Improved context-aware adaptive contrastive decoding method

Abstract

An improved context-aware adaptive contrastive decoding method, relating to the technical field of artificial intelligence. The method comprises: for relevant retrieved content obtained by a retrieval-augmented generation (RAG) framework, using a dynamic context-weighted adaptive contrastive decoding (DC-ACD) method to perform DC-ACD management on the relevant retrieved content: step 1, establishing a dynamic weight computation model for contextual semantic relevance, and ensuring, by means of the dynamic weight computation model for semantic relevance, that a higher weight is assigned to retrieved content context having a higher semantic similarity; step 2, performing multi-document fusion; step 3, evaluating context quality; step 4, computing a Token-level dynamic weight on the basis of whether prior knowledge of an intelligent language model and / or contextual information are / is relied on; and step 5, analyzing a final decoding probability distribution: combining all weight proportions to obtain the final decoding probability distribution.

Description

Improved context-aware adaptive contrast decoding method

[0001] The present application claims priority to the Chinese patent application No. 202411291133.9, filed on September 14, 2024, and entitled "Improved context-aware adaptive contrast decoding method", the whole content of which is incorporated herein by reference. TECHNICAL FIELD

[0002] The present application discloses an improved context-aware adaptive contrast decoding method, and relates to the technical field of artificial intelligence. BACKGROUND

[0003] 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 (LLM). RAG combines external knowledge bases to enable models to access more extensive and updated information, thereby generating more accurate and relevant content. In practical applications, the retrieved context information often contains noise or irrelevant content, which negatively affects the quality of the generated text and is not conducive to text optimization. SUMMARY

[0004] The present application provides an improved context-aware adaptive contrast decoding method to effectively balance the prior knowledge and context information of the text generation model, and realizes a more accurate and flexible text generation process.

[0005] The specific solutions provided by the present application are as follows:

[0006] The present application provides an improved context-aware adaptive contrast decoding method, comprising:

[0007] For the relevant retrieval content obtained by the Retrieval-Augmented Generation (RAG) framework, the DC-ACD method is used to perform adaptive contrast decoding management of the dynamic context weight of the relevant retrieval content:

[0008] Step 1: Establish a dynamic weight calculation model of context semantic relevance to ensure that the retrieval content with higher semantic similarity is assigned a higher weight through the dynamic weight calculation model of semantic relevance;

[0009] Step 2: Perform multi-document fusion: use the top k documents of the relevant retrieval content to perform multi-document fusion according to the relevance score, and combine the multi-document information to compensate for the context of incomplete information;

[0010] Step 3: Evaluate the quality of the context: introduce a context quality evaluation network to evaluate the quality of the retrieved context, and give higher weight to the context with higher quality;

[0011] Step 4: Calculate the dynamic weight at the Token level according to whether to rely on the prior knowledge of the intelligent language model and / or the context information;

[0012] Step 5: Analyze the final decoding probability distribution: integrate the weight proportions of each part to obtain the final decoding probability distribution.

[0013] Further, the dynamic weight calculation model for establishing the context semantic correlation in step 1 of the improved context-aware adaptive contrast decoding method is as follows: 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 the question q and the context c, and λ is a balance factor adjusted through the validation set.

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

[0016] Multi-document fusion is performed, where: is the fused context logit, w i represents the weight of the i-th document, is the logit based on the i-th document.

[0017] Further, in step 3 of the improved context-aware adaptive contrast decoding method, the formula is used: α DC-ACD (t)=α DC-ACD (t)·σ(S(c))

[0018] The quality of the context is evaluated, and S(c) is a double-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 a pre-trained model, and the output q score =S(c) is mapped to the interval (0,1) through the sigmoid function σ.

[0019] Further, in step 4 of the improved context-aware adaptive contrast decoding method, the formula is used: αtoken (t) = f(e t h t ,c)

[0020] Compute the dynamic weight of 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.

[0021] Further, the step 5 of the improved context-aware adaptive contrast decoding method utilizes the formula:

[0022] Obtain the final decoding probability distribution, where z t is the logit output of the original intelligent language model, representing the prediction based on the model's prior knowledge, a token (t) is the dynamic weight of Token level, a DC-ACD (t) is the dynamically calculated context weight, is the fused context logit, represents the additional contribution of context information relative to the prior knowledge of the intelligent language model.

[0023] The application also provides an improved context-aware adaptive contrast decoding device, comprising a decoding management module,

[0024] The decoding management module utilizes the DC-ACD method to perform adaptive contrast decoding management of the relevant search content obtained by the RAG framework for search enhancement:

[0025] Step 1: Establish a dynamic weight calculation model for context semantic relevance, which ensures that search content with higher semantic similarity is assigned a higher weight through the dynamic weight calculation model for semantic relevance;

[0026] Step 2: Perform multi-document fusion: use the top k documents of the relevant search content to perform multi-document fusion according to the relevance score, which can compensate for incomplete information context after combining multi-document information;

[0027] Step 3: Evaluate the quality of the context: introduce a context quality evaluation network to evaluate the quality of the search content context, so that the context with higher quality obtains a greater weight;

[0028] Step 4: Calculate the dynamic weight of Token level according to whether to rely on the prior knowledge of the intelligent language model and / or the context information;

[0029] Step 5: Analyze the final decoding probability distribution: combine the weight proportions of each part to obtain the final decoding probability distribution.

[0030] The application also provides a computer readable medium having computer instructions stored thereon, which, when executed by a processor, cause the processor to perform the improved context-aware adaptive contrastive decoding method.

[0031] The application has the following advantages:

[0032] Fine-grained context utilization: By combining semantic similarity and dynamic token-level weights, more fine-grained and accurate context information utilization is achieved.

[0033] Enhanced robustness: Multi-document fusion and context quality evaluation mechanism improve the robustness of the model to noise and irrelevant information.

[0034] Flexible adaptability: Dynamic weight calculation enables the model to flexibly cope with different types and qualities of context.

[0035] Improved information utilization efficiency: The multi-document fusion mechanism fully utilizes the rich information provided by the retrieval system.

[0036] Enhanced interpretability: Through explicit context quality evaluation and token-level weights, the interpretability of model decision-making is improved.

[0037] The DC-ACD method in the application, namely the dynamic context-weighted adaptive contrastive decoding method (Dynamic Context-Weighted Adaptive Contrastive Decoding, DC-ACD), is applicable to various natural language processing tasks that require the combination of external knowledge, especially in the face of potential noise or incomplete context, such as open-domain question answering, knowledge-intensive dialogue systems, fact checking and verification, personalized content generation, etc. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 is a flowchart of the method of the application. DETAILED DESCRIPTION

[0039] The application will be further described below in conjunction with the drawings and specific examples, so that those skilled in the art can better understand the application and implement it, but the examples are not limiting the application.

[0040] Example 1

[0041] The application provides an improved context-aware adaptive contrastive decoding method, comprising:

[0042] For the relevant search content obtained by the retrieval enhancement generation RAG framework, the DC-ACD method is used to adaptively manage the dynamic context weight of the relevant search content:

[0043] Step 1: Establish a dynamic weight calculation model of context semantic correlation, and ensure that the search content with higher semantic similarity is assigned a higher weight through the dynamic weight calculation model of semantic correlation.

[0044] The dynamic weight calculation model of context semantic correlation is established, and the formula is as follows: α DC-ACD (t)=λ·α ACD (t)+(1-λ)·sim(q,c);

[0045] 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 the question q and the context c, and λ is a balance factor adjusted through the validation set.

[0046] In specific implementation, the calculation of λ·ɑ ACD (t) follows the following method, which is calculated based on entropy. The entropy calculation process of the ACD method is as follows:

[0047] For a given input x and the previously generated sequence y<t, the output probability distribution of the language model at time t is calculated under the conditions of having context c and not having context: 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 answering task), y ＜t represents the sequence generated before time t, and c is the context information.

[0048] The entropies of the two probability distributions are calculated: 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 )·logP θ (Yt | x, y ＜t ), where the summation is over all tokens in the vocabulary.

[0049] Compute the difference in entropy: AH = H(Y t ) - H(Y t | c).

[0050] Map the entropy difference to the interval (0, 1) using a sigmoid function to get a ACD (t): This a ACD (t) is the weight computed at time t. It reflects the degree to which the context information contributes to reducing the model's uncertainty. When AH is large, it means that the context significantly reduces the model's uncertainty, and a ACD (t) is close to 1; otherwise, it is close to 0.

[0051] And sim(q, c) can be calculated using a pre-trained semantic matching model, such as BERT, RoBERTa, etc. The calculation steps are as follows:

[0052] 1. Input the question q and the context c into the pre-trained model to obtain their vector representations;

[0053] 2. Calculate the cosine similarity of the two vectors as the value of sim(q, c).

[0054] To better understand the role of this formula, assume there is a question "Who invented the telephone?", and the retrieval system returns two contexts:

[0055] c1: "Alexander Graham Bell invented the telephone in 1876."

[0056] c2: "Thomas Edison invented many important electronic devices."

[0057] Using a semantic matching model to calculate: sim(q, c1) = 0.85 and sim(q, c2) = 0.60, assuming λ = 0.7, a ACD (t) = 0.5, then:

[0058] For c1, a DC-ACD (t) = 0.7 * 0.5 + 0.3 * 0.85 = 0.605,

[0059] For c2, a DC-ACD (t) = 0.7 * 0.5 + 0.3 * 0.60 = 0.530,

[0060] The above example shows that even if the two contexts have different a ACD(t) The same value, but since c1 has a higher semantic similarity to the problem, it will have a greater weight. This method ensures that more relevant information can play a greater role in the generation process.

[0061] Step 2: Perform multi-document fusion: Use the top k documents of the relevant search content to perform multi-document fusion based on relevance scores. Combining multi-document information can compensate for incomplete information context.

[0062] Where the formula is:

[0063] 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 i-th document.

[0064] Specifically:

[0065] a) Document weight w i Can be determined based on the relevance score returned by the search system, for example, using the softmax function to normalize the relevance score: Where exp represents the softmax function, score i represents the relevance score of the i-th document, and the relevance score calculation can be based on the calculation logic of traditional search engines.

[0066] b) For each document, calculate the score of ACD

[0067] c) Finally, sum the weighted logit of all documents to get the fused

[0068] For example, assume the search system returns 3 documents with relevance scores of 10, 8, and 6. First, calculate the weight:

[0069] Then, assume that at a certain time step t, the logit corresponding to these three documents is: The fused logit is:

[0070] This fusion method can consider the information of multiple relevant documents, improving the comprehensiveness and accuracy of information utilization, while enhancing the robustness of the system to noise and incomplete information.

[0071] Step 3: Evaluate context quality: Introduce a context quality evaluation network to evaluate the quality of the search content context, so that higher quality contexts have a greater weight.

[0072] where the formula is: a DC-ACD (t) = a DC-ACD (t) · s(S(c))

[0073] The context quality is evaluated, and 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 of the word vectors of the context or the vector representation obtained using a pre-trained model, and the output q score = S(c) is mapped to the interval (0, 1) through the sigmoid function s.

[0074] For example, assuming for the previous two contexts:

[0075] c1: “Alexander Graham Bell invented the telephone in 1876.”

[0076] c2: “Thomas Edison invented many important electronic devices.”

[0077] The outputs of the quality evaluation network are respectively: q score1 = 0.85, q score2 = 0.60, then the final weight adjustment is:

[0078] For c1, a DC-ACD (t) = 0.605·sigmod(0.85) ≈ 0.605·0.701 ≈ 0.424; for c2, a DC-ACD (t) = 0.530·sigmod(0.60) ≈ 0.530·0.646 ≈ 0.342.

[0079] This shows that the context c1 with higher quality obtains a greater weight, while the influence of the context c2 with lower quality is further reduced. In this way, the influence of high-quality information can be further highlighted, while the interference of low-quality or irrelevant information is reduced.

[0080] Step 4: Calculate the dynamic weight of Token level according to whether it depends on the prior knowledge of intelligent language model and / or context information.

[0081] where the formula is: a token (t) = f(e t , h t , c)

[0082] Calculate the dynamic weight of 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.

[0083] The specific process is as follows:

[0084] a) f can be a multi-layer perceptron (MLP), for example:

[0085] f(e t ,h t ,c) = sigmod(W2·ReLU(W1·[e t ; h t ; c] + b1) + b2), where [e t ; h t ; c] represents concatenating the three vectors;

[0086] b) The output of the network is mapped to the interval (0, 1) by the sigmoid function σ, representing the degree of dependence of the current token on the context;

[0087] 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":

[0088] For the token "Alexander", α token (t) can be close to 1, because it highly depends on the context information;

[0089] For the common grammatical word "of", α token (t) can be close to 0, because it mainly depends on the language knowledge of the model;

[0090] For the word "telephone", α token (t) can be around 0.5, because it depends on both the context and the prior knowledge of the model. This fine-grained weight adjustment allows the model to more accurately utilize context information, resulting in more accurate and coherent text generation.

[0091] Step 5: Analyze the final decoding probability distribution: integrate the proportion of each part weight to obtain the final decoding probability distribution.

[0092] where the formula is:

[0093] 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 token-level dynamic weight, α DC-ACD (t) is the dynamically calculated context weight, is the fused context logit, represents the additional contribution of the context information relative to the prior knowledge of the intelligent language model.

[0094] In this way, using the DC-ACD method can dynamically balance the prior knowledge and context information of the model at the token level, and realize a more accurate decoding process.

[0095] Embodiment 2

[0096] The application also provides an improved context-aware adaptive contrast decoding device, comprising a decoding management module,

[0097] The decoding management module uses the DC-ACD method to perform adaptive contrast decoding management of the relevant search content obtained by the RAG framework for search enhancement:

[0098] Step 1: Establish a dynamic weight calculation model of context semantic relevance, and ensure that the search content context with higher semantic similarity is assigned a higher weight through the dynamic weight calculation model of semantic relevance;

[0099] Step 2: Multi-document fusion: use the top k documents of the relevant search content to perform multi-document fusion according to the relevance score, and combine the multi-document information to make up for the incomplete context information,

[0100] Step 3: Evaluate the context quality: introduce a context quality evaluation network to evaluate the context quality of the search content, so that the context with higher quality obtains a greater weight,

[0101] Step 4: Calculate the dynamic weight at the token level according to whether to rely on the prior knowledge and / or context information of the intelligent language model,

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

[0103] The information interaction and execution process between the modules in the above device are based on the same concept as the method embodiments of the application, and the specific content can be referred to the description in the method embodiments of the application, which will not be repeated here.

[0104] Similarly, the benefits of the device of the application are:

[0105] Fine-grained context utilization: by combining semantic similarity and dynamic token-level weight, more fine and accurate context information utilization is realized.

[0106] Enhanced robustness: multi-document fusion and context quality evaluation mechanism improve the robustness of the model to noise and irrelevant information.

[0107] Flexibility: Dynamic weight calculation allows the model to adapt to different types and qualities of context.

[0108] Efficiency of information utilization: The multi-document fusion mechanism fully utilizes the rich information provided by the retrieval system.

[0109] Enhanced interpretability: Through explicit context quality evaluation and token-level weights, the model's decision-making interpretability is improved.

[0110] It should be noted that not all steps and modules in the above processes and device structures are necessary, and some steps or modules can be omitted 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 embodiments can be a physical structure or a logical structure, i.e., some modules can be implemented by the same physical entity, or some modules can be implemented by multiple physical entities, or they can be implemented by some components in multiple independent devices.

[0111] Embodiment 3

[0112] The present application also provides a computer readable medium having computer instructions stored thereon, which, when executed by a processor, cause the processor to perform the improved context-aware adaptive contrast decoding method. Specifically, a system or device equipped with a storage medium can be provided, which stores software program code for implementing the functions of any of the above embodiments, and causes the computer (or CPU or MPU) of the system or device to read and execute the program code stored in the storage medium.

[0113] In this case, the program code read from the storage medium itself can implement the functions of any of the above embodiments, so the program code and the storage medium storing the program code constitute part of the present application.

[0114] Embodiments of storage media for providing program code include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD+RW), magnetic tapes, non-volatile memory cards, and ROM. Alternatively, program code can be downloaded from a server computer via a communication network.

[0115] In addition, it should be clear that not only can the actual operation be completed by executing the program code read by the computer, but also part or all of the actual operation can be completed by the operating system or the like operating on the computer based on the instructions of the program code, thereby realizing the functions of any of the above embodiments.

[0116] Further, it is understood that the programs read out from the storage medium can be written to the memories of an expansion board inserted into the computer or an expansion unit connected to the computer, and then, on the basis of instructions of the programs, the CPU or the like mounted on the expansion board or the expansion unit can execute some or all of the actual operations, thereby implementing the functions of any of the above-described embodiments.

[0117] The above-described embodiments are merely preferred embodiments of the present application and are not intended to limit the protection scope of the present application. Equivalent replacements or modifications made by those skilled in the art based on the present application fall within the protection scope of the present application. The protection scope of the present application is defined by the claims.

Claims

1. An improved context aware adaptive offset decoding method, characterized by The method comprises the following steps: The DC-ACD method is used to manage the dynamic context weight of the relevant search content obtained by the retrieval enhancement generation RAG framework: Step 1: A dynamic weight calculation model of context semantic correlation is established, and the dynamic weight calculation model of context semantic correlation is used to ensure that the search content with higher semantic similarity is assigned with higher weight; Step 2: Multi-document fusion is performed: the first k documents of the relevant search content are used to perform multi-document fusion according to the relevance score, and the multi-document information can make up for the incomplete context; Step 3: The context quality is evaluated: a context quality evaluation network is introduced to evaluate the context quality of the search content, so that the context with higher quality obtains greater weight; Step 4: The dynamic weight of the Token level is calculated according to whether the prior knowledge of the intelligent language model and / or the context information is relied on; Step 5: The final decoding probability distribution is analyzed: the final decoding probability distribution is obtained by comprehensively considering the weight proportions of the parts.

2. The improved context-aware adaptive contrast decoding method according to claim 1, characterized by the dynamic weight calculation model for establishing the context semantic correlation in step 1, the formula being as follows: α DC-ACD (t) = λ · α ACD (t) + (1 - λ) · sim(q, c) wherein: a DC-ACD (t) denotes the dynamic weight at time t, λ · a 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 a balance factor tuned by the validation set.

3. The improved context aware adaptive bit-plane decoding method according to claim 1, characterized in that In step 2 the formula is used: Multi-document fusion is performed, wherein: is the fused context logit, w i denotes the weight of the i-th document, is the logit based on the i-th document.

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

5. The improved context aware adaptive bit-plane decoding method according to claim 1, characterized in that The step 4 uses the formula: a token (t) = f(e t , h t , c) Compute dynamic weights at 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-adaptive entropy decoding method according to claim 1, wherein In step 5 the formula: obtaining a final decoding probability distribution, where z t is the logit output of the original intelligent language model, representing the model's prediction based on parametric knowledge, a token (t) is a dynamic weight at the token level, a DC-ACD (t) is a dynamically computed contextual weight, is the fused context logit, represents the additional contribution of the context information to the prior knowledge of the intelligent language model.

7. An improved context aware adaptive bit-plane decoding apparatus characterized by The decoding management module comprises a DC-ACD method, The DC-ACD method is used to manage the dynamic context weight of the relevant search content obtained by the retrieval enhancement generation RAG framework: Step 1: A dynamic weight calculation model of context semantic correlation is established, and the dynamic weight calculation model of context semantic correlation is used to ensure that the search content with higher semantic similarity is assigned with higher weight; Step 2: Multi-document fusion is performed: the first k documents of the relevant search content are used to perform multi-document fusion according to the relevance score, and the multi-document information can make up for the incomplete context; Step 3: The context quality is evaluated: a context quality evaluation network is introduced to evaluate the context quality of the search content, so that the context with higher quality obtains greater weight; Step 4: The dynamic weight of the Token level is calculated according to whether the prior knowledge of the intelligent language model and / or the context information is relied on; Step 5: The final decoding probability distribution is analyzed: the final decoding probability distribution is obtained by comprehensively considering the weight proportions of the parts.

8. A computer readable medium characterized by The computer readable medium stores computer instructions, and the computer instructions make the processor execute the improved context-aware adaptive contrast decoding method in any one of claims 1 to 6 when executed by the processor.

Images