A low-resource sequence error detection method based on multi-dimensional explicit features and deep representations

Abstract

In order to guarantee the smooth running of cross-country information interaction and high-level cognitive tasks, it is urgent to combine intelligent text computing with low-resource sequence processing. Sequence error detection for low-resource languages is the basis for multilingual interaction. The present invention aims to identify long-tail anomalies in complex ligature sequences, and constructs an adaptive gated dual-flow fusion error detection model. A multi-dimensional explicit feature fusion method is proposed, that is, in addition to deep semantic mainstream, four-dimensional features such as membership, edit distance, rarity and script legality are extracted at the word level to build a rule auxiliary flow, overcoming the morphological perception blind area caused by sub-word segmentation. In addition, the present invention designs an adaptive gating unit to dynamically allocate dual-flow reliance weights, and combines probability alignment and cost-sensitive loss for optimization, constructing a morphological perception error detection model that adapts to complex variants, achieving better results in long-tail anomaly detection.

Description

Technical Background

[0001] With the rapid development of global digitalization, massive text sequences have become the core carrier of cross-border information exchange and knowledge transfer. In the field of intelligent text computing, performing high-precision anomaly detection and verification on the underlying sequences is an important prerequisite for ensuring the smooth operation of subsequent advanced cognitive tasks such as complex semantic analysis and machine translation. Early studies, such as [1], systematically summarized the anomaly recognition technology of automatic text sequences, while subsequent studies, such as [2], further demonstrated that the noise of the underlying sequence variants can cause serious cascading effects. Especially in rigorous translation tasks, subtle morphological deviations can significantly damage the feature alignment ability of the model, leading to the collapse of the performance of the entire computing system.

[0002] Research on text verification for specific low-resource languages ​​is of paramount practical necessity. Strategically, the expanding scale of digital trade between my country and South Asian countries means that the accuracy of underlying texts, such as Bengali, directly impacts the absolute compliance of cross-border financial settlements, customs data integration, and international legal contracts. In terms of cultural exchange, high-quality underlying sequence verification is a core foundation for the protection of digital classics and multilingual translation. However, the writing systems of low-resource languages ​​like Bengali are extremely complex, and the lack of high-quality annotated corpora presents a severe bottleneck for anomaly detection technology. As a vowel-tagged script, Bengali's physical rendering heavily relies on the implicit connector Hasant (encoded as U+09CD) for gating and fusion. Cognitive biases during the input phase can easily lead to the physical disintegration of ligature structures, resulting in degenerate variants of illegal rendering. These structural anomalies caused by defects in the underlying encoding mechanism exhibit a long-tail distribution, constituting a core challenge in sequence error detection tasks.

[0003] Looking back at the technological evolution, sequence anomaly detection has mainly undergone a transformation from rule-driven to statistical modeling and then to deep representation. Initially, the field mainly relied on hard-coded rules and static dictionary matching. For example, [3] proposed a speech coding-based inspection architecture to identify illegal isolated variants by modifying the pronunciation similarity algorithm. This method is detached from context and is difficult to cope with the high cost of feature comparison. Subsequently, [4] used the N-gram model to calculate the joint probability distribution of continuous word sequences and effectively marked the suspected anomaly positions in complex sequences by setting an empirical threshold for probability transition.

[0004] With the breakthrough of deep learning, the academic community has built complex neural networks to solve the problem of fine-grained verification. For example, [5] proposed a detection architecture that combines convolutional neural networks to extract local topological features, and [6] designed a verification framework based on deep representation, which has made significant progress in sequence localization and anomaly detection accuracy. However, when faced with extremely complex special sequences, there is still a bottleneck at the bottom layer: on the one hand, the absolute dependence of general pre-trained models on the sub-word segmentation mechanism [7] is very likely to cause morphological perception blind spots, causing the network to lose local character shape perception and making it difficult to capture long-tail anomalies such as the failure of underlying implicit connections and structural disintegration [8]; on the other hand, single deep semantic flow models are very likely to be misled by global coherence when performing anomaly calculations, and produce excessive feature smoothing when facing highly hidden variants in legal contexts, resulting in missed detections and degradation of detection performance [9].

[0005] To address the aforementioned challenges, this invention proposes a dual-stream feature fusion and verification method for complex concatenated word sequences. This invention constructs a "deep semantics-external rules" dual-stream feature extraction architecture, introduces four-dimensional explicit prior features including script legality in parallel, and designs an adaptive gated game unit to achieve dynamic trust allocation between macro-context and micro-physical rules. This invention aims to overcome the blind spots of pre-trained word segmentation perception and suppress conservative biases in anomaly detection, thereby significantly improving the detection efficiency for complex long-tailed anomaly sequences.

[0006] 2. Related work

[0007] 2.1. Methods for Text Sequence Validation and Error Detection

[0008] Text sequence validation and error detection is a fundamental task in Natural Language Processing (NLP), aiming to identify and locate anomalous variations in text sequences that violate grammatical rules, semantic logic, or underlying physical structure. It forms the basis for downstream NLP tasks such as relation extraction, machine translation, knowledge graphs, and question answering systems to obtain high-quality input. Early error detection tasks primarily relied on matching dictionaries and manually defined rules, making them highly dependent on expert-defined rule sets. With the development of machine learning, researchers have gradually applied traditional probabilistic graphical models to capture sequence features, significantly reducing manual costs while improving recognition efficiency. In recent years, deep learning methods, with their powerful feature representation capabilities, have become the mainstream solution for solving text validation tasks.

[0009] Among them, rule-based methods directly match accurate anomaly locations using a set of grammatical constraints manually defined by experts [10, 11], achieving extremely high accuracy in specific domains. Traditional machine learning methods treat error detection as a classic sequence labeling problem, aiming to assign a label to each unit in the sequence. The most commonly used models at present include Hidden Markov Models (HMM)

[12] and Random Field Models (CRF)

[13] . With the improvement of hardware computing power and the maturity of word embedding technology, neural networks have been shown to be able to handle complex sequence verification tasks more effectively. The most classic architectures include Deep Neural Networks (DNN)

[14] , Convolutional Neural Networks (CNN)

[15] , Recurrent Neural Networks (RNN)

[16] , and Long Short-Term Memory Networks (LSTM)

[17] . Inspired by these methods, many improved models have been widely used, such as Bi-LSTM models that introduce bidirectional feature flow [18, 19], models that add attention mechanisms [20, 21], and large-scale pre-trained models that introduce transfer learning [22, 23].

[0010] 2.2. Fusion of Multi-Source Explicit Features in Error Detection Tasks

[0011] English and other languages ​​have natural space separators, while for some complex low-resource languages ​​(such as Bengali), text sequences lack obvious physical boundaries, and their underlying rendering logic often contains implicit connection mechanisms. If word segmentation or word division is performed directly, a large number of features containing physical structure information will be fragmented, thus losing their verification value. Given that external explicit features (such as dictionary information, glyph rules, and script legality constraints) contain a lot of useful knowledge, the main idea of ​​verification algorithms in recent years has focused on how to integrate these multi-source information into the training process of neural networks to help the model obtain accurate semantic boundaries and physical bottom lines. For example,

[24] proposed the famous Lattice-LSTM model using the lattice structure between word sequences to solve the problem of information loss in sequence labeling. Subsequently, some studies have tried to integrate pre-trained context embedding with lexical information and achieved better results. For example,

[25] introduced the FLAT model, which combines the parallel capability of Transformer with external features through a planar lattice structure;

[26] proposed a simple and effective Soft-Lexicon method, which effectively improved the combination effect of the model and the pre-trained language model. However, existing methods mostly use shallow or randomly initialized linear layers to fuse features, without integrating them into the deep architecture of the pre-trained model. To address this,

[27] designed a lexical augmented BERT model for sequence labeling (LEBERT) that directly utilizes deep external knowledge by constructing specific adapters. Inspired by this kind of feature fusion idea, this invention proposes a dual-stream feature fusion scheme, which aims to integrate complex low-level physical rules with high-level semantic depth, thereby improving the detection efficiency of complex long-tailed anomaly sequences.

[0012] 3. Shortcomings of existing technologies

[0013] 3.1. Existing feature fusion verification methods suffer from architectural redundancy and interaction limitations.

[0014] In complex sequence error detection tasks, pure character-level networks typically require the introduction of external physical rules and dictionary priors to assist in boundary recognition. Existing feature fusion methods are mainly divided into two categories: shallow concatenation and deep reconstruction. Shallow concatenation methods simply stack discrete rules and continuous semantic vectors at the model input, lacking deep nonlinear interaction mechanisms, which can easily lead to difficulties in effectively aligning multi-source feature spaces. Deep reconstruction methods attempt to achieve feature embedding by modifying the core operators of the pre-training base (such as reconstructing the self-attention matrix). This alters the structural integrity of the pre-trained model, resulting in redundant model parameters, increased computational overhead during inference, and limited cross-language transfer and generalization capabilities. Furthermore, existing fusion strategies mostly employ static fusion mechanisms, lacking the ability to dynamically allocate weights based on contextual confidence. Therefore, existing methods face a constraint between performance improvement and architectural flexibility, urgently requiring a new, efficient fusion architecture with adaptive feature adjustment capabilities.

[0015] 3.2. General pre-trained models that rely on word segmentation lack awareness of underlying features.

[0016] While large-scale pre-trained language models (PLMs) excel in macroscopic semantic representation, they still have fundamental limitations when dealing with low-resource sequences with complex underlying topologies. First, general models heavily rely on word segmentation algorithms (such as BPE). This mechanism is primarily driven by corpus statistical frequency and struggles to account for the morphological constraints of specific languages. For example, when processing compound ligatures containing implicit connectors (such as the Hasant character in Bengali), the word segmentation mechanism easily leads to over-segmentation of complete local morphemes. This low-level segmentation disrupts the original implicit topological connections between characters, causing deep networks to lose their ability to perceive structural legitimacy during feature extraction, making it difficult to effectively capture long-tail variants such as ligature degradation or structural disintegration. Second, traditional single-stream semantic networks are susceptible to over-smoothing of features due to the influence of global coherence context when performing anomaly detection, making it difficult for the system to distinguish between highly concealed anomalies and legitimate semantic variants, leading to missed detections. This indicates that architectures relying on a single semantic stream and word segmentation have technical bottlenecks in microscopic variant detection. Summary of the Invention

[0017] 4.1. Design a two-stream network architecture that integrates multi-dimensional explicit prior features.

[0018] In error detection tasks involving complex hyphenated sequences, underlying script rules and physical structure constraints can provide crucial local prior information, assisting the model in accurately defining the boundaries of legitimate sequences. In this invention, we integrate multi-dimensional external explicit features into the detection model as an independent computational flow, enabling the model to effectively address the problem of identifying long-tailed anomalous variants caused by implicit connector failure. To deeply integrate this underlying information, this invention constructs a multi-dimensional feature extraction matrix including membership, rarity, and script legitimacy to perform low-level feature mapping on the input sequence. This method avoids the structural damage caused by traditional word segmentation, supplementing the network with physical topological priors in a simple and efficient manner.

[0019] 4.2. Constructing an adaptive gated feature fusion mechanism for complex variants

[0020] Different types of sequence anomalies exhibit significantly different degrees of dependence on macroscopic contextual semantics and microscopic physical rules. Traditional static feature concatenation methods struggle to achieve effective information alignment and weight balance when dealing with highly concealed variants, limiting the error detection performance of pre-trained bases in specific domains. To maximize the effectiveness of multi-source features in complex sequence verification tasks, this invention further designs an adaptive gated game unit based on the aforementioned two-stream network structure. This unit aims to dynamically evaluate and allocate the trust weights of the deep semantic stream and the external rule stream according to the current distribution of contextual features, effectively suppressing the false negative phenomenon caused by excessive feature smoothing in a single deep network, thereby constructing a high-precision and robust sequence error detection model. Attached Figure Description

[0022] Figure 1 Error detection framework based on adaptive gated dual-stream fusion network

[0023] 5.1. Basic Semantic Feature Extraction Model

[0024] In order to extract high-dimensional macro-global contextual features from complex hyphenated sequences, this invention abandons the traditional masked language model (MLM) and instead adopts the ELECTRA architecture based on the Replaced Token Detection (RTD) mechanism as the foundation for semantic feature extraction.

[0025] The core advantage of the ELECTRA architecture lies in its distributed "generator-discriminator" adversarial mechanism. During the pre-training phase, the model requires the discriminator to perform a binary classification on each unit in the sequence, determining whether it is the original, genuine input or a replaced, forged word. This mechanism has a high degree of isomorphism with the anomaly detection task of text sequences in terms of mathematical modeling and physical logic, giving it a naturally more acute ability to perceive error features.

[0026] Based on the pre-trained ELECTRA platform, after the input sequence is segmented into word fragments, each unit vector is constructed by superimposing token embeddings and position embeddings. Given an input sequence... The sequence is completely input into a multi-layer Transformer encoder. Through layer-by-layer abstraction and interaction via a bidirectional multi-head self-attention mechanism, a hidden state feature matrix containing rich macro-contextual dependencies is finally extracted. :

[0027] (1)

[0028] in, This represents the dimension of the hidden layer. For the output feature vector, the base model performs feature mapping and dimensionality reduction on it through a fully connected network, initially outputting a judgment result that depends only on the global context semantics.

[0029] 5.2. Sequence Error Detection Model Based on Adaptive Gated Two-Stream Fusion Network

[0030] This invention proposes a fine-grained dual-stream feature-enhanced sequence error detection model. It integrates the underlying script validity and physical boundary rules into a deep network via independent computational flows, effectively overcoming the loss of morphological awareness caused by word segmentation and accurately detecting hidden anomalies such as complex ligature degradation. The specific computational flow of the model is as follows:

[0031] (1) Deep semantic mainstream feature encoding layer: After obtaining the deep context hidden layer matrix output by the above ELECTRA encoder Subsequently, to map this to a specific error detection task space, this module designs a basic classification head based on a multilayer perceptron (MLP). This module performs feature processing using a non-linear activation function with Gaussian error linear units (GELUs) to generate a basic logit matrix that depends solely on deep semantics. :

[0032] (2)

[0033] in, The weight matrix is ​​a learnable matrix. This is the bias term. The output tensor corresponds to the initial space mapping for the three-class classification detection task (correct, general anomaly, and hyphenated degenerate structure anomaly).

[0034] (2) External Multidimensional Rule Auxiliary Flow Extraction Layer: In addition to the mainstream deep semantic analysis, this invention constructs an external rule auxiliary flow that bypasses the sub-word segmentation mechanism. For each target word in the sequence... This module directly extracts explicit priors at the complete physical vocabulary level from four dimensions: orthography, morphology, statistical distribution, and underlying rendering, constructing a 4-dimensional weak signal feature vector. :

[0035] (3)

[0036] in, This represents a hard membership constraint based on a static dictionary; To normalize the distance penalty term; This is a log-smoothed rarity feature based on the global word frequency distribution; This method renders legality constraints for scripts based on regular expressions and underlying Unicode encoding validation. The process transforms discrete external physical rules into continuous tensors and projects them as explicit feature correction terms using a single-layer network. .

[0037] (3) Adaptive Gated Game Theory Fusion Layer: To overcome the fusion bottleneck of heterogeneous features and avoid gradient conflicts or excessive feature smoothing, this invention designs an adaptive gating unit to perform dynamic trust allocation. First, the backbone semantic representation is... With 4D regular features Cascaded, input to residual calibration network to extract joint representation :

[0038] (4)

[0039] Subsequently, the adaptive gating scalar weight of the current word between the two streams is calculated through a dynamic routing mechanism. And an implicit residual correction term for micro-fine calibration. :

[0040] (5)

[0041] (6)

[0042] Ultimately, in the log-odds space, the model uses gated weights. Using this as a fulcrum, basic semantic prediction Explicit rule modification items and implicit residuals Perform weighted aggregation to generate the final predicted log odds:

[0043] (7)

[0044] (4) Probability Alignment and Cost-Sensitive Optimization Layer: Since the upstream pre-training base uses word segmentation, and the final error detection output requires word-level localization, this layer adopts a probabilistic inference-based Noisy-OR alignment strategy. The anomalous probability of a word fragment is mapped back to the anomalous probability of the complete word. :

[0045] (8)

[0046] Furthermore, to address the extreme long-tail imbalance distribution exhibited by legitimate correct units and degenerate hybrid units at the bottom layer in the sequence, this invention reconstructs a cost-sensitive cross-entropy loss function with penalty weights. :

[0047] (9)

[0048] in, This is a static penalty weight set according to the complexity of the anomaly category. This strategy forces the optimizer to favor rare and structurally broken anomaly samples during backpropagation, effectively improving the system's long-tail recall capability and global detection robustness.

[0049] 6. Specific training process

[0050] (1) Preprocessing and Multidimensional Feature Mapping: A hybrid training set containing real error-correcting corpus and morphological adversarial pseudo-data is constructed. For the input original text sequence, the underlying script parser is first used to extract four-dimensional explicit prior features including dictionary membership, normalized edit distance, word frequency rarity, and script rendering legitimacy, and these features are mapped into continuous rule feature vectors to construct an external rule auxiliary flow. At the same time, preprocessing for sub-word segmentation and word-level boundary alignment is performed to provide a unified physical coordinate system for subsequent dual-flow feature alignment.

[0051] (2) First-stage training (morphological feature enhancement): The preprocessed adversarial morphological enhancement pseudo-data is input into the adaptive gated dual-stream fusion network. In this stage, high-dimensional deep semantic features are extracted using the underlying pre-training base based on the RTD mechanism, and preliminary game-theoretic fusion is performed with the external rule-based auxiliary stream. By injecting high-density ligature and key break noise in this stage, the model is forced to learn the complex underlying morphological variant representations of Bengali, thereby establishing a preliminary ability to perceive illegal rendering structures.

[0052] (3) Second-stage fine-tuning and iterative optimization: A "backbone retention - head reset" transfer learning strategy is adopted to input the error-correction corpus with the real distribution into the model. First, the Noisy-OR probabilistic inference mechanism is used to map the log odds at the sub-word level back to the probability space at the macro-word level, so as to achieve logical alignment between the output and the original sequence. Then, a cost-sensitive weighted cross-entropy loss function is introduced, and asymmetric penalty weights are assigned according to the distribution scarcity of different error categories to calculate the deviation between the predicted label sequence and the real label sequence. The parameter weights of the adaptive gating unit and the classification head are continuously updated through the backpropagation algorithm, so as to suppress the gradient submersion of long-tail categories and achieve iterative optimization of the fine-grained error detection model.

[0053] (4) Testing, Evaluation and Performance Analysis: For the trained fine-grained sequence error detection model, performance was evaluated using a pre-defined and expert-verified test set. Precision, recall and F1 score were calculated for each category (correct words, general spelling errors, and ligature degeneration errors), and the macro-average index was analyzed in particular to quantitatively verify the effectiveness of the dual-stream fusion architecture in addressing morphological perception blind spots and improving the recall capability of long-tail anomalies.

[0054] 7. Compared with the prior art, what are the advantages of the technical solution of the present invention?

[0055] 7.1. An efficient and parallel multidimensional explicit feature fusion method is proposed.

[0056] Integrating external rule information into deep learning models typically relies on complex model architecture reconstruction or multi-task joint learning. Such methods are computationally intensive and have significant practical complexity in low-resource scenarios. To address this, this invention proposes an efficient and parallel feature fusion method: independently constructing an external rule auxiliary flow based on word-level granularity outside the backbone deep semantic network. By directly extracting explicit prior features from four dimensions—membership, normalized edit distance, statistical rarity, and script validity—this invention bypasses the morphological perception blind spot caused by word segmentation mechanisms, injecting underlying physical structure constraints into the model in a more direct and lightweight manner. This enables the error detection model to accurately capture deep structural anomalies such as ligature degradation.

[0057] 7.2. A morphological perception error detection model with adaptive decision-making capabilities was constructed.

[0058] General deep representation models are built on large-scale general corpora. When faced with specific sequences with complex rendering rules and morphological features, they are prone to over-smoothing due to interference from global semantic coherence, making it difficult to achieve optimal error detection results. To maximize the efficiency of the verification model in complex variant recognition, this invention designs an adaptive gating game unit based on a two-stream network. This unit can dynamically allocate trust weights between "deep global semantics" and "local physical rule baseline" based on the real-time confidence features of the input sequence. Combined with a cost-sensitive loss optimization strategy, this model effectively overcomes the false negative problem of single-stream models in extreme long-tail anomaly categories, significantly improving the accuracy and long-tail recall balance of the system in complex variant detection tasks.

[0059] 8. Other

[0060] References:

[0061] [1]Kukich K. Techniques for automatically correcting words in text[J]. ACM computing surveys (CSUR), 1992, 24(4): 377-439.

[0062] [2]Belinkov Y, Bisk Y. Synthetic and Natural Noise Both Break NeuralMachine Translation[C] / / ICLR. 2018.

[0063] [3]UzZaman N, Khan M. A Comprehensive Bengali Spelling Checker[C] / / ICCPB. 2006.

[0064] [4]Islam MIK, et al. A context-sensitive approach to find optimum language model for automatic Bangla spelling correction[J]. IJASCA, 2018, 9(11): 184-191.

[0065] [5]Rahman C, et al. BSpell: A CNN-blended BERT based Bangla spellchecker[C] / / BLP-2023. 2023: 7-17.

[0066] [6]Bijoy M H, et al. A transformer-based spelling error correctionframework for Bangla and resource scarce Indic languages[J]. Computer Speech& Language, 2025, 89: 101703.

[0067] [7]Sennrich R, et al. Neural machine translation of rare words withsubword units[C] / / ACL. 2016: 1715-1725.

[0068] [8]Khuyagbaatar B, et al. Evaluating subword tokenization: Aliensubword composition and oov generalization challenge[J]. arXiv: 2404.13292,2024.

[0069] [9]Kong Z, et al. Scholargec: Enhancing controllability of largelanguage model for chinese academic grammatical error correction[C] / / AAAI.2025.

[0070]

[10] Forney G D. The viterbi algorithm[J]. Proceedings of the IEEE,2005, 61(3): 268-278.

[0071]

[11] Brill E. A simple rule-based part of speech tagger[C] / / Speech andNatural Language: Proceedings of a Workshop Held at Harriman, New York,February 23-26, 1992. 1992.

[0072]

[12] Rabiner L R. A tutorial on hidden Markov models and selectedapplications in speech recognition[J]. Proceedings of the IEEE, 2002, 77(2):257-286.

[0073]

[13] Lafferty J, McCallum A, Pereira F C N. Conditional random fields:Probabilistic models for segmenting and labeling sequence data[J]. 2001.

[0074]

[14] Collobert R, Weston J, Bottou L, et al. Natural languageprocessing (almost) from scratch[J]. 2011.

[0075]

[15] Kim Y. Convolutional neural networks for sentence classification[C] / / Proceedings of the 2014 conference on empirical methods in naturallanguage processing (EMNLP). 2014: 1746-1751.

[0076]

[16] Mikolov T, Karafiát M, Burget L, et al. Recurrent neural networkbased language model[C] / / Interspeech. 2010, 2(3): 1045-1048.

[0077]

[17] Graves A. Long short-term memory[J]. Supervised sequencelabelling with recurrent neural networks, 2012: 37-45.

[0078]

[18] Graves A, Fernández S, Schmidhuber J. Bidirectional LSTM networksfor improved phoneme classification and recognition[C] / / Internationalconference on artificial neural networks. Berlin, Heidelberg: Springer BerlinHeidelberg, 2005: 799-804.

[0079]

[19] Huang Z, Xu W, Yu K. Bidirectional LSTM-CRF models for sequencetagging[J]. arXiv preprint arXiv:1508.01991, 2015.

[0080]

[20] Vaswani A, Shazeer N, Parmar N, et al. Attention is all you need[J]. Advances in neural information processing systems, 2017, 30.

[0081]

[21] Bahdanau D, Cho K, Bengio Y. Neural machine translation byjointly learning to align and translate[J]. arXiv preprint arXiv:1409.0473,2014.

[0082]

[22] Devlin J, Chang M W, Lee K, et al. Bert: Pre-training of deepbidirectional transformers for language understanding[C] / / Proceedings of the2019 conference of the North American chapter of the association forcomputational linguistics: human language technologies, volume 1 (long andshort papers). 2019: 4171-4186.

[0083]

[23] Sarzynska-Wawer J, Wawer A, Pawlak A, et al. Detecting formalthought disorder by deep contextualized word representations[J]. Psychiatryresearch, 2021, 304: 114135.

[0084]

[24] Zhang Y, Yang J. Chinese NER using lattice LSTM[C] / / Proceedingsof the 56th Annual Meeting of the Association for Computational Linguistics(Volume 1: Long Papers). 2018: 1554-1564.

[0085]

[25] Li X, Yan H, Qiu X, et al. FLAT: Chinese NER using flat-latticetransformer[C] / / Proceedings of the 58th annual meeting of the association forcomputational linguistics. 2020: 6836-6842.

[0086]

[26] Ma R, Peng M, Zhang Q, et al. Simplify the usage of lexicon inChinese NER[C] / / Proceedings of the 58th annual meeting of the association forcomputational linguistics. 2020: 5951-5960.

[0087]

[27] Liu W, Fu X, Zhang Y, et al. Lexicon enhanced Chinese sequencelabeling using BERT adapter[C] / / Proceedings of the 59th annual meeting of theassociation for computational linguistics and the 11th international jointconference on natural language processing (volume 1: long papers). 2021:5847-5858。

Claims

1. A low-resource sequence error detection method based on multidimensional explicit features and deep representation, characterized in that, A dual-stream feature extraction architecture of "deep semantics-external rules" was constructed. By introducing multi-dimensional explicit prior features, including script legality, the model is assisted in perceiving the boundary of the underlying physical structure of the sequence. Adaptive gated game units are used to realize dynamic trust allocation between macro context and micro physical rules.

2. The low-resource sequence error detection method based on multidimensional explicit features and deep representation according to claim 1, characterized in that, An external multidimensional rule auxiliary flow extraction method is proposed, which directly extracts four explicit features of dictionary membership, normalized edit distance, statistical rarity and script validity at the granularity of complete physical vocabulary, constructs external rule auxiliary flow and projects it as explicit feature correction terms to make up for the lack of morphological perception caused by the sub-word segmentation mechanism.

3. The low-resource sequence error detection method based on multidimensional explicit features and deep representation according to claim 1, characterized in that, To enhance the ability to detect long-tail anomalies, ELECTRA based on the substitution marker detection (RTD) mechanism is used as the semantic extraction base, and a two-stage training strategy is implemented: the first stage uses morphological adversarial augmentation data for training to strengthen the model's perception of illegal rendering structures. The second stage uses a real-distribution error-correcting corpus for fine-tuning to eliminate data distribution bias and improve the model's engineering robustness.

4. The low-resource sequence error detection method based on multidimensional explicit features and deep representation according to claim 1, characterized in that, Includes the following steps: Step 1: Preprocess the sequence to be detected and extract multidimensional features to obtain the deep global semantic mainstream representation and the word-granular external multidimensional rule auxiliary flow representation of the input sequence respectively; Step 2: Input the dual-stream features into the adaptive gated game unit, and dynamically calculate the gate weights based on the semantic confidence and local glyph anomaly at the current time to achieve adaptive fusion of heterogeneous features; Step 3: Use the Noisy-OR alignment strategy based on probabilistic reasoning to map the sub-word level probability to the word level probability space, and backpropagate the prediction bias through the reconstructed cost-sensitive weighted loss function to iteratively optimize the model's recognition accuracy on long-tail categories; Step 4: For the trained error detection model, use an independent test set to evaluate its performance and analyze its precision, recall and global balance under different anomaly categories.

Images