Improved context-aware adaptive contrast decoding method

Abstract

The invention discloses an improved context-aware adaptive contrast decoding method, and relates to the technical field of artificial intelligence. Comprising the following steps of: aiming at related retrieval contents obtained by a retrieval enhancement generation (RAG) framework, carrying out adaptive comparison decoding management of dynamic context weights on the related retrieval contents by utilizing a DC-ACD method: step 1, establishing a dynamic weight calculation model of context semantic relevance; through a dynamic weight calculation model of semantic correlation, it is ensured that retrieval content contexts with higher semantic similarity are allocated with higher weights; the method comprises the following steps of: (1) carrying out multi-document fusion, (2) evaluating context quality, (4) calculating Token-level dynamic weight according to whether to depend on priori knowledge and / or context information of an intelligent language model or not, and (5) analyzing final decoding probability distribution: synthesizing the weight proportion of each part to obtain the final decoding probability distribution.

Description

Technical Field

[0001] The invention discloses an improved context-aware adaptive contrast decoding method, and relates to the technical field of artificial intelligence. Background Art

[0002] Retrieval-Augmented Generation (RAG) has made significant progress in recent years and has become an effective method to improve the performance of large language models (LLMs). By combining external knowledge bases, RAG enables the model to access more extensive and updated information, thereby generating more accurate and relevant content. In practical applications, the retrieved contextual information often contains noise or irrelevant content, which has a negative impact on the quality of the generated text of the model and is not conducive to text optimization. Summary of the invention

[0003] In view of the problems in the prior art, the present invention provides an improved context-aware adaptive contrast decoding method, which effectively balances the prior knowledge and context information of a text generation model and realizes a more accurate and flexible text generation process.

[0004] The specific scheme proposed by the present invention is:

[0005] The present invention provides an improved context-aware adaptive contrast decoding method, comprising:

[0006] For the relevant retrieval content obtained by the retrieval enhancement generation RAG framework, the DC-ACD method is used to perform adaptive comparative decoding management of the dynamic context weights on the relevant retrieval content:

[0007] Step 1: Establish a dynamic weight calculation model for contextual semantic relevance, and ensure that the retrieval content context with higher semantic similarity is assigned a higher weight through the dynamic weight calculation model for semantic relevance;

[0008] Step 2: Perform multi-document fusion: Use the top k documents of the relevant search content to perform multi-document fusion based on the relevance score. Combining multi-document information can make up for the context of incomplete information.

[0009] Step 3: Evaluate context quality: Introduce a context quality evaluation network to evaluate the quality of the retrieved content context, so that higher quality contexts are given greater weight.

[0010] Step 4: Calculate the dynamic weight at the token level based on whether it relies on the prior knowledge and / or contextual information of the intelligent language model.

[0011] Step 5: Analyze the final decoding probability distribution: Comprehensively consider the weight ratios of each part to obtain the final decoding probability distribution.

[0012] Furthermore, in step 1 of the improved context-aware adaptive contrast decoding method, a dynamic weight calculation model of context semantic relevance is established, and the formula is as follows:

[0013] α DC-ACD (t) = λ·α ACD (t)+(1-λ)·sim(q,c)

[0014] Where: α DC-ACD (t) represents the dynamic weight at time t, λ·α ACD (t) is the weight calculated by the original ACD method at time t, sim(q,c) represents the semantic similarity between question q and context c, and λ is the balancing factor tuned by the validation set.

[0015] Further, in step 2 of the improved context-aware adaptive contrast decoding method, the formula is used:

[0016]

[0017] Perform multi-document fusion, where: is the fused context logit, w i represents the weight of the i-th document, is the logit based on the ith document.

[0018] Further, in step 3 of the improved context-aware adaptive contrast decoding method, the formula is used:

[0019] α DC-ACD (t) = α DC-ACD (t)·σ(S(c))

[0020] To evaluate the quality of the context, S(c) is a two-layer feedforward network: S(c) = W2·ReLU(W1·c+b1)+b2, where W1, W2, b1, b2 are learnable parameters. The input c of the network is the average word vector of the context or the vector representation obtained using the pre-trained model. The output q of the network is score =S(c) is mapped to the (0,1) interval through the sigmoid function σ.

[0021] Further, in step 4 of the improved context-aware adaptive contrast decoding method, the formula is used:

[0022] α token (t) = f(e t ,h t ,c)

[0023] Calculate the dynamic weight at the Token level, where e tis the token embedding at time t, h t is the hidden state at time t, c is the context representation, and f is a small neural network.

[0024] Further, in step 5 of the improved context-aware adaptive contrast decoding method, the formula is used:

[0025]

[0026] Get the final decoding probability distribution, where z t is the logit output of the original intelligent language model, representing the model's prediction based on parameter knowledge, α token (t) is the dynamic weight at the token level, α DC-ACD (t) is the dynamically calculated context weight, is the fused context logit, Represents the additional contribution of contextual information relative to the prior knowledge of the intelligent language model.

[0027] The present invention also provides an improved context-aware adaptive contrast decoding device, comprising a decoding management module,

[0028] The decoding management module uses the DC-ACD method to perform adaptive comparative decoding management of dynamic context weights on the relevant search content obtained by the retrieval enhancement RAG framework:

[0029] Step 1: Establish a dynamic weight calculation model for contextual semantic relevance, and ensure that the retrieval content context with higher semantic similarity is assigned a higher weight through the dynamic weight calculation model for semantic relevance;

[0030] Step 2: Perform multi-document fusion: Use the top k documents of the relevant search content to perform multi-document fusion based on the relevance score. Combining multi-document information can make up for the context of incomplete information.

[0031] Step 3: Evaluate context quality: Introduce a context quality evaluation network to evaluate the quality of the retrieved content context, so that higher quality contexts are given greater weight.

[0032] Step 4: Calculate the dynamic weight at the token level based on whether it relies on the prior knowledge and / or contextual information of the intelligent language model.

[0033] Step 5: Analyze the final decoding probability distribution: Comprehensively consider the weight ratios of each part to obtain the final decoding probability distribution.

[0034] The present invention also provides a computer-readable medium, on which computer instructions are stored. When the computer instructions are executed by a processor, the processor is caused to execute the improved context-aware adaptive contrast decoding method.

[0035] The benefits of the present invention are:

[0036] Refined context utilization: By combining semantic similarity and dynamic token-level weights, more refined and accurate context information utilization is achieved.

[0037] Enhanced robustness: Multi-document fusion and contextual quality assessment mechanisms improve the model’s robustness to noise and irrelevant information.

[0038] Flexible adaptability: Dynamic weight calculation enables the model to flexibly respond to contexts of different types and qualities.

[0039] Improved information utilization efficiency: The multi-document fusion mechanism makes full use of the rich information provided by the retrieval system.

[0040] Enhanced interpretability: Improved interpretability of model decisions through explicit contextual quality assessment and token-level weights.

[0041] The DC-ACD method in the present invention, namely the Dynamic Context-Weighted Adaptive Contrastive Decoding (DC-ACD), is suitable for various natural language processing tasks that need to combine external knowledge, especially when facing potential noise or incomplete context, such as open domain question answering, knowledge-intensive dialogue systems, fact checking and verification, personalized content generation and other scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a schematic flow chart of the method of the present invention. DETAILED DESCRIPTION

[0043] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it, but the embodiments are not intended to limit the present invention.

[0044] Example 1

[0045] The present invention provides an improved context-aware adaptive contrast decoding method, comprising:

[0046] For the relevant retrieval content obtained by the retrieval enhancement generation RAG framework, the DC-ACD method is used to perform adaptive comparative decoding management of the dynamic context weights on the relevant retrieval content:

[0047] Step 1: Establish a dynamic weight calculation model for context semantic relevance, and ensure that the retrieval content context with higher semantic similarity is assigned a higher weight through the dynamic weight calculation model for context semantic relevance.

[0048] Among them, the formula for establishing the dynamic weight calculation model for context semantic relevance is as follows:

[0049] α DC-ACD (t) = λ·α ACD (t) + (1 - λ)·sim(q, c)

[0050] Among them: α DC-ACD (t) represents the dynamic weight at time t, λ·α ACD (t) is the weight calculated by the original ACD method at time t, sim(q, c) represents the semantic similarity between the question q and the context c, and λ is a balance factor tuned through the validation set.

[0051] In specific implementation, the calculation of λ·α ACD (t) follows the following method and is calculated based on entropy value. The entropy value calculation process of the ACD method is as follows:

[0052] For the given input x and the previously generated sequence y<t, calculate the output probability distributions of the language model at time t in the cases with and without context c respectively: P θ (Y t |x, y <t , c) and P θ (Y t |x, y <t ), θ represents the parameters of the language model, Y t represents the output token at time t, x is the input (usually the question in the question-and-answer task), y <t represents the sequence generated before time t, and c is the context information.

[0053] Calculate the entropy of these two probability distributions: H(Y t |c) = -∑P θ (Y t |x, y <t , c)·logP θ (Y t |x, y <t , c) and H(Y t ) = -∑P θ (Y t |x, y <t ), where the summation is performed over all tokens in the vocabulary.

[0054] Calculate the difference in entropy: ΔH = H(Y t )-H(Y t |c)

[0055] Use the sigmoid function to map the entropy difference to the (0,1) interval and get α ACD (t): This α ACD (t) is the weight calculated at time t. It reflects the contribution of context information to reducing model uncertainty. When ΔH is large, it means that the context significantly reduces the uncertainty of the model. ACD (t) is close to 1; otherwise, it is close to 0.

[0056] Sim(q,c) can be calculated using pre-trained semantic matching models such as BERT, RoBERTa, etc. The calculation steps are as follows:

[0057] 1. Input the question q and context c into the pre-trained model to obtain their vector representations.

[0058] 2. Calculate the cosine similarity of the two vectors as the value of sim(q,c).

[0059] To better understand how this formula works, imagine that for the question “Who invented the telephone?”, the search system returns two pieces of context:

[0060] c1: "Alexander Graham Bell invented the telephone in 1876."

[0061] c2: "Thomas Edison invented many important electronic devices."

[0062] Using the semantic matching model, we calculated: sim(q,c1)=0.85 and sim(q,c2)=0.60, assuming λ=0.7, α ACD (t) = 0.5, then:

[0063] For c1, α DC-ACD (t)=0.7·0.5+0.3·0.85=0.605,

[0064] For c2, α DC-ACD (t)=0.7·0.5+0.3·0.60=0.530,

[0065] The above example shows that even if the α of two contexts ACD (t) has the same value, but c1 will get a larger weight due to its higher semantic similarity to the question. This approach ensures that more relevant information can play a greater role in the generation process.

[0066] Step 2: Perform multi-document fusion: Use the top k documents of the relevant search content to perform multi-document fusion based on the relevance score. Combining multi-document information can make up for the context of incomplete information.

[0067] The formula used is:

[0068]

[0069] Perform multi-document fusion, where: is the fused context logit, w i represents the weight of the i-th document, is the logit based on the ith document.

[0070] Specifically:

[0071] a) Document weight w i It can be determined based on the relevance score returned by the retrieval system, for example, by normalizing the relevance score using a softmax function: Where exp represents the softmax function, score i Represents the relevance score of the i-th document. The relevance score calculation can be based on the calculation logic of traditional search engines.

[0072] b) For each document, calculate the ACD score

[0073] c) Finally, the logit weighted sum of all documents is obtained

[0074] For example, suppose the retrieval system returns 3 documents with relevance scores of 10, 8, and 6. First, calculate the weights:

[0075]

[0076] Then, suppose that at a certain time step t, the logits corresponding to these three documents are: The fused logit is:

[0077]

[0078] This fusion method can comprehensively consider the information of multiple related documents, improve the comprehensiveness and accuracy of information utilization, and enhance the robustness of the system to noise and incomplete information.

[0079] Step 3: Evaluate context quality: Introduce a context quality assessment network to evaluate the quality of the retrieved content context, so that higher quality contexts receive greater weight.

[0080] The formula used is:

[0081] α DC-ACD (t) = α DC-ACD (t)·σ(S(c))

[0082] To evaluate the quality of the context, S(c) is a two-layer feedforward network: S(c) = W2·ReLU(W1·c+b1)+b2, where W1, W2, b1, b2 are learnable parameters. The input c of the network is the average word vector of the context or the vector representation obtained using the pre-trained model. The output q of the network is score =S(c) is mapped to the (0,1) interval through the sigmoid function σ.

[0083] For example, given the two previous contexts:

[0084] c1: "Alexander Graham Bell invented the telephone in 1876."

[0085] c2: "Thomas Edison invented many important electronic devices."

[0086] The outputs of the quality assessment network are: q score 1 =0.85,q score 2 =0.60, the final weight adjustment is:

[0087] For c1, α DC-ACD (t) = 0.605·sigmod(0.85)≈0.605·0.701≈0.424; for c2, α DC-ACD (t)=0.530·sigmod(0.60)≈0.530·0.646≈0.342.

[0088] This shows that the higher quality context c1 gets a greater weight, while the lower quality context c2 is further reduced in influence. In this way, the influence of high-quality information can be further highlighted while reducing the interference of low-quality or irrelevant information.

[0089] Step 4: Calculate the dynamic weight at the token level based on whether it relies on the prior knowledge and / or contextual information of the intelligent language model.

[0090] The formula used is:

[0091] α token (t) = f(e t ,h t ,c)

[0092] Calculate the dynamic weight at the Token level, where e t is the token embedding at time t, h tis the hidden state at time t, c is the context representation, and f is a small neural network.

[0093] The specific process is as follows:

[0094] a) f can be a multi-layer perceptron (MLP). For example:

[0095] f(e t ,h t ,c) = sigmod(W2·ReLU(W1·[e t ;h t ;c]+b1)+b2), where [e t ;h t ;c]

[0096] means concatenating these three vectors.

[0097] b) The output of the network is mapped to the interval (0, 1) through the sigmoid function σ, representing the degree of dependence of the current token on the context.

[0098] c) In the training phase, the f network can be trained in an end-to-end manner so that it learns the degree of dependence of different tokens on the context. For example, assuming that when generating the sentence "Alexander Graham Bell invented the telephone":

[0099] For the token "Alexander", α token (t) may be close to 1 because it highly depends on context information.

[0100] For the common grammar word "了", α token (t) may be close to 0 because it mainly depends on the language knowledge of the model.

[0101] For the word "telephone", α token (t) may be around 0.5 because it depends on both context and the prior knowledge of the model. This fine-grained weight adjustment enables the model to utilize context information more precisely, thereby generating more accurate and coherent text.

[0102] Step 5: Analyze the final decoding probability distribution: Combine the weight ratios of each part to obtain the final decoding probability distribution.

[0103] Among them, use the formula:

[0104]

[0105] to obtain the final decoding probability distribution, where z t is the logit output of the original intelligent language model, representing the prediction of the model based on parameter knowledge, and α token(t) is the dynamic weight at the token level, α DC-ACD (t) is the dynamically calculated context weight, is the fused context logit, Represents the additional contribution of contextual information relative to the prior knowledge of the intelligent language model.

[0106] In this way, the DC-ACD method can dynamically balance the model's prior knowledge and contextual information at the token level, achieving a more accurate decoding process.

[0107] Example 2

[0108] The present invention also provides an improved context-aware adaptive contrast decoding device, comprising a decoding management module,

[0109] The decoding management module uses the DC-ACD method to perform adaptive comparative decoding management of dynamic context weights on the relevant search content obtained by the retrieval enhancement RAG framework:

[0110] Step 1: Establish a dynamic weight calculation model for contextual semantic relevance, and ensure that the retrieval content context with higher semantic similarity is assigned a higher weight through the dynamic weight calculation model for semantic relevance;

[0111] Step 2: Perform multi-document fusion: Use the top k documents of the relevant search content to perform multi-document fusion based on the relevance score. Combining multi-document information can make up for the context of incomplete information.

[0112] Step 3: Evaluate context quality: Introduce a context quality evaluation network to evaluate the quality of the retrieved content context, so that higher quality contexts are given greater weight.

[0113] Step 4: Calculate the dynamic weight at the token level based on whether it relies on the prior knowledge and / or contextual information of the intelligent language model.

[0114] Step 5: Analyze the final decoding probability distribution: Comprehensively consider the weight ratios of each part to obtain the final decoding probability distribution.

[0115] As the information interaction and execution process between the modules in the above-mentioned device are based on the same concept as the embodiment of the method of the present invention, the specific contents can be found in the description of the embodiment of the method of the present invention and will not be repeated here.

[0116] Likewise, the benefits of the device of the present invention are:

[0117] Refined context utilization: By combining semantic similarity and dynamic token-level weights, more refined and accurate context information utilization is achieved.

[0118] Enhanced robustness: Multi-document fusion and contextual quality assessment mechanisms improve the model’s robustness to noise and irrelevant information.

[0119] Flexible adaptability: Dynamic weight calculation enables the model to flexibly respond to contexts of different types and qualities.

[0120] Improved information utilization efficiency: The multi-document fusion mechanism makes full use of the rich information provided by the retrieval system.

[0121] Enhanced interpretability: Improved interpretability of model decisions through explicit contextual quality assessment and token-level weights.

[0122] It should be noted that not all steps and modules in the above-mentioned processes and device structures are necessary, and some steps or modules can be ignored according to actual needs. The execution order of each step is not fixed and can be adjusted as needed. The system structure described in the above-mentioned embodiments can be a physical structure or a logical structure, that is, some modules may be implemented by the same physical entity, or some modules may be implemented by multiple physical entities, or some components in multiple independent devices may be implemented together.

[0123] Example 3

[0124] The present invention also provides a computer-readable medium, on which a computer instruction is stored, and when the computer instruction is executed by a processor, the processor executes the improved context-aware adaptive contrast decoding method. Specifically, a system or device equipped with a storage medium can be provided, on which a software program code implementing the functions of any of the above embodiments is stored, and a computer (or CPU or MPU) of the system or device reads and executes the program code stored in the storage medium.

[0125] In this case, the program code itself read from the storage medium can realize the function of any one of the above-mentioned embodiments, and thus the program code and the storage medium storing the program code constitute a part of the present invention.

[0126] The storage medium embodiments for providing the program code include a floppy disk, a hard disk, a magneto-optical disk, an optical disk (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD+RW), a magnetic tape, a non-volatile memory card, and a ROM. Alternatively, the program code can be downloaded from a server computer by a communication network.

[0127] In addition, it should be clear that the functions of any of the above embodiments can be implemented not only by executing the program code read by the computer, but also by enabling an operating system operating on the computer to complete part or all of the actual operations based on instructions from the program code.

[0128] In addition, it can be understood that the program code read from the storage medium is written to a memory provided in an expansion board inserted into the computer or written to a memory provided in an expansion unit connected to the computer, and then based on the instructions of the program code, a CPU installed on the expansion board or the expansion unit is enabled to perform part or all of the actual operations, thereby realizing the functions of any of the above-mentioned embodiments.

[0129] The above-described embodiments are only preferred embodiments for fully illustrating the present invention, and the protection scope of the present invention is not limited thereto. Equivalent substitutions or changes made by those skilled in the art based on the present invention are within the protection scope of the present invention. The protection scope of the present invention shall be subject to the claims.

Claims

1. An improved context-aware adaptive contrast decoding method, characterized in that include: For the relevant retrieval content obtained by the retrieval enhancement generation RAG framework, the DC-ACD method is used to perform adaptive comparative decoding management of the dynamic context weights on the relevant retrieval content: Step 1: Establish a dynamic weight calculation model for contextual semantic relevance, and ensure that the retrieval content context with higher semantic similarity is assigned a higher weight through the dynamic weight calculation model for semantic relevance; Step 2: Perform multi-document fusion: Use the top k documents of the relevant search content to perform multi-document fusion based on the relevance score. Combining multi-document information can make up for the context of incomplete information. Step 3: Evaluate context quality: Introduce a context quality evaluation network to evaluate the quality of the retrieved content context, so that higher quality contexts are given greater weight. Step 4: Calculate the dynamic weight at the token level based on whether it relies on the prior knowledge and / or contextual information of the intelligent language model. Step 5: Analyze the final decoding probability distribution: Comprehensively consider the weight ratios of each part to obtain the final decoding probability distribution.

2. According to the improved context-aware adaptive contrast decoding method of claim 1, it is characterized in that a dynamic weight calculation model of context semantic relevance is established in step 1, and the formula is as follows: a DC-ACD (t)=λ·a ACD (t)+(1-λ)·sim(q,c) in: α DC-ACD (t) represents the dynamic weight at time t, λ·α ACD (t) is the weight calculated by the original ACD method at time t, sim(q,c) represents the semantic similarity between question q and context c, and λ is the balancing factor tuned by the validation set.

3. The improved context-aware adaptive contrast decoding method according to claim 1, characterized in that In step 2, use the formula: Perform multi-document fusion, where: is the fused context logit, w i represents the weight of the i-th document, is the logit based on the ith document.

4. The improved context-aware adaptive contrast decoding method according to claim 1, characterized in that the formula is used in step 3: a DC-ACD (t)=a DC-ACD (t)·σ(S(c)) To evaluate the quality of the context, S(c) is a two-layer feedforward network: S(c) = W2·ReLU(W1·c+b1)+b2, where W1, W2, b1, b2 are learnable parameters. The input c of the network is the average word vector of the context or the vector representation obtained using the pre-trained model. The output q of the network is score =S(c) is mapped to the (0,1) interval through the sigmoid function σ.

5. The improved context-aware adaptive contrast decoding method according to claim 1, characterized in that The formula used in step 4 is: α token (t)=f(e t ,h t ,c) Calculate the dynamic weight at the Token level, where e t is the token embedding at time t, h t is the hidden state at time t, c is the context representation, and f is a small neural network.

6. The improved context-aware adaptive contrast decoding method according to claim 1, characterized in that In step 5, use the formula: Get the final decoding probability distribution, where z t is the logit output of the original intelligent language model, representing the model's prediction based on parameter knowledge, α token (t) is the dynamic weight at the token level, α DC-ACD (t) is the dynamically calculated context weight, is the fused context logit, Represents the additional contribution of contextual information relative to the prior knowledge of the intelligent language model.

7. An improved context-aware adaptive contrast decoding device, characterized in that Including decoding management module, The decoding management module uses the DC-ACD method to perform adaptive comparative decoding management of dynamic context weights on the relevant search content obtained by the retrieval enhancement RAG framework: Step 1: Establish a dynamic weight calculation model for contextual semantic relevance, and ensure that the retrieval content context with higher semantic similarity is assigned a higher weight through the dynamic weight calculation model for semantic relevance; Step 2: Perform multi-document fusion: Use the top k documents of the relevant search content to perform multi-document fusion based on the relevance score. Combining multi-document information can make up for the context of incomplete information. Step 3: Evaluate context quality: Introduce a context quality evaluation network to evaluate the quality of the retrieved content context, so that higher quality contexts are given greater weight. Step 4: Calculate the dynamic weight at the token level based on whether it relies on the prior knowledge and / or contextual information of the intelligent language model. Step 5: Analyze the final decoding probability distribution: Comprehensively consider the weight ratios of each part to obtain the final decoding probability distribution.

8. A computer readable medium, characterized in that The computer-readable medium stores computer instructions, and when the computer instructions are executed by a processor, the processor executes an improved context-aware adaptive contrast decoding method as described in any one of claims 1 to 6.

Images