A strong correlation control method for long text streaming conversion

Abstract

The present application provides a strong correlation control method for long text streaming conversion, which belongs to the field of long text processing and solves the problems of uncontrollable correlation, semantic deviation and loss of key information in long text conversion. The method includes: receiving streaming input fragments, calculating multi-dimensional dynamic strong correlation relationships including semantic retention, structural correspondence, key information extraction, and style consistency; generating key prompt words based on this, and forcing the output to focus on the core elements through dynamic weight adjustment and large model generation probability correction; when the correlation is insufficient, weight enhancement and keyword supplementation form a feedback loop. Prompt words are screened based on the weak correlation dimensions, and the pre-trained model and domain vocabulary are combined to quantitatively evaluate each dimension. The solution realizes precise correlation control of input and output, and is suitable for scenarios such as summary generation, structured extraction, style transfer, etc., and effectively improves the accuracy of semantic retention, structural alignment and multi-task adaptability.

Claims

1. A strong correlation control method for long text streaming conversion, characterized in that: The following steps are involved: Receive streaming input fragments of long text; Calculating a dynamic strong correlation between the streaming input segment and the corresponding converted output segment, wherein the dynamic strong correlation is a multi-dimensional quantized correlation strength vector; Generate a key prompt word set according to the dynamic strong association relationship, wherein the key prompt word set is used to constrain the large model to generate the conversion output segment; Correcting the generation probability distribution of the large model by using the key prompt word set to obtain a constrained conversion output segment; The dynamic strong association relationship and the key prompt word set are adjusted based on the feedback of the constrained converted output segment until a preset association threshold is met.

2. The strong correlation control method for long text streaming conversion according to claim 1 is characterized in that: The multiple dimensions of the dynamic strong correlation relationship include at least one of semantic retention, structural correspondence, key information extraction and style consistency.

3. The strong correlation control method for long text streaming conversion according to claim 2 is characterized in that: The semantic retention is obtained by calculating the cosine similarity of the semantic embedding vectors of the streaming input segment and the converted output segment using a pre-trained semantic model; the structural correspondence is obtained by calculating the cosine similarity of the semantic embedding vectors of the structural partitioned segments of the input and output; The key information extraction degree is obtained by the input and output co-occurrence ratio of elements in the key information library; the style consistency is obtained by calculating the style matching probability of the converted output segment through a style classification model.

4. The strong correlation control method for long text streaming conversion according to claim 2 is characterized in that: The step of generating a key prompt word set according to the dynamic strong association relationship includes: Extracting a candidate word set from the streaming input segment and the historical context window; Determining a weak dimension in the dynamic strong correlation relationship, and screening candidate words that are strongly correlated with the weak dimension; The weights of the filtered candidate words are dynamically adjusted according to the current scores of the weak dimensions to obtain the key prompt word set.

5. The strong correlation control method for long text streaming conversion according to claim 4 is characterized in that: The formula for dynamically adjusting the candidate word weight is: ， in, is the initial weight of the candidate word, is the adjustment coefficient, is the current score of the weak dimension.

6. The strong correlation control method for long text streaming conversion according to claim 1 is characterized in that: The formula for generating the probability distribution of the modified large model is: ,in, is the key prompt word set, is the weight of the prompt word, is the indicator function, when k appears in 1 if it is in, 0 otherwise. is the generation probability after the prompt word constraint, is the conversion output fragment at time t; is the streaming input segment at time t; T is the historical context sequence containing input-output pairs before time t; is the set of key prompt words at time t.

7. The strong correlation control method for long text streaming conversion according to claim 1 is characterized in that: The feedback adjustment step includes: If the overall strength of the dynamic strong correlation relationship is lower than a preset correlation threshold, the weight of the prompt word that does not appear in the converted output segment is increased; New prompt words related to the current weak dimension are added from the streaming input segment to update the key prompt word set.

8. The strong correlation control method for long text streaming conversion according to claim 3 is characterized in that: The key information library contains preset entities, values, and time information, and the calculation of the key information extraction degree gives priority to matching elements in the key information library.

9. The strong correlation control method for long text streaming conversion according to claim 3 is characterized in that: The style classification model is a pre-trained text style recognition model, and the calculation of style consistency is based on the posterior probability of the target style label.

10. The strong correlation control method for long text streaming conversion according to claim 4 is characterized in that: The historical context window is maintained by a sliding window mechanism and includes a recent preset number of streaming input segments and corresponding converted output segments.

Images