for the LSTM baseline.

  • The task does not traditionally parse the question separately. This likely improves accuracy and is a good example of merging two RNNs.

  • The word vector embeddings are not shared between the story and question RNNs.

  • See how the accuracy changes given 10,000 training samples (en-10k) instead of only 1000. 1000 was used in order to be comparable to the original paper.

  • Experiment with GRU, LSTM, and JZS1-3 as they give subtly different results.

  • The length and noise (i.e. ‘useless’ story components) impact the ability for LSTMs / GRUs to provide the correct answer. Given only the supporting facts, these RNNs can achieve 100% accuracy on many tasks. Memory networks and neural networks that use attentional processes can efficiently search through this noise to find the relevant statements, improving performance substantially. This becomes especially obvious on QA2 and QA3, both far longer than QA1.

library(keras)
library(readr)
library(stringr)
library(purrr)
library(tibble)
library(dplyr)

# Function definition -----------------------------------------------------

tokenize_words <- function(x){
  x <- x %>% 
    str_replace_all('([[:punct:]]+)', ' \\1') %>% 
    str_split(' ') %>%
    unlist()
  x[x != ""]
}

parse_stories <- function(lines, only_supporting = FALSE){
  lines <- lines %>% 
    str_split(" ", n = 2) %>%
    map_df(~tibble(nid = as.integer(.x[[1]]), line = .x[[2]]))
  
  lines <- lines %>%
    mutate(
      split = map(line, ~str_split(.x, "\t")[[1]]),
      q = map_chr(split, ~.x[1]),
      a = map_chr(split, ~.x[2]),
      supporting = map(split, ~.x[3] %>% str_split(" ") %>% unlist() %>% as.integer()),
      story_id = c(0, cumsum(nid[-nrow(.)] > nid[-1]))
    ) %>%
    select(-split)
  
  stories <- lines %>%
    filter(is.na(a)) %>%
    select(nid_story = nid, story_id, story = q)
  
  questions <- lines %>%
    filter(!is.na(a)) %>%
    select(-line) %>%
    left_join(stories, by = "story_id") %>%
    filter(nid_story < nid)

  if(only_supporting){
    questions <- questions %>%
      filter(map2_lgl(nid_story, supporting, ~.x %in% .y))
  }
    
  questions %>%
    group_by(story_id, nid, question = q, answer = a) %>%
    summarise(story = paste(story, collapse = " ")) %>%
    ungroup() %>% 
    mutate(
      question = map(question, ~tokenize_words(.x)),
      story = map(story, ~tokenize_words(.x)),
      id = row_number()
    ) %>%
    select(id, question, answer, story)
}

vectorize_stories <- function(data, vocab, story_maxlen, query_maxlen){
  
  questions <- map(data$question, function(x){
    map_int(x, ~which(.x == vocab))
  })
  
  stories <- map(data$story, function(x){
    map_int(x, ~which(.x == vocab))
  })
  
  # "" represents padding
  answers <- sapply(c("", vocab), function(x){
    as.integer(x == data$answer)
  })
  

  list(
    questions = pad_sequences(questions, maxlen = query_maxlen),
    stories   = pad_sequences(stories, maxlen = story_maxlen),
    answers   = answers
  )
}

# Parameters --------------------------------------------------------------

max_length <- 99999
embed_hidden_size <- 50
batch_size <- 32
epochs <- 40

# Data Preparation --------------------------------------------------------

path <- get_file(
  fname = "babi-tasks-v1-2.tar.gz",
  origin = "https://s3.amazonaws.com/text-datasets/babi_tasks_1-20_v1-2.tar.gz"
)
untar(path, exdir = str_replace(path, fixed(".tar.gz"), "/"))
path <- str_replace(path, fixed(".tar.gz"), "/")

# Default QA1 with 1000 samples
# challenge = '%stasks_1-20_v1-2/en/qa1_single-supporting-fact_%s.txt'
# QA1 with 10,000 samples
# challenge = '%stasks_1-20_v1-2/en-10k/qa1_single-supporting-fact_%s.txt'
# QA2 with 1000 samples
challenge <- "%stasks_1-20_v1-2/en/qa2_two-supporting-facts_%s.txt"
# QA2 with 10,000 samples
# challenge = '%stasks_1-20_v1-2/en-10k/qa2_two-supporting-facts_%s.txt'

train <- read_lines(sprintf(challenge, path, "train")) %>%
  parse_stories() %>%
  filter(map_int(story, ~length(.x)) <= max_length)

test <- read_lines(sprintf(challenge, path, "test")) %>%
  parse_stories() %>%
  filter(map_int(story, ~length(.x)) <= max_length)

# extract the vocabulary
all_data <- bind_rows(train, test)
vocab <- c(unlist(all_data$question), all_data$answer, 
           unlist(all_data$story)) %>%
  unique() %>%
  sort()

# Reserve 0 for masking via pad_sequences
vocab_size <- length(vocab) + 1
story_maxlen <- map_int(all_data$story, ~length(.x)) %>% max()
query_maxlen <- map_int(all_data$question, ~length(.x)) %>% max()

# vectorized versions of training and test sets
train_vec <- vectorize_stories(train, vocab, story_maxlen, query_maxlen)
test_vec <- vectorize_stories(test, vocab, story_maxlen, query_maxlen)

# Defining the model ------------------------------------------------------

sentence <- layer_input(shape = c(story_maxlen), dtype = "int32")
encoded_sentence <- sentence %>% 
  layer_embedding(input_dim = vocab_size, output_dim = embed_hidden_size) %>%
  layer_dropout(rate = 0.3)

question <- layer_input(shape = c(query_maxlen), dtype = "int32")
encoded_question <- question %>%
  layer_embedding(input_dim = vocab_size, output_dim = embed_hidden_size) %>%
  layer_dropout(rate = 0.3) %>%
  layer_lstm(units = embed_hidden_size) %>%
  layer_repeat_vector(n = story_maxlen)

merged <- list(encoded_sentence, encoded_question) %>%
  layer_add() %>%
  layer_lstm(units = embed_hidden_size) %>%
  layer_dropout(rate = 0.3)

preds <- merged %>%
  layer_dense(units = vocab_size, activation = "softmax")

model <- keras_model(inputs = list(sentence, question), outputs = preds)
model %>% compile(
  optimizer = "adam",
  loss = "categorical_crossentropy",
  metrics = "accuracy"
)

model

# Training ----------------------------------------------------------------

model %>% fit(
  x = list(train_vec$stories, train_vec$questions),
  y = train_vec$answers,
  batch_size = batch_size,
  epochs = epochs,
  validation_split=0.05
)

evaluation <- model %>% evaluate(
  x = list(test_vec$stories, test_vec$questions),
  y = test_vec$answers,
  batch_size = batch_size
)

evaluation