CN111985220A - End-to-end judicial literature automatic proofreading method based on deep learning - Google Patents

End-to-end judicial literature automatic proofreading method based on deep learning Download PDF

Info

Publication number
CN111985220A
CN111985220A CN202010751655.8A CN202010751655A CN111985220A CN 111985220 A CN111985220 A CN 111985220A CN 202010751655 A CN202010751655 A CN 202010751655A CN 111985220 A CN111985220 A CN 111985220A
Authority
CN
China
Prior art keywords
model
encoder
decoder
likelihood
neural network
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202010751655.8A
Other languages
Chinese (zh)
Inventor
朱海麒
姜峰
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Harbin Institute of Technology
Original Assignee
Harbin Institute of Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Harbin Institute of Technology filed Critical Harbin Institute of Technology
Priority to CN202010751655.8A priority Critical patent/CN111985220A/en
Publication of CN111985220A publication Critical patent/CN111985220A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/20—Natural language analysis
    • G06F40/253—Grammatical analysis; Style critique
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Abstract

The invention discloses an end-to-end judicial literature automatic proofreading method based on deep learning, and belongs to the technical field of natural language processing. The automatic proofreading method comprises the following steps: step one, a Transformer model structure is provided; training the Transformer model to obtain the likelihood of the maximized model on training data S; and step three, introducing a length penalty term into the likelihood obtained in the step two to obtain a decoding strategy. The invention uses the encoder-decoder model-transform based on the self-attention mechanism, effectively avoids the defects of the cyclic neural network and the convolutional neural network, and the proposed method is far superior to the performance of the encoder-decoder model based on the cyclic neural network and the convolutional neural network.

Description

End-to-end judicial literature automatic proofreading method based on deep learning
Technical Field
The invention relates to an end-to-end judicial literature automatic proofreading method based on deep learning, and belongs to the technical field of natural language processing.
Background
With the gradual improvement of informatization in the judicial field, a large number of judicial documents are generated, and in the face of massive judicial document texts, the manually written judicial documents have certain implicit grammatical errors, which provides a serious challenge for the traditional manual-based proofreading. The method can make the text more fluent and easy to read if the grammar error is not corrected, and the judicial documents are used as carriers for law enforcement, which will bring great influence if there is grammar error or logic error, but it is obviously unrealistic to process a large amount of texts based on manual proofreading, so that the text correction technology is more and more concerned in recent years.
Compared with tasks such as machine translation, automatic question answering and dialogue, the grammar error correction task is relatively small and the research community scale is not large, and generally speaking, the grammar error correction research is processed through three stages in a methodology: artificial rule based methods, statistical classifier based methods, machine translation based methods. In recent years, the development of deep learning has led to the application of a series of end-to-end learning methods in the field of Natural Language Processing (NLP), and Machine Translation has also gradually shifted from the original Statistical Machine Translation (SMT) method to a series of Neural Machine Translation (NMT) methods, such as RNN seq2seq models, attention Machine models, ConvS2S models, and self-attention Machine model-based Transformer models, as a hot research problem in the field of NLP. Therefore, we roughly divide the syntax error correction method into four kinds: rule-based methods, statistical classifier-based methods, SMT-based methods, and NMT-based methods.
The manual rule-based approach has a key drawback in that it cannot cover all error types in the text. For some syntax errors involving complex contexts, it is almost impossible to give all the error correction rules, e.g. word collocation errors. Furthermore, the error correction rules often need to be given by human language experts, which undoubtedly consumes a lot of costs.
The classifier-based method is not tried in the chinese text error correction task, and the main reasons for this are the word boundaries and the huge character set of chinese. Strictly speaking, Chinese has no word boundaries, there is no definite interval of words in Chinese, and each word has a very short length, and there are not a large number of fixed phrases like in English, so in the process of correcting Chinese grammar error, the context factor must be considered, which will cause great trouble to the classifier method.
The rule-based method is one of the most widely used methods at present, and although the method has the advantages of high accuracy, no requirement for corpus labeling, good interpretability and the like, the key defects of the method cannot be compensated, and considerable manpower and material resources are consumed in the process of rule making. The method based on the classifier needs to construct different classifiers and confusion sets aiming at different types of errors, the method based on the SMT can automatically learn the confusion sets from parallel data without other linguistic input, and multiple error types can be corrected by using one SMT model, so that the method is better at correcting complex errors. Despite the above advantages, the SMT based approach relies on large scale manual labeling of parallel corpora, whereas the classifier based approach is able to learn models from unlabeled corpora. In addition, the SMT-based GEC system is limited by generalization capability and cannot effectively access broader source and target contexts, so researchers apply the NMT method to the task of text automatic proofreading, and a series of RNN seq2 seq-based models are proposed.
In addition, at present, there is no automatic Chinese text proofreading method specially for judicial documents, and for the informatization development of the judicial field, the automatic Chinese text proofreading method specially for the judicial documents is exactly required by the automatic Chinese text proofreading method. The general automatic proofreading method for Chinese text often cannot identify proper nouns and legal terms in the judicial field, thus causing great trouble to the automatic proofreading of the text. Therefore, the calibration accuracy is not ideal and the device cannot be normally used under the actual condition.
At present, the mainstream method is to regard grammar error correction as a monolingual translation task, and the error correction process is to "translate" an error sentence into a correct sentence. With the progress of neural machine translation, a large number of methods and models are migrated into a grammar error correction task, however, before 2018, the seq2seq model based on RNN fails to exceed the system based on statistical machine translation, and the seq2seq model based on multilayer CNN adopted by Chollampatt et al becomes a neural network method exceeding the statistical machine translation system for the first time. On the basis, the invention uses a more advanced model Transformer to a Chinese grammar error correction task and obtains a better result.
Disclosure of Invention
The invention aims to provide an end-to-end judicial literature automatic proofreading method based on deep learning, so as to solve the defects in the existing text automatic proofreading method.
An end-to-end judicial literature automatic proofreading method based on deep learning comprises the following steps:
step one, a Transformer model structure is provided;
training the Transformer model to obtain the likelihood of the maximized model on training data S;
and step three, introducing a length penalty term into the likelihood obtained in the step two to obtain a decoding strategy.
Further, the Transformer model structure comprises an encoder and a decoder, wherein the encoder and the decoder each comprise 6 identical layers, and each layer comprises a self-attention sublayer and a forward neural network sublayer.
Further, there is an encoder-decoder attention sublayer between the encoder and decoder, and the respective sublayers are connected by applying residual errors at the output and are subjected to layer normalization in the same layer.
Furthermore, in the second step, in order to utilize the position information of the symbols in the sequence, a position code is blended into the input embedding, and the dimension of the position code and the implicit dimension d of the transform modelmodelSimilarly, the specific calculation formula is:
Figure BDA0002610192720000031
Figure BDA0002610192720000032
where pos is a position index of a symbol in the sequence, i indicates a component of a position-coded vector, and embedding is the output of the previous sublayer, including a query vector, a key vector, and a value vector.
Further, in step two, in training the Transformer model, the maximum likelihood estimation is used, and the goal is to maximize the likelihood of the model on the training data S:
θ＝argmax∑(x，y)∈S logp(y|x；θ)。
further, in step three, specifically, given an input error sentence x, the Transformer model generates a target segment correction sentence y using Beam Search decodingcorAt each time step, the first k candidate sentences with the highest score are reserved, a length penalty term is introduced into the original likelihood score to obtain a decoding strategy, and the calculation formula is as follows:
Figure BDA0002610192720000033
Figure BDA0002610192720000034
where LP is the length penalty, α is a hyper-parameter, and score is the likelihood score.
The main advantages of the invention are: the invention provides an end-to-end judicial literature automatic proofreading method based on deep learning, which uses an encoder-decoder model-Transformer based on an attention-self mechanism to effectively avoid the defects of a cyclic neural network and a convolutional neural network. In the Transformer, any position in the sequence can interact with all other positions through a self-attention mechanism, and the characteristic makes the long-distance dependence of the Transformer modeling become very direct; furthermore, the degree of parallelization of the computations in the transform is very high, and the computations at each position in the sequence can be done simultaneously. Currently, in machine translation, the performance of a Transformer is far better than that of a coder-decoder model based on a circular neural network or a convolutional neural network. Experimental results show that the method of the present invention is far superior to the performance of encoder-decoder models based on recurrent neural networks and convolutional neural networks.
Drawings
FIG. 1 shows the structure of a Transformer model.
Detailed Description
The technical solutions in the embodiments of the present invention will be described clearly and completely with reference to the accompanying drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
The invention refers to a Transformer encoder-decoder model which is obviously superior to other structures (a cyclic neural network and a convolutional neural network) in the machine translation task, converts grammar error correction into a monolingual machine translation task, and introduces the specific structure of the Transformer model, a model loss function and a training criterion in detail.
An end-to-end judicial literature automatic proofreading method based on deep learning comprises the following steps:
step one, a Transformer model structure is provided;
training the Transformer model to obtain the likelihood of the maximized model on training data S;
and step three, introducing a length penalty term into the likelihood obtained in the step two to obtain a decoding strategy.
Further, the Transformer model structure comprises an encoder and a decoder, wherein the encoder and the decoder each comprise 6 identical layers, and each layer comprises a self-attention sublayer and a forward neural network sublayer.
In particular, the Transformer model is similar to a conventional encoder-decoder, and comprises an encoder and a decoderFIG. 1 shows a structure diagram of a Transformer model. Given source-side error sentence x ═ x1，x2，...，xmB, the transform encoder encodes x as a set of implicit state representations e ═ e in continuous space1，e2，...，emBased on this representation, the transform decoder generates a target-side corrected sentence y ═ y { y } time-step by time-step1，y2，...，yw，}。
Further, there is an encoder-decoder attention sublayer between the encoder and decoder, and the sublayers are connected by applying residual errors at the output and are subjected to layer normalization.
Specifically, like a normal Transformer, the encoder and the decoder respectively include 6 identical layers, and the layers in the encoder are composed of a self-attention sublayer and a forward neural network sublayer. The input first passes through the self-attention sublayer, after which the same forward neural network is applied to the outputs from different locations of the attention sublayer. The layers at the decoder side also contain a self-attention sublayer and a forward neural network sublayer, and between these two, there is also an encoder-decoder attention sublayer, similar in structure to the attention layers in the typical recurrent neural network encoder-decoder model. At the output of all sub-layers of the encoder, decoder, residual concatenation is applied and layer normalization is done.
When encoding a symbol at one position in a sequence, the self-attention mechanism allows it to interact with symbols at other positions in the sequence to help generate a more optimal encoding for that symbol. In the self-attention sublayer, multiple attention points are applied, and a query vector, a key vector and a value vector are all output (or directly input embedding) from a previous sublayer.
The encoder-decoder attention sublayer is similar to the attention layer in a typical recurrent neural network encoder-decoder model, with the query vector q coming from the output of the previous sublayer in the decoder, and the set of key vectors K and the set of value vectors V both coming from the output of the encoder.
Furthermore, in the second step, since the Transformer model does not include any loop structure, in order to utilize the position information of the symbols in the sequence, a position code is merged into the input embedding, and the dimension of the position code and the implicit dimension d of the Transformer model are mergedmodelSimilarly, the specific calculation formula is:
Figure BDA0002610192720000051
Figure BDA0002610192720000052
where pos is a position index of a symbol in the sequence, i indicates a component of a position-coded vector, and embedding is the output of the previous sublayer, including a query vector, a key vector, and a value vector.
Further, in step two, in training the Transformer model, the maximum likelihood estimation is used, and the goal is to maximize the likelihood of the model on the training data S:
θ＝argmax∑(x，y)∈S logp(y|x；θ)。
further, in step three, specifically, given an input error sentence x, the Transformer model generates a target segment correction sentence y using Beam Search decodingcorAt each time step, the first k candidate sentences with the highest score are reserved, a length penalty term is introduced into the original likelihood score to obtain a decoding strategy, and the calculation formula is as follows:
Figure BDA0002610192720000061
Figure BDA0002610192720000062
wherein, LP is a length penalty item, and a sentence that is too long will cause grammar quality problems such as redundancy, so that a penalty is given to a sentence that is too long. Alpha is a hyperparameter and is a reasonable range obtained through experiments. Socre is a likelihood score, a likelihood score is calculated for all the obtained candidate sentences, then the sentences are sorted according to the likelihood score, and the first k sentences are taken as our results.
Specifically, in the experimental process, a series of methods based on different models are selected for comparison, and for fairness, results of a plurality of methods using additional corpus resources are omitted. The introduction of each baseline system is as follows: chollampatt et al methods of stacking multilayer convolutional neural networks as encoder-decoder models (MLConv), and systems in which they perform model integration (MLConv (4ens)) and reorder candidate correction sentences output by the syntactic error correction model using edit distance features (MLConv (4ens) + EO); chollampatt et al statistical machine translation method based systems (SMT + NNJM), their systems using a word-level statistical machine translation model and integrating a neural network model and a language model; yuan et al are based on a system of recurrent neural network encoder-decoder models with attention mechanism (RNNSearch + align), which additionally deals with spelling errors by means of an unsupervised word alignment model. Based on the model methods, the experimental data selected by the user is used for retraining, and the experimental result similar to that in the original text is obtained.
In the aspect of data fusion model selection, corresponding tests are carried out on a Transformer model, and the best effect is obtained in an NLPCC + HSK + judicial data fusion mode. Experiments show that the scale of the data can influence the effect of the grammar error correction method to a certain extent. Therefore, in subsequent model comparison experiments, an NLPCC + HSK + judicial data fusion mode is selected as a training data set, judicial test data which is divided in advance is completely used as a test set, and the following experiment results are obtained.
Figure BDA0002610192720000071
Table 1 model comparative experimental results
From experimental results, our method using Transformer outperformed others to some extent, MLConv (4ens) + EO being closest in performance to Transformer, but MLConv being less than about 5 points on a single model than Transformer. Other methods based on the recurrent neural network are inferior to those of the multi-layer convolutional neural network, and we guess that this is probably because the convolutional neural network has excellent ability to capture local features, and most of syntax errors occur in a part of a sentence.
In summary, considering the syntax error correction of the judicial writing as a machine translation problem, it is feasible to apply the Transformer model, and the effect can indeed exceed the existing syntax error correction method. Moreover, the problem that a large number of rules need to be manually made in the rule-based grammar error correction method is avoided, but the corpus scale problem in the special field can still be one of the main reasons for limiting the improvement of the effect of the method.
One specific embodiment is set forth below:
in the experimental process, a series of methods based on different models are selected for comparison, and for fairness, results of a plurality of methods using additional corpus resources are omitted. The introduction of each baseline system is as follows: chollampatt et al methods of stacking multilayer convolutional neural networks as encoder-decoder models (MLConv), and systems in which they perform model integration (MLConv (4ens)) and reorder candidate correction sentences output by the syntactic error correction model using edit distance features (MLConv (4ens) + EO); chollampatt et al statistical machine translation method based systems (SMT + NNJM), their systems using a word-level statistical machine translation model and integrating a neural network model and a language model; yuan et al are based on a system of recurrent neural network encoder-decoder models with attention mechanism (RNNSearch + align), which additionally deals with spelling errors by means of an unsupervised word alignment model. Based on the model methods, the experimental data selected by the user is used for retraining, and the experimental result similar to that in the original text is obtained.
In the aspect of data fusion model selection, corresponding tests are carried out on a Transformer model, and for a professional document, the best effect is obtained in an NLPCC + HSK + judicial data fusion mode. Experiments show that the scale of the data can influence the effect of the grammar error correction method to a certain extent. Therefore, in subsequent model comparison experiments, an NLPCC + HSK + judicial data fusion mode is selected as a training data set, judicial test data which is divided in advance is completely used as a test set, and the following experiment results are obtained.
Figure BDA0002610192720000081
TABLE 2 model comparison test results (common Chinese text)
Figure BDA0002610192720000082
TABLE 3 model comparison test results (professional paperwork)
From experimental results, our method using Transformer outperformed others to some extent, MLConv (4ens) + EO being closest in performance to Transformer, but MLConv being less than about 5 points on a single model than Transformer. Other methods based on the recurrent neural network are inferior to those of the multi-layer convolutional neural network, and we guess that this is probably because the convolutional neural network has excellent ability to capture local features, and most of syntax errors occur in a part of a sentence.

Claims (6)

1. An end-to-end judicial literature automatic proofreading method based on deep learning is characterized by comprising the following steps:
step one, a Transformer model structure is provided;
training the Transformer model to obtain the likelihood of the maximized model on training data S;
and step three, introducing a length penalty term into the likelihood obtained in the step two to obtain a decoding strategy.
2. The method of claim 1, wherein the transform model structure comprises an encoder and a decoder, each of the encoder and the decoder comprises 6 identical layers, and each layer comprises a self-attention sublayer and a forward neural network sublayer.
3. The method for automatic proofreading of end-to-end judicial documents based on deep learning according to claim 2, characterized in that there is also a sub-layer of encoder-decoder attention between encoder and decoder, each sub-layer is connected by applying residual error at output, and layer normalization is performed in the same layer.
4. The method for automatically proofreading an end-to-end judicial literature based on deep learning of claim 1, wherein in the second step, in order to utilize the position information of the symbols in the sequence, a position code is merged into the input embedding, and the dimension of the position code is equal to the implicit dimension d of the transform modelmodelSimilarly, the specific calculation formula is:
Figure FDA0002610192710000011
Figure FDA0002610192710000012
where pos is a position index of a symbol in the sequence, i indicates a component of a position-coded vector, and embedding is the output of the previous sublayer, including a query vector, a key vector, and a value vector.
5. The method for automatically proofreading the end-to-end judicial documents based on deep learning of claim 1, wherein in the step two, when training the Transformer model, the maximum likelihood estimation is used, and the goal is to maximize the likelihood of the model on the training data S:
θ＝argmax∑(x，y)∈Slogp(y|x；θ)。
6. the method of claim 1, wherein in step three, specifically, given an input error sentence x, the Transformer model generates a target segment correction sentence y by using Beam Search decodingcorAt each time step, the first k candidate sentences with the highest score are reserved, a length penalty term is introduced into the original likelihood score to obtain a decoding strategy, and the calculation formula is as follows:
Figure FDA0002610192710000021
Figure FDA0002610192710000022
where LP is the length penalty, α is a hyper-parameter, and score is the likelihood score.
CN202010751655.8A 2020-07-30 2020-07-30 End-to-end judicial literature automatic proofreading method based on deep learning Pending CN111985220A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010751655.8A CN111985220A (en) 2020-07-30 2020-07-30 End-to-end judicial literature automatic proofreading method based on deep learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010751655.8A CN111985220A (en) 2020-07-30 2020-07-30 End-to-end judicial literature automatic proofreading method based on deep learning

Publications (1)

Publication Number Publication Date
CN111985220A true CN111985220A (en) 2020-11-24

Family

ID=73445616

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010751655.8A Pending CN111985220A (en) 2020-07-30 2020-07-30 End-to-end judicial literature automatic proofreading method based on deep learning

Country Status (1)

Country Link
CN (1) CN111985220A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11747970B2 (en) 2021-09-23 2023-09-05 International Business Machines Corporation Interactive graphical display of multiple overlapping hypotheses or document versions

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101710985A (en) * 2009-12-11 2010-05-19 哈尔滨工业大学 Image brightness compensation method for image coding
CN110489521A (en) * 2019-07-15 2019-11-22 北京三快在线科技有限公司 Text categories detection method, device, electronic equipment and computer-readable medium
CN110737764A (en) * 2019-10-24 2020-01-31 西北工业大学 personalized dialogue content generating method
CN111160050A (en) * 2019-12-20 2020-05-15 沈阳雅译网络技术有限公司 Chapter-level neural machine translation method based on context memory network

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101710985A (en) * 2009-12-11 2010-05-19 哈尔滨工业大学 Image brightness compensation method for image coding
CN110489521A (en) * 2019-07-15 2019-11-22 北京三快在线科技有限公司 Text categories detection method, device, electronic equipment and computer-readable medium
CN110737764A (en) * 2019-10-24 2020-01-31 西北工业大学 personalized dialogue content generating method
CN111160050A (en) * 2019-12-20 2020-05-15 沈阳雅译网络技术有限公司 Chapter-level neural machine translation method based on context memory network

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
亢晓勉等: "融合篇章结构位置编码的神经机器翻译", 《智能科学与技术学报》 *
朱晨光著: "《机器阅读理解》", 30 March 2020, 机械工业出版社 *
邓俊锋等: "基于 Back－translation的语法错误纠正", 《智能计算机与应用》 *

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11747970B2 (en) 2021-09-23 2023-09-05 International Business Machines Corporation Interactive graphical display of multiple overlapping hypotheses or document versions

Similar Documents

Publication Publication Date Title
CN108920473A (en) A kind of data enhancing machine translation method based on similar word and synonym replacement
Brown et al. Analysis, statistical transfer, and synthesis in machine translation
US20050289463A1 (en) Systems and methods for spell correction of non-roman characters and words
Szarvas et al. A highly accurate Named Entity corpus for Hungarian
CN110276069A (en) A kind of Chinese braille mistake automatic testing method, system and storage medium
Dien et al. POS-tagger for English-Vietnamese bilingual corpus
Zhang et al. Named entity recognition method in health preserving field based on BERT
CN114662476A (en) Character sequence recognition method fusing dictionary and character features
CN115658898A (en) Chinese and English book entity relation extraction method, system and equipment
Roy et al. Unsupervised context-sensitive bangla spelling correction with character n-gram
Zhang et al. Design and implementation of Chinese Common Braille translation system integrating Braille word segmentation and concatenation rules
Darwish et al. Effective multi-dialectal arabic POS tagging
CN111274827B (en) Suffix translation method based on multi-target learning of word bag
CN111985220A (en) End-to-end judicial literature automatic proofreading method based on deep learning
Feng et al. Multi-level cross-lingual attentive neural architecture for low resource name tagging
CN111382583A (en) Chinese-Uygur name translation system with mixed multiple strategies
Deksne Bidirectional lstm tagger for latvian grammatical error detection
Saengthongpattana et al. Thai-english and english-thai translation performance of transformer machine translation
Duan et al. Pinyin as a feature of neural machine translation for Chinese speech recognition error correction
Narayan et al. Machine translation using quantum neural network for simple sentences
Athanaselis et al. A corpus based technique for repairing ill-formed sentences with word order errors using co-occurrences of n-grams
Rapp A Part-of-Speech-Based Search Algorithm for Translation Memories.
Getachew et al. Gex'ez-English Bi-Directional Neural Machine Translation Using Transformer
Yazar et al. Low-Resource Neural Machine Translation: A Systematic Literature Review
CN114328848B (en) Text processing method and device

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
RJ01 Rejection of invention patent application after publication

Application publication date: 20201124

RJ01 Rejection of invention patent application after publication