Method for decoding semantic communication based on space-based computation

Abstract

The application discloses a semantic communication decoding method based on space-based computing, comprising the following steps: receiving a semantic encoding signal transmitted through a channel, and generating an initial decoding distribution through a non-autoregressive semantic decoder; constructing an empty information observation adapting to the current channel characteristics, inputting the observation into the decoder to estimate an implicit language prior, and correcting the initial decoding distribution to obtain a corrected channel likelihood distribution; combining the semantic prior of a pre-trained language model, quantifying the channel uncertainty and semantic uncertainty of each word position respectively, and generating a refinement priority based on the source ratio of the channel uncertainty in the total uncertainty; and iteratively refining the text sequence according to the priority. The application can explicitly eliminate the internal prior bias of a large model by using an empty observation, suppress the illusion phenomenon under a low signal-to-noise ratio, and realize the on-demand allocation of computing resources by distinguishing the uncertainty sources, thereby improving the accuracy and efficiency of decoding.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of wireless communication and artificial intelligence, and in particular, it is a semantic communication decoding method based on space-based computing. Background Technology

[0002] In satellite-ground collaborative computing and satellite internet scenarios, spectrum resources are scarce and channel environments are complex and variable, posing significant challenges to bandwidth and power consumption for traditional bit-transparent transmission. Semantic communication, by extracting and transmitting semantic features behind data, can significantly reduce data transmission volume and improve robustness in low signal-to-noise ratio environments, becoming a key technological path for achieving efficient space-based information exchange. Utilizing the powerful contextual understanding and generation capabilities of Large Language Models (LLMs) for receiver decoding is currently an important means to improve the quality of semantic recovery.

[0003] Existing semantic communication systems mostly employ a deep learning-based source-channel joint coding (JSCC) architecture. At the receiver, a pre-trained language model is typically used directly as the decoder to map the noisy semantic features transmitted through the channel into a text sequence. These methods primarily rely on the probability distribution learned by the model from massive corpora to fill the information gaps caused by channel fading, i.e., utilizing the model's internal knowledge base to assist signal recovery. The decoding process usually employs autoregressive or non-autoregressive methods, directly outputting the sequence of words with the highest probability based on the received feature vectors.

[0004] However, in noisy electromagnetic environments such as space-based computing, several issues remain to be addressed. Specifically, the complex noise of the physical channel weakens the effective features received by the decoder, causing the implicit bias in the linguistic statistical priors (logits domain) embedded within the large language model to overshadow the actual features. Existing mechanisms fail to decouple this implicit prior from the physical channel likelihood, leading the model to often ignore channel facts and overconfidently predict errors based solely on linguistic probabilities. This results in the illusion of semantic coherence that contradicts the intended transmission, and conventional mathematical correction methods struggle to capture this nonlinear bias. Furthermore, existing iterative decoding strategies typically trigger backtracking based on a single confidence threshold, failing to distinguish whether low confidence stems from noise degradation of the physical channel (channel uncertainty requiring correction) or from inherent ambiguity in the language itself (semantic uncertainty requiring no correction). This results in ineffective iterative calculations of numerous inherently ambiguous but uncorrectable terms (such as open semantics) under extremely limited computing power and latency budgets in satellite-to-ground links, causing congestion or delays in critical instruction processing. Summary of the Invention

[0005] The purpose of this invention is to provide a semantic communication decoding method based on space-based computing to solve the aforementioned problems in the existing technology.

[0006] The technical solution, a semantic communication decoding method based on space-based computing, includes:

[0007] Receive the semantic encoded signal transmitted through the channel, input the semantic encoded signal into the pre-configured non-autoregressive semantic decoder, and generate an initial decoding distribution covering the candidate word set;

[0008] Construct a spatial information observation adapted to the current channel characteristics, input the spatial information observation into a non-autoregressive semantic decoder to estimate the implicit language prior, and perform debiasing correction on the initial decoding distribution based on the implicit language prior to obtain the corrected channel likelihood distribution;

[0009] Based on the corrected channel likelihood distribution and the semantic prior distribution from the pre-trained language model, the channel uncertainty and semantic uncertainty of each word position are quantified respectively, and the refinement priority of each word position is generated based on the uncertainty source ratio of the channel uncertainty in the total uncertainty.

[0010] The text sequence obtained from the initial decoding of the pre-stored text is iteratively refined according to the refinement priority until the preset convergence condition is met, and the final recovered text is obtained.

[0011] Beneficial effects: This invention can explicitly eliminate the internal prior bias of large models by using empty observations, suppress the illusion phenomenon under low signal-to-noise ratio, and improve the accuracy and efficiency of decoding by distinguishing the sources of uncertainty and realizing on-demand allocation of computing resources. Attached Figure Description

[0012] Figure 1 A flowchart illustrating the steps of a space-based computing-based semantic communication decoding method provided in this application embodiment.

[0013] Figure 2 A flowchart illustrating the steps for constructing spatial information observations adapted to the current channel characteristics, as provided in this application embodiment.

[0014] Figure 3 This is a flowchart illustrating the steps for correcting bias in the initial decoding distribution based on implicit language priors, as provided in an embodiment of this application.

[0015] Figure 4 A flowchart illustrating the steps for obtaining the corrected channel likelihood distribution provided in this application embodiment. Detailed Implementation

[0016] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0017] It should be noted that the terms include and have, and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or device.

[0018] like Figure 1 As shown, a semantic communication decoding method based on space-based computing includes the following steps:

[0019] The semantic encoded signal transmitted through the channel is received and input into a pre-configured non-autoregressive semantic decoder to generate an initial decoding distribution covering the candidate word set.

[0020] In this embodiment, the receiving end acquires the signal transmitted via the wireless channel through the radio frequency front-end and the baseband processing module. The semantically encoded signal refers to the signal sequence generated at the transmitting end after semantic feature extraction and joint encoding of the source channel. It typically contains continuous numerical vectors carrying semantic residual features and auxiliary identification information indicating the semantic hierarchy. Specifically, the non-autoregressive semantic decoder can employ a neural network model based on a Transformer architecture, capable of outputting the probability distribution of all positions in the sequence in parallel, without needing to generate each word sequentially as in an autoregressive model. The initial decoding distribution is the decoder's original output after receiving the current signal, such as the output of the normalized exponential function (Softmax) layer, reflecting the decoder's predicted probability of candidate words at each word position before bias correction.

[0021] A spatial information observation adapted to the current channel characteristics is constructed. The spatial information observation is input into a non-autoregressive semantic decoder to estimate the implicit language prior. Based on the implicit language prior, the initial decoding distribution is debiased and corrected to obtain the corrected channel likelihood distribution.

[0022] Specifically, the empty-information observation is a constructed virtual input signal that simulates the receiver's state when the transmitter does not send any valid semantic information. For example, in an additive white Gaussian noise channel, the transmitted signal is a zero vector, and the receiver only observes noise; the expected value of this observation is the empty-information observation. When the empty-information observation is fed into the decoder, since the input contains no semantic information, the decoder's output is determined by the language bias fixed in its parameters; this output is the implicit language prior. Through mathematical operations, such as subtraction in the logarithmic domain, the implicit language prior component is removed from the initial decoded distribution, restoring the corrected channel likelihood distribution that purely reflects the physical transmission quality of the channel. This effectively suppresses the large model illusion, where the model tends to generate fluent but erroneous text when the channel noise is high.

[0023] For example, during end-to-end training, the non-autoregressive semantic decoder optimizes its parameters with the goal of minimizing the reconstructed cross-entropy from the received signal to the source text. The training corpus consists of natural language text, which inherently contains linguistic patterns such as word frequency statistics, word order collocation, and grammatical structure. After extensive training, the decoder's network parameters inevitably encode these linguistic statistical patterns, forming a language bias distribution implicitly within the model. This is defined as the implicit language prior, denoted as P. imp (v), where v is a candidate word in the vocabulary. According to Bayes' theorem, the decoder receives a noisy signal y. t Then, the probability distribution P of its output at the t-th word position. dec (v|y t In fact, it approximates the posterior probability, rather than pure channel likelihood. The relationship between the two can be expressed by the following formula:

[0024] P dec (v|y t Proportional to P(y) t |v)*P imp (v);

[0025] Where P dec (v|y t P(y) represents the probability value output by the decoder for candidate word v at the t-th word position. t |v) represents the true channel likelihood, i.e., the observed received signal y assuming the transmitted content is v. t The physical probability, P imp (v) represents the language prior distribution implicitly learned by the decoder.

[0026] If we consider the above P containing implicit priors... dec It is directly fed as the channel likelihood into the subsequent Bayesian fusion process, along with the semantic prior P provided by the pre-trained language model. sem Multiplying them again, the fused posterior distribution is equivalent to:

[0027] P post (v) is proportional to P(y) t |v)*P imp (v) αt *P sem (v) (1-αt );

[0028] Where P post (v) represents the posterior probability after fusion, α t To integrate weights, P imp (v) αt Representing implicit priors with α t The power-law form remains in the channel likelihood term. It can be seen that language prior information is factored in twice in the expression: once through P... imp Implicit in the channel likelihood term, one pass through P sem Explicit introduction. Prior double inclusion causes the posterior distribution to be overly biased towards high-frequency common terms, systematically suppressing correct terms that are low-frequency but have sufficient channel evidence. This is particularly pronounced in low signal-to-noise ratio environments, manifesting as a confidence illusion where the decoder output is semantically fluent but contradicts the sending intent. The debiasing correction operation in this embodiment aims to eliminate the aforementioned prior double inclusion problem by removing implicit linguistic priors from the decoder output, restoring a correction distribution that approximates the true channel likelihood.

[0029] Based on the corrected channel likelihood distribution and the semantic prior distribution from the pre-trained language model, the channel uncertainty and semantic uncertainty of each lexical position are quantified respectively, and the refinement priority of each lexical position is generated based on the uncertainty source ratio of channel uncertainty in the total uncertainty.

[0030] In this embodiment, a pathogenesis diagnosis mechanism is introduced. Channel uncertainty can be measured using the information entropy of the corrected channel likelihood distribution; a higher entropy value indicates more severe channel transmission impairment. Semantic uncertainty is measured using the prediction distribution entropy of an externally pre-trained language model, reflecting the degree of ambiguity in the context itself. The uncertainty source ratio is a quantitative indicator used to distinguish whether the uncertainty at the current location is mainly caused by channel noise or by linguistic ambiguity. The refinement priority generated based on the uncertainty source ratio can guide subsequent steps to prioritize processing locations with poor channel quality but originally clear semantics, while skipping locations with good channel quality but inherent semantic ambiguity, thus achieving precise allocation of computing resources.

[0031] The text sequence obtained from the initial decoding of the pre-stored text is iteratively refined according to the refinement priority until the preset convergence condition is met, and the final recovered text is obtained.

[0032] Specifically, the system uses an iterative algorithm to correct the text based on the calculated priority order. Preferably, a constraint-binding search strategy based on attention masks is adopted, that is, attention windows are opened only for high-priority areas to be refined for resampling, while low-priority areas remain frozen. The iterative process continues until a preset convergence condition is met, such as the semantic distance between two adjacent decoding results being lower than a certain threshold, or the maximum allowable latency being reached. The final recovered text output is the high-quality decoding result after bias correction and targeted refinement.

[0033] Alternatively, the semantic communication decoding method based on a large language model can be described as follows: receiving a semantically encoded signal transmitted through a channel; inputting the semantically encoded signal into a non-autoregressive semantic decoder to generate an initial decoding distribution covering a candidate set of lexical terms; determining candidate lexical terms at each lexical term position based on the initial decoding distribution to obtain an initial text sequence; constructing a spatial information observation adapted to the current channel characteristics; inputting the spatial information observation into the non-autoregressive semantic decoder to estimate the implicit language prior; and correcting the initial decoding distribution based on the implicit language prior to obtain a corrected channel likelihood distribution; inputting the initial text sequence into a pre-trained language model to obtain the semantic prior distribution at each lexical term position; quantifying the channel uncertainty and semantic uncertainty at each lexical term position based on the corrected channel likelihood distribution and the semantic prior distribution; and generating a refinement priority for each lexical term position based on the ratio of channel uncertainty to uncertainty sources in the total uncertainty; and iteratively refining the initial text sequence according to the refinement priority until a preset convergence condition is met to obtain the final recovered text.

[0034] In one possible implementation, the semantically encoded signal is generated by the transmitter through a cascaded residual variational information bottleneck architecture, which contains a predetermined number of cascaded encoding stages.

[0035] Specifically, the sending end maps the input text sequence X to a deep semantic feature representation Z. Z then enters a concatenated coding system containing L stages. Unlike traditional parallel coding, this embodiment employs a serial concatenated structure, allowing semantic information to increase layer by layer in granularity. The information bottleneck theory seeks the optimal trade-off between compressing information and preserving task relevance. Given the original semantic features Z and the semantic task objective S to be preserved, the information bottleneck theory optimizes the coding mapping to obtain the compressed semantic representation T, retaining as little original information about Z as possible (i.e., sufficient compression) while preserving as much predictive ability about S as possible (i.e., semantic integrity). Its optimization objective can be expressed as:

[0036] min I(Z;T)-λ*I(T;S;

[0037] Where I(Z;T) is the mutual information between the original features Z and the compressed representation T, used to measure how much information the compressed representation retains relative to the original features, and the smaller the value, the more complete the compression; I(T;S) is the mutual information between the compressed representation T and the semantic task objective S, used to measure the predictive ability of the compressed representation for the semantic objective, and the larger the value, the more complete the semantic preservation; λ is a positive-zero Lagrange multiplier used to control the strength of the trade-off between the two objectives of compression and preservation; min represents minimization. The value of λ directly determines the semantic granularity of the compressed representation: when λ is small, the optimization result tends to be highly compressed, retaining only the core semantic skeleton; when λ is large, the optimization result allows for the retention of more detailed information. In this embodiment, an increasing sequence of Lagrange multipliers λ1 < λ2 < ... < λ is set. L This is to achieve multi-level semantic decomposition from coarse to fine, where L is the total number of semantic levels.

[0038] In the l-th encoding stage, the original semantic feature representation and the cumulative reconstructed semantic representation output from the previous l-1 encoding stages are used as joint inputs to generate the semantic residual increment of the l-th layer. The semantic residual increment only encodes the new semantic information relative to the cumulative reconstructed semantic representation.

[0039] In this embodiment, the encoder Enc of the l-th layer l Receive the original feature Z and the cumulative reconstructed feature T from the previous layer. l-1 As input, its output is used to generate the semantic residual increment ΔT of the l-th layer through a reparameterization technique. l The specific calculation process is as follows:

[0040] The encoder outputs the mean vector μ l and standard deviation vector σ l μ l , σ l =Enc l (Z, T) l-1 );

[0041] Where μ l Let σ be the mean distribution of the residuals in the l-th layer. l Let Enc be the standard deviation distribution of the residuals in the l-th layer. l This is the encoding function for the neural network in layer l.

[0042] The residual increment ΔT is generated by sampling standard Gaussian noise ε~N(0,I). l =μ l +σ l *ε;

[0043] Where * denotes element-wise multiplication, ε is a randomly sampled standard Gaussian noise vector, and N(0, I) represents a multivariate standard normal distribution. The residual increment ΔT lUsed to supplement semantic details that were not covered by the first l-1 layers.

[0044] The cumulative reconstructed features of the current layer are updated to: T l =T l-1 +ΔT l ;

[0045] The residual structure mathematically enforces nesting between levels, so that lower-level encoding (e.g., l=1) contains only the global theme, while higher-level encoding (e.g., l=L) gradually supplements the minor decorative details.

[0046] In the cascaded architecture described above, the optimization objective for each layer is a conditional information bottleneck. For the l-th layer (l≥2), the optimization objective is:

[0047] min I(Z;ΔT l | T l-1 )-λ l *I(ΔT l ;S| T l-1 );

[0048] Where I(Z; ΔT) l | T l-1 To reconstruct T from the known previous layers l-1 Under the condition of original feature Z and residual increment ΔT of layer l l The conditional mutual information between them is used to constrain the l-th layer increment to carry only the necessary new information based on the previous layer; I(ΔT) l ;S| T l-1 λ represents the contribution of the residual increment to the prediction of the semantic target S under the known reconstruction conditions of the previous layer, and is used to measure how much semantic detail the increment effectively supplements based on the previous layer; l The Lagrange multipliers for the l-th layer increase with each layer. The optimization objective for the l-th layer is the unconditional form min I(Z;ΔT1)-λ1*I(ΔT1;S).

[0049] Since the mutual information term cannot be directly computed in a high-dimensional continuous space, a variational approximation method is preferred to transform it into a loss function that can be optimized by stochastic gradient descent. Specifically, the encoding distribution is parameterized as a diagonal Gaussian distribution, and a standard normal prior distribution p0(ΔT) is set for each stage. l The regularization criterion is N(0, I). Conditional mutual information I(Z; ΔT) is used. l | T l-1 The variational upper bound of the distribution is obtained by calculating the KL divergence between the coding distribution and the standard normal prior. Since both are Gaussian in form, the KL divergence has a closed-form solution for the t-th lexical position in the l-th stage:

[0050] D KL_l_t=(1 / 2)*Σ j (σ l_t_j 2 +μ l_t_j 2 -1-log(σ l_t_j 2 ));

[0051] Where D KL_l_t Let Σ be the KL divergence value of the t-th word position in the l-th layer. j This represents the summation of the components j representing dimension d, σ l_t_j Let μ be the standard deviation parameter of the j-th dimension at the t-th position in the l-th layer. l_t_j Let be the corresponding mean parameter, and log be the natural logarithm function. The variational rate term R for stage l is obtained by averaging over all word positions. hat_IB (l):

[0052] R hat_IB (l)=(1 / T)*Σ t D KL_l_t ;

[0053] Where R hat_IB (l) represents the variational rate term of the l-th layer, i.e., the average information carried by the increment of this layer relative to the prior distribution; T is the word sequence length; Σ t This represents the summation over all word positions. R hat_IB (l) is a non-negative real number; a larger value indicates that the increment of this layer contains more information about the original feature Z. The variational rate term will replace information entropy as a metric for the amount of information in each layer during subsequent bitrate allocation. The reparameterized sampling operation ΔT in this embodiment... l =μ l +σ l *ε separates randomness from network parameters into an external noise variable ε, allowing gradients to propagate back through the sampling operation to the encoded network parameters, thus enabling end-to-end training.

[0054] The semantically encoded signal contains the semantic residual increments corresponding to each level.

[0055] Specifically, the final transmitted semantically encoded signal is generated by ΔT from each layer. l The sequence is composed after quantization. Furthermore, to support source-aware analysis at the receiver, the sender also calculates and transmits semantic importance weights ω(l). Preferably, the semantic importance weights are based on the task relevance loss L of each layer on the validation set. rel The calculation yielded:

[0056] ω(l)=exp(-L rel (l) / τ) / ∑exp(-L rel (k) / τ);

[0057] Where ω(l) is the semantic importance weight of the l-th layer, L rel (l) represents the correlation loss of the l-th layer, τ is the temperature parameter, and k is the layer index.

[0058] The semantic importance weight reflects the contribution of the l-th layer residual to semantic understanding and will be used by the receiver to calculate the refinement priority. Simultaneously, the transmitter dynamically adjusts the quantization precision of each layer based on the current channel state information (CSI) through a continuous mapping network, achieving joint source-channel adaptation.

[0059] Furthermore, the task relevance loss L in the semantic importance weight ω(l) rel (l) The training process originates from the cascaded variational information bottleneck (VIB). During the training phase, the relevance term I(ΔT) in the information bottleneck optimization objective... l ;S| T l-1 Variational approximation is achieved by introducing an auxiliary decoding network. The auxiliary decoding network uses the residual increment ΔT... l and the previous layer cumulative reconstruction T l-1 The concatenated vector is taken as input, and the label space of the semantic task objective S is taken as output. During training, the semantic information contained in the residual increments of each layer is learned to be predicted. The prediction loss of the auxiliary decoding network is the correlation loss L. rel (l) When S is a discrete classification label, cross-entropy loss is used, and when S is a continuous semantic embedding, mean squared error loss is used.

[0060] Optionally, the semantic task objective S employs a multi-granularity semantic labeling system, organized hierarchically from coarse to fine, with each level labeling semantic components of different granularities in the original text. Taking natural language text as an example, this labeling system includes four levels: the first level is topic semantic labels, labeling the topic category of each text paragraph or sentence, with the coarsest granularity, reflecting the overall theme; the second level is entity and key concept labels, labeling the location and type of each named entity and core concept phrase in the text, with medium granularity, identifying the skeletal information carriers; the third level is syntactic role labels, labeling the syntactic function of each word, with relatively fine granularity; and the fourth level is semantic relation labels, labeling the fine-grained semantic dependencies between words, with the finest granularity. All labels at each level are automatically obtained by pre-trained natural language processing tools, such as text classification models, named entity recognition models, dependency parsing machines, and semantic role labeling models, from the training corpus, without requiring manual labeling. The correspondence between labels at each level and information bottleneck levels is naturally established through an increasing sequence of Lagrange multipliers: when λ1 is minimized, information bottleneck optimization retains only information related to the coarsest-grained label under highly compressed constraints; λ L At its maximum, optimization allows for the retention of more information, gradually covering the semantic components corresponding to fine-grained tags.

[0061] Combining the above components, the variational training objective for each encoding stage l is:

[0062] L VIB (l)=L rate (l)+λ l *L rel (l);

[0063] Where L VIB (l) represents the variational information bottleneck loss in stage l; L rate (l) represents the variational rate term loss in stage l, i.e., the batch expectation of the KL divergence between the encoding distribution and the standard normal prior, used to constrain the degree of information compression; λ l L is the Lagrange multiplier of the l-th layer, used to balance compression and preservation; rel (l) represents the relevance loss of the auxiliary decoding network in stage l, i.e., the negative log-likelihood of the predicted semantic labels, used to ensure semantic preservation. The joint VIB loss of all L stages is:

[0064] L VIB =Σ l L VIB (l);

[0065] Where L VIB For the VIB joint loss across all stages, Σ l This represents the summation of l from 1 to L. L VIB With reconstruction loss L recon Together, they constitute a complete end-to-end training objective.

[0066] According to one aspect of this application, the semantic importance weight ω(l) is used as a fixed constant after training, and is correlated with the hierarchical index l of each word position. t Together, they are encapsulated into semantic-level identifier information. This semantic-level identifier information, in the form of low-overhead side information, is transmitted to the receiving end along with the semantically encoded signal. After the receiving end recovers the semantic-level identifier information, it directly reads ω(l). t ω(l) is used to calculate the refinement priority. ω(l) does not change with the input data during the inference phase and is a static parameter determined during the system initialization phase.

[0067] Furthermore, during the encoding process at the transmitting end, the transmission bitrate for each semantic level is allocated according to a joint weighted average of semantic importance and information content. The total available transmission bitrate R... total Under the constraints, the code rate R allocated in the l-th layer l The calculation is as follows:

[0068] R l =R total *(ω(l)*R hat_IB (l)) / Σ l_prime (ω(lprime )*R hat_IB (l prime ));

[0069] Where R l The transmission code rate allocated to layer l; R total R represents the total available code rate determined by the current channel conditions; ω(l) represents the semantic importance weight of the l-th layer; hat_IB (l) represents the variational rate term of the l-th layer, used to measure the information content scale of the semantic content of this layer; Σ l_prime This indicates that normalized summation is performed over all L levels; prime To sum and iterate through variables. In the numerator, ω(l) and R... hat_IB The product of (l) takes into account both semantic importance and information content: even if a certain layer has a large amount of information, if its semantic importance is low, the allocated code rate will be limited accordingly; conversely, even if the core semantic layer has a moderate amount of information, it will be guaranteed priority code rate due to its high importance weight.

[0070] like Figure 2 As shown, in an exemplary embodiment, constructing a spatial information observation adapted to the current channel characteristics includes:

[0071] Obtain the feature dimensions of the semantically encoded signal;

[0072] A vector sequence consistent with the semantically encoded signal in the feature dimension is constructed, and the values ​​of each element of the vector sequence are set as the expected values ​​of the received signal under the condition that there is no effective signal transmission in the current channel, which are used as spatial information observations.

[0073] In this embodiment, to accurately estimate the implicit prior of the decoder, an input signal with no information at the physical layer needs to be constructed. The feature dimension typically refers to the dimension d of the semantic coding vector. The system reads the dimension parameters of the received semantic coding signal, such as the sequence length T and the feature dimension d. A vector sequence with shape T multiplied by d is constructed. To ensure that this vector sequence does not carry any semantic information from the transmitter, its value is set to the statistical expectation of the channel. Specifically, assuming the transmitter does not send any signal (i.e., the transmitted vector is zero), and the receiver only receives channel noise or interference, the mathematical expectation of the received signal is the statistical expectation value. Any non-zero expectation value may be misread by the decoder as a valid semantic feature; only the statistical expectation value can represent an empty state.

[0074] In a preferred implementation, when the current channel is an additive white Gaussian noise channel, a Rayleigh fading channel, or an orthogonal frequency division multiplexing channel, the statistical expectation value of the received signal under the condition of no effective signal transmission is a zero vector; the spatial information observation is a sequence of all zero vectors.

[0075] Specifically, in additive white Gaussian noise (AWGN) channels, the noise mean is zero; in Rayleigh fading channels, the fading coefficients are independent of the transmitted signal and have random phases, and their statistical expectation is also zero; in orthogonal frequency division multiplexing (OFDM) systems, the statistical characteristics of the subcarrier signals in the frequency domain also satisfy the zero-mean assumption. Therefore, for these channels, the spatial information observation is specifically implemented as an all-zero vector sequence. For example, if the feature dimension d is 512, then the spatial observation vector corresponding to each word position is a vector containing 512 zeros. The construction method of all-zero vectors has extremely low computational cost and can be efficiently generated through matrix operations, making it the preferred solution for engineering implementation.

[0076] Furthermore, the derivation of the above conclusion is illustrated using a Rayleigh fading channel as an example. In a Rayleigh fading channel, the received signal can be expressed as:

[0077] y t =h t *x t +n t ;

[0078] Where y t h is the received signal vector corresponding to the t-th word position. t Let x be the complex Gaussian fading coefficient experienced by the t-th word position, following a Rayleigh distribution with a mean of zero. t Let n be the encoded signal vector from the transmitting end. t Let x be a vector of zero-mean additive white Gaussian noise, and * denote element-wise multiplication of a scalar with a vector. The objective of constructing a spatial information observation is to ensure that it does not carry information about the transmitted content x. t Any information is taken as null information observation, which is the conditional expectation of the received signal under the assumption that the transmitted content follows a uniform distribution. For systems using symmetrical modulation schemes such as quadrature amplitude modulation or quadrature phase shift keying, the constellation points of the transmitted symbols are symmetrically distributed around the origin, satisfying the statistical expectation E[x] of the transmitted signal. t The zero vector is equal to the Rayleigh fading coefficient h. t Statistical expectation E[h t The value is also zero. Since the transmitted signal and the fading coefficient are statistically independent, therefore:

[0079] E[y t ]=E[h t ]*E[x t ]+E[n t ] = 0 * 0 + 0 = 0;

[0080] Where E[ ] represents the statistical expectation operation. The above derivation shows that, under Rayleigh fading channels, the spatial information observations are unified as all-zero vectors. A similar derivation process can be applied to additive white Gaussian noise channels (where there is no fading term, and it is directly derived from E[x]). tThe conclusion is drawn from the fact that ]=0, and from the subcarriers in the orthogonal frequency division multiplexing system (each subcarrier independently yields a zero vector when its expectation is taken). Therefore, for all mainstream channel models using symmetric modulation schemes and with zero-mean noise, the spatial information observation is unified as a sequence of all-zero vectors with the same dimension as the semantically coded signal, independent of specific channel types and equalization strategies. A sufficient condition for the above conclusion is that the transmitted signal constellation diagram satisfies zero-mean symmetry, a condition that generally holds in modern digital modulation schemes.

[0081] In one possible embodiment, an offline construction step is included before inputting the null information observations into the non-autoregressive semantic decoder to estimate the implicit language prior:

[0082] For several commonly used sequence lengths, corresponding empty information observations are constructed and input into a non-autoregressive semantic decoder to pre-compute implicit language priors of the corresponding lengths.

[0083] Establish and store the mapping relationship between sequence length and implicit language priors;

[0084] Estimating implicit language priors specifically includes: detecting the actual sequence length of the semantically encoded signal; and directly retrieving the corresponding implicit language prior from the stored mapping relationship based on the actual sequence length.

[0085] This embodiment describes an offline optimization strategy to reduce latency in real-time inference. Since the content of spatial information observations, such as all-zero vectors, depends only on the sequence length and is independent of the specific received signal content, the decoder's response to spatial observations, i.e., the implicit language prior, is deterministic. In the offline phase, the system can pre-run the decoder and record the output probability distribution for commonly used lengths preset in the communication protocol, such as 64, 128, 256, and 512. The probability distribution is stored in a lookup table (LUT). In the online inference phase, the system only needs to detect the length of the currently received signal to directly look up the corresponding implicit language prior P in the table. null This eliminates the need to perform time-consuming decoder forward inference again, thus reducing computational overhead.

[0086] like Figure 3 As shown, according to one aspect of this application, bias correction of the initial decoding distribution based on implicit language priors includes:

[0087] Obtain the instantaneous signal-to-noise ratio of the semantically encoded signal at each word position;

[0088] The adaptive debiasing intensity corresponding to each word position is generated based on the instantaneous signal-to-noise ratio, wherein the adaptive debiasing intensity is negatively correlated with the instantaneous signal-to-noise ratio;

[0089] By utilizing adaptive debiasing intensity, the correction magnitude of implicit language priors on the initial decoding distribution is controlled.

[0090] Specifically, a channel-aware adaptive mechanism is introduced. In semantic communication, the decoder's reliance on implicit priors is closely related to channel quality: when channel quality is extremely poor (low signal-to-noise ratio), the decoder receives less effective information and tends to generate text by guessing. In this case, the interference from implicit priors is greatest, requiring strong debiasing. Conversely, when channel quality is good (high signal-to-noise ratio), the decoder mainly relies on channel evidence, implicit prior interference is small, and the debiasing intensity should be weakened. Therefore, the system estimates the instantaneous signal-to-noise ratio at each lexical position, which reflects the local signal reliability. Adaptive debiasing intensity is generated based on the negative correlation principle: the lower the signal-to-noise ratio, the greater the debiasing intensity; the higher the signal-to-noise ratio, the smaller the debiasing intensity. This dynamic adjustment ensures that while suppressing illusions, beneficial linguistic knowledge is not incorrectly subtracted under high-quality channel conditions.

[0091] In a preferred embodiment, the adaptive debiasing intensity corresponding to each word position is generated based on the instantaneous signal-to-noise ratio. Specifically, a mapping relationship is established using the Sigmoid function, and the calculation formula is as follows:

[0092] η t =η max / (1+exp(β*(γ t -γ η )));

[0093] Where η t To adaptive depolarization intensity, γ t For instantaneous signal-to-noise ratio, η max β is the pre-configured upper limit of depolarization intensity, β is the pre-configured temperature parameter, and γ is the pre-configured temperature parameter. η This is a pre-configured reference threshold.

[0094] Specifically, a variant of the Sigmoid function is used to achieve a smooth negative correlation mapping. Where η max The maximum correction force is controlled, typically ranging from 0 to 1, for example, 0.8; γ η The sensitivity point of the depolarization operation is defined, that is, when the signal-to-noise ratio is lower than the threshold, the depolarization intensity is increased; β controls the steepness of the transition, and the larger the β value, the more drastic the intensity change.

[0095] For example: Let η be... max =1.0, β=0.5, γ η =0dB. Case A (deep fading): If the instantaneous signal-to-noise ratio γ at a certain location is 0dB. t =-10dB. Therefore, exp(0.5*(-10-0))=exp(-5)≈0.0067. η t≈1.0 / (1+0.0067)≈0.99. At this point, the debiasing intensity is close to 1, meaning that the implicit prior is almost completely subtracted. Case B (Strong Signal): If the instantaneous signal-to-noise ratio γ at a certain location... t =10dB. Therefore, exp(0.5*(10-0))=exp(5)≈148.4. η t ≈1.0 / (1+148.4)≈0.0067. At this point, the depolarization intensity is close to 0, meaning that almost no depolarization is performed.

[0096] like Figure 4 As shown, in one embodiment of this application, obtaining the corrected channel likelihood distribution includes:

[0097] The initial decoding distribution and the implicit language prior are converted into a logarithmic field representation to obtain the logarithmic values ​​of the initial decoding distribution and the implicit language prior.

[0098] For each lexical candidate, the corrected log score is obtained by subtracting the weighted log value of the implicit language prior from the log value of the initial decoding distribution; where the weighting is the adaptive debiasing intensity.

[0099] Normalize the logarithm score of the correction to obtain the likelihood distribution of the corrected channel.

[0100] In this embodiment, to avoid numerical instability caused by probability domain division, the operation is performed in the logarithmic domain. The specific calculation formula is as follows:

[0101] log P ch_star (v)=log P dec (v)-η t *log P null (v);

[0102] Where P ch_star (v) represents the corrected channel likelihood of candidate term v, P dec (v) represents the probability of v in the initial decoding distribution, P null (v) represents the probability of v in the implicit language prior. To prevent numerical errors caused by log(0), a very small smoothing term, such as 1e, is usually added to the probability value. -9 .

[0103] In a numerical computation example: assume the vocabulary contains only [word A, word B]. Initial decoding P dec =[0.9, 0.1] (The model favors word A). Implicit prior P null =[0.8, 0.2] (the model itself favors word A). If the calculated debiasing intensity η t =1.0. Then log P ch_star(A)=log(0.9)-1.0*log(0.8)≈-0.105-(-0.223)=0.118. log P ch_star (B) = log(0.1) - 1.0 * log(0.2) ≈ -1.0 - (-0.699) = -0.301. After performing Softmax normalization, P ch_star (A)≈exp(0.118) / (exp(0.118)+exp(-0.301))≈1.125 / (1.125+0.740)≈0.60. It can be seen that the probability of word A after bias removal decreased from 0.9 to 0.60, indicating that the initial high probability largely stemmed from the inherent bias of the model rather than channel evidence. The corrected distribution more accurately reflects the channel information.

[0104] Furthermore, the above logarithmic field subtraction operation is equivalent to division in the probability field. Specifically, the probability field relationship between the corrected channel likelihood, the initial decoding distribution, and the implicit language prior can be expressed as:

[0105] P ch_star (v|y t Proportional to P dec (v|y t ) / P null (v|t)^η t ;

[0106] Where P ch_star (v|y t P represents the corrected channel likelihood probability. dec (v|y t P represents the original output probability of the decoder, i.e., the approximate posterior probability. null (v|t) represents the implicit language prior estimate, η t For adaptive debiasing intensity, ^ denotes exponentiation. The decoder output (approximate posterior) is divided by the implicit prior η. t The power of η is the quotient, which is the corrected channel likelihood. t When P equals 1, implicit priors are completely removed, and the corrected P ch_star Approximates the true channel likelihood P(y) t |v);when η t When the value equals 0, no correction is performed, and the system reverts to the original scheme. After this debiasing correction, the channel likelihood P is corrected. ch_star Substituting this into the subsequent Bayesian fusion, the fusion result becomes:

[0107] P post (v) is proportional to P ch_star (v) αt *P sem (v) 1-αt ≈P(yt |v) αt *P sem (v) 1-αt ;

[0108] Where P post (v) represents the posterior probability after fusion, α t For channel-semantic fusion weights, P sem (v) represents semantic prior. It can be seen that language prior is only obtained through P. sem The one-time introduction eliminates the problem of prior double inclusion, making the fusion weight α t The correct channel-semantic balance meaning is ensured.

[0109] It should be noted that the above adaptive debiasing intensity η t With channel-semantic fusion weight α t Although both are based on instantaneous signal-to-noise ratio generation, they represent two independent degrees of freedom addressing different levels of problems and cannot be substituted for one another. η t The problem addressed is the purity of channel likelihood extraction, specifically how to extract implicit linguistic priors from the decoder's mixed output, allowing the channel likelihood to return to its physical meaning. α t This problem addresses the optimal balance between channel likelihood and semantic prior, specifically, how to dynamically adjust the relative contributions of the two information sources based on channel quality, given a clean channel likelihood. The two operations have a strict sequential dependency: the bias removal operation precedes it, producing a clean, corrected channel likelihood distribution; the fusion operation follows, combining the corrected channel likelihood and semantic prior with adaptive weights. If the bias removal step is omitted and only α is relied upon... t If fusion regulation is performed, then regardless of α t Regardless of the value chosen, implicit prior biases always remain in the channel likelihood term, and the theoretical self-consistency of the fusion result cannot be guaranteed.

[0110] For example, regarding the debiasing operation, assume the vocabulary contains only three words: w1, w2, and w3. Obtain the received raw logits, i.e., the logarithmic values ​​of the decoder output before debiasing. Assume the raw output logarithmic values ​​obtained through forward inference from the non-autoregressive semantic decoder are (2.0, 0.5, 0.1). In this case, w1 has the highest probability. Perform implicit language prior estimation. Feed the empty information observation signal into the same decoder to obtain the implicit language prior logarithmic value estimate as (0.5, -0.2, 0.0). Perform log-domain debiasing. The calculation formula is:

[0111] log P ch *(v|y t )=log P dec (v|y t )-η t *log Pnull (v|t).

[0112] Where P ch *(v|y t ) represents the corrected channel likelihood probability. When the debiasing intensity η t When the value is 1, the specific calculation results are as follows: for w1, 2.0 - 0.5 = 1.5; for w2, 0.5 - (-0.2) = 0.7; for w3, 0.1 - 0.0 = 0.1. After correction, w1 is still the candidate with the highest probability, but the relative confidence of w2 is significantly improved. In the probability domain, the above logarithmic subtraction operation is equivalent to dividing the original output probability of the decoder by η of the implicit language prior probability. t The power, i.e., P ch *(v|y t Proportional to P dec (v|y t ) / P null (v|t)^η t This method removes the high-frequency word probability bias that is systematically inflated due to the linguistic statistical patterns learned during the training phase from the decoder output, providing a cleaner basis for channel likelihood estimation for subsequent Bayesian fusion.

[0113] This embodiment constructs an empty input without semantic content, allowing the decoder to expose its internally fixed language bias. Combined with the instantaneous channel quality, the bias reduction intensity is dynamically adjusted to obtain a corrected channel likelihood distribution that reflects the true physical transmission quality.

[0114] In a further embodiment, after generating the corrected channel likelihood distribution and before quantizing the channel uncertainty and semantic uncertainty, the method further includes:

[0115] Calculate the joint preliminary screening score for each lexical candidate. The joint preliminary screening score is the weighted sum of the corrected channel likelihood distribution and the semantic prior distribution in the logarithmic domain.

[0116] A predetermined number of candidate words with the highest joint initial screening scores are selected to form a Top-K candidate set;

[0117] The channel uncertainty and semantic uncertainty of each word position are quantified, specifically on the Top-K candidate set.

[0118] This embodiment describes a Top-K candidate space reduction strategy. Since a complete vocabulary typically contains tens of thousands of lexical units, directly calculating the posterior probability and entropy across the entire space is extremely time-consuming. Therefore, the system performs a rapid initial screening. The formula for calculating the joint initial screening score is:

[0119] s t (v)=log P ch_star (v)+log P sem(v);

[0120] Where s t (v) is the initial screening score for candidate word v, P ch_star (v) is used to correct the channel likelihood, P sem (v) represents semantic prior. The joint initial screening score considers both channel evidence and semantic plausibility. Based on the score ranking, the top K terms (e.g., K=50) are selected to form the Top-K candidate set. Subsequent Bayesian fusion and entropy calculation are performed only on the reduced set, improving processing speed.

[0121] Furthermore, after constructing the Top-K candidate set, truncation and re-normalization operations need to be performed on the corrected channel likelihood distribution and the semantic prior distribution. Specifically, for the t-th lexical position, a sub-distribution covering only the Top-K candidate set is extracted from the complete vocabulary distribution of the corrected channel likelihood, and the extracted K probability values ​​are re-normalized so that their sum is 1, resulting in the truncated corrected channel likelihood distribution. The same truncation and normalization operations are performed on the semantic prior distribution to obtain the truncated semantic prior distribution. Subsequent logarithmic domain fusion and the calculation of channel entropy and semantic entropy are all performed on the above-mentioned truncated and normalized distribution. This ensures that the probability distribution still satisfies the valid normalization conditions in the reduced candidate space, guaranteeing the mathematical correctness of the entropy calculation and posterior fusion results.

[0122] According to one aspect of this application, the channel uncertainty and semantic uncertainty at each lexical position are quantified, including:

[0123] For each lexical position, calculate the information entropy of the corrected channel likelihood distribution, and use it as the channel entropy;

[0124] Calculate the information entropy of the semantic prior distribution, and use it as the semantic entropy.

[0125] Specifically, information entropy is a standard indicator for measuring the degree of disorder in a distribution. Channel entropy H ch The formula for calculating (t) is:

[0126] H ch (t)=-Σ(P ch_star (v)*log P ch_star (v));

[0127] The higher the channel entropy, the flatter the channel output, and the weaker the channel's ability to distinguish that location.

[0128] Semantic entropy H sem The formula for calculating (t) is:

[0129] H sem (t)=-Σ(P sem (v)*log P sem (v));

[0130] The higher the semantic entropy, the more likely the language model believes that there are multiple reasonable words to fill in at that position, and the stronger the ambiguity of the context itself.

[0131] It should be noted that the channel entropy H in this embodiment... ch (t) must be based on the corrected channel likelihood P ch_star The calculation must be performed, and it cannot be based on the decoder's raw output P. dec Calculation. Because P without bias correction. dec It contains implicit language priors P imp The information, the probability of high-frequency common words is P imp Systematic elevation artificially increases the concentration of the distribution and artificially decreases the information entropy. In this case, H ch (t) does not reflect pure channel uncertainty, but rather a mixture of channel uncertainty and linguistic a priori certainty, which leads to an uncertainty source ratio ξ t The diagnosis failed. Only after bias correction was P... ch_star The implicit language prior components have been removed, and its entropy value now purely reflects the reliability of information transmission at the physical layer of the channel, making it possible to base ξ on... t The diagnostic results for uncertainty sources possess physical interpretability and practical validity. This constitutes a longitudinal causal dependency: debiasing correction provides a reliable input data foundation for entropy ratio analysis, while the source-aware refinement strategy transforms the channel likelihood advantage after debiasing into improved computational efficiency in the iterative refinement process. In each round of the iterative refinement process, both a more accurate fusion confidence score and a more reasonable refinement target selection are obtained, simultaneously improving convergence speed and final decoding quality, with the overall gain exceeding the sum of individual improvements.

[0132] In a further embodiment, the ratio of channel uncertainty to total uncertainty sources is specifically calculated using the following formula:

[0133] ξ t =H ch (t) / (H ch (t)+H sem (t)+ε ξ );

[0134] Where ξ t H represents the uncertainty source ratio at the t-th word position. ch (t) is the channel entropy, H sem (t) represents semantic entropy, ε ξ These are pre-configured positive numbers, typically very small positive numbers.

[0135] In this embodiment, the uncertainty source ratio ξ tThis intuitively reflects the proportion of channel factors in the total uncertainty. When ξ t When it approaches 1, it means that H ch >>H sem This refers to a situation where the channel is chaotic, but the semantics are well-defined. This corresponds to transmission errors, such as burst noise in the channel disrupting words in a fixed collocation. Such positions are ideal for iterative refinement because the language model has a definite solution to correct channel errors. When ξ t When it approaches 0, it means that H sem >>H ch This means the information source is well-defined, but the semantics are chaotic. This situation corresponds to open-ended questions in names, place names, or literary works, where whatever the information source indicates is what it is, and the language model cannot provide effective correction. Resources should not be wasted on refining such positions.

[0136] In a further embodiment, the semantically encoded signal carries semantic hierarchy identifier information to generate a refinement priority for each lexical position, including:

[0137] Based on the semantic hierarchy identifier information, determine the semantic importance weight of each word position;

[0138] The joint confidence of each lexical position is obtained by fusing the corrected channel likelihood distribution and the semantic prior distribution.

[0139] The source-aware modulation factor is generated based on the uncertainty source ratio, and the refinement priority is calculated using the following formula: r t =(1-c t )*ω(l t )*ψ(ξ t );

[0140] Where r t To refine priority, c t For joint confidence, ω(l) t ) represents the semantic importance weight, ψ(ξ) t ) is the source-sensing modulation factor, which is positively correlated with the uncertainty source ratio.

[0141] Specifically, the final priority r t The generative model integrates information from three dimensions: (1-c t ): The underlying uncertainty in decoding; the lower the confidence level, the higher the priority; ω(l) t ): Semantic importance; the more central the semantic level to which this position belongs, such as the main theme level, the higher the priority; ψ(ξ) t Source-aware modulation, where uncertainty primarily originates from the channel (ξ). t If the value is large, then the priority is increased; if it originates from semantics (ξ), then the priority is increased. t If the priority is low (e.g., small), then the priority is lowered.

[0142] In some alternative implementations, the source-sensing modulation factor is calculated as a linear function:

[0143] ψ(ξ t ) = ψ0 + (1 - ψ0) * ​​ξ t ;

[0144] Where ψ0 is a pre-configured lower bound of the basic modulation, for example, 0.2, and its value ranges from 0 to 1, ensuring that even if there is semantic ambiguity, there is still a slight chance of refinement at positions with extremely low confidence.

[0145] In a possible numerical example: Scenario A: The channel is severely impaired. t =0.3 (low), ω=1.0 (important), ξ t =0.9 (channel dominant). Therefore, ψ≈0.2+0.8*0.9=0.92. t =0.7 * 1.0 * 0.92 = 0.644. This has extremely high priority and will be included in the refinement process. Scenario B: Inherent semantic ambiguity. t =0.3 (low), ω=1.0 (important), ξ t =0.1 (semantic dominance). Therefore, ψ≈0.2+0.8*0.1=0.28. t =0.7 * 1.0 * 0.28 = 0.196. The priority is lowered to avoid ineffective refinement.

[0146] In a preferred implementation, the corrected channel likelihood distribution and the semantic prior distribution are fused, including:

[0147] Based on the instantaneous signal-to-noise ratio of the semantically encoded signal, channel-semantic adaptive fusion weights are generated;

[0148] The weighted posterior score for each word position in the logarithmic field is calculated using the following formula:

[0149] l t *(v)=α t *log P ch_star (v)+(1-α t )*log P sem (v);

[0150] Among them l t * (v) represents the weighted posterior score, α t For channel-semantic adaptive fusion weights, P ch_star To correct the channel likelihood distribution, P sem For semantic prior distribution;

[0151] The weighted posterior scores are normalized, and the scores corresponding to the peak probabilities are determined as the joint confidence levels.

[0152] This embodiment details the Bayesian posterior fusion process. Channel-semantic adaptive fusion weight α t It reflects the system's level of trust in channel information. Similar to debiasing strength, α t Also based on instantaneous signal-to-noise ratio γ t The calculations show a positive correlation: the higher the signal-to-noise ratio, the higher α. t The closer to 1 (trust channel); the lower the signal-to-noise ratio, α t The closer to 0 (trust semantics), the better. The specific weight calculation formula can be: α t =1 / (1+exp(-β fuse *(γ t -γ0))). The fusion formula uses a logarithmic domain weighted sum, which is equivalent to the geometric mean in the probability domain. This effectively avoids the numerical underflow problem caused by the low probability of a single source. The peak value of the final normalized probability distribution is the joint confidence level c. t This quantifies the overall reliability of the current decoding results.

[0153] It should be noted that the fusion weight α t The specific generation method adopts a Sigmoid forward mapping based on instantaneous signal-to-noise ratio:

[0154] α t =1 / (1+exp(-β fuse *(γ t -γ0)));

[0155] Where α t γ is the channel-semantic fusion weight at the t-th lexical position; t γ0 is the instantaneous signal-to-noise ratio estimate at this location; γ0 is the pre-configured signal-to-noise ratio reference threshold, i.e., the balance point between channel and semantic weighted contributions; β0 fuse γ is a positive temperature parameter used to control the steepness of the weight transition; exp represents the natural exponential function. t When α is significantly higher than γ0, t When γ approaches 1, posterior inference is mainly dominated by channel likelihood; when γ... t When α is much lower than γ0, t Approaching 0, posterior inference becomes semantic prior-driven. Fusion weight α t With depolarization intensity η t The signal-to-noise ratio dependence is in the opposite direction: η t Negatively correlated with signal-to-noise ratio (strong debiasing at low signal-to-noise ratio), α t It is positively correlated with the signal-to-noise ratio (high signal-to-noise ratio indicates reliable channel), and the two work together to achieve the optimal decoding strategy under different channel conditions.

[0156] According to one aspect of this application, before calculating the joint initial screening score for each lexical candidate, the method further includes:

[0157] Obtain the semantic prior distribution of the pre-trained language model based on the initial decoded text context.

[0158] In this embodiment, the system feeds the initial decoded text sequence position by position into the pre-trained language model in an autoregressive manner. For the t-th word position, the pre-trained language model uses the sequence of all decoded words preceding that position x. ＜t As contextual input, the output is the predicted probability distribution for the t-th position, denoted as the semantic prior distribution P for that position. sem (v|x ＜t ), where v traverses all candidate lexical units in the vocabulary. The pre-trained language model can employ autoregressive language models such as the GPT series or LLaMA series. The parameters obtained through pre-training on massive text corpora remain frozen during the inference phase and do not participate in end-to-end training. Semantic prior distributions encode the lexical, syntactic, and semantic rules of natural language, providing semantic constraints independent of channel observations for subsequent fusion.

[0159] It should be noted that the semantic prior distribution and the corrected channel likelihood distribution differ fundamentally in their information sources: the corrected channel likelihood uses the channel received signal y. t As a condition, it reflects the probability of each candidate lexical unit at the physical transmission layer; the semantic prior is based on the decoded text context x. ＜t This serves as a condition, reflecting the rationality of each candidate lexical unit at the linguistic level. Both provide complementary perspectives on the same lexical unit position, forming the theoretical premise for the subsequent Bayesian fusion framework.

[0160] In one embodiment, suppose the context of the sentence being decoded by the receiver is "The dog chased the...". At position 3, the candidate word pointed to by the signal received after channel transmission is ambiguous between "Cat" and "Hat". The corrected channel likelihood distribution P at this position is... ch *Information entropy H ch The high (t) indicates that the channel's physical layer cannot provide reliable distinguishing criteria. Simultaneously, the pre-trained language model, based on the context "The dog chased the...", predicts with extremely high probability that the position should be "Cat", leading to a semantic prior distribution P... sem Information entropy H sem (t) is extremely low. Based on this, the uncertainty source ratio ξ is calculated. t =H ch (t) / (H ch (t)+H sem (t)+ε ξ ), due to Hch (t) is much larger than H sem (t), ξ t When the value approaches 1, the system classifies this position as a channel-degraded uncertainty. Source-aware modulation factor ψ(ξ) t If the value is close to 1, no suppression is applied to the refinement priority at that position. The refinement priority score r at that position... t =(1-c t )* ω(l t )* ψ(ξ t If the value remains high, it will be given priority for inclusion in the set of locations to be refined.

[0161] Conversely, if the channel observation quality at a certain location is good (H ch (t) low), but this position naturally has ambiguity in the linguistic context. For example, "bank" can refer to a bank or a riverbank, leading to H sem If (t) is higher, then ξ t When the value approaches 0, the system classifies it as semantically ambiguous uncertainty. At this point, ψ(ξ) t The refinement priority is effectively suppressed at the lower bound ψ0 of the basic modulation, avoiding the waste of limited iterative refinement computing power in the region where the channel observation itself is the most reliable source of evidence.

[0162] This embodiment solves the problem of blindly refining all low-confidence locations in the prior art. By introducing the entropy ratio index, it accurately identifies those locations that are severely damaged by the channel and can be repaired through semantics.

[0163] In one embodiment of this application, the text sequence obtained from the initial decoding and pre-stored data is iteratively refined according to a refinement priority, including:

[0164] The positions of lexical elements with a refinement priority higher than the pre-configured refinement threshold are assigned to the set of positions to be refined, and the remaining positions are assigned to the set of frozen positions.

[0165] In this embodiment, the system reads the refinement priority r of each lexical position. t Pre-configured refinement threshold θ r This is the criterion used to determine whether refinement is needed. It can be set as an absolute value, such as 0.6, or as a relative proportion, such as taking the values ​​corresponding to the top 30% of positions in the priority sort. Based on the refinement threshold, the index of the entire text sequence is divided into two mutually exclusive subsets: those satisfying r t ≥θ r The index is assigned to the set R of positions to be refined, meaning that the decoding results of these positions are unreliable and have high refinement value; satisfying r t <θ rThe index is included in the frozen position set F, which means that the decoding results of these positions are already reliable enough or the semantics themselves are ambiguous and do not need to be refined.

[0166] As a preferred implementation, the refinement threshold θ r It is not fixed, but rather related to the current remaining latency budget. When the system detects that there is ample remaining available time, the threshold can be appropriately lowered to expand the size of the set R to be refined in order to pursue higher quality; when the remaining time is tight, the threshold is raised to retain only the most critical few positions for refinement, thereby achieving a dynamic balance between performance and latency.

[0167] Construct an asymmetric attention mask matrix. The asymmetric attention mask matrix opens the attention connections of the set of positions to be refined to all word positions, and closes the attention update channels of the frozen set of positions.

[0168] In this embodiment, a specific attention mask matrix M needs to be constructed to achieve local refinement in the Transformer decoder. The asymmetry is reflected in the inequality between the positions to be refined and the frozen positions in terms of information flow. Specifically, for each position i in set R, its corresponding mask row M[i,:] is set to all 1s or an attention-allowed state, meaning that when position i is regenerated, it can see the contextual information of all positions in the sequence (including R and F) and use the global context to correct its own errors. Conversely, for each position j in set F, its corresponding update channel is physically closed. For example, in the self-attention calculation, the query operation on j is skipped, and the output of the previous round is directly copied, or its mask row is set to only focus on itself, so that its output remains unchanged in the current iteration. This makes the frozen positions act as stable anchor points, providing reliable contextual support for the positions to be refined, while avoiding repeated calculations of already converged regions.

[0169] Under the constraint of the asymmetric attention mask matrix, a non-autoregressive semantic decoder is used to perform constraint bundle search on the lexical units in the set of positions to be refined, generating the refined text for the current round.

[0170] Specifically, the decoder performs forward inference, but generates new probability distributions only for positions in set R. The system employs a constrained bundle search strategy, maintaining a candidate sequence bundle of width B. During search tree expansion, the tokens in set F are forced to remain fixed to the decoding results of the previous round, and the search algorithm explores different candidate words in the vocabulary only for positions in set R.

[0171] Building upon this foundation, a latency-aware adaptive simplification strategy can be introduced to further control latency. As the number of iterations increases, or when the remaining computation time is detected to be below a specific warning threshold, the system automatically reduces the beam search width B and the Top-K candidate set size K. For example, K=50 and B=5 are used in the first iteration; in subsequent iterations or when a timeout warning is issued, K=10 and B=1 are adjusted. This allows the system to complete the core refinement task even under extreme latency constraints.

[0172] During the constraint-bound search process, the scoring function for candidate sequences comprehensively considers two aspects. The first is the conditional probability based on channel evidence, i.e., the probability value of each candidate word output by the semantic decoder at the position to be refined, using the preprocessed received signal as input. This probability value reflects the consistency between the candidate sequence and the channel observation. The second aspect is the sequence-level semantic coherence constraint, i.e., using a pre-trained language model to evaluate the linguistic fluency and contextual consistency of the candidate sequence. The final score of the candidate sequence is a weighted sum of the above two aspects in the logarithmic domain, where the weighting coefficients can reuse the fusion weight α. t Alternatively, a fixed empirical value can be used. In each step of the search expansion, the system retains the B candidate sequences with the highest scores as the bundle width and discards the remaining low-scoring candidates, thereby achieving a balance between search quality and computational cost.

[0173] In another embodiment of this application, the preset convergence condition includes that the semantic gain is lower than the adaptive convergence threshold, and the calculation steps of the semantic gain include:

[0174] Extract sentence-level embedding vectors from the current round of refined text and the previous round of decoded text, respectively.

[0175] In this embodiment, a termination condition for the iteration is defined. To determine whether the refinement has reached a saturation point, the system needs to quantify the semantic changes in the decoding results between adjacent rounds. The refined text of the current round is denoted as x. n The text decoded in the previous round is denoted as x. prev The system utilizes a lightweight pre-trained semantic encoder to map the two text sequences into sentence-level embedding vectors e in a high-dimensional space. n and e prev It condenses the global semantic information of the corresponding text.

[0176] Calculate the normalized semantic distance between two sentence-level embedding vectors as the semantic gain.

[0177] Specifically, semantic changes are quantified by calculating distances in the vector space. To eliminate the influence of vector magnitude, cosine distance is preferred. Semantic gain Δ sem The calculation formula is as follows:

[0178] Δ sem=1-(e n *e prev ) / (||e n ||*||e prev ||);

[0179] Where Δ sem For semantic gain, e n e is the sentence-level embedding vector for the current round. prev Let be the sentence-level embedding vector from the previous round, where * denotes the vector dot product, and || denotes the Euclidean norm (modulus) of the vector. When the semantics of the two rounds of text are highly consistent, the cosine similarity is close to 1, and the semantic gain is close to 0, indicating that the iteration has reached a stable state. When the semantics of the two rounds of text are significantly different, the semantic gain is significantly greater than 0, indicating that the refinement operation has brought about substantial semantic adjustments.

[0180] When the semantic gain is lower than the adaptive convergence threshold that decays with each iteration, the convergence condition is considered met.

[0181] In this embodiment, the system compares the calculated semantic gain with a dynamic threshold. The adaptive convergence threshold ε' n It is not a fixed value, but rather exhibits an exponential decay trend as the number of iterations n increases. Its calculation formula can be:

[0182] ε' n =ε'0*ρ n-1 ;

[0183] Where ε' n ε'0 is the convergence threshold for the nth round; ε'0 is the initial threshold, for example, 0.1; ρ is the decay factor, for example, 0.8, ranging from 0 to 1. In the early stages of iteration, the text may contain significant errors, and the system tolerates large semantic adjustments (high threshold), allowing for substantial corrections. As iterations progress and the text becomes more refined, the system requires stricter convergence criteria (low threshold). Any minor semantic fluctuations may be considered non-convergent and require further refinement, or conversely, if even minor fluctuations are not observed, the process terminates rapidly. Once Δ... sem <ε' n If the number of iterations reaches the preset limit, the system will determine that the iteration has converged, terminate the iteration, and output the result.

[0184] In one embodiment, suppose the original text sent by the sender is "Hello World". After transmission through the channel, the initial text decoded by the receiver is "Hella World", indicating two transmission errors.

[0185] In the first iteration, the system performs refinement priority calculation. It calculates the refinement priority score *r* for each lexical position in the sequence. tThe system found that "Hella" is a low-frequency word in English, and its corrected channel likelihood distribution P ch The information entropy of * is high, while the semantic prior distribution information entropy of the pre-trained language model at this position is low (the language model believes that this position should be a common greeting). Therefore, the uncertainty source ratio ξ at this position is high. t The score is too high, indicating a channel degradation problem; therefore, the priority score r is refined. t The highest priority is given to Warld, but with slightly lower priority than Hella. The system is based on the refinement threshold θ. r The highest-priority Hella position is added to the refinement set R, and this position is marked as open in the attention mask. The system performs constraint-based search refinement decoding at this position, and the pre-trained language model predicts the most probable word at this position as Hello based on the context and training memory. The updated sequence is Hello Ward.

[0186] Proceeding to the second iteration. The current sequence is Hello Ward. The system recalculates the semantic prior distribution and joint confidence of each position based on the updated context. The joint confidence c of the Hello position. t It is close to 1, refinement priority score r t Approaching zero, it is assigned to the frozen position set F. The refinement priority score of the Warld position becomes the highest in this round. The system detects that the semantic prior probability of Warld paired with Hello is extremely low, while the prior probability of World is extremely high. Therefore, refinement decoding is performed at this position, correcting it to World. The updated sequence is Hello World.

[0187] At this point, semantic gain-driven dynamic convergence determination is performed. The semantic gain Δ of the decoded text in the current round and the previous round is calculated. sem n =1-(e n ·e n-1 ) / (||e n ||*||e n-1 ||). The sentence-level embedding vectors for "Hello World" and "Hello World" are extracted by the pre-trained semantic encoder, and Δ is calculated. sem 2 The value is less than the adaptive convergence threshold ε'2=ε'0*ρ 2-1 The system determines that the refinement process has semantically converged and terminates the iteration. In this example, ε'0 is set to 1e. -2 ρ is set to 0.5. The final output text is "Hello World", which is exactly the same as the original text sent.

[0188] In another embodiment of this application, the refinement priority score can also be calculated additively. The priority score for each lexical position is... i The calculation is as follows:

[0189] Score i =(1-C joint )+α p *(H channel -H source )+β p *I level ;

[0190] Among them, Score i C is the priority score for the i-th word position; joint The joint confidence score for that position; (1-C joint H represents the degree of decoding uncertainty at that position. channel The information entropy of the channel likelihood distribution at this location is corrected; H source The information entropy of the semantic prior distribution at this location; (H) channel -H source The difference between channel entropy and semantic entropy is represented by a positive value, indicating that channel uncertainty is higher than semantic uncertainty (i.e., channel degradation), while a negative value indicates that semantic uncertainty is higher than channel uncertainty (i.e., semantic ambiguity). level α represents the importance weight of the semantic hierarchy to which this position belongs, determined by the semantic hierarchy identifier information. p β is the balance coefficient of the channel-semantic entropy difference term; p This is the balance coefficient for the semantic hierarchy importance term.

[0191] With the aforementioned multiplication form r t =(1-c t )* ω(l t )* ψ(ξ t In contrast, the three factors in the additive form contribute independently, each with different physical dimensions, and require α to be used. p and β p Perform dimensional normalization and importance adjustment. The advantage of the additive form is that each contribution can be interpreted and adjusted independently, making it suitable for evaluating the impact of each factor item by item in the early stages of system design. The advantage of the multiplicative form is that the three factors naturally constitute a joint constraint, and if any factor is zero, the priority is zero, eliminating the need for an additional truncation mechanism. In actual deployment, the appropriate formula form can be selected based on the computational constraints and debugging requirements of the application scenario.

[0192] This embodiment uses asymmetric attention masking to concentrate computing resources in the region with the highest uncertainty, and introduces a dynamic termination criterion based on semantic gain to solve the problems of computational waste and uncontrollable latency caused by traditional fixed-number iterations.

[0193] According to one aspect of this application, during the system construction phase, an end-to-end neural network architecture comprising a transmitter encoder and a receiver decoder needs to be built. The transmitter includes a Transformer-based semantic feature extractor and a cascaded residual VIB module; the receiver includes a non-autoregressive semantic decoder. To achieve end-to-end differentiable training, a differentiable channel simulation layer is inserted between the transmitter and receiver.

[0194] The function of the WeChat channel simulation layer is to simulate the signal distortion of the physical channel during training, while maintaining the backpropagation path of the gradient. For additive white Gaussian noise channels, the WeChat channel simulation layer achieves this through a reparameterization technique: y = x + n = x + σ noise *ε; where y is the received signal, x is the transmitted signal, and σ noise Let ε be the noise standard deviation determined by the current training signal-to-noise ratio, and let ε be a random variable sampled from a standard normal distribution. Since ε is independent of the network parameters, the gradient can be propagated back to x through the addition operation, thereby updating the encoder parameters.

[0195] The system's joint training objective function L total It consists of two parts: reconstruction loss L recon And information bottleneck regularization loss L VIB L total =L recon +λ vib *L VIB ; where L recon Cross-entropy loss is typically used to measure the difference between the decoded text and the original text; L VIB Constraining the mutual information between the encoded semantic representation Z and the input X prevents overfitting; λ vib This is the balance coefficient.

[0196] It should be noted that the implicit language prior originates from the parameter distribution learned by the decoder after training on a large amount of text corpus. During the training phase, it is not necessary to explicitly construct empty information observations for debiasing training, as debiasing is a correction strategy for the inference phase. However, to ensure the effectiveness of spatiotemporal information observations during inference, the decoder's positional encoding module must possess good generalization ability, capable of handling all-zero inputs without numerical collapse.

[0197] Information bottleneck regularization loss L in joint training loss VIB It consists of the sum of the losses of L encoding stages. The loss of each stage includes two terms: the rate term loss (KL divergence, degree of constraint compression) and the Lagrange multiplier λ of that stage. lA weighted relevance loss (the semantic label prediction loss of the auxiliary decoding network, ensuring semantic preservation) is applied. Since the input of stage l in the cascaded architecture depends on the output of stage (l-1), the parameters of the encoding network and the auxiliary decoding network in each stage are jointly optimized through backpropagation. During forward computation, encoding, sampling, and cumulative reconstruction of each stage are performed sequentially in the cascaded order from l to L; during backpropagation, gradients are propagated back from stage L to stage 1 along the cascaded structure. The auxiliary decoding network only participates in computation during the training phase and is not used during the inference phase.

[0198] Furthermore, the semantic importance weight ω(l) was determined through statistical analysis on the validation set after model training. Specifically, on the validation set, the task-related loss, such as reconstruction error or semantic classification error, was calculated when decoding using only the residuals of layer l, denoted as L. rel (l). The smaller the loss, the more crucial the semantic information contained in that layer. The final weight ω(l) is calculated by applying L... rel The negative value of (l) is obtained by Softmax normalization and is fixed in the system as a static parameter, which is called when generating the refinement priority.

[0199] Unlike autoregressive decoders, which depend on the output of the previous step at each step and thus must generate output sequentially position by position, the non-autoregressive decoder used in this application does not impose causal masking constraints at the self-attention layer. The output at each term position depends only on the global context of the received signal and not on the decoding results of other positions. Therefore, all positions can generate prediction distributions synchronously in a single forward inference. This architecture offers two advantages: first, the channel likelihood distributions at each position can be extracted in parallel, providing an efficient computational foundation for position-by-position debiasing correction and position-by-position entropy calculation; second, the output at any position is not affected by the cascading effects of decoding errors at other positions, making it possible to selectively freeze some positions and refine the remaining positions through attention masking. During the training phase, the non-autoregressive decoder uses the reconstruction cross-entropy loss L... recon Loss is calculated synchronously at all locations and backpropagated, without the need for autoregression.

[0200] According to one aspect of this application, an electronic device includes:

[0201] Memory, used to store computer programs;

[0202] In this embodiment, the memory specifically refers to the physical device inside the electronic device used to store data and instructions. It may include high-speed random access memory (RAM) for temporarily storing intermediate calculation results during processor operation, such as initial decoding distributions, null information observation vectors, and refined priority maps. Furthermore, the memory may also include non-volatile memory, such as flash memory, hard disk drives, or solid-state drives.

[0203] It should be noted that, to support offline optimization strategies, the memory pre-stores a mapping table between sequence length and implicit language priors. This mapping table stores pre-calculated implicit language prior distribution data for different commonly used sequence lengths, enabling the processor to quickly obtain the parameters needed for bias removal during real-time decoding by looking up the table, without having to repeatedly perform forward inference with empty observations. Simultaneously, the memory also stores parameter files of the pre-trained language model, such as weight matrices and bias vectors, for the processor to call to calculate semantic prior distributions and semantic entropy.

[0204] A processor is configured to execute the computer program to implement the space-based computing-based semantic communication decoding method as described in any of the above embodiments.

[0205] In this embodiment, the processor is the core of the electronic device's computation and control. Its specific implementation can be diverse to adapt to different application scenarios and performance requirements. In some general computing scenarios, the processor can be a central processing unit (CPU). In scenarios requiring high-concurrency processing, given that this application involves a large number of neural network matrix operations, a graphics processing unit (GPU), tensor processing unit (TPU), or a dedicated neural network processor is preferred. Furthermore, for embedded communication devices with extremely high latency requirements, the processor can also be implemented as a field-programmable gate array (FPGA) or an application-specific integrated circuit (ASIC).

[0206] When the processor executes the computer program in memory, it performs the following operations: It receives semantically encoded signals transmitted via a channel through the communication interface of the control device (such as a radio frequency transceiver or antenna array). Internally, it constructs a spatial information observation consistent with the dimension of the received signal, such as generating an all-zero vector, and uses a neural network inference engine to compute the initial decoding distribution and implicit language priors in parallel, or directly reads pre-stored implicit priors from memory. It uses an arithmetic logic unit to perform mathematical operations as described in the above embodiments, including calculating the instantaneous signal-to-noise ratio, generating adaptive debiasing intensity using the sigmoid function, performing subtraction debiasing in the logarithmic domain, calculating channel entropy and semantic entropy, calculating the uncertainty source ratio, and generating a refinement priority. Based on the generated priority, the processor controls the decoding algorithm to enter an iterative refinement loop, using a caching mechanism to maintain the state of the positions to be refined and the frozen positions until the semantic gain convergence condition is met, and finally presents the recovered text to the user or downstream task through an output interface (such as a display screen, audio synthesizer, or next-level application interface).

[0207] According to another aspect of this application, a computer-readable storage medium having a computer program stored thereon is provided.

[0208] In this embodiment, a computer-readable storage medium refers to any non-transient medium capable of storing computer program code and readable by a computer device. Its scope covers electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices. Specific preferred examples include, but are not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices (such as magnetic tape), or any suitable combination thereof. This medium can be distributed as a standalone software product (such as a firmware upgrade package or an app installation package) or integrated onto the hardware circuit board of an electronic device.

[0209] When the computer program is executed by the processor, it implements the space-based computing-based semantic communication decoding method described in any of the above embodiments.

[0210] In this embodiment, the computer program includes a series of machine-executable instructions. These instructions are organized using a modular programming approach and may specifically include:

[0211] Signal interface module: used to parse semantically encoded signals and channel state information transmitted at the physical layer;

[0212] The debiasing module is used to perform spatial information observation construction, implicit prior table lookup or estimation, adaptive intensity calculation, and logarithmic domain debiasing operations.

[0213] Source awareness analysis module: used to perform Top-K candidate screening, dual-source probability fusion, entropy calculation, uncertainty source ratio analysis, and priority generation;

[0214] Iteration control module: used to construct asymmetric masks according to priorities, perform constraint bundle search, calculate semantic gain and determine the termination of iteration.

[0215] When these instruction modules are loaded into the processor and run, they work together to convert the noisy input signal into a high-quality text sequence, fully reproducing the semantic communication decoding method described in this application at the machine level.

[0216] In one embodiment of this application, in the scenario of low Earth orbit (LEO) satellite Earth observation data backhaul and satellite-ground collaborative computing, the transmitting end is located on a resource-constrained LEO satellite, and the receiving end is located at a ground station or edge computing gateway. The process of the space-based computing-based semantic communication decoding method is as follows:

[0217] The transmitting end is a low-Earth orbit remote sensing satellite equipped with a lightweight semantic encoder. The satellite encodes multimodal remote sensing description text acquired onboard, such as a moving heat source target detected at 30 degrees North latitude, using a cascaded residual variational information bottleneck architecture. Due to power consumption limitations at the satellite end, the encoder only generates the most critical semantic residual features to minimize downlink bandwidth consumption. The ground receiver receives radio frequency signals transmitted via a satellite-to-ground link (e.g., Ka-band) through a high-gain antenna. Due to significant atmospheric rain attenuation and Doppler shift in the satellite-to-ground link, the received semantically encoded signal is typically accompanied by severe amplitude fluctuations and phase noise. The baseband processing unit at the ground station first performs Doppler compensation and channel equalization, then inputs the recovered semantic feature vector into a non-autoregressive semantic decoder deployed on a ground edge server. Although this decoder possesses strong language recovery capabilities, it is prone to forcibly constructing erroneous observation conclusions based on context during periods of deep fading (i.e., extremely low signal-to-noise ratio) in the satellite-to-ground link (i.e., model hallucinations).

[0218] To address the unique characteristics of space-based scenarios, the system performs cross-layer bias correction between the physical and semantic layers. The ground receiver uses pilot signals from the satellite-to-ground link for real-time channel estimation to identify the current channel state, such as whether it is in a rain attenuation zone. Spatial information observations are constructed. For satellite downlinks using orthogonal frequency division multiplexing waveforms, the system constructs a frequency-domain subcarrier sequence with all zeros, or a baseband signal containing only Gaussian white noise, to simulate the reception state when the satellite is not transmitting any effective semantics. These spatial information observations are input into a large-scale model at the ground end to obtain implicit language priors. For example, the model might, without any input, tend to predict high-frequency words such as ocean and clouds based solely on training habits. Correction is then performed using a logarithmic domain bias correction formula: log P ch_star =log P dec -η t *log P null ; where the depolarization intensity η t The signal-to-noise ratio (SNR) of the satellite-to-ground link is determined in real time. When the satellite passes through a rain zone, causing a sharp drop in SNR, η... t Automatically approaching 1, it forcefully eliminates the model's blind guesses about clouds, forcing the decoder to dig out the real information about the heat source in the weak signal.

[0219] While ground-based computing centers possess abundant computing resources, they face immense pressure from concurrent satellite data transmissions, necessitating the avoidance of inefficient computations. The system calculates the uncertainty source ratio ξ for each decoded token. t ξ t =H ch / (H ch +H sem +ε ξ Case A (channel impairment): If the ξ of a certain term (such as coordinate data) tThe high priority indicates that burst noise in the satellite-to-ground link caused the data ambiguity, which is a repairable transmission error. The system marks it as high priority and allocates computing power for refinement. Case B (semantic ambiguity): If a certain term, such as the description of ground texture ξ... t Very low, i.e., semantic entropy H sem The "dominant" ambiguity indicates that the satellite's observation resolution is insufficient to clearly perceive the information, or that the text description itself is inherently ambiguous. This is considered raw ambiguity that doesn't require correction, and the system marks it as low priority, omitting any refinement. This ensures that ground-based computing power is used solely to combat the poor performance of satellite-to-ground communication channels, rather than being wasted on guessing information that the satellite itself is uncertain about.

[0220] For channel impairment areas marked as high priority, such as the coordinates of critical targets obscured by noise, the decoder utilizes clear surrounding contextual information (such as location names and time) to perform multiple rounds of correction through constraint-based search. To further adapt to the short time window of satellite overpasses, i.e., strict time delay constraints, this embodiment employs a time delay-aware termination strategy: ε' n =ε'0*(1+T elapsed / T budget ); where T elapsed T represents the decoding time already consumed. budget The estimated link time for satellite overpass. As the satellite moves away, the available time decreases, and the convergence threshold ε'... n The system dynamically improves and automatically reduces the refinement precision to ensure that the complete final recovery text is output before the link is disconnected.

[0221] Alternatively, as an alternative to multi-granularity semantic decomposition based on a cascaded residual variational information bottleneck architecture, multi-granularity feature extraction can also be achieved based on syntactic analysis. Specifically, the system uses a pre-trained dependency parser to perform syntactic tree parsing on the input text sequence, extracting core syntactic nodes such as subject, predicate, and object from the syntactic tree as core semantic components for encoding, and mapping them to core semantic feature vectors; while modifying words (such as adjectives, adverbs, and relative clauses, etc., which are not core components) are mapped to detailed semantic feature vectors, or are directly ignored when the channel bandwidth is extremely limited. For example, for the sentence "A red apple fell from a height," the core semantic nodes extracted by syntactic analysis are apple (subject) and falling (predicate), while red (attributive modifier) ​​and from a height (adverbial modifier) ​​are classified as detailed semantics. The multi-granularity partitioning method based on syntactic structure does not require an additional variational information bottleneck training process, has low computational overhead, and is suitable for scenarios where the computing power of the transmitting end is extremely limited, such as low-power IoT sensor nodes or on-board embedded processors.

[0222] This embodiment solves the problem that traditional satellite communication suffers from low signal-to-noise ratio environments caused by rain attenuation or large dynamic Doppler frequency shift, where large models at the receiver are prone to hallucinations and the computing power scheduling efficiency is low.

[0223] In one alternative implementation, the system can, in addition to the cascaded residual variational information bottleneck architecture, further assist in distinguishing core features from detailed features by calculating the global semantic saliency score for each word. The global semantic saliency score can be used for initial annotation and verification of multi-granularity semantic labels in the VIB architecture, or as an independent, lightweight multi-granularity partitioning method. The formula for calculating the global semantic saliency score is as follows:

[0224] S score (t)=(1 / L)*Σ(α i *A self (i,t))+β idf *IDF(t);

[0225] Where S score (t) represents the semantic saliency score of the t-th word in the text sequence, L is the total number of layers in the semantic encoder, Σ() represents the summation of layer indices i from 1 to L, and α i Let A be the weight coefficient of the i-th layer encoder. self (i, t) represents the sum of self-attention weights obtained for the t-th word in the i-th layer, β idf IDF(t) is the weighting factor for inverse document frequency, and IDF(t) is the inverse document frequency value of the t-th word in the preset corpus.

[0226] Based on the calculated S score (t), the system sets a dynamic significance threshold. The significance threshold can be set to 1.2 times the average score of all words in the current sequence. When the S of a certain word... score When the saliency score (t) is greater than the threshold, the feature vector corresponding to the word is classified as a core semantic feature vector; otherwise, it is classified as a detailed semantic feature vector. For example, in the sentence "red apples are falling," the saliency scores for "apple" and "falling" are usually high and are identified as core semantics; while the score for "red" is low and is identified as detailed semantics. The saliency score comprehensively utilizes the global aggregation capability of the Transformer's multi-layer self-attention mechanism and the lexical sparseness information of inverse document frequency, so that the judgment of semantic importance takes into account both contextual relevance and lexical statistical features.

[0227] According to one aspect of this application, the semantic residual increment ΔT at each level is... l When mapping channel transmission symbols to the complex domain, the system employs a joint source-channel coding technique based on a deep neural network. To prevent signal power from exceeding hardware limitations and to meet the dynamic range constraints of the actual RF transceiver unit, the system needs to normalize the complex symbols output by the neural network. The normalization formula is as follows:

[0228] z norm =sqrt(ndim )*(z raw -μ batch ) / sqrt(σ var +ε*);

[0229] Where z norm For the normalized channel transmission symbols, n dim z is the dimension of the feature vector. raw μ is the original complex number sign of the neural network output. batch σ is the mean of the symbols in the current batch. var ε* represents the variance of the current batch symbol, and ε* is a small constant to prevent division by zero errors; specifically, it can take the value 1e. -8 .

[0230] The normalization operation constrains the average power of the transmitted signal to a fixed level, enabling the non-uniform resource allocation scheme in the subsequent power allocation strategy, which allocates higher transmit power to core semantic feature vectors and lower transmit power to detailed semantic feature vectors, to be executed on a unified power benchmark.

[0231] Furthermore, in some optional implementations, the system can also dynamically adjust the power allocation ratio between core semantics and detailed semantics based on the current Channel State Information (CSI). When the channel signal-to-noise ratio is detected to be lower than a preset threshold, the system will adjust the power allocation coefficient to concentrate more than 80% of the transmit power on transmitting the normalized symbols corresponding to the core semantic feature vectors, ensuring the reachability of basic semantic information under extremely poor channel conditions.

[0232] This invention employs an implicit prior debiasing mechanism based on spatial information observations. By constructing physical spatial observations (such as all-zero vectors) devoid of semantic information, it forces the decoder to expose its internally fixed linguistic statistical biases. The debiasing intensity is dynamically calculated using instantaneous signal-to-noise ratio (SNR), and this implicit prior is explicitly subtracted from the decoding result in the logarithmic domain. This eliminates the path of blindly guessing words while ignoring channel evidence under low SNR conditions, effectively restoring the true channel likelihood and improving signal recovery fidelity. Using source-aware uncertainty analysis and an asymmetric refinement strategy, the uncertainty source ratio of channel entropy to semantic entropy is quantified to accurately distinguish whether decoding errors originate from physical transmission impairments (requiring repair) or inherent semantic ambiguity (not requiring repair). Combined with asymmetric attention masks, the system only iteratively searches high-priority regions causing channel errors, while freezing semantically ambiguous regions. This avoids computational investment in invalid locations, reducing end-side energy consumption and processing latency while ensuring decoding quality, thus adapting to the resource constraints of space-based computing scenarios.

[0233] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. A semantic communication decoding method based on space-based computing, characterized in that, include: Receive the semantic encoded signal transmitted through the channel, input the semantic encoded signal into the pre-configured non-autoregressive semantic decoder, and generate an initial decoding distribution covering the candidate word set; Construct a spatial information observation adapted to the current channel characteristics, input the spatial information observation into a non-autoregressive semantic decoder to estimate the implicit language prior, and perform debiasing correction on the initial decoding distribution based on the implicit language prior to obtain the corrected channel likelihood distribution; Based on the corrected channel likelihood distribution and the semantic prior distribution from the pre-trained language model, the channel uncertainty and semantic uncertainty of each word position are quantified respectively, and the refinement priority of each word position is generated based on the uncertainty source ratio of the channel uncertainty in the total uncertainty. The text sequence obtained from the initial decoding of the pre-stored text is iteratively refined according to the refinement priority until the preset convergence condition is met, and the final recovered text is obtained. The initial decoding distribution is corrected based on implicit language priors, including: Obtain the instantaneous signal-to-noise ratio of the semantically encoded signal at each word position; The adaptive debiasing intensity corresponding to each word position is generated based on the instantaneous signal-to-noise ratio, wherein the adaptive debiasing intensity is negatively correlated with the instantaneous signal-to-noise ratio; By utilizing adaptive debiasing intensity, the correction magnitude of implicit language priors on the initial decoding distribution is controlled; The adaptive debiasing intensity corresponding to each word position is generated based on the instantaneous signal-to-noise ratio. Specifically, the mapping relationship is established using the Sigmoid function, and the calculation formula is as follows: or t =h max / (1+exp(β*(γ t -c η ))); Where η t To adaptive depolarization intensity, γ t For instantaneous signal-to-noise ratio, η max β is the pre-configured upper limit of depolarization intensity, β is the pre-configured temperature parameter, and γ is the pre-configured temperature parameter. η This is a pre-configured reference threshold; The channel uncertainty and semantic uncertainty at each lexical position are quantified separately, including: For each lexical position, calculate the information entropy of the corrected channel likelihood distribution, and use it as the channel entropy; Calculate the information entropy of the semantic prior distribution, and use it as the semantic entropy; The specific formula for calculating the ratio of channel uncertainty to total uncertainty sources is as follows: ξ t =H ch (t) / (H ch (t)+H sem (t)+ε ξ ); Where ξ t H represents the uncertainty source ratio at the t-th word position. ch (t) is the channel entropy, H sem (t) represents semantic entropy, ε ξ These are pre-configured positive numbers.

2. The method according to claim 1, characterized in that, Constructing spatial information observations adapted to the current channel characteristics includes: Obtain the feature dimensions of the semantically encoded signal; A vector sequence consistent with the semantically encoded signal in the feature dimension is constructed, and the values ​​of each element of the vector sequence are set as the expected values ​​of the received signal under the condition that there is no effective signal transmission in the current channel, which are used as the spatial information observation.

3. The method according to claim 2, characterized in that, When the current channel is an additive white Gaussian noise channel, a Rayleigh fading channel, or an orthogonal frequency division multiplexing channel, the statistical expectation value of the received signal under the condition of no effective signal transmission is a zero vector; the spatial information observation is a sequence of all zero vectors.

4. The method according to claim 1, characterized in that, The corrected channel likelihood distribution is obtained as follows: The initial decoding distribution and the implicit language prior are converted into a logarithmic field representation to obtain the logarithmic values ​​of the initial decoding distribution and the implicit language prior. For each lexical candidate, the corrected log score is obtained by subtracting the weighted log value of the implicit language prior from the log value of the initial decoding distribution; where the weighting is the adaptive debiasing intensity. Normalize the logarithm score of the correction to obtain the likelihood distribution of the corrected channel.

5. The method according to claim 1, characterized in that, The semantically encoded signal carries semantic hierarchy identifier information, generating the refinement priority of each lexical position, including: Based on the semantic hierarchy identifier information, determine the semantic importance weight of each word position; The joint confidence of each lexical position is obtained by fusing the corrected channel likelihood distribution and the semantic prior distribution. The source-aware modulation factor is generated based on the uncertainty source ratio, and the refinement priority is calculated using the following formula: r t =(1-c t )*ω(l t )*ψ(ξ t ); Where r t To refine priority, c t For joint confidence, ω(l) t ) represents the semantic importance weight, ψ(ξ) t ) is the source-sensing modulation factor, which is positively correlated with the uncertainty source ratio.

6. The method according to claim 1, characterized in that, The pre-stored text sequence obtained from the initial decoding is iteratively refined according to the refinement priority, including: The positions of words whose refinement priority is higher than the pre-configured refinement threshold are assigned to the set of positions to be refined, and the remaining positions are assigned to the set of frozen positions. Construct an asymmetric attention mask matrix. The asymmetric attention mask matrix opens the attention connections of the set of positions to be refined to all word positions, and closes the attention update channels of the frozen set of positions. Under the constraint of the asymmetric attention mask matrix, a non-autoregressive semantic decoder is used to perform constraint bundle search on the lexical units in the set of positions to be refined, generating the refined text for the current round.

Images