A method and system for SEO content detection and publishing gating based on information gain

Abstract

This invention discloses an SEO content detection and publishing gating method and system based on information gain. The method includes: for a target query, performing paragraph-level opinion sentence decomposition on collected search engine result page documents and candidate documents to construct a consensus semantic field; semantically aligning and filtering the opinion sentences of candidate documents with the consensus semantic field; performing multi-source evidence verification and reliability aggregation on the filtered effective new opinion sentences to obtain a macro-level semantic novelty measure of the candidate documents; constructing a consensus knowledge graph to identify structural holes and calculate structural hole gain; performing joint scoring and publishing decisions; and outputting a structured evidence package. This invention quantifies the evaluation of novelty and information gain by constructing a ranking-weighted consensus semantic field; ensures the reliability of the measured gain through evidence-conditional macro-divergence calculation and opinion sentence-level multi-source verification; and generates a structured evidence package organized by opinion sentences, making the decision verifiable and auditable.

Description

Technical Field

[0001] This invention relates to the field of information retrieval technology, specifically to an SEO content detection and publishing gating method and system based on information gain. Background Technology

[0002] In the current digital content ecosystem, search engine optimization (SEO) is a key means to improve the visibility of online content. Traditional content production and SEO strategies mainly revolve around technical indicators such as keyword density, backlinks, and page metadata. However, with the continuous evolution of search engine algorithms, simple keyword matching and templated content are no longer sufficient to achieve good search rankings. The widespread adoption of large language models has made it easier to mass-produce text highly relevant to queries; however, this also brings new challenges: the generated content is often highly homogenized with existing search results, lacking genuine information increment, or introducing unverified or even false information in pursuit of novelty. Existing content quality assessment systems mostly focus on grammar checks, plagiarism detection, or basic fact-checking, lacking a systematic method to quantify the information gain of content relative to the consensus of current search engine results pages, while simultaneously ensuring its factual reliability and logical value. Specifically, existing technologies have the following problems:

[0003] The inability to dynamically construct and quantify current search engine consensus (including macro-level topic distribution and micro-level specific viewpoints) for a specific query leads to a lack of accurate benchmarks for novelty assessment. Existing textual dissimilarity measures (such as simple vector distance and traditional topic model dissimilarity) are easily interfered with by irrelevant text, padding, or factually incorrect novel statements, failing to effectively distinguish between valuable incremental changes and speculative or low-quality differences. Fact-checking is often treated as a separate step, not deeply coupled with the informational value of the content (such as whether it fills logical gaps in the knowledge graph); the system struggles to determine whether a new viewpoint is an important discovery or a trivial detail. Summary of the Invention

[0004] The purpose of this invention is to provide an SEO content detection and publishing gating method and system based on information gain, so as to solve the problems mentioned in the background art, such as the lack of fine-grained consensus modeling, unreliable information gain measurement, and the disconnect between verification and value assessment.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] According to one aspect of the present invention, an SEO content detection and publishing gating method based on information gain is provided, the method comprising:

[0007] For the target query, collect the search engine result page documents, perform content cleaning and paragraph-level opinion sentence decomposition on the search engine result page documents and candidate documents, and construct a consensus semantic field that includes consensus semantic distribution and consensus opinion sentence library;

[0008] The opinion sentences of the candidate documents are semantically aligned and filtered with the consensus semantic field. The filtered effective new opinion sentences are verified by multiple sources and their reliability is aggregated to obtain the macro-semantic novelty measure of the candidate documents.

[0009] A consensus knowledge graph is constructed based on the consensus opinion sentence library to identify structural holes, and the structural hole gain generated by the candidate documents in effectively filling the structural holes is calculated.

[0010] By integrating the macro-semantic novelty metric, the structural hole gain, and the semantic entropy of the candidate documents, a joint scoring and publication decision is made, and a structured evidence package is output.

[0011] Based on the aforementioned scheme, the consensus semantic field is constructed by: assigning attenuation weights based on search ranking to the documents on the search engine results pages; performing a weighted average of the semantic concept distributions of the documents on each search engine results page based on the attenuation weights to obtain the consensus semantic distribution; and collecting the opinion sentences of the documents on each search engine results page and adding weights to form the consensus opinion sentence library.

[0012] Based on the aforementioned scheme, the paragraph-level opinion sentence decomposition includes generating standardized candidate opinion sentences from the paragraph text using a large language model with fine-tuned instructions; verifying the structural integrity of the candidate opinion sentences through dependency parsing; and performing semantic fidelity filtering by calculating the natural language inference implications score between the candidate opinion sentences and the original text.

[0013] Based on the aforementioned scheme, the semantic alignment and filtering includes calculating the relevance between candidate opinion sentences and the query intent, filtering off-topic opinions; calculating the semantic similarity between candidate opinion sentences and sentences in the consensus opinion sentence library, classifying opinion sentences into redundant opinions, conflicting candidates, or new candidates according to preset redundancy thresholds and conflict candidate thresholds; and calculating residual novelty for new candidate opinion sentences based on the semantic similarity.

[0014] Based on the aforementioned scheme, the multi-source evidence verification includes retrieving evidence fragments from multiple evidence sources that support or refute the viewpoint statement; using a natural language inference model to calculate the probability of each piece of evidence supporting, remaining neutral, or contradicting the viewpoint statement; and calculating the comprehensive weight of the evidence based on the authority level of the source, timeliness, and relevance to the viewpoint statement.

[0015] Based on the aforementioned scheme, the macro-semantic novelty measure is the evidence-conditional Jensen-Shannon divergence, calculated using methods including only employing the semantic novelty measure.Figure 1 A set of valid new opinion sentences that are consistent, non-redundant, non-hard-conflict, and meet reliability standards is obtained; based on the semantic concept distribution and reliability score of the valid new opinion sentences, the conditional semantic distribution of candidate documents is reconstructed; the Jensen-Shannon divergence between the conditional semantic distribution and the consensus semantic distribution is calculated.

[0016] Based on the aforementioned scheme, the structural hole gain is calculated by identifying node pairs in the consensus knowledge graph that do not have explicit links but have high link prediction scores as structural holes; assigning weights to each structural hole; when the opinion sentence in a candidate document can connect the node pairs of the structural hole with a relationship verified by the multi-source evidence, it is determined to be a valid filling, and the weights of the filled structural holes are accumulated to calculate the gain.

[0017] Based on the aforementioned scheme, the calculation model for the joint score is as follows:

[0018] ;

[0019] in, Candidate documents Overall score For structural hole gain, For macro-semantic novelty measurement, The average effective information gain of opinion sentences within candidate documents. Let be the semantic entropy, and γ, α, β, δ be configurable weight coefficients.

[0020] Based on the aforementioned scheme, the release decision includes approval, rewriting, or rejection; the structured evidence package is organized by opinion sentence and includes opinion sentence text, alignment status, matching consensus sentence, evidence chain details, reliability indicators, structural hole associations, and triggering decision rules.

[0021] According to another aspect of the present invention, an SEO content detection and publishing gating system based on information gain is provided. The system includes: a data acquisition and preprocessing module, an opinion sentence decomposition and semantic representation module, a consensus semantic field modeling module, a multi-source evidence verification and reliability assessment module, a core gain calculation module, and a gating decision and feedback module.

[0022] The data acquisition and preprocessing module is used to acquire SERP documents and receive candidate documents, and to perform content cleaning.

[0023] The opinion sentence decomposition and semantic representation module is used to perform paragraph-level opinion sentence decomposition and map opinion sentences into semantic vectors.

[0024] The consensus semantic field modeling module is used to construct a ranking-weighted consensus semantic distribution and a consensus opinion sentence library;

[0025] The multi-source evidence verification and reliability assessment module is used to retrieve evidence, assess logical consistency, and calculate the reliability score of the opinion statement.

[0026] The core gain calculation module is used to calculate macroscopic semantic novelty measure, structural hole gain, and semantic entropy;

[0027] The gated decision and feedback module is used to perform joint scoring, output release decisions, and generate structured evidence packages.

[0028] To make the description of this invention clearer, some key terms and concepts used in this specification and claims are explained as follows:

[0029] Regarding "consensus," for a specific query, the collective knowledge state presented by the current search engine results page is not a single viewpoint, but a structured and computable data object constructed through the method of this invention. The consensus semantic field includes a macro-consensus semantic distribution and a micro-consensus viewpoint sentence library. The macro-consensus semantic distribution is a probability distribution in a preset semantic concept space obtained by weighted aggregation of search engine result page (SERP) document rankings. The micro-consensus viewpoint sentence library is a set of standardized viewpoint sentences extracted from SERP documents and assigned weights, along with their weights. "Consensus" is dynamic, query-related, and serves as a benchmark for evaluating any candidate content.

[0030] Information gain is used to quantify the added value of candidate documents relative to the consensus mentioned above; it includes macro-semantic novelty gain, structural hole gain, and micro-effective information gain; among which, macro-semantic novelty gain is measured by the distribution divergence through evidence conditionation, structural hole gain is the value brought by filling logical gaps in the consensus knowledge graph, and micro-effective information gain is a fine-grained aggregation based on the novelty and reliability of opinion sentences.

[0031] Regarding SEO content detection, the core of this invention lies in assessing whether content is relevant to the target query and provides both authentic and reliable information with added value, thus meeting the core ranking principles of modern search engines for high-quality content. Regarding publication gating decisions, this refers to automated and interpretable publication control decisions made on candidate content based on gain measurement, reliability verification, and preset rules; further, it includes a structured evidence package. Regarding structural holes, these refer to node pairs in a knowledge graph built based on consensus content that are not directly connected but should have strong semantic relationships. These missing links represent potential logical relationships or knowledge gaps not yet covered by the current SERP content; filling high-weight structural holes is considered a core manifestation of providing higher-order information gain.

[0032] As can be seen from the above technical solution, this invention has at least the following advantages and positive effects compared with the prior art: By constructing a ranking-weighted consensus semantic field, the ambiguous SEO competitive environment is transformed into a precise mathematical representation containing macro-distribution and micro-viewpoint databases, making the evaluation of novelty and gain have an objective, dynamic, and quantifiable benchmark. Through evidence-conditional macro-divergence calculation and opinion sentence-level multi-source verification, the measurement of information gain is strongly bound to the support of external authoritative evidence from a mechanism perspective; fundamentally suppressing speculative behavior of falsely inflating novelty scores by piling up unreliable information or marginal off-topic content, ensuring that the measured gain is a credible gain. Introducing graph-based structural hole identification and filling gain calculation, the system can go beyond keyword coverage and semantic similarity, evaluate the value of content from the level of knowledge structure and logical connection, and discover higher-order knowledge value. By generating structured evidence packages organized by opinion sentences, the overall decision is decomposed into a transparent display of the alignment status, evidence support, conflict situation, and higher-order value contribution of each core claim; this not only makes the pass / rejection decision verifiable and auditable, but also provides clear and fine-grained action guidelines for content optimization.

[0033] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0034] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort. In the drawings:

[0035] Figure 1 This is a schematic diagram of an SEO content detection and publishing gating method based on information gain according to the present invention;

[0036] Figure 2 This is a schematic diagram of the process for determining the filling of holes in the atlas structure according to the present invention;

[0037] Figure 3 This is a schematic diagram of an SEO content detection and publishing gating system based on information gain according to the present invention. Detailed Implementation

[0038] To more clearly illustrate the purpose, technical solutions, and advantages of 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. The exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein. On the contrary, these embodiments are provided so that the present invention will be more comprehensive and complete, and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0039] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a full understanding of embodiments of the invention. However, those skilled in the art will recognize that the technical solutions of the invention can be practiced without one or more of the specific details, or other methods, components, apparatuses, steps, etc., can be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of the invention.

[0040] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0041] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0042] The present invention will now be described in detail with reference to specific embodiments.

[0043] Example 1

[0044] like Figure 1 As shown in the figure, this embodiment provides an SEO content detection and publishing gating method based on information gain. The specific steps of the method are as follows:

[0045] S1: For the target query, collect the search engine result page documents, perform content cleaning and paragraph-level opinion sentence decomposition on the search engine result page documents and candidate documents, and construct a consensus semantic field containing consensus semantic distribution and consensus opinion sentence library.

[0046] Upon receiving the user's target query q, the system automatically triggers the data collection process from the search engine results page (SERP) for that query; it obtains structured search results through web crawlers or legitimate API interfaces provided by the search engine; and it collects the top N web page documents from the search engine results page (SERP) to form the initial SERP document set. The number of documents to be collected, N, is a configurable parameter that can be dynamically adjusted based on system computing power, network stability, and task requirements. Each document is recorded synchronously during the collection process. Ranking position The original URL, page title, summary fragment returned by the search engine, and crawl timestamp.

[0047] For the collected original SERP documents and user-submitted candidate documents Perform unified content cleaning and structuring. Based on HTML tag tree (DOM) traversal, remove... <script>、<style>、<meta>、注释及常见广告与跟踪代码标签，以移除噪声标签。通过统计方法识别跨多页面重复出现的模板区块（如导航栏、页脚、侧边栏、相关推荐模块），以识别并过滤模板化结构；示例性地，计算不同页面间相同DOM路径的文本相似度，若某区块在超过阈值比例的页面中出现且内容高度一致，则将其标记为模板并移除。根据HTML标签语义与文本分布特征，使用如Readability类算法初步定位正文区域；可选地，采用预训练的段落分类模型（如基于BERT的序列分类器）对DOM节点进行二分类（正文 / 非正文），进一步提升提取精度，尤其适用于结构复杂或非标准的网页。对提取出的正文进行编码统一（转换为UTF-8）、冗余空格与换行符清理、以及特殊字符转义处理，确保文本格式一致。根据换行符、段落标签（）及句子结束标点等，将正文分割为逻辑段落序列，并保留标题层级（<h1>-<h6>）信息作为段落重要性权重参考；移除过短段落（如字符数少于阈值）及明显无关内容（如"版权声明”"隐私政策”等通过关键词规则匹配的文本）。对经过数据清洗与结构化的SERP文档集合记为。

[0048] 进一步地，将非结构化的段落文本转化为规范化的观点句集合，以作为适合精准计算的规范化单元，执行段落级观点句分解，采用生成-校验-锚定的混合架构。利用指令微调的大语言模型将段落文本生成规范化的候选观点句；通过依存句法分析对候选观点句进行结构完整性校验；通过计算候选观点句与原文的自然语言推理蕴含分数，进行语义忠实性过滤。具体地，在约束生成阶段，构建结构化提示指令，输入文档段落，驱动经过指令微调的大语言模型识别段落中的核心事实断言；提示指令中包含如"输出标准化陈述句”、"保留量化指标”、"拆解复合逻辑”等明确指令约束，并强制模型以JSON格式输出生成的候选观点句。在句法校验阶段，为抑制模型生成片段化或不完整的短语，将候选观点句输入至轻量级依存句法分析器；定义合法性判定函数：当且仅当该句的依存句法树的根节点为动词或形容词，且存在指向主语nsubj（名词性主语）依存关系以及指向宾语或补语的核心论元（如obj,attr,comp等）依存关系时，判定该句结构合格，予以保留。为确保通过句法校验的观点句在语义上忠实于原文段落，计算该观点句对原段落的语义忠实度。在原意锚定阶段，使用预训练的自然语言推理模型（NLI），以原始段落为前提，以生成的观点句为假设，计算蕴含分数。设定一个语义忠实度阈值（如0.85），仅保留蕴含分数高于该阈值的观点句，记入该段落对应的观点句集合。通过以上三个阶段，每个文档段落的文本被转化为一个规范化的观点句集合；整个文档d的观点句集合为所有段落观点句集合的并集，即；该流程同时且一致地应用于所有SERP文档及待检测的候选文档，以确保后续比对与分析在统一的语义粒度上进行。

[0049] 进一步地，基于清洗和分解后的SERP数据，构建作为基准的共识语义场，该场从宏观统计分布与微观知识单元两个层面，对当前搜索引擎关于查询q的共识进行建模。具体地，为体现搜索结果排序中蕴含的置信度或重要性差异，为每个SERP文档分配一个排名衰减权重。本实施例中，排名衰减权重定义为：

[0050] ；

[0051] 其中，为可配置的衰减系数，用于控制权重随排名下降的速率；为文档在SERP中的排名；将权重归一化：。对于宏观共识语义分布构建，将每个SERP文档映射到一个语义概念分布。对于语义概念分布映射，具体地，为构建可计算的宏观语义分布，需要将每个文档（包括SERP文档及候选文档）表示为在一个预定义的语义概念空间上的概率分布，设存在一个语义概念集合，其可通过领域语料的聚类、主题模型（如LDA / BERTopic）或预训练模型的嵌入空间聚类中心获得；每个概念对应一个具有语义代表性的向量。对于文档中任意一个观点句v，首先获取其语义嵌入向量h(v)。然后，计算该向量与每个概念向量的语义相似度（如余弦相似度），并通过Softmax函数归一化，得到观点句v在概念空间上的软分配分布Q(v)：

[0052] ；

[0053] 其中为温度参数，控制分布的尖锐程度。文档的语义概念分布，由其所有观点句的分布Q(v) 聚合而成。一种基础的聚合方式为平均池化，并可引入段落位置权重：

[0054] ；

[0055] 其中为归一化常数；为防止后续散度计算异常，对进行平滑处理，得到平滑后的文档分布。通过加权平均聚合所有SERP文档的平滑分布，构建全局的宏观共识语义分布：

[0056]

[0057] 该分布反映在当前搜索结果中，不同语义概念的整体加权覆盖情况。对于微观共识观点句库构建，汇集所有SERP文档中分解得到的观点句，形成微观的共识观点句库：

[0058] ；

[0059] 为库中每条共识观点句附加权重，该权重综合了其来源文档的权威性（通过排名权重体现）及其在原文中的重要性（通过段落权重体现）：

[0060] ；

[0061] 其中为观点句v 的来源文档索引，表示其所在的段落，为段落权重函数。

[0062] 通过从原始查询到构建结构化基准，输出结构化的候选文档观点句集合；共识语义场，由宏观的排名加权的共识语义分布、微观的带权重的共识观点句库（）；该共识语义场将作为后续所有新颖性、冗余性、冲突性及结构洞分析的计算基准。

[0063] S2：将所述候选文档的观点句与所述共识语义场进行语义对齐与筛选，对筛选出的有效新增观点句进行多源证据验证与可靠性聚合，得到所述候选文档的宏观语义新颖性度量。

[0064] 将候选文档dc中分解得到的观点句集合与SERP共识观点句库进行细粒度对齐与分类。计算候选观点句与查询的意图相关度，过滤离题观点；计算候选观点句与共识观点句库中句子的语义相似度，根据预设的冗余阈值与冲突候选阈值，将观点句分类为冗余观点、冲突候选或新增候选；基于语义相似度为新增候选观点句计算残差新颖性。

[0065] 具体地，对于每个候选观点句，计算其与查询q的意图相关度，并进行意图一致性过滤：若，则将该观点句标记为离题，并将其排除在后续有效增益计算之外。对于通过意图过滤的观点句，计算其与共识观点句库中每一条观点句的语义相似度sim(v,u)；优选采用基于Sentence-BERT或类似模型的句子嵌入向量的余弦相似度实现；确定最相似共识观点句及其对应的最大相似度得分。

[0066] 其中，语义相似度sim(v,u)的计算，通过将观点句v和u的语义表示在向量空间中进行比对来实现。利用预训练的句子编码模型，将观点句v和共识观点句u分别映射为固定维度的语义向量（即句嵌入），记为h(v)和h(u)；句子编码模型优选为在大规模语料上训练、能捕获深层语义的模型。计算这两个语义向量之间的相似度，一种实施方式采用余弦相似度，其计算如下：

[0067] ；

[0068] 最相似共识观点句及其最大相似度得分可通过遍历计算并取最大值得到：

[0069] ；

[0070] 同理，通过与观点句相似度计算相同的句子编码模型获取查询q和观点句v的语义向量表示，得到查询向量h(q)和观点句向量h(v)；计算两个语义向量之间的相似度：

[0071] ；

[0072] 根据可配置阈值对意图一致的观点句进行分类：1）冗余观点，若（如），则认为该观点已在SERP中被充分覆盖，标记为冗余，其信息增益在后续计算中视为零；2）冲突候选，若（如），则认为该观点与现有共识具有可比性，标记为冲突候选，进入后续冲突验证流程（如多源证据检索以确认是否存在实际矛盾）；3）新增候选，若，则标记为新增候选，进入证据验证流程。被标记为新增候选的观点句即为有效新增观点句，为每个有效新增观点句计算其残差新颖性，该值量化该观点句相对于SERP共识的差异程度，作为其微观层面新颖性的初步度量指标。

[0073] 进一步地，对有效新增观点句进行多源证据验证，包括多证据源检索与逻辑一致性验证。从多个证据源中检索支持或反驳观点句的证据片段；利用自然语言推理模型计算每条证据对观点句的支持、中立或矛盾概率；根据证据的来源权威等级、时效性及与观点句的相关性，计算证据的综合权重。

[0074] 具体地，从候选观点句v中提取核心语义成分，包括主体实体、属性断言及量化修饰，自动构造结构化的布尔检索式或自然语言查询请求。在多个证据源中并发执行检索操作，如权威外部知识源、内部知识库、SERP内容对照集（可作为参照基准）等；对每个证据源返回的候选片段根据其与查询的语义相关性得分进行排序，选取每个证据源中排名前K个片段（K可配置），构成初始证据片段集合；对每条证据，校验其发布时间戳是否超过预设的有效期窗口，是则过滤；通过时效性校验的证据片段构成有效证据集合。使用预训练的自然语言推理模型，对每一组证据片段-观点句对(e,v)进行逻辑关系判定。该模型以前述证据片段e作为前提，以候选观点句v作为假设，输出一个三分类的概率分布，用以量化e对v的逻辑支持程度：

[0075] ；

[0076] 其中，表示证据e支持观点v的概率；表示证据e与观点v为中立（即不相关或无法推断）的概率；表示证据e反驳或矛盾于观点v的概率；。自然语言推理模型可优选采用大规模蕴含数据集上微调的RoBERTa-large或DeBERTa-v3-large架构。为每条证据e定义三维权重因子：来源权威等级，根据预设的权威性映射表赋值，时效衰减因子，为证据发布时间与当前时间间隔，为衰减系数；语义相关性，基于句子嵌入向量计算；计算每条证据的原始权重，归一化得到证据的综合权重。

[0077] 基于多源证据的评估结果，计算每个候选观点句（即上述有效新增观点句）的可靠性分数。对于每条候选观点句v，基于其关联的有效证据集合，通过加权聚合计算其整体支持度与冲突度：

[0078] ；

[0079] ；

[0080] 其中，为证据聚合支持度，为证据聚合冲突度，为证据e的归一化权重（综合考虑了来源权威性、时效性及与观点句的相关性）。

[0081] 为评估候选文档内容生成的不确定性与稳定性，计算其语义熵；以相同的原始指令或上下文，通过改变随机种子、解码参数或提示变体等方式，独立生成M个候选文档的变体；利用自然语言推理模型或语义相似度计算，将所有M个生成变体进行两两比较，将其划分为个语义等价簇，同一簇内的变体在核心主张和逻辑上高度一致。计算每个语义等价簇的概率；为第k个簇中的变体数量；计算归一化语义熵：

[0082] ；

[0083] 趋近于0表示生成结果高度一致、确定性高；趋近于1表示生成结果极为分散、不确定性高，属于高风险内容。

[0084] 结合支持度、冲突度、证据覆盖率以及语义熵，通过可调节的Sigmoid函数计算最终可靠性分数：

[0085] ；

[0086] 其中，为Sigmoid函数，为缩放因子，用于调整Sigmoid函数输入值的尺度；为可配置权重系数，用于控制各项的影响程度；为高置信证据覆盖率（可选指标），例如定义为，用于奖励有强证据支撑的观点；为候选文档的语义熵，作为对内容不确定性的全局惩罚项。

[0087] 采用证据条件化的Jensen-Shannon散度（EC-JSD）进行宏观新颖性度量，避免堆砌无关内容或不可靠内容虚假提升新颖性的行为。仅采用通过意图一致、非冗余、非硬冲突且可靠性达标的有效新增观点句集合；基于有效新增观点句的语义概念分布及其可靠性分数，重建候选文档的条件化语义分布；计算条件化语义分布与共识语义分布之间的Jensen-Shannon散度。

[0088] 具体地，构建有效新增观点句集合，筛选同时满足以下所有条件的观点句参与宏观增益计算，构成有效新增观点句集合，筛选条件：意图一致，；非冗余：；可靠性达标：；无非验证硬冲突，未触发"相似度高且冲突度高”的硬冲突规则。基于集合中的观点句的语义概念分布Q(v)及其可靠性分数R(v)，重建候选文档的语义概念分布：

[0089] ；

[0090] 其中，为观点句v的语义概念分布，为可选的段落或位置权重，为归一化常数；对进行平滑处理得到。计算平滑后的条件化候选分布与SERP共识语义分布之间的Jensen-Shannon散度：

[0091] ；

[0092] 其中，为两个分布的中间分布，为Kullback-Leibler散度，为候选文档相对于共识语义分布的宏观语义新颖性度量。宏观语义新颖性仅由那些意图相关、非冗余、可靠且无冲突的有效新增观点驱动，从机制上抑制了低质或投机内容对增益计算的干扰。

[0093] 步骤S2输出结构化中间数据，包括分类后的候选观点句列表（含冗余、冲突、新增标签），每个新增观点句的残差新颖性Δ(v)、可靠性 R(v)、证据片段集及支持 / 冲突度，候选文档的宏观EC-JSD值，有效新增观点句集合。

[0094] S3：基于所述共识观点句库构建共识知识图谱以识别结构洞，并计算所述候选文档的通过有效填补所述结构洞所产生的结构洞增益。

[0095] 结合图2，将非结构化的文本共识转化为结构化的知识图谱表示；共识知识图谱的构建，其数据源可直接来源于前述步骤构建的共识观点句库中的规范化断言；亦可从SERP文档集合重新进行深度分析以获得更丰富的实体关系。

[0096] 具体地，一种实施方式中，对共识观点句库中的每条观点句v，根据其规范化表达的句法结构，通过预定义的规则模板或轻量级序列标注模型，将其解析为形式化的三元组(h,r,t)。其中，h表示头实体，r表示关系谓词，t表示尾实体或属性值。对解析得到的所有三元组中的实体提及进行消歧与归一化处理，将指向现实世界同一对象的多个提及（如"TensorFlow”、"TF框架”）链接至唯一的实体标识符，形成全局一致的实体节点集合；将所有归一化后的三元组作为边加入图谱，构建共识知识图谱。每条边可继承其来源观点句v的权重，并可附加来源文档、出现频次等元数据。

[0097] 在另一种实施方式中，对SERP文档集合中的每篇文档进行句子级分析，包括命名实体识别、开放域关系抽取。对于命名实体识别，采用基于Transformer架构的预训练模型，识别文本中的实体提及，其类型可包括技术术语、产品型号、组织机构、人物、地点、时间、量化指标与数值等。对于开放域关系抽取，利用基于预训练语言模型的关系抽取方法，从包含实体的句子中抽取出以三元组形式表示的结构化断言，即（头实体，关系，尾实体）；关系类型包括但不限于因果（导致、促进）、比较（优于、类似于）、属性（具有、尺寸为）、包含（部分为、隶属于）、时序步骤（之后、第一步）等。对不同文档中识别出的实体提及进行消歧与链接，将指向现实世界同一对象的多个提及归一化至一个标准的实体标识符；例如，通过实体链接技术将"TensorFlow”、"TF”和"GoogleTensorFlow”统一映射到标准实体"TensorFlow（机器学习框架）”；形成全局唯一的实体节点集合。基于归一化后的实体集合，将步骤S1抽取的所有关系三元组合并，构建统一的共识知识图谱，其中节点集为实体集合，边集为已识别的关系集合，每条边连接两个实体节点，并标注具体的关系类型；每条边（关系）可附加丰富的元数据属性，例如：支撑该关系出现的来源文档列表、在所有SERP文档中的共现频次、以及首次 / 末次出现的时间戳等。

[0098] 对共识知识图谱进行图表示学习，得到其低维向量表示，并定义链接预测评分函数。具体地，采用图表示学习，如图神经网络模型或知识图谱嵌入模型对图谱进行训练，为每个实体节点学习一个低维稠密向量表示。定义链接预测评分函数s(u,v)∈(0,1]，用于量化图中任意两个节点u与v之间存在直接语义关联（即属于一个未被显式抽取出的关系三元组）的预测概率。该评分函数不区分具体关系类型，评估链接存在的总体置信度，可采用以下任一形式：基于距离的评分函数：

[0099] ；

[0100] 其中，定义距离：，r为关系向量；为尺度参数，用于保证分数落在(0,1]且具备可用的阈值区间；该式衡量节点嵌入在向量空间中的接近程度，距离越小评分越大；

[0101] 双线性评分函数：

[0102] ；

[0103] 通过一个可学习的参数矩阵捕获节点对间的交互；其中，为节点u和v的d维嵌入向量；W为可学习矩阵，为Sigmoid函数，将评分映射至(0,1)区间；

[0104] 神经网络评分函数：

[0105] ；

[0106] 使用神经网络建模节点对之间的复杂关联模式；其中MLP为多层感知机，concat表示向量拼接，通过Sigmoid函数生成评分。

[0107] 基于链接预测评分，识别共识图谱中有待填补的高价值空白，即结构洞。具体地，将结构洞定义为在现有图谱中不存在显式关系边，但链接预测评分较高的节点对，表明二者间存在未被当前SERP内容覆盖但具有潜在高关联价值的语义缺口，其表达式为：

[0108] ；

[0109] 其中为可配置的置信度阈值，用于筛选出预测置信度足够高的潜在链接作为高价值结构洞候选。为每个结构洞候选分配权重，以量化其潜在的填补价值。该权重的计算综合考虑：预测置信度，链接预测评分反映该潜在链接的强度，置信度越高，权重越大；节点重要性，基于图谱中心性指标衡量节点u和v自身在图谱中的重要性，重要节点间缺失的链接通常价值更高；查询相关性，节点对与原始查询q的语义相关性，相关性越高，意味着填补该缺口对满足用户查询意图的贡献可能越大；信息稀缺性（可选的强化因子），引入类似逆文档频率思想，对在SERP文档集合中被较少共同提及的节点对赋予更高权重，以鼓励系统发现非显而易见、具备信息增量的新颖关联，避免偏向于频繁出现的显性主题。在一种权重分配的实施方式中，可计算为：

[0110] ；

[0111] 其中，IDFpair(u,v)为节点对的逆共现频率，例如定义为log(Nnu,v+1)，N为SERP文档总数，nu,v为节点u和v共同出现在同一文档中的次数；η(u,v)为融合了节点重要性与查询相关性的因子。

[0112] 进一步地，判断候选文档dc是否有效填补识别出的结构洞，并计算其带来的结构洞增益。对候选文档dc进行与前述相同的实体识别与关系抽取，得到其新增关系集合，该集合的元素通常以（头实体，关系，尾实体）三元组形式表示。对于每个结构洞进行检查以判定是否被有效填补，包括结构匹配检查、证据支持验证。对于结构匹配检查，检查中是否存在一条关系r能够直接连接节点u和v，即存在形如(u,r,v)或(v,r,u)的三元组，可选地，允许关系 r 在语义上等同于或高度相似于图谱中预测的隐含关系类型，而无需严格字面匹配。对于证据支持验证，对通过结构匹配检查的候选关系，必须确保在候选文档dc中表达该特定填补关系的观点句通过了前述步骤S2的多源证据验证，即该观点句的可靠性分数且证据聚合支持度。当且仅当结构洞(u,v)同时满足结构匹配与证据支持两个条件时，才被判定为被候选文档有效填补，否则判定为未填补；将此判定结果记为指示函数，否则为0。

[0113] 量化候选文档填补结构洞所带来的知识增量价值，对所有被有效填补的结构洞的权重进行求和，得到基础增益值：

[0114] ；

[0115] 其中为示性函数。在一种实施方式中，为避免增益值因候选文档篇幅或图谱中结构洞总数等外部因素而产生偏置，可对基础增益进行归一化处理，使其范围稳定在[0,1)区间；计算表达式为：

[0116] ；

[0117] 其中，为极小的正常数，以防止分母为零。在另一种实施方式中，为鼓励内容生产聚焦于填补最高价值的逻辑缺口，而非追求填补数量，可仅计算对权重最高的前K个结构洞（记为集合）的填补贡献：

[0118] ；

[0119] 其中，表示按权重降序排列后取前K个结构洞构成的子集。可选择输出、或中的任意一项或多项，作为最终用于门控决策的结构洞增益指标。

[0120] S4：融合所述宏观语义新颖性度量、所述结构洞增益以及所述候选文档的语义熵，进行联合评分与发布决策，并输出结构化证据包。

[0121] 候选文档的综合评分由以下四项指标加权合成，即联合评分的计算模型为：

[0122] ；

[0123] 其中，为结构洞增益，可根据实施需求选择使用基础增益、归一化增益或Top-K增益中的一种；为证据条件化的Jensen-Shannon散度（步骤S2）；为文档级平均有效信息增益，即文档中所有观点句的平均有效信息增益：，为单条观点句的有效信息增益，定义为：，其中为有效性指示函数，当观点句v同时满足：意图一致（）、非冗余（）、未触发硬冲突规则时，取值为1，否则为0；为观点句的残差新颖性，为观点句的可靠性分数（步骤S2）；为候选文档的语义熵（步骤S2）；γ,α,β,δ≥0为各项指标对应的可配置的权重系数，优选设置γ≥max(α,β)以体现对填补高价值结构洞的核心激励。

[0124] 根据综合评分与预设规则，输出明确的发布决策标签（PASS / REWRITE / REJECT）及依据。具体地，对于驳回（REJECT）判定规则，若候选文档触发以下任一条件，则判定为驳回发布：1）离题，文档整体与查询的意图相关性低于阈值，即；2）硬冲突，存在至少一个观点句v，其与SERP共识的相似度且证据聚合冲突度；3）缺乏核心价值，结构洞增益低于要求，即（或其选用的变体低于对应阈值）；4）质量过低，文档综合评分未达到发布要求，即。

[0125] 对于重写（REWRITE）判定规则，若候选文档未触发驳回规则，但存在以下可优化情形，则判定为建议重写：1）需补充验证，存在观点句与共识差异大（残差新颖性Δ(v) 高），但其证据支持度Sup(v) 处于中等水平且未构成硬冲突，或可靠性分数R(v) 处于临界区间（证据不足但未被证伪），将提示补充权威证据或调整表述以提升可靠性；2）需强化结构填补，文档整体未离题且无硬冲突，但结构洞增益低于预设的优秀阈值（该阈值高于驳回阈值），可提示围绕若干高权重、未被填补的结构洞进行内容补充。

[0126] 对于通过（PASS）判定规则，需同时满足以下所有条件：1）无离题，文档与查询意图相关，即；2）无硬冲突，不存在任何满足硬冲突条件的观点句；3）结构洞价值达标，结构洞增益满足核心价值要求，即（或其变体达到对应阈值）；4）综合评分达标，文档综合评分达到发布门槛，即；5）所有被判定为有效新增（）或用于填补核心结构洞的观点句，其可靠性分数均满足最低要求，即。

[0127] 进一步地，输出结构化证据包，以确保决策过程透明、可审计。证据包以候选文档的观点句为基本组织单元，对于每条观点句v，生成一条结构化记录，每条记录包含：观点句文本，原始文本及其规范化表达；对齐状态，根据与SERP共识的比对结果，标记为"冗余”、"冲突候选”、"新增候选”或"离题”；共识匹配，当观点句状态为"冗余”、"冲突候选”或"新增候选”时，提供其最相似的SERP共识观点句的文本及语义相似度；证据链详情，该观点句检索到的主要证据片段列表，每条证据注明来源、发布时间、权威等级、归一化权重以及自然语言推理模型判定的支持 / 中立 / 矛盾概率（）；可靠性指标，该观点句的支持度Sup(v)、冲突度Con(v)、证据覆盖率及可靠性分数R(v)；结构洞关联，若该观点句用于填补结构洞，则记录所填补的结构洞(u,v)、其预测评分s(u,v)、权重及填补状态（如"有效填补”、"匹配但证据不足”）；触发规则，明确列出导致当前状态所触发的具体规则条目。

[0128] 在证据包末尾提供全局摘要信息，包括：最终决策（PASS / REJECT / REWRITE）；各项核心指标值，包括：宏观语义新颖性度量、结构洞增益（或或）、平均有效增益、语义熵、综合评分；关键决策依据的文本简述。

[0129] 实施例2

[0130] 如图3所示，本实施例示例性的呈现了一种基于信息增益的SEO内容检测与发布门控系统，包括数据采集与预处理模块、观点句分解与语义表示模块、共识语义场建模模块、多源证据校验与可靠性评估模块、核心增益计算模块、门控决策与反馈模块；

[0131] 数据采集与预处理模块，用于采集SERP文档及接收候选文档，并进行内容清洗。作为系统的输入接口，负责接收用户查询，并通过内置网络爬虫或搜索引擎API自动采集相应的SERP文档。同时，该模块集成去噪、正文抽取与段落分割算法，对原始网页及用户提交的候选文档进行统一清洗与结构化处理。

[0132] 观点句分解与语义表示模块，用于执行段落级观点句分解，并将观点句映射为语义向量。采用生成-校验-锚定的混合策略，将非结构化的段落文本自动分解为规范化的观点句集合；利用预训练的句子嵌入模型将每个观点句映射为高维语义向量，为后续的语义计算提供基础表示。

[0133] 共识语义场建模模块，用于构建排名加权的共识语义分布及共识观点句库。基于预处理后的SERP文档，构建系统进行价值判断的基准；其主要功能包括：为SERP文档分配基于排名的衰减权重，并以此加权聚合各文档的语义概念分布，形成宏观的共识语义分布；同时，合并所有SERP观点句并附加权重，构建微观的共识观点句库。

[0134] 多源证据校验与可靠性评估模块，用于检索证据、评估逻辑一致性并计算观点句可靠性分数。连接内部知识库与外部权威数据源，为候选观点句自动检索相关证据片段；通过集成自然语言推理模型评估证据与观点句的逻辑关系，并结合证据来源、时效等因素计算每条观点句的可靠性分数。

[0135] 核心增益计算模块，用于计算宏观语义新颖性度量、结构洞增益及语义熵。该模块包含三个协同工作的子单元：宏观散度计算单元负责计算证据条件化的语义新颖性度量；结构洞挖掘单元负责在共识知识图谱中识别高价值缺失链接并计算填补增益；语义熵检测单元则负责评估生成内容的不确定性，作为惩罚项。

[0136] 门控决策与反馈模块，用于执行联合评分、输出发布决策并生成结构化证据包。根据预设的阈值规则与联合评分模型，综合各项指标作出通过、重写或驳回的决策；该模块封装所有中间计算结果，生成可追溯、可审计的结构化证据包，并可根据发布后的反馈数据动态优化系统参数。

[0137] 本领域技术人员在考虑说明书及实践这里公开的发明后，将容易想到本发明的其它实施方案。本申请旨在涵盖本发明的任何变型、用途或者适应性变化，这些变型、用途或者适应性变化遵循本发明的一般性原理并包括本发明未公开的本技术领域中的公知常识或惯用技术手段。说明书和实施例仅被视为示例性的，本发明的真正范围和精神由权利要求指出。应当理解的是，本发明并不局限于上面已经描述并在附图中示出的精确结构，并且可以在不脱离其范围进行各种修改和改变。本发明的范围仅由所附的权利要求来限制。< / script>

Claims

1. An SEO content detection and publishing gating method based on information gain, characterized in that, The method includes: For the target query, search engine result page documents are collected. Content cleaning and paragraph-level opinion sentence decomposition are performed on the search engine result page documents and candidate documents to construct a consensus semantic field containing a consensus semantic distribution and a consensus opinion sentence library. Constructing the consensus semantic field includes: assigning decay weights based on search ranking to the search engine result page documents; performing a weighted average of the semantic concept distributions of each search engine result page document based on the decay weights to obtain the consensus semantic distribution; and collecting the opinion sentences from each search engine result page document and adding weights to form the consensus opinion sentence library. The candidate documents' viewpoint sentences are semantically aligned and filtered with the consensus semantic field. The filtered, valid new viewpoint sentences are then subjected to multi-source evidence verification and reliability aggregation to obtain a macro-level semantic novelty measure for the candidate documents. This macro-level semantic novelty measure is the Jensen-Shannon divergence of evidence-conditionalized viewpoint sentences, calculated using only a set of valid new viewpoint sentences that meet the requirements of consistent intent, non-redundancy, non-hard conflict, and reliability. Based on the semantic concept distribution and reliability score of the valid new viewpoint sentences, the conditionalized semantic distribution of the candidate documents is reconstructed. Finally, the Jensen-Shannon divergence between the conditionalized semantic distribution and the consensus semantic distribution is calculated. A consensus knowledge graph is constructed based on the consensus opinion sentence library to identify structural holes, and the structural hole gain generated by the candidate document by effectively filling the structural holes is calculated; the calculation of the structural hole gain includes identifying node pairs in the consensus knowledge graph that do not have explicit links but have high link prediction scores as structural holes; assigning weights to each structural hole; when the opinion sentence in the candidate document can connect the node pairs of the structural hole with the relationship verified by the multi-source evidence, it is determined to be effectively filled, and the weights of the filled structural holes are accumulated to calculate the gain; By integrating the macro-semantic novelty metric, the structural hole gain, and the semantic entropy of the candidate documents, a joint scoring and publication decision is made, and a structured evidence package is output.

2. The SEO content detection and publishing gating method based on information gain according to claim 1, characterized in that, The paragraph-level opinion sentence decomposition includes generating standardized candidate opinion sentences from the paragraph text using a large language model with fine-tuned instructions; verifying the structural integrity of the candidate opinion sentences through dependency parsing; and performing semantic fidelity filtering by calculating the natural language inference implications scores of the candidate opinion sentences and the original text.

3. The information gain based SEO content detection and publishing gating method according to claim 1, wherein, The semantic alignment and filtering include calculating the relevance between candidate opinion sentences and the query intent, filtering off-topic opinions; calculating the semantic similarity between candidate opinion sentences and sentences in the consensus opinion sentence library, and classifying opinion sentences into redundant opinions, conflicting candidates, or new candidates based on preset redundancy thresholds and conflict candidate thresholds. The residual novelty is calculated for newly added candidate opinion sentences based on the semantic similarity.

4. The SEO content detection and publishing gating method based on information gain according to claim 1, characterized in that, The multi-source evidence verification includes retrieving evidence fragments from multiple evidence sources that support or refute the viewpoint statement; using a natural language inference model to calculate the probability of each piece of evidence supporting, remaining neutral, or contradicting the viewpoint statement; and calculating the comprehensive weight of the evidence based on the authority level of the source, its timeliness, and its relevance to the viewpoint statement.

5. The SEO content detection and publishing gating method based on information gain according to claim 1, characterized in that, The calculation model for the joint score is as follows: ； in, Candidate documents Overall score For structural hole gain, For macro-semantic novelty measurement, The average effective information gain of opinion sentences within candidate documents. Let be the semantic entropy, and γ, α, β, δ be configurable weight coefficients.

6. The SEO content detection and publishing gating method based on information gain according to claim 1, characterized in that, The release decision includes approval, rewriting, or rejection; the structured evidence package is organized by opinion sentence and includes opinion sentence text, alignment status, matching consensus sentences, evidence chain details, reliability indicators, structural hole associations, and triggering decision rules.

7. An SEO content detection and publishing gating system based on information gain, used to implement the method as described in any one of claims 1-6, characterized in that, It includes modules for data acquisition and preprocessing, opinion sentence decomposition and semantic representation, consensus semantic field modeling, multi-source evidence verification and reliability assessment, core gain calculation, and gating decision and feedback. The data acquisition and preprocessing module is used to acquire SERP documents and receive candidate documents, and to perform content cleaning. The opinion sentence decomposition and semantic representation module is used to perform paragraph-level opinion sentence decomposition and map opinion sentences into semantic vectors. The consensus semantic field modeling module is used to construct a ranking-weighted consensus semantic distribution and a consensus opinion sentence library; The multi-source evidence verification and reliability assessment module is used to retrieve evidence, assess logical consistency, and calculate the reliability score of the opinion statement. The core gain calculation module is used to calculate macroscopic semantic novelty measure, structural hole gain, and semantic entropy; The gated decision and feedback module is used to perform joint scoring, output release decisions, and generate structured evidence packages.

Images