python - Normalizar puntaje de clasificación con pesas

nlp nltk

Question

Estoy trabajando en un problema de búsqueda de documentos donde, dado un conjunto de documentos y una consulta de búsqueda, quiero encontrar el documento más cercano a la consulta. El modelo que estoy usando se basa en TfidfVectorizer en scikit. Creé 4 diferentes vectores tf_idf para todos los documentos usando 4 tipos diferentes de tokenizers. Cada tokenizador divide la cadena en n-grams donde n está en el rango 1 ... 4.

Por ejemplo:

doc_1 = "Singularity is still a confusing phenomenon in physics"
doc_2 = "Quantum theory still wins over String theory"

Entonces model_1 usará un tokenizer de 1 gramo, model_2 usará un tokenizer de 2 gramos.

Luego, para una consulta de búsqueda dada, calculo la similitud del coseno entre el término de búsqueda y todos los otros documentos que usan estos 4 modelos.

Por ejemplo, consulta de búsqueda: Singularidad en física cuántica. La consulta de búsqueda se divide en n-grams y los valores de tf_idf se calculan a partir del modelo de n-gramas correspondiente.

Por lo tanto, para cada par de consulta-documento tengo 4 valores de similitud basados ​​en el modelo n-gram utilizado. Por ejemplo:

1-gram similarity = 0.4370303325246957
2-gram similarity = 0.36617374546988996
3-gram similarity = 0.29519246156322099
4-gram similarity = 0.2902998188509896

Todos estos puntajes de similitud se normalizan en una escala de 0 a 1. Ahora quiero calcular una puntuación normalizada agregada de modo que para cualquier par de documento de consulta, la similitud de n-gramas más alta tenga un peso realmente alto. Básicamente, cuanto mayor sea la similitud de ngram, mayor será su impacto en el puntaje general.

¿Alguien puede sugerir una solución?

Answer 1

Hay muchas formas de jugar con los números:

>>> onegram_sim = 0.43
>>> twogram_sim = 0.36
>>> threegram_sim = 0.29
>>> fourgram_sim = 0.29
# Sum(x) / len(list)
>>> all_sim = sum([onegram_sim, twogram_sim, threegram_sim, fourgram_sim]) / 4
>>> all_sim
0.3425
# Sum(x*x) / len(list)
>>> all_sim = sum(map(lambda x: x**2, [onegram_sim, twogram_sim, threegram_sim, fourgram_sim])) / 4
>>> all_sim
0.120675
# Product(x)
>>> from operator import mul
>>> onetofour_sim = [onegram_sim, twogram_sim, threegram_sim, fourgram_sim]
>>> reduce(mul, onetofour_sim, 1)
0.013018679999999998

Eventualmente, lo que te lleve a un mejor puntaje de precisión para tu tarea final es la mejor solución.

Más allá de tu pregunta:

Para calcular la similitud de documentos, hay una llamada a la tarea SemEval de larga duración. Similitud textual semántica https://groups.google.com/forum/#!forum/sts-semeval

Las estrategias comunes incluyen (no exhaustivamente):

1. Use un corpus anotado con puntuaciones de similitud para pares de oraciones, extraiga algunas características, entrene un regresor y emite un puntaje de similitud

2. Use algún tipo de semántica de espacio vectorial (lectura muy recomendada: http://www.jair.org/media/2934/live-2934-4846-jair.pdf ) y luego haga algunos puntajes de similitud de vectores (eche un vistazo a Cómo calcular la similitud del coseno dado 2 cadenas de frases? - Python )

   yo. Un subconjunto de la jerga de la semántica del espacio vectorial será útil (a veces conocido como incrustaciones de palabras), a veces las personas entrenan un espacio vectorial con modelos de tema / redes neuronales / aprendizaje profundo (otras palabras de moda relacionadas), consulte http: //u.cs. biu.ac.il/~yogo/cvsc2015.pdf

   ii. También podría usar vectores de bolsa de palabras más tradicionales y comprimir el espacio con TF-IDF o cualquier otra reducción de dimensionalidad "latente" y luego usar alguna función de similitud de vectores para obtener la similitud.

   iii. Cree una función de similitud vectorial elegante (por ejemplo, cosmul , consulte https://radimrehurek.com/gensim/models/word2vec.html ) y luego modifique la función y cosmul en espacios diferentes.

3. Utilice algunos recursos léxicos que contengan una ontología de conceptos (por ejemplo, WordNet, Cyc, etc.) y luego compare la similitud atravesando los gráficos conceptuales (consulte http://www.nltk.org/howto/wordnet.html ). Un ejemplo sería https://github.com/alvations/pywsd/blob/master/pywsd/similarity.py

Teniendo en cuenta lo anterior como fondo y sin anotaciones, intentemos hackear algunos ejemplos de espacio vectorial:

Primero intentemos ngrams simples con vectores binarios simples:

import numpy as np
from nltk import ngrams

doc1 = "Singularity is still a confusing phenomenon in physics".split()
doc2 = "Quantum theory still wins over String theory".split()
_vec1 = list(ngrams(doc1, 3))
_vec2 = list(ngrams(doc2, 3))
# Create a full dictionary of all possible ngrams.
vec_dict = list(set(_vec1).union(_vec2))
print ''Vector Dict:'', vec_dict
# Now vectorize the documents
vec1 = [1 if ng in _vec1 else 0 for ng in vec_dict]
vec2 = [1 if ng in _vec2 else 0 for ng in vec_dict]
print ''Vectorzied:'', vec1, vec2
print ''Similarity:'', np.dot(vec1, vec2)

[fuera]:

Vector Dict: [(''still'', ''a'', ''confusing''), (''confusing'', ''phenomenon'', ''in''), (''theory'', ''still'', ''wins''), (''is'', ''still'', ''a''), (''over'', ''String'', ''theory''), (''a'', ''confusing'', ''phenomenon''), (''wins'', ''over'', ''String''), (''Singularity'', ''is'', ''still''), (''still'', ''wins'', ''over''), (''phenomenon'', ''in'', ''physics''), (''Quantum'', ''theory'', ''still'')] 

Vectorzied: [1, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0] [0, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1] 

Similarity: 0 

Ahora intentemos incluir desde 1 gramo a ngrams (donde n = len(sent) ) y poner todo en el diccionario vectorial con los ngramos binarios:

import numpy as np
from nltk import ngrams

def everygrams(sequence):
    """
    This function returns all possible ngrams for n 
    ranging from 1 to len(sequence).
    >>> list(everygrams(''a b c''.split()))
    [(''a'',), (''b'',), (''c'',), (''a'', ''b''), (''b'', ''c''), (''a'', ''b'', ''c'')]
    """
    for n in range(1, len(sequence)+1):
        for ng in ngrams(sequence, n):
            yield ng

doc1 = "Singularity is still a confusing phenomenon in physics".split()
doc2 = "Quantum theory still wins over String theory".split()
_vec1 = list(everygrams(doc1))
_vec2 = list(everygrams(doc2))
# Create a full dictionary of all possible ngrams.
vec_dict = list(set(_vec1).union(_vec2))
print ''Vector Dict:'', vec_dict, ''/n''
# Now vectorize the documents
vec1 = [1 if ng in _vec1 else 0 for ng in vec_dict]
vec2 = [1 if ng in _vec2 else 0 for ng in vec_dict]
print ''Vectorzied:'', vec1, vec2, ''/n''
print ''Similarity:'', np.dot(vec1, vec2), ''/n''

[fuera]:

Vector Dict: [(''still'', ''a''), (''over'', ''String''), (''theory'', ''still'', ''wins'', ''over'', ''String'', ''theory''), (''String'', ''theory''), (''physics'',), (''in'',), (''wins'', ''over'', ''String'', ''theory''), (''is'', ''still'', ''a'', ''confusing'', ''phenomenon'', ''in''), (''theory'', ''still'', ''wins''), (''Singularity'', ''is'', ''still'', ''a'', ''confusing'', ''phenomenon''), (''a'',), (''wins'',), (''is'', ''still'', ''a''), (''Singularity'', ''is''), (''phenomenon'', ''in''), (''still'', ''wins'', ''over'', ''String''), (''Singularity'', ''is'', ''still'', ''a'', ''confusing'', ''phenomenon'', ''in'', ''physics''), (''Quantum'', ''theory'', ''still'', ''wins'', ''over''), (''a'', ''confusing'', ''phenomenon''), (''Singularity'', ''is'', ''still'', ''a''), (''confusing'', ''phenomenon''), (''confusing'', ''phenomenon'', ''in'', ''physics''), (''Singularity'', ''is'', ''still''), (''is'', ''still'', ''a'', ''confusing'', ''phenomenon'', ''in'', ''physics''), (''wins'', ''over''), (''theory'', ''still'', ''wins'', ''over''), (''phenomenon'',), (''Quantum'', ''theory'', ''still'', ''wins'', ''over'', ''String''), (''is'', ''still''), (''still'', ''wins'', ''over''), (''is'', ''still'', ''a'', ''confusing'', ''phenomenon''), (''phenomenon'', ''in'', ''physics''), (''Quantum'', ''theory'', ''still'', ''wins''), (''Quantum'', ''theory'', ''still''), (''a'', ''confusing'', ''phenomenon'', ''in'', ''physics''), (''Singularity'', ''is'', ''still'', ''a'', ''confusing''), (''still'', ''a'', ''confusing'', ''phenomenon'', ''in''), (''still'', ''a'', ''confusing''), (''is'', ''still'', ''a'', ''confusing''), (''in'', ''physics''), (''Quantum'', ''theory'', ''still'', ''wins'', ''over'', ''String'', ''theory''), (''confusing'', ''phenomenon'', ''in''), (''theory'', ''still''), (''Quantum'', ''theory''), (''is'',), (''String'',), (''over'', ''String'', ''theory''), (''still'', ''a'', ''confusing'', ''phenomenon'', ''in'', ''physics''), (''a'', ''confusing''), (''still'', ''wins''), (''still'',), (''over'',), (''still'', ''a'', ''confusing'', ''phenomenon''), (''wins'', ''over'', ''String''), (''Singularity'',), (''confusing'',), (''theory'',), (''Singularity'', ''is'', ''still'', ''a'', ''confusing'', ''phenomenon'', ''in''), (''still'', ''wins'', ''over'', ''String'', ''theory''), (''a'', ''confusing'', ''phenomenon'', ''in''), (''Quantum'',), (''theory'', ''still'', ''wins'', ''over'', ''String'')] 

Vectorzied: [1, 0, 0, 0, 1, 1, 0, 1, 0, 1, 1, 0, 1, 1, 1, 0, 1, 0, 1, 1, 1, 1, 1, 1, 0, 0, 1, 0, 1, 0, 1, 1, 0, 0, 1, 1, 1, 1, 1, 1, 0, 1, 0, 0, 1, 0, 0, 1, 1, 0, 1, 0, 1, 0, 1, 1, 0, 1, 0, 1, 0, 0] [0, 1, 1, 1, 0, 0, 1, 0, 1, 0, 0, 1, 0, 0, 0, 1, 0, 1, 0, 0, 0, 0, 0, 0, 1, 1, 0, 1, 0, 1, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 1, 0, 1, 1, 0, 1, 1, 0, 0, 1, 1, 1, 0, 1, 0, 0, 1, 0, 1, 0, 1, 1] 

Similarity: 1 

Ahora intentemos normalizar por no. de posibles ngrams:

import numpy as np
from nltk import ngrams

def everygrams(sequence):
    """
    This function returns all possible ngrams for n 
    ranging from 1 to len(sequence).
    >>> list(everygrams(''a b c''.split()))
    [(''a'',), (''b'',), (''c'',), (''a'', ''b''), (''b'', ''c''), (''a'', ''b'', ''c'')]
    """
    for n in range(1, len(sequence)+1):
        for ng in ngrams(sequence, n):
            yield ng

doc1 = "Singularity is still a confusing phenomenon in physics".split()
doc2 = "Quantum theory still wins over String theory".split()
_vec1 = list(everygrams(doc1))
_vec2 = list(everygrams(doc2))
# Create a full dictionary of all possible ngrams.
vec_dict = list(set(_vec1).union(_vec2))
print ''Vector Dict:'', vec_dict, ''/n''
# Now vectorize the documents
vec1 = [1/float(len(_vec1)) if ng in _vec1 else 0 for ng in vec_dict]
vec2 = [1/float(len(_vec2)) if ng in _vec2 else 0 for ng in vec_dict]
print ''Vectorzied:'', vec1, vec2, ''/n''
print ''Similarity:'', np.dot(vec1, vec2), ''/n''

Se ve mejor, afuera:

Vector Dict: [(''still'', ''a''), (''over'', ''String''), (''theory'', ''still'', ''wins'', ''over'', ''String'', ''theory''), (''String'', ''theory''), (''physics'',), (''in'',), (''wins'', ''over'', ''String'', ''theory''), (''is'', ''still'', ''a'', ''confusing'', ''phenomenon'', ''in''), (''theory'', ''still'', ''wins''), (''Singularity'', ''is'', ''still'', ''a'', ''confusing'', ''phenomenon''), (''a'',), (''wins'',), (''is'', ''still'', ''a''), (''Singularity'', ''is''), (''phenomenon'', ''in''), (''still'', ''wins'', ''over'', ''String''), (''Singularity'', ''is'', ''still'', ''a'', ''confusing'', ''phenomenon'', ''in'', ''physics''), (''Quantum'', ''theory'', ''still'', ''wins'', ''over''), (''a'', ''confusing'', ''phenomenon''), (''Singularity'', ''is'', ''still'', ''a''), (''confusing'', ''phenomenon''), (''confusing'', ''phenomenon'', ''in'', ''physics''), (''Singularity'', ''is'', ''still''), (''is'', ''still'', ''a'', ''confusing'', ''phenomenon'', ''in'', ''physics''), (''wins'', ''over''), (''theory'', ''still'', ''wins'', ''over''), (''phenomenon'',), (''Quantum'', ''theory'', ''still'', ''wins'', ''over'', ''String''), (''is'', ''still''), (''still'', ''wins'', ''over''), (''is'', ''still'', ''a'', ''confusing'', ''phenomenon''), (''phenomenon'', ''in'', ''physics''), (''Quantum'', ''theory'', ''still'', ''wins''), (''Quantum'', ''theory'', ''still''), (''a'', ''confusing'', ''phenomenon'', ''in'', ''physics''), (''Singularity'', ''is'', ''still'', ''a'', ''confusing''), (''still'', ''a'', ''confusing'', ''phenomenon'', ''in''), (''still'', ''a'', ''confusing''), (''is'', ''still'', ''a'', ''confusing''), (''in'', ''physics''), (''Quantum'', ''theory'', ''still'', ''wins'', ''over'', ''String'', ''theory''), (''confusing'', ''phenomenon'', ''in''), (''theory'', ''still''), (''Quantum'', ''theory''), (''is'',), (''String'',), (''over'', ''String'', ''theory''), (''still'', ''a'', ''confusing'', ''phenomenon'', ''in'', ''physics''), (''a'', ''confusing''), (''still'', ''wins''), (''still'',), (''over'',), (''still'', ''a'', ''confusing'', ''phenomenon''), (''wins'', ''over'', ''String''), (''Singularity'',), (''confusing'',), (''theory'',), (''Singularity'', ''is'', ''still'', ''a'', ''confusing'', ''phenomenon'', ''in''), (''still'', ''wins'', ''over'', ''String'', ''theory''), (''a'', ''confusing'', ''phenomenon'', ''in''), (''Quantum'',), (''theory'', ''still'', ''wins'', ''over'', ''String'')] 

Vectorzied: [0.027777777777777776, 0, 0, 0, 0.027777777777777776, 0.027777777777777776, 0, 0.027777777777777776, 0, 0.027777777777777776, 0.027777777777777776, 0, 0.027777777777777776, 0.027777777777777776, 0.027777777777777776, 0, 0.027777777777777776, 0, 0.027777777777777776, 0.027777777777777776, 0.027777777777777776, 0.027777777777777776, 0.027777777777777776, 0.027777777777777776, 0, 0, 0.027777777777777776, 0, 0.027777777777777776, 0, 0.027777777777777776, 0.027777777777777776, 0, 0, 0.027777777777777776, 0.027777777777777776, 0.027777777777777776, 0.027777777777777776, 0.027777777777777776, 0.027777777777777776, 0, 0.027777777777777776, 0, 0, 0.027777777777777776, 0, 0, 0.027777777777777776, 0.027777777777777776, 0, 0.027777777777777776, 0, 0.027777777777777776, 0, 0.027777777777777776, 0.027777777777777776, 0, 0.027777777777777776, 0, 0.027777777777777776, 0, 0] [0, 0.03571428571428571, 0.03571428571428571, 0.03571428571428571, 0, 0, 0.03571428571428571, 0, 0.03571428571428571, 0, 0, 0.03571428571428571, 0, 0, 0, 0.03571428571428571, 0, 0.03571428571428571, 0, 0, 0, 0, 0, 0, 0.03571428571428571, 0.03571428571428571, 0, 0.03571428571428571, 0, 0.03571428571428571, 0, 0, 0.03571428571428571, 0.03571428571428571, 0, 0, 0, 0, 0, 0, 0.03571428571428571, 0, 0.03571428571428571, 0.03571428571428571, 0, 0.03571428571428571, 0.03571428571428571, 0, 0, 0.03571428571428571, 0.03571428571428571, 0.03571428571428571, 0, 0.03571428571428571, 0, 0, 0.03571428571428571, 0, 0.03571428571428571, 0, 0.03571428571428571, 0.03571428571428571] 

Similarity: 0.000992063492063 

Ahora intentemos contar los ngramos en lugar de tomar 1/len(_vec) , es decir, _vec.count(ng) / len(_vec) :

import numpy as np
from nltk import ngrams

def everygrams(sequence):
    """
    This function returns all possible ngrams for n 
    ranging from 1 to len(sequence).
    >>> list(everygrams(''a b c''.split()))
    [(''a'',), (''b'',), (''c'',), (''a'', ''b''), (''b'', ''c''), (''a'', ''b'', ''c'')]
    """
    for n in range(1, len(sequence)+1):
        for ng in ngrams(sequence, n):
            yield ng

doc1 = "Singularity is still a confusing phenomenon in physics".split()
doc2 = "Quantum theory still wins over String theory".split()
_vec1 = list(everygrams(doc1))
_vec2 = list(everygrams(doc2))
# Create a full dictionary of all possible ngrams.
vec_dict = list(set(_vec1).union(_vec2))
print ''Vector Dict:'', vec_dict, ''/n''
# Now vectorize the documents
vec1 = [_vec1.count(ng)/float(len(_vec1)) if ng in _vec1 else 0 for ng in vec_dict]
vec2 = [_vec2.count(ng)/float(len(_vec2)) if ng in _vec2 else 0 for ng in vec_dict]
print ''Vectorzied:'', vec1, vec2, ''/n''
print ''Similarity:'', np.dot(vec1, vec2), ''/n''

Como era de esperar, dado que los recuentos son todos 1, es el mismo puntaje de similitud:

Vector Dict: [(''still'', ''a''), (''over'', ''String''), (''theory'', ''still'', ''wins'', ''over'', ''String'', ''theory''), (''String'', ''theory''), (''physics'',), (''in'',), (''wins'', ''over'', ''String'', ''theory''), (''is'', ''still'', ''a'', ''confusing'', ''phenomenon'', ''in''), (''theory'', ''still'', ''wins''), (''Singularity'', ''is'', ''still'', ''a'', ''confusing'', ''phenomenon''), (''a'',), (''wins'',), (''is'', ''still'', ''a''), (''Singularity'', ''is''), (''phenomenon'', ''in''), (''still'', ''wins'', ''over'', ''String''), (''Singularity'', ''is'', ''still'', ''a'', ''confusing'', ''phenomenon'', ''in'', ''physics''), (''Quantum'', ''theory'', ''still'', ''wins'', ''over''), (''a'', ''confusing'', ''phenomenon''), (''Singularity'', ''is'', ''still'', ''a''), (''confusing'', ''phenomenon''), (''confusing'', ''phenomenon'', ''in'', ''physics''), (''Singularity'', ''is'', ''still''), (''is'', ''still'', ''a'', ''confusing'', ''phenomenon'', ''in'', ''physics''), (''wins'', ''over''), (''theory'', ''still'', ''wins'', ''over''), (''phenomenon'',), (''Quantum'', ''theory'', ''still'', ''wins'', ''over'', ''String''), (''is'', ''still''), (''still'', ''wins'', ''over''), (''is'', ''still'', ''a'', ''confusing'', ''phenomenon''), (''phenomenon'', ''in'', ''physics''), (''Quantum'', ''theory'', ''still'', ''wins''), (''Quantum'', ''theory'', ''still''), (''a'', ''confusing'', ''phenomenon'', ''in'', ''physics''), (''Singularity'', ''is'', ''still'', ''a'', ''confusing''), (''still'', ''a'', ''confusing'', ''phenomenon'', ''in''), (''still'', ''a'', ''confusing''), (''is'', ''still'', ''a'', ''confusing''), (''in'', ''physics''), (''Quantum'', ''theory'', ''still'', ''wins'', ''over'', ''String'', ''theory''), (''confusing'', ''phenomenon'', ''in''), (''theory'', ''still''), (''Quantum'', ''theory''), (''is'',), (''String'',), (''over'', ''String'', ''theory''), (''still'', ''a'', ''confusing'', ''phenomenon'', ''in'', ''physics''), (''a'', ''confusing''), (''still'', ''wins''), (''still'',), (''over'',), (''still'', ''a'', ''confusing'', ''phenomenon''), (''wins'', ''over'', ''String''), (''Singularity'',), (''confusing'',), (''theory'',), (''Singularity'', ''is'', ''still'', ''a'', ''confusing'', ''phenomenon'', ''in''), (''still'', ''wins'', ''over'', ''String'', ''theory''), (''a'', ''confusing'', ''phenomenon'', ''in''), (''Quantum'',), (''theory'', ''still'', ''wins'', ''over'', ''String'')] 

Vectorzied: [0.027777777777777776, 0, 0, 0, 0.027777777777777776, 0.027777777777777776, 0, 0.027777777777777776, 0, 0.027777777777777776, 0.027777777777777776, 0, 0.027777777777777776, 0.027777777777777776, 0.027777777777777776, 0, 0.027777777777777776, 0, 0.027777777777777776, 0.027777777777777776, 0.027777777777777776, 0.027777777777777776, 0.027777777777777776, 0.027777777777777776, 0, 0, 0.027777777777777776, 0, 0.027777777777777776, 0, 0.027777777777777776, 0.027777777777777776, 0, 0, 0.027777777777777776, 0.027777777777777776, 0.027777777777777776, 0.027777777777777776, 0.027777777777777776, 0.027777777777777776, 0, 0.027777777777777776, 0, 0, 0.027777777777777776, 0, 0, 0.027777777777777776, 0.027777777777777776, 0, 0.027777777777777776, 0, 0.027777777777777776, 0, 0.027777777777777776, 0.027777777777777776, 0, 0.027777777777777776, 0, 0.027777777777777776, 0, 0] [0, 0.03571428571428571, 0.03571428571428571, 0.03571428571428571, 0, 0, 0.03571428571428571, 0, 0.03571428571428571, 0, 0, 0.03571428571428571, 0, 0, 0, 0.03571428571428571, 0, 0.03571428571428571, 0, 0, 0, 0, 0, 0, 0.03571428571428571, 0.03571428571428571, 0, 0.03571428571428571, 0, 0.03571428571428571, 0, 0, 0.03571428571428571, 0.03571428571428571, 0, 0, 0, 0, 0, 0, 0.03571428571428571, 0, 0.03571428571428571, 0.03571428571428571, 0, 0.03571428571428571, 0.03571428571428571, 0, 0, 0.03571428571428571, 0.03571428571428571, 0.03571428571428571, 0, 0.03571428571428571, 0, 0, 0.07142857142857142, 0, 0.03571428571428571, 0, 0.03571428571428571, 0.03571428571428571] 

Similarity: 0.000992063492063 

Además de los ngrams, también puedes probar los skipgrams: ¿cómo calcular los skipgrams en python?