51 results about "Perplexity" patented technology

The technology disclosed provides a so-called “pointer sentinel mixture architecture” for neural network sequence models that has the ability to either reproduce a token from a recent context or produce a token from a predefined vocabulary. In one implementation, a pointer sentinel-LSTM architecture achieves state of the art language modeling performance of 70.9 perplexity on the Penn Treebank dataset, while using far fewer parameters than a standard softmax LSTM.

Two statistics are disclosed for determining the quality of language models. These statistics are called acoustic perplexity and the synthetic acoustic word error rate (SAWER), and they depend upon methods for computing the acoustic confusability of words. It is possible to substitute models of acoustic data in place of real acoustic data in order to determine acoustic confusability. An evaluation model is created, a synthesizer model is created, and a matrix is determined from the evaluation and synthesizer models. Each of the evaluation and synthesizer models is a hidden Markov model. Once the matrix is determined, a confusability calculation may be performed. Different methods are used to determine synthetic likelihoods. The confusability may be normalized and smoothed and methods are disclosed that increase the speed of performing the matrix inversion and the confusability calculation. A method for caching and reusing computations for similar words is disclosed. Acoustic perplexity and SAWER are determined and applied.

Described herein are systems and methods for compressing or otherwise reducing the memory requirements for storing and computing the model parameters in recurrent neural language models. Embodiments include space compression methodologies that share the structured parameters at the input embedding layer, the output embedding layers, or both of a recurrent neural language model to significantly reduce the size of model parameters, but still compactly represent the original input and output embedding layers. Embodiments of the methodology are easy to implement and tune. Experiments on several data sets show that embodiments achieved similar perplexity and BLEU score results while only using a fraction of the parameters.

The invention belongs to the field of data mining, and relates to an information interactive network-based criminal individual recognition method. The method comprises the following steps of: (1) obtaining a data set which comprises criminal activity contents, and pre-processing the data set; (2) extracting keyword descriptions of criminal topics; (3) determining the number of subjects of a subject model LDA on the basis of perplexity; (4) extracting interactive subjects between individuals in the pre-processed data set on the basis of LDA, wherein the interactive subjects are as follows: an association probability matrix of the interactive subjects and keywords, and an association probability matrix of interactive edges of the interactive subjects; (5) calculating weights of the interactive edges; (6) calculating local criminal suspects of the individuals on the basis of a structure of a weighted information interactive network; and (7) calculating global criminal suspects of the individuals on the basis of a fuzzy K-means cluster and distance density cluster combined method, and recognizing the criminal individuals. The method is independent of prior information, and can be used for analyzing the most possible suspected person according to communication contents so that the case handling efficiency is improved.

The invention provides a method for optimizing a corpus and a device for optimizing the corpus. The device for optimizing the corpus comprises a filter unit, a replacement unit and a perplexity calculation unit; the filter unit performs filtration on sentences in the corpus on the basis of an optimization parameter to obtain sentence pairs to be optimized; the replacement unit performs replacement on at least part of the sentence pairs to be optimized; the perplexity calculation unit calculates the perplexity of the replaced sentence pairs, and the replaced sentence pairs can serve as optimization results of the sentence pairs to be optimized on the condition that the perplexity of the replaced sentence pairs is smaller than that of the sentence pairs to be optimized.

The embodiment of the invention provides a training method and system for a language model, wherein the method specifically comprises the following steps: acquiring seed corpuses of all fields, and training the seed models in the corresponding fields based on the seed corpuses of all the fields; screening big data corpuses by utilizing the perplexity of the corpuses on the seed models of all the fields to obtain seed screening corpuses of in the corresponding fields; respectively training by utilizing the seed screening corpuses of all fields to obtain the screening model of the corresponding field; fusing the screening models of all the fields to obtain corresponding screening fusion model. According to the training method and system for the language model disclosed by the embodiment of the invention, the parameter reasonableness of the language model can be improved on the premise of reducing a computation burden and saving time.

A system and method are disclosed for detecting and labeling places in a video stream using change-points detection. The system comprises a place label generation module configured to assign place labels probabilistically to places in the video stream based on the measurements of the measurement stream representing the video. For each measurement in the segment, the place label generation module classifies the measurement by computing the probability of the measurement being classified by a learned Gaussian Process classifier. Based on the probabilities generated with respect to all the measurements in the segment, the place label generation module determines the place label for the segment. In cases where a Gaussian Process classifier cannot positively classify a segment, the place label generation module determines whether the segment corresponds to an unknown place based on the perplexity statistics of the classification and a threshold value.

The invention discloses a method for screening out the content of a music player. As for some websites of music players at present, netizens upload a great amount of music and songs, which greatly enriches the content of the player but brings about perplexity for users during the playing operation due to numerous repeated contents. In the invention, a small program plugin is used to screen out the content in a playlist, and the found repeated contents can be deleted.

The invention relates to the technical field of artificial intelligence, and discloses an error correction method. The method comprises the following steps: detecting that a to-be-corrected object exists in a text; extracting context contents of the to-be-corrected objects based on the positions of the to-be-corrected objects, inputting the corresponding similar objects into the error correction model according to the context contents and the similar objects to obtain corresponding alternative probabilities of the to-be-corrected objects, and selecting one corresponding object from the alternative probabilities as a replacement object to perform replacement processing on the to-be-corrected objects based on the alternative probabilities. The invention further provides an error correction device and equipment and a storage medium. Information of a to-be-corrected object is predicted based on the to-be-corrected object and context content at the same time; the confusion degree of a language model during semantic recognition can be reduced, so that relatively accurate similar objects are extracted, then the alternative probability of each similar object is calculated based on an errorcorrection model in combination with context content, and a relatively large object is selected from the alternative probability, so that the probability of each character or word is improved, and the final error correction accuracy is also improved.

The invention relates to a biomedicine technology subject mining method based on LDA, belonging to the technical field of information retrieval. The method of the present invention first employs an LDA to view a document as a combination of vectors of a contained word, then the number of semantic topics K is determined by using the evaluation function Perplexity; and finally, a probability p of each document di on all Topics is calculated, and two matrices, one doc-Topic matrix and the other word-Topic matrix, are obtained, so that LDA projects documents and words onto a set of Topics, tryingto find out the potential relationship between documents and words, documents and documents, words and words. LDA is an unsupervised algorithm, and every Topic does not require a specified condition.However, after clustering, the probability distribution of the words on each Topic is calculated, and the words with high probability on the Topic can describe the meaning of the Topic very well.

The invention discloses a new-type network path detection method realized based on a TCP protocol. Detection messages with different TTL values are sent to a target; and response messages triggered bythe detection messages are identified, thereby obtaining a routing paths from local to the target. ICMP response messages and TCP response messages can be triggered by the detection messages. According to the method, perplexity brought by forbidden PING setting to path detection can be effectively avoided, and validity of a detected network path to a target node is improved.

The invention provides a statement correction method and device based on a bert model and an ngram model, and belongs to the technical field of data processing. The method comprises the following steps: determining a plurality of candidate target words used for replacing wrong words in statements, and forming a plurality of candidate statements; for each candidate statement, calculating a first confusion degree of the statement based on a preset bert model, and calculating a second confusion degree of the statement based on a preset ngram model; giving a first weighting coefficient of the first confusion degree, and giving a second weighting coefficient of the second confusion degree; determining the total perplexity of each candidate statement by weighting the first perplexity and the second perplexity; and determining a final statement based on the total perplexity. According to the method, the bert model and the ngram model are complementary, so that the accuracy of statement correction is improved.

The invention discloses a language model fusion method and device, a medium and a computer program product, and the method comprises the steps: obtaining voice data of a target language scene, and extracting voice-to-text scoring data for converting the voice data into text data, and word error rate data corresponding to the text data; acquiring confusion score data of a to-be-fused language model corresponding to each preset language scene on the text data, and combining each confusion score data, the voice-to-text score data and the word error rate data to obtain combined training data; based on the combined training data, constructing a classification sorting model; and based on the model weight parameters of the classification sorting model, performing weighted fusion on the to-be-fused language models to obtain a fused language model. The technical problem that the language model fusion effect is poor is solved.

There is provided a system configured to appropriately determine a topic count in accordance with LDA to estimate latent meanings of a document. For a plurality of documents d, a perplexity PPL of each document d is evaluated in accordance with a document generation probability in which the document d is generated when topic counts N for defining a topic model based on the LDA as a document generation model are hypothetically specified as different values and word groups are specified by different random numbers. The topic model is defined by a reference topic count N₀ determined by combining a first topic count N₁ (the number of topics indicating a highest cumulative frequency at which the perplexity PPL first indicates a minimum value) and a second topic count N₂ (the number of topics indicating a highest cumulative frequency at which the perplexity PPL indicates a smallest value).