L- the type of language model used to generate text from categories
M- the multivariate distribution over categories
public class LMClassifier<L extends LanguageModel,M extends MultivariateDistribution> extends Object implements JointClassifier<CharSequence>
LMClassifierperforms joint probability-based classification of character sequences into non-overlapping categories based on language models for each category and a multivariate distribution over categories. Thus the subclass of
Classificationreturned by the classify method is
JointClassification. In addition to joint and conditional probabilities of categories given the input, the score of the returned joint classification is the character plus category sample entropy rate.
A language-model classifier is constructed from a fixed, finite set of categories which are assumed to have disjoint (non-overlapping) sets of members. The categories are represented as simple strings. Each category is assigned a language model. Furthermore, a multivariate distribution over the set of categories assigns marginal category probabilities.
Joint log probabilities are determined in the usual way:
log2 P(cs,cat) = log2 P(cs|cat) + log2 P(cat)
P(cs|cat)is the probability of the character sequence
csin the language model for category
P(cat)is the probability assigned by the multivariate distribution over categories. Scores are defined to be adjusted sample cross-entropy rates:
Note that the contribution of the category probability to the score approaches zero as the sample size grows and the data overwhelms the pre-data expectation. Also note that each category has its estimate divided by the same amount, so the probabilistic ordering is preserved. If the language models are process models, the cross-entropy rate is just
= (log2 P(cs,cat)) / (cs.length() + 2)
= (log2 P(cs|cat) + log2 P(cat)) / (cs.length() + 2)
(log2P(cs|cat))/cs.length(); for process models, add one to the denominator to account for figuratively generating the end-of-character-sequence symbol.
Note that maximizing joint probabilities is the same as
maximizing conditional probabilities because the character sequence
cs is constant:
A computation of conditional estimates
= ARGMAXcat P(cs,cat) / P(cs)
= ARGMAXcat P(cs,cat)
P(cat|cs)given the joint estimates is defined in
To ensure consistent estimates, all of the language models should either be process language models or sequence language models over the same set of characters, depending on whether probability normalization is over fixed length sequences or over all strings. On the other hand, the models themselves may be a mixture of n-gram lengths and smoothing parameters, or even in the case of sequence models, tokenized models and sequence character models.
Boolean classifiers for membership can be constructed with this
class by means of a positive language model and a negative model.
A character sequence is considered an instance of the category if
they are more likely in the positive model than the negative model.
There are several strategies for constructing anti-models. The
most common methodology is to build an anti-model from an unbiased
sample of negative cases, but this requires supervision for
negative cases and tends to bias toward the model with more
training data given the way language model cross-entropy tends to
go down with more training data in general. Another approach is to
build a weaker model from the same training data as the positive
model, for instance by using lower order n-grams for the negative
model. The simplest approach is to use a uniform negative model,
which amounts to a cross-entropy rejection threshold; this is the
Language model classifiers may be trained using
DynamicLMClassifier using a trainable multivariate estimator and
dynamic language models.
|Constructor and Description|
Construct a joint classifier for character sequences classifying over a specified set of categories, with a multivariate distribution over those categories and a language model per category.
|Modifier and Type||Method and Description|
Returns the array of categories for this classifier.
Returns a multivariate distribution over categories for this classifier.
Returns the joint classification of the specified character sequence.
A convenience method returning a joint classification over a character array slice.
Returns the language model for the specified category.
The category distribution is the marginal over categories, and each language model provides a conditional estimate given its category. Categorization is as described in the class documentation.
categories- Array of categories for classification.
languageModels- A parallel array of language models for the categories.
categoryDistribution- The marginal distribution over the categories for classification.
IllegalArgumentException- If there are not at least two categories, or if the category and language model arrays are not the same length, or if there are duplicate categories.
public String categories()
This method copies the array and thus changes to it do not affect the categories for this classifier. Thus if a client needs repeated access to a classifier's categores, this method should be called once and the resulting array reused.
P(cSeq|category), the conditional probability of a character sequence given the specified category as described in the class documentation above.
Changes to the returned model affect this classifier's behavior.
category- The specified category.
IllegalArgumentException- If the category is not known.
public M categoryDistribution()
P(category), the marginal distribution over categories used during classification as described in the class documentation.
Changes to the returned distribution affect this classifier's behavior.
public JointClassification classify(CharSequence cSeq)
cSeq- Character sequence being classified.
IllegalArgumentException- If the specified object is not a character sequence.
public JointClassification classifyJoint(char cs, int start, int end)
estimateJoint(cs,start,end)returns the same result as
cs- Underlying character array.
start- Index of first character in slice.
end- One plus the index of the last character in the slice.
IllegalArgumentException- If the start index is less than zero or greater than the end index or if the end index is not within bounds.