Tencent Technology (shenzhen) Company Limited

Tsinghua University

Jinchao Zhang

Fandong Meng

Jie Zhou

Qian Zhou

Hao Dong

Kai Ni

Jinchao Pang

Rui Tian

Mingfei Xu

Hui Zhang

Cheng Niu

Yang Feng

Yao Qiu

Jiao Ou

Jinan Xu

Lanlan Wang

Xiangwen Zhu

Huanhuan Wang

Haifei Hu

Xinghui Li

This disclosure relates to a method and an apparatus for training a text classification model. The method may include determining a semantic representation of the training sample using the text classification model and determining a predicted classification result of the training sample based on the semantic representation. The method may further include generating an adversarial sample corresponding to the training sample based on the training sample and perturbation information and determining a semantic representation of the adversarial sample corresponding to the training sample using the text classification model. The method may further include determining a classification loss of the text classification model based on the predicted classification result, determining a contrastive loss of the text classification model based on the semantic representation and the semantic representation of the adversarial sample, and training the text classification model based on the classification loss and the contrastive loss.

Method and apparatus for training text classification model

A computer device acquires, based on a prior network and a posterior network of a dialogue model, at least two first dialogue features and at least two second dialogue features of a first dialogue. The computer device outputs, using the prior network, a probability distribution of dialogue features in accordance with the acquired dialogue features and estimates, using the posterior network, the probability distribution of the dialogue features output by the prior network. The computer device updates the dialogue model and updates the posterior network based on dialogue features of the first dialogue. The computer device also updates a discriminator in the dialogue model based on at least two first dialogue features and at least two second dialogue features of a second dialogue. In accordance with a determination that the dialogue model has satisfied a training end condition, the computer device determines that the dialogue model is a trained model.

Method and apparatus for training dialog model, computer device, and storage medium

A method and apparatus for generating a hidden state in a recurrent neural network for language processing are provided. The method includes: generating regional word vectors of at least two dimensions of a target word vector inputted at a first moment; combining the regional word vectors to obtain combined regional word vectors of at least two dimensions; performing aggregation transformation processing on the combined regional word vectors based on a feedforward neural network, to obtain an aggregated word vector corresponding to the target word vector; and generating a target hidden state corresponding to the target word vector based on the aggregated word vector.

Method and apparatus for generating hidden state in recurrent neural network for language processing

Embodiments of this disclosure disclose a sentence processing method and device. The method may include performing word segmentation operation on a source sentence to be encoded to obtain m words. The method may further include obtaining an iword in the m words using an iencoding processing node in the n encoding processing nodes, and obtaining an (i−1)word vector from an (i−1)encoding processing node. The method may further include performing linear operation and non-linear operation on the iword and the (i−1)word vector using the first unit of the iencoding processing node to obtain an ioperation result, and outputting the ioperation result to the at least one second unit for processing to obtain an iword vector. The method may further include generating, in response to obtaining m word vectors, a sentence vector according to the m word vectors.

Sentence encoding and decoding method, storage medium, and device

A data processing method includes: obtaining a to-be-detected text, and determining a context word set and a candidate substitute word set corresponding to a to-be-detected word in the to-be-detected text to be inputted into a pronoun resolution neural network for feature extraction; performing positive-example iteration processing and negative-example iteration processing on the features corresponding to the context word set and the candidate substitute word set, to obtain a positive-example feature vector length and a negative-example feature vector length, and calculating a substitute probability corresponding to each candidate substitute word in the candidate substitute word set according to the positive-example feature vector length and the negative-example feature vector length; determining a target substitute word according to the substitute probability corresponding to the each candidate substitute word; and inserting the target substitute word into the to-be-detected text according to a position corresponding to the to-be-detected word, to obtain a target text.

Data processing method and pronoun resolution neural network training method

A language sequence labeling method and includes: reading a first embedding representation of a language sequence, the first embedding representation including a character-level word embedding representation, a pre-trained word embedding representation, and a global word embedding representation of the language sequence, the global word embedding representation referring to a global context representation of the language sequence; performing first depth transformation (DT) encoding on the first embedding representation based on a first DT recurrent neural network (RNN), to output a first hidden-layer state representation corresponding to each word in the language sequence; and decoding the first hidden-layer state representations of the language sequence, to obtain a labeling result of one or more elements in the language sequence.

Language sequence labeling method and apparatus, storage medium, and computing device

A design method of a spectrometer and a spectrometer are disclosed, including the following steps: 1) determining an incident angle of a second incident slit and a groove-shaped cycle of a concave grating; 2) estimating a blaze angle of the concave grating, determining a surface material and a groove-shaped structure of the concave grating; 3) acquiring wavelength-diffraction efficiency curves; 4) determining values of incident angles &#x3b8;and &#x3b8;and values of wavelengths &#x3bb;and &#x3bb;, and setting &#x3bb;to equal &#x3bb;; 5) acquiring a record structural parameter and a use structural parameter; 6) determining a manufacture parameter of the concave grating; 7) determining locations of the three incident slits and the three photodetectors relative to the concave grating. The spectrometer acquired by using this method has relatively high diffraction efficiency in most spectrum regions and effectively alleviates the problem of relatively low diffraction efficiency in a broad spectrum region.

Method of spectrometer and spectrometer

The invention provides an exposure method of a holographic grating and an exposure light path. The exposure method includes: (1) determining initial positions (C, D) of the two exposure light sources (S, S); (2) calculating imaging quality parameters of the grating; (3) setting a compensating mirror (A) in the initial light path; (4) adjusting a position of the exposure light source (S) to a new position (D) according to a position of the compensating mirror (A); (5) calculating imaging quality parameters of the grating; (6) judging whether the imaging quality parameters in the step (5) and the imaging quality parameters in the step (2) are equal, and if yes, using the new position (D) as a final position of the exposure light source (S). The exposure method and exposure light path may effectively solve a problem of a much too close distance between exposure light sources.

Exposure method of holographic grating and exposure light path

An optimized design method for manufacturing a Fresnel grating is disclosed, including the following steps: (1) making a Fresnel surface type of the Fresnel grating equivalent to a curved surface type, and determining, based on the curved surface type to which the Fresnel surface type is equivalent, an optical path difference function of the Fresnel grating: &#x3a6; (&#x3bb;)=&#x3c;APPB&#x3e;&#x2212;&#x3c;AOB&#x3e;+Nm&#x3bb;; and (2) determining a Fresnel grating parameter that minimizes a function value of the optical path difference function, so as to manufacture a Fresnel grating that has an aberration elimination effect. The manufactured Fresnel grating may effectively eliminate a portion of aberrations of the Fresnel grating, and increase a resolution of a spectrometer.

Optimized design method for manufacturing fresnel grating

A manufacture method for a blazed concave grating is provided, including: replicating a blazed concave grating by means of a replication process, after the replication is completed, conducting heat preservation and cooling, then separating two gratings (G, G) by using different pulling manners respectively, and finally splicing the replicated and separated concave grating blanks, to obtain a required blazed concave grating. This manufacture method can solve the problem of a big error in an existing replication method for a blazed concave grating.

Growing community of inventors

Guangdong, China

Jinchao Zhang