273 lines
9.9 KiB
Nim
273 lines
9.9 KiB
Nim
# Copyright 2021 Mattia Giambirtone
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
|
|
# Module to extract features from textual datasets
|
|
import strutils
|
|
import sequtils
|
|
import tables
|
|
import math
|
|
import sets
|
|
import normalize
|
|
|
|
|
|
import ../matrix
|
|
import emoji
|
|
|
|
|
|
type
|
|
TextPreprocessor* = ref object of RootObj
|
|
## A preprocessor for textual datasets.
|
|
## Assumes the input is a sequence of
|
|
## sentences with space as a separator
|
|
corpus: HashSet[string]
|
|
stopwords: HashSet[string]
|
|
stripPunctuation: bool
|
|
toLower: bool
|
|
normalize: bool
|
|
TextVectorizer* = ref object of RootObj
|
|
## A base type for all text vectorizers, defines
|
|
## the common interface
|
|
corpus: HashSet[string]
|
|
features: HashSet[string]
|
|
vocab: Table[string, int]
|
|
maxDf: float
|
|
minDf: float
|
|
sublinearTf: bool
|
|
preprocessor: TextPreprocessor
|
|
CountVectorizer* = ref object of TextVectorizer
|
|
## Vectorizes a textual dataset using word
|
|
## counts as weights
|
|
TFIDFVectorizer* = ref object of TextVectorizer
|
|
## Vectorizes a textual dataset using smoothed
|
|
## TF-IDF values as weights
|
|
smoothIdf: bool
|
|
|
|
|
|
proc toHashSet[T](m: Matrix[T]): HashSet[T] =
|
|
result = initHashSet[T]()
|
|
for row in m:
|
|
for element in row:
|
|
result.incl(element)
|
|
|
|
|
|
proc newTextPreprocessor*(stopwords: Matrix[string], stripPunctuation: bool, toLower: bool, normalize: bool): TextPreprocessor =
|
|
## Initializes a new TextPreprocessor object
|
|
new(result)
|
|
result.stopwords = toHashSet(stopwords)
|
|
result.stripPunctuation = stripPunctuation
|
|
result.toLower = toLower
|
|
result.normalize = normalize
|
|
|
|
|
|
proc newTextVectorizer*(preprocessor: TextPreprocessor, minDf, maxDf: float, sublinearTf: bool): TextVectorizer =
|
|
## Initializes a new TextVectorizer object
|
|
new(result)
|
|
result.minDf = minDf
|
|
result.maxDf = maxDf
|
|
result.sublinearTf = sublinearTf
|
|
result.preprocessor = preprocessor
|
|
|
|
|
|
proc newCountVectorizer*(preprocessor: TextPreprocessor, minDf, maxDf: float, sublinearTf: bool): CountVectorizer =
|
|
## Initializes a new CountVectorizer object
|
|
new(result)
|
|
result.minDf = minDf
|
|
result.maxDf = maxDf
|
|
result.sublinearTf = sublinearTf
|
|
result.preprocessor = preprocessor
|
|
|
|
|
|
proc newTFIDFVectorizer*(preprocessor: TextPreprocessor, minDf, maxDf: float, smoothIdf: bool, sublinearTf: bool): TFIDFVectorizer =
|
|
## Initializes a new TFIDFVectorizer object
|
|
new(result)
|
|
result.minDf = minDf
|
|
result.maxDf = maxDf
|
|
result.smooth_idf = smooth_idf
|
|
result.sublinearTf = sublinearTf
|
|
result.preprocessor = preprocessor
|
|
|
|
|
|
proc fit*(self: TextPreprocessor, corpus: Matrix[string]) =
|
|
## Fits the preprocessor to the given corpus
|
|
self.corpus = toHashSet(corpus)
|
|
|
|
|
|
proc transform*(self: TextPreprocessor, x: Matrix[string]): Matrix[string] =
|
|
## Tranforms the data in the vector X according to
|
|
## the given initialization parameters
|
|
var res: seq[string] = @[]
|
|
var stripped: string = ""
|
|
var emojizer = newEmojizer()
|
|
for document in x[0]:
|
|
stripped = document
|
|
if self.normalize:
|
|
stripped = emojizer.demojize(stripped, strip=true)
|
|
stripped = toNFKC(stripped)
|
|
if self.toLower:
|
|
stripped = stripped.toLowerAscii()
|
|
if self.stripPunctuation:
|
|
stripped = stripped.multiReplace(("'", ""), ("!", ""), ("\"", ""), ("#", ""), ("$", ""), ("%", ""), ("&", ""), ("\\", ""), ("(", ""), (")", ""), ("*", ""), ("+", ""), (",", ""), ("-", ""), (".", ""), ("/", ""), (":", ""), (";", ""), ("<", ""), ("=", ""), (">", ""), ("?", ""), ("@", ""), ("[", ""), ("]", ""), ("_", ""), ("`", ""), ("{", ""), ("|", ""), ("}", ""), ("~", ""))
|
|
res.add(join(filter(stripped.strip().split(), proc (s: string): bool = s != "" and s notin self.stopwords), " "))
|
|
result = newMatrix(res)
|
|
|
|
|
|
proc transform*(self: TextPreprocessor, x: string): string =
|
|
## Tranforms the string X according to
|
|
## the given initialization parameters
|
|
var stripped: string = x
|
|
var emojizer = newEmojizer()
|
|
if self.normalize and not isNFKC(stripped):
|
|
stripped = emojizer.demojize(stripped, strip=true)
|
|
stripped = toNFKC(stripped)
|
|
if self.toLower:
|
|
stripped = stripped.toLowerAscii()
|
|
if self.stripPunctuation:
|
|
stripped = stripped.multiReplace(("'", ""), ("!", ""), ("\"", ""), ("#", ""), ("$", ""), ("%", ""), ("&", ""), ("\\", ""), ("(", ""), (")", ""), ("*", ""), ("+", ""), (",", ""), ("-", ""), (".", ""), ("/", ""), (":", ""), (";", ""), ("<", ""), ("=", ""), (">", ""), ("?", ""), ("@", ""), ("[", ""), ("]", ""), ("_", ""), ("`", ""), ("{", ""), ("|", ""), ("}", ""), ("~", ""))
|
|
result.add(join(filter(stripped.strip().split(), proc (s: string): bool = s != "" and s notin self.stopwords), " "))
|
|
|
|
|
|
proc fitTransform*(self: TextPreprocessor, corpus, x: Matrix[string]): Matrix[string] =
|
|
## Shorthand for fit() and transform()
|
|
self.fit(corpus)
|
|
result = self.transform(x)
|
|
|
|
|
|
proc termFrequency(self: TextVectorizer, term, document: string): float =
|
|
## Calculates the frequency of a given term
|
|
## in our corpus
|
|
|
|
result = document.count(term) / document.len()
|
|
if self.sublinearTf:
|
|
result = 1 + ln(result)
|
|
|
|
|
|
proc documentFrequency(self: TextVectorizer, term: string): float =
|
|
## Computes the document frequency of a given term in our corpus
|
|
var x = 0
|
|
for document in self.corpus:
|
|
if term in document:
|
|
x += 1
|
|
result = x / self.corpus.len()
|
|
|
|
|
|
proc inverseDocumentFrequency(self: TFIDFVectorizer, term: string): float =
|
|
## Computes the inverse document frequency of a given term in the corpus
|
|
var n = self.corpus.len()
|
|
var df = self.documentFrequency(term)
|
|
if self.smooth_idf:
|
|
n += 1
|
|
df += 1
|
|
result = ln(float(n) / df) + 1 # This constant addition makes sure that
|
|
# even words appearing in all documents are not ignored completely
|
|
|
|
|
|
proc fit*(self: TextVectorizer, corpus: Matrix[string]) =
|
|
## Fits the vectorizer according to the given corpus
|
|
self.corpus = toHashSet(corpus)
|
|
var preprocessor = TextPreprocessor(self.preprocessor)
|
|
preprocessor.fit(corpus)
|
|
for document in corpus[0]:
|
|
for word in preprocessor.transform(document).split():
|
|
self.features.incl(word)
|
|
# self.vocab is needed to build the matrix later!
|
|
# It maps a word to the column where it's supposed to go
|
|
# in our matrix when transform() is called
|
|
var index: int
|
|
var df: float
|
|
self.vocab = initTable[string, int]()
|
|
for word in self.features:
|
|
df = self.documentFrequency(word)
|
|
if df > self.minDf and df < self.maxDf:
|
|
# Here we exclude words that occur too
|
|
# frequently and the ones that occur too
|
|
# rarely to further reduce potential biases.
|
|
# This can be seen as a way of "detecting" stopwords
|
|
# beyond the provided ones
|
|
self.vocab[word] = index
|
|
index += 1
|
|
|
|
|
|
proc getFeatureNames*(self: TextVectorizer): Matrix[string] =
|
|
## Returns the list of analyzed features
|
|
var res: seq[string] = @[]
|
|
for feature in self.features:
|
|
res.add(feature)
|
|
result = newMatrix(res)
|
|
|
|
|
|
proc getVocabulary*(self: TextVectorizer): Matrix[string] =
|
|
## Returns the vocabulary of the vectorizer
|
|
var res: seq[string] = @[]
|
|
for word in self.vocab.keys():
|
|
res.add(word)
|
|
result = newMatrix(res)
|
|
|
|
|
|
proc count(self: CountVectorizer, word: string): int =
|
|
## Counts the occurrences of a word in our corpus
|
|
for document in self.corpus:
|
|
result += document.count(word)
|
|
|
|
|
|
proc transform*(self: CountVectorizer, x: Matrix[string]): Matrix[float] =
|
|
## Transforms the corpus into a bidimensional matrix
|
|
## of shape (len(X), len(self.vocab))
|
|
# We initialize the matrix with zeros!
|
|
# This is basically a sparse matrix
|
|
var res = newSeqOfCap[seq[float]](len(x) * len(self.vocab))
|
|
var x = TextPreprocessor(self.preprocessor).transform(x)
|
|
for r in 0..<len(x):
|
|
res.add(@[])
|
|
for c in 0..<len(self.vocab):
|
|
res[^1].add(0.0)
|
|
for i, document in x[0]:
|
|
for word in document.split(" "):
|
|
if word in self.vocab:
|
|
res[i][self.vocab[word]] = float(self.count(word))
|
|
result = newMatrix(res)
|
|
|
|
|
|
proc fitTransform*(self: CountVectorizer, corpus, x: Matrix[string]): Matrix[float] =
|
|
## Shorthand for fit() and transform()
|
|
self.fit(corpus)
|
|
result = self.transform(x)
|
|
|
|
|
|
proc transform*(self: TFIDFVectorizer, x: Matrix[string]): Matrix[float] =
|
|
## Transforms the corpus into a bidimensional matrix
|
|
## of shape (len(X), len(self.vocab))
|
|
# We initialize the matrix with zeros!
|
|
# This is basically a sparse matrix
|
|
var x = TextPreprocessor(self.preprocessor).transform(x)
|
|
var tf: float
|
|
var idf: float
|
|
var res = newSeqOfCap[seq[float]](len(x) * len(self.vocab))
|
|
for r in 0..<len(x):
|
|
res.add(@[])
|
|
for c in 0..<len(self.vocab):
|
|
res[^1].add(0.0)
|
|
for i, document in x[0]:
|
|
for word in document.split(" "):
|
|
if word in self.vocab:
|
|
tf = self.termFrequency(word, document)
|
|
idf = self.inverseDocumentFrequency(word)
|
|
res[i][self.vocab[word]] = tf * idf
|
|
result = newMatrix(res)
|
|
|
|
|
|
proc fitTransform*(self: TFIDFVectorizer, corpus, x: Matrix[string]): Matrix[float] =
|
|
## Shorthand for fit() and transform()
|
|
self.fit(corpus)
|
|
result = self.transform(x)
|