Article screening method, apparatus, device, and medium

Abstract

The application provides an article screening method, device, equipment and medium. The method comprises: obtaining multi-round historical dialogue texts of a user in an academic retrieval process through a large language model; performing attenuation processing on attenuation coefficients corresponding to all rounds of historical dialogues before topic drift occurs; weighting and accumulating term frequencies of each word contained in the multi-round historical dialogue texts according to dialogue weights corresponding to each round of historical dialogue to generate a global weighted term frequency vector, and screening a plurality of target words from each word based on the global weighted term frequency vector; the dialogue weight corresponding to each round of historical dialogue is determined based on the attenuation coefficient and the round; a plurality of candidate articles are retrieved based on the plurality of target words; the multi-round historical dialogue texts are analyzed to determine a target academic intent of the user in the academic retrieval, and the plurality of candidate articles are screened according to the target academic intent to obtain a target article, thereby improving the accuracy and demand fitting degree of the retrieval result.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and more specifically, to an article screening method, apparatus, device, and medium. Background Technology

[0002] With the rapid development of large language models and intelligent dialogue technologies, academic article screening is gradually evolving towards intelligence and interactivity. As researchers' requirements for the accuracy and adaptability of retrieval systems continue to increase, various academic retrieval tools have been widely applied in research scenarios such as project research, review writing, and results verification. Existing academic article screening methods are mainly divided into two categories: one is a retrieval method based on single-round keyword matching, where the user enters a set of keywords each time, and the system directly performs a matching search in the database; the other is a retrieval enhancement method based on large models, where the keywords entered by the user are freely rewritten by a large model before the search is performed, and the search results are ranked with fixed weights.

[0003] The methods described above have achieved some success in simple search scenarios, but they still have shortcomings when facing the in-depth academic needs of researchers in long-term, continuous, and multi-round interactive dialogues. Because the search and filtering are based solely on keywords entered by the user in a single round, they may not match the academic needs that users continuously focus on in long-term, continuous, and multi-round interactive dialogues, resulting in search results that are out of touch with the user's actual needs. Summary of the Invention

[0004] In view of this, this application provides an article screening method, apparatus, device and medium that can mine the multi-round historical dialogue text of users in the academic search process, extract search terms through adaptive round decay weighting, and analyze the user's academic intent to screen the search results and match them with the user's real needs.

[0005] Specifically, this application is implemented through the following technical solution: According to a first aspect of this application, an article screening method is provided, the method comprising: Acquire the historical dialogue text of users during academic retrieval using a large language model; For each round of historical dialogue text except the first round, determine whether the topic has shifted relative to the previous round of historical dialogue. If the topic has shifted relative to the previous round of historical dialogue, attenuation processing is applied to the attenuation coefficients of all rounds of historical dialogue before the current round of historical dialogue. According to the dialogue weight corresponding to each round of historical dialogue, the word frequencies of each word contained in the multi-round historical dialogue text are weighted and accumulated to generate a global weighted word frequency vector. Based on the global weighted word frequency vector, multiple target words are selected from each word. The dialogue weight corresponding to each round of historical dialogue is determined based on the decay coefficient and round number corresponding to each round of historical dialogue. Based on the aforementioned target terms, a search is performed to obtain multiple candidate articles; The multi-round historical dialogue text is parsed to determine the user's target academic intent for academic retrieval. The multiple candidate articles are then filtered according to the target academic intent to obtain the target article.

[0006] In one optional implementation, parsing the multi-turn historical dialogue text to determine the user's target academic intent for academic retrieval includes: The multi-round historical dialogue text is parsed to determine multiple candidate academic intentions of the user in conducting academic searches; For each candidate academic intent, based on the mapping relationship between academic intent and representative words, the representative word coverage and representative word density of the multi-turn historical dialogue text relative to the candidate academic intent are determined; Based on the representative word coverage and the representative word density, the confidence level of the candidate academic intent is determined; Candidate academic intents with a confidence level greater than a confidence threshold are identified as the target academic intents for the user to conduct academic searches.

[0007] In one optional implementation, the step of filtering multiple target words from the various words based on the globally weighted word frequency vector includes: Based on the globally weighted term frequency vector, the term frequency-inverse document frequency (TF-IDF) score of each word is determined; For a score sequence composed of the various TF-IDF scores, a first filtering threshold is determined based on the mean and standard deviation of the score sequence; Based on words whose TF-IDF scores are greater than the first filtering threshold, multiple target words are determined.

[0008] In one optional implementation, the target academic intent carries a confidence level; the step of filtering the plurality of candidate articles according to the target academic intent to obtain the target article includes: For each candidate article, a ranking factor score is determined for the candidate article under multiple preset dimensions, and the ranking factor scores under multiple preset dimensions are normalized to obtain a normalized ranking factor score under multiple preset dimensions; the multiple preset dimensions include the number of article citations, publication year, author authority, article accessibility and journal influence. For each of the target academic intentions, a baseline weight for the target academic intention under multiple preset dimensions is determined based on the mapping relationship between academic intention and weight. Based on the confidence level of the target academic intent, the baseline weights under multiple preset dimensions are adjusted and normalized to obtain the adjusted weights of the target academic intent under multiple preset dimensions. Based on the normalized ranking factor scores under the multiple preset dimensions and the adjusted weights under the multiple preset dimensions, the ranking score corresponding to the candidate article is determined. The candidate articles with the highest ranking scores (first preset number) are selected as the target articles.

[0009] In one optional implementation, determining the ranking score corresponding to the candidate article based on the normalized ranking factor scores under the multiple preset dimensions and the adjusted weights under the multiple preset dimensions includes: When the target academic intent is singular, the normalized ranking factor scores under the multiple preset dimensions are weighted and summed based on the adjusted weights under the multiple preset dimensions to obtain the ranking score corresponding to the candidate article. When there are multiple target academic intentions, target weights under multiple preset dimensions are determined based on the confidence level of each target academic intention and the adjusted weights under multiple preset dimensions. The normalized ranking factor scores under the multiple preset dimensions are then weighted and summed based on the target weights under the multiple preset dimensions to obtain the ranking score corresponding to the candidate article.

[0010] In one optional implementation, the retrieval based on the plurality of target words to obtain a plurality of candidate articles includes: For each target word, the target word is expanded with synonyms and abbreviations to obtain a first set of expanded words, and the semantic similarity and hierarchical distance between each expanded word in the first set of expanded words and the target word are determined; The target word is expanded by parent and child classes to obtain a second set of expanded words, and the semantic similarity and hierarchical distance between each expanded word in the second set of expanded words and the target word are determined. Based on each target word and the first extended word set, the second extended word set corresponding to the target word, the semantic similarity, and the hierarchical distance, a retrieval formula is constructed. A search using the search query yields multiple candidate articles.

[0011] In one optional implementation, the step of constructing a retrieval formula based on each of the target words and the first extended word set, the second extended word set corresponding to the target words, the semantic similarity, and the hierarchical distance includes: A second filtering threshold is determined based on the semantic similarity corresponding to each of the target words; For each target word, the union of the first extended word set and the second extended word set is determined, and the target extended words corresponding to the target word are selected from the union based on the second filtering threshold. Based on the semantic similarity and hierarchical distance corresponding to the target extended words, the retrieval weight corresponding to the target extended words is determined; Based on the retrieval weights, a retrieval formula is constructed using the target expanded terms.

[0012] According to a second aspect of this application, an article filtering device is provided, the device comprising: The dialogue acquisition module is used to acquire the historical dialogue text of users during academic retrieval using a large language model; The drift detection module is used to determine whether the topic of each round of historical dialogue has drifted relative to the previous round of historical dialogue for each round of historical dialogue text except the first round. If topic drift occurs in any round of historical dialogue relative to the previous round of historical dialogue, the attenuation coefficient corresponding to all rounds of historical dialogue before this round of historical dialogue is attenuated. The word filtering module is used to weight and accumulate the word frequencies of each word in the multi-round historical dialogue text according to the dialogue weight corresponding to each round of historical dialogue, generate a global weighted word frequency vector, and filter out multiple target words from each word based on the global weighted word frequency vector; the dialogue weight corresponding to each round of historical dialogue is determined based on the decay coefficient and round number corresponding to each round of historical dialogue. The article retrieval module is used to retrieve multiple candidate articles based on the multiple target words; The intent mining module is used to parse the multi-turn historical dialogue text, determine the user's target academic intent for academic retrieval, and filter the multiple candidate articles according to the target academic intent to obtain the target article.

[0013] According to a third aspect of this application, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the article filtering method described in the first aspect above.

[0014] According to a fourth aspect of this application, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the article filtering method described in the first aspect above.

[0015] The article filtering method, apparatus, device, and medium provided in this application can mine multi-round historical dialogue texts during a user's academic retrieval process. By dynamically adjusting the attenuation coefficients corresponding to each round of historical dialogue through topic drift detection, and determining the dialogue weights based on the attenuation coefficients and rounds, a globally weighted word frequency vector is dynamically generated by combining word frequency distribution. This allows words in recent and topically coherent dialogues to have higher dialogue weights, thereby accurately filtering out target words for retrieval and more accurately reflecting the user's academic retrieval focus. Furthermore, it can parse the user's true and in-depth academic intent from multi-round historical dialogue texts, retrieve candidate articles based on target words, and perform secondary filtering according to the target academic intent to obtain the target article. This adapts to changes in multi-round dialogue context and continuous retrieval needs, reduces reliance on the free output of large language models, and ensures that target articles match the user's real needs. It effectively reduces the problems of one-sided demand mining and low matching degree of retrieval results caused by single-round retrieval filtering, improving the accuracy and demand relevance of retrieval results.

[0016] It should be understood that the above general description and the following detailed description are merely exemplary and explanatory, and are not intended to limit the technical solutions of this disclosure.

[0017] To make the above-mentioned objects, features and advantages of this disclosure more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating an exemplary embodiment of the present application of an article filtering method; Figure 2 This is a schematic diagram illustrating a process for determining target words according to an exemplary embodiment of this application; Figure 3 This is a schematic diagram illustrating an article filtering process according to an exemplary embodiment of this application; Figure 4 This is a schematic diagram of an article filtering device shown in an exemplary embodiment of this application; Figure 5This is a schematic diagram of the structure of a computer device shown in an exemplary embodiment of this application. Detailed Implementation

[0019] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0020] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0021] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0022] Research has found that when screening academic articles, searches are often based solely on keywords entered by users in a single round, lacking an understanding of multiple rounds of user interaction. This may not align with the academic needs that users consistently focus on during long-term, continuous, and multi-round interactions, and it fails to identify the core issues that users are consistently concerned with throughout the conversations. Furthermore, if users do not explicitly provide keywords in a particular round of conversation (e.g., simply asking for relevant literature), it is difficult to determine the user's intent, leading to search results that are disconnected from the user's actual needs.

[0023] Based on the above research, this application provides an article screening method that can mine multi-round historical dialogue texts during users' academic retrieval process. It dynamically adjusts the attenuation coefficients corresponding to each round of historical dialogue through topic drift detection, determines the dialogue weights based on the attenuation coefficients and rounds, and dynamically generates a globally weighted word frequency vector by combining word frequency distribution. This allows words in recent and topically coherent dialogues to have higher dialogue weights, thereby accurately screening target words for retrieval and more accurately reflecting the focus of users' academic retrieval. Furthermore, it can parse the true and in-depth target academic intent of users in their academic retrieval from multi-round historical dialogue texts, retrieve candidate articles based on target words, and perform secondary screening according to the target academic intent to obtain the target articles. This method can adapt to changes in the context of multi-round dialogues and the need for continuous retrieval, improving the accuracy and relevance of retrieval results.

[0024] To facilitate understanding of this embodiment, a detailed description of the article filtering method disclosed in this application is provided first. The executing entity of the article filtering method provided in this application is generally a computer device with certain computing power. This computer device can be a server, which can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud storage, big data, and artificial intelligence platforms. In some possible implementations, the computer device can also be a terminal device, which can be a mobile device, terminal, handheld device, computing device, vehicle-mounted device, etc. In other implementations, this article filtering method can be applied to an implementation environment composed of a terminal device and a server. Furthermore, this article filtering method can also be implemented by a processor calling computer-readable instructions stored in memory.

[0025] The following description, in conjunction with the accompanying drawings, illustrates an article screening method provided by an embodiment of this application.

[0026] See Figure 1 The diagram shown is a flowchart illustrating an article filtering method according to an exemplary embodiment of this application. Figure 1 As shown in the figure, the article filtering method provided in this embodiment includes steps S101 to S105, wherein: S101: Obtain the historical dialogue text of the user during the academic retrieval process using a large language model.

[0027] Here, the embodiments of this disclosure focus on the multi-round historical dialogue between the user and the model system during the academic retrieval process through the large language model, so as to obtain the user's exclusive target words and target academic needs through in-depth mining of the multi-round historical dialogue text.

[0028] S102: For each round of historical dialogue text except the first round, determine whether the topic of the current round of historical dialogue has shifted relative to the previous round of historical dialogue. If the topic of the current round of historical dialogue has shifted relative to the previous round of historical dialogue, perform attenuation processing on the attenuation coefficients corresponding to all rounds of historical dialogue before the current round of historical dialogue.

[0029] In this step, the historical dialogue text for each round can be converted into a TF-IDF vector. For each round of historical dialogue text except the first round, the cosine similarity between the TF-IDF vector corresponding to the current round of historical dialogue text and the TF-IDF vector corresponding to the previous round of historical dialogue text can be determined. The degree of topic drift is measured by the cosine similarity between the TF-IDF vector corresponding to the current round of historical dialogue text and the TF-IDF vector corresponding to the previous round of historical dialogue text. The cosine similarity between the TF-IDF vector corresponding to the current round of historical dialogue text and the TF-IDF vector corresponding to the previous round of historical dialogue text is compared with a topic coherence threshold. The specific value of the topic coherence threshold depends on the needs of topic drift judgment and is not specifically limited here. For example, the topic coherence threshold is 0.4.

[0030] If the cosine similarity between the TF-IDF vector corresponding to the current round of historical dialogue text and the TF-IDF vector corresponding to the previous round of historical dialogue text is less than the topic coherence threshold, it can be determined that the current round of historical dialogue has experienced topic drift relative to the previous round of historical dialogue.

[0031] For historical dialogue rounds where topic drift occurred, attenuation processing can be applied to the attenuation coefficients of all previous historical dialogue rounds. It is understood that this attenuation processing can be cumulative. For example, taking a 6-round dialogue, if both the third and fifth rounds are determined to have experienced topic drift, the attenuation coefficients corresponding to the first and second rounds will be attenuated twice, the attenuation coefficients corresponding to the third and fourth rounds will be attenuated twice, and the attenuation coefficients corresponding to the fifth and sixth rounds will not be attenuated.

[0032] Here, for rounds that do not undergo attenuation processing, the corresponding attenuation coefficient is... It can be set to a preset attenuation coefficient. The specific value of the preset attenuation coefficient depends on the attenuation processing requirements and is not specifically limited here. For example, the preset attenuation coefficient... Each time attenuation is performed, a penalty factor can be added to the attenuation coefficient. The specific value of the penalty factor is determined based on the needs of the decay process, and is not specifically limited here. For example, the penalty factor... For example, for a round that undergoes one attenuation process, the corresponding attenuation coefficient is... For each round undergoing attenuation processing, the corresponding attenuation coefficient is... .

[0033] S103: According to the dialogue weight corresponding to each round of historical dialogue, the word frequencies of each word contained in the multi-round historical dialogue text are weighted and accumulated to generate a global weighted word frequency vector. Based on the global weighted word frequency vector, multiple target words are selected from each word. The dialogue weight corresponding to each round of historical dialogue is determined based on the attenuation coefficient and round number corresponding to each round of historical dialogue.

[0034] In this step, the dialogue weight for each round of historical dialogue is determined based on the decay coefficient and round number corresponding to each round. Specifically, the dialogue weight for each round of historical dialogue can be determined using the following formula (1): (1) in, Indicates the rounds corresponding to historical dialogues. Indicates the first The dialogue weight corresponding to each round of historical dialogue. Indicates the first The attenuation coefficient corresponding to each round of historical dialogue. This indicates the total number of rounds of historical dialogue.

[0035] Based on the dialogue weight corresponding to each round of historical dialogue, the word frequencies of each word contained in the multi-round historical dialogue text are weighted and accumulated to generate a globally weighted word frequency vector. This allows words in recent and topic-coherent dialogues to have higher dialogue weights, thus more accurately reflecting the focus of the user's academic search. The globally weighted word frequency vector is a sparse vector with a vocabulary dimension, and each component corresponds to the weighted cumulative word frequency of each word.

[0036] In some possible implementations, the step of filtering multiple target words from the various words based on the globally weighted word frequency vector includes: Based on the globally weighted word frequency vector, the word frequency-inverse document frequency (TF-IDF) score of each word is determined; For a score sequence composed of the various TF-IDF scores, a first filtering threshold is determined based on the mean and standard deviation of the score sequence; Based on words whose TF-IDF scores are greater than the first filtering threshold, multiple target words are determined.

[0037] In the above steps, the term frequency-inverse document frequency (TF-IDF) score of each word can be determined based on the global weighted term frequency vector. Specifically, the weighted cumulative term frequency corresponding to each word can be extracted from the global weighted term frequency vector, and this weighted cumulative term frequency is used as the effective term frequency statistical value of the corresponding word. Then, the inverse document frequency calculation rule is introduced, and combined with the distribution of each word in the overall corpus, the TF-IDF calculation formula is substituted into the formula to solve for the TF value and IDF value corresponding to each word in turn. By multiplying the TF value and IDF value corresponding to each word, the TF-IDF score corresponding to each word is obtained.

[0038] The TF-IDF scores corresponding to each word are used to construct a score sequence, and the mean and standard deviation of the score sequence are determined. Based on the mean and standard deviation of the score sequence, a first filtering threshold is determined. Specifically, the first filtering threshold can be determined by the following formula (2): (2) in, This represents the first filtering threshold. This represents the mean of the fraction sequence. The standard deviation of a fractional sequence This represents the sensitivity coefficient.

[0039] The specific value of the sensitivity coefficient is determined based on the needs of search term filtering, and is not specifically limited here. For example, the sensitivity coefficient... .

[0040] Here, the first filtering threshold is adaptively adjusted according to the word frequency distribution in the historical dialogue text of multiple rounds, rather than being a fixed value. This can adapt to the data fluctuations caused by differences in dialogue length, word count, word range, and topic content, flexibly adapt to the dialogue text characteristics of each academic search, reduce the situation where a fixed threshold over-filters effective keywords when the dialogue vocabulary is sparse, or fails to remove a large number of invalid and redundant words when the vocabulary is complex, accurately adapt to the word distribution pattern in the current context, ensure that the keyword selection criteria are dynamically reasonable, and improve the flexibility and adaptability of target word selection.

[0041] A candidate word set is formed based on words whose TF-IDF scores are greater than the first filtering threshold. Redundant information can be filtered from the multi-turn historical dialogue text using large language models, intelligent agents, etc., to extract words related to academic research, such as research fields, technical methods, authors, and institutions, resulting in an extracted word set. The intersection of the candidate word set and the extracted word set is determined to filter out words irrelevant to the user's academic search needs. A second preset number of words with the highest TF-IDF scores in the intersection are identified as the target words. For example, the words in the intersection can be sorted in descending order of TF-IDF scores, and the top second preset number of words in this descending TF-IDF score order can be selected as the target words. The specific value of the second preset number depends on the search term filtering needs and is not specifically limited here; for example, the second preset number could be 5.

[0042] In this way, the TF-IDF score of each word can be determined based on the globally weighted word frequency vector. The first filtering threshold can be adaptively set by combining the mean and standard deviation of the score sequence to select target words. This ensures that the selected target words have high discriminative power and strong scene relevance, providing support for the subsequent retrieval of candidate articles and improving the effectiveness and accuracy of article retrieval.

[0043] To more clearly demonstrate the process of obtaining the target words, see [link to documentation]. Figure 2 This is a schematic diagram illustrating a process for determining target words, as shown in an exemplary embodiment of this application. Figure 2 As shown, multi-round historical dialogue text is acquired, topic drift detection is performed, and the attenuation coefficient corresponding to each round of historical dialogue is determined. Based on the attenuation coefficient and round number corresponding to each round of historical dialogue, the dialogue weight corresponding to each round of historical dialogue is determined. According to the dialogue weight corresponding to each round of historical dialogue, the word frequencies of each word contained in the multi-round historical dialogue text are weighted and accumulated to generate a global weighted word frequency vector. Based on the global weighted word frequency vector, a first filtering threshold is determined. Based on the first filtering threshold, a candidate word set is determined, and an extraction word set can also be obtained. Based on the candidate word set and the extraction word set, the target word is determined. The specific steps are described in the aforementioned embodiment and will not be repeated here.

[0044] S104: Based on the multiple target words, perform a search to obtain multiple candidate articles.

[0045] In this step, multiple target terms can be deterministically expanded and filtered without relying on large models for word rewriting. Traditional academic retrieval relies on large models for keyword expansion, lacking a deterministic mechanism. Completely entrusting keyword rewriting to large models results in an opaque and unverifiable process. Large models may introduce terms irrelevant to the user's research field (i.e., illusion words), causing search results to deviate from the user's needs. Furthermore, traditional methods do not utilize academic thesaurus and domain ontology for expansion, failing to cover domain synonyms, abbreviations, full names, and hyponyms / hypernyms of the same concept, leading to insufficient search recall.

[0046] In some possible implementations, the retrieval based on the plurality of target words yields a plurality of candidate articles, including: For each target word, the target word is expanded with synonyms and abbreviations to obtain a first set of expanded words, and the semantic similarity and hierarchical distance between each expanded word in the first set of expanded words and the target word are determined; The target word is expanded by parent and child classes to obtain a second set of expanded words, and the semantic similarity and hierarchical distance between each expanded word in the second set of expanded words and the target word are determined. Based on each target word and the first extended word set, the second extended word set corresponding to the target word, the semantic similarity, and the hierarchical distance, a retrieval formula is constructed. A search using the search query yields multiple candidate articles.

[0047] In the above steps, for each target word, a pre-set thesaurus of academic terms can be used, such as the Medical Subject Headings (MeSH) and the Association for Computing Machinery (ACM) Computer Science Classification System (CCS), to expand the target word with synonyms and abbreviations. Specifically, synonyms, domain abbreviations, and full names of the target word can be searched, and the resulting expanded words constitute a first set of expanded words. For example, for the target word "machine learning," the corresponding first set of expanded words may include "ML," "machine learning," and "statistical learning." The semantic similarity and hierarchical distance between each expanded word in the first set of expanded words and the target word can be determined. Optionally, the determination of the hierarchical distance can be based on a domain ontology tree, such as the ACM CCS hierarchical structure. The hierarchical distance between the expanded words and the target word in the domain ontology tree can be determined according to the hierarchical structure of the domain ontology tree.

[0048] For each target word, based on the domain ontology tree, parent and child class expansion can be performed. Specifically, the upper-level parent class and lower-level child class of the target word can be determined, and the resulting expanded words constitute a second expanded word set. For example, for the target word "random forest," its upper-level parent class is "ensemble learning," and its lower-level child classes include "extremely random tree," etc. Furthermore, the semantic similarity and hierarchical distance between each expanded word in the second expanded word set and the target word can be determined.

[0049] Based on each target word and its corresponding first extended word set, second extended word set, semantic similarity, and hierarchical distance, a search query is constructed. Following this query, searches are performed on mainstream academic paper databases such as Semantic Scholar and Arxiv. The retrieved articles are then aggregated and deduplicated to obtain multiple candidate articles. These candidate articles can come from academic journals, conference papers, etc., and are examples of academic papers and literature.

[0050] In this way, the target words are expanded by synonyms and abbreviations, as well as by parent and child categories. By combining the semantic similarity and hierarchical distance between the expanded words and the target words, a refined composite search formula is constructed. This reduces the limitations of single keyword searches, fully covers synonyms, abbreviations and full names in the academic field, as well as hierarchical professional terms. It makes up for the problems of single target word expressions and incomplete professional term references obtained from user dialogues, broadens the search coverage, and effectively reduces missed detections.

[0051] In some possible implementations, constructing a retrieval formula based on each of the target words and the first extended word set, the second extended word set corresponding to the target words, the semantic similarity, and the hierarchical distance includes: A second filtering threshold is determined based on the semantic similarity corresponding to each of the target words; For each target word, the union of the first extended word set and the second extended word set is determined, and the target extended words corresponding to the target word are selected from the union based on the second filtering threshold. Based on the semantic similarity and hierarchical distance corresponding to the target extended words, the retrieval weight corresponding to the target extended words is determined; Based on the retrieval weights, a retrieval formula is constructed using the target expanded terms.

[0052] In the above steps, a semantic similarity sequence is constructed based on the semantic similarity corresponding to each target word, and the mean and standard deviation of the semantic similarity sequence are determined. A second filtering threshold is determined based on the mean and standard deviation of the semantic similarity sequence. Specifically, the difference between the mean and standard deviation of the semantic similarity sequence can be used as the second filtering threshold. Specifically, the second filtering threshold can be determined using the following formula (3): (3) in, This indicates the second filtering threshold. This represents the mean of the semantic similarity sequence. The standard deviation of the semantic similarity sequence is represented.

[0053] Here, the second filtering threshold is adaptively adjusted based on semantic similarity, rather than being a fixed value. This allows for the accurate selection of semantically relevant and effective extended words, while filtering out redundant extended words with excessive semantic deviation or weak relevance, thereby improving the accuracy of target extended word selection.

[0054] For each target word, the union of the first set of extended words and the second set of extended words is determined. From this union, extended words with semantic similarity less than the second filtering threshold, extended words with character length less than a preset length, and extended words not included in the academic lexicon are removed. The remaining extended words are then identified as the target extended words corresponding to the target word. The specific value of the preset length depends on the needs of extended word filtering and is not specifically limited here; for example, the preset length may be 2.

[0055] For all target words corresponding to the target extended words, a retrieval weight is determined for each target extended word based on the semantic similarity and the hierarchical distance. Specifically, the semantic similarity is used as the first weight, and the hierarchical distance is used to determine the second weight. Here, the smaller the hierarchical distance, the higher the second weight. Specifically, the second weight can be determined by the following formula (4): (4) in, Indicates the second weight. Indicates the distance between hierarchical levels.

[0056] Based on the first weight and the second weight, the retrieval weight corresponding to the target expanded term is determined. Specifically, the retrieval weight can be determined by the following formula (5): (5) in, Indicates retrieval weight. This represents the retrieval coefficient corresponding to the first weight. Indicates the first weight. This represents the retrieval coefficient corresponding to the second weight. This indicates the second weight.

[0057] The sum of the retrieval coefficients corresponding to the first weight and the second weight is 1. The specific values ​​of the retrieval coefficients corresponding to the first weight and the second weight are determined based on the target expanded term filtering needs, and are not specifically limited here. For example, the retrieval coefficient corresponding to the first weight is 0.6, and the retrieval coefficient corresponding to the second weight is 0.4.

[0058] Based on the retrieval weights, the target extended terms are used to construct the search formula. For example, the target extended terms corresponding to high retrieval weights are used to construct precise search terms, and the target extended terms corresponding to low retrieval weights are used to construct extended search formulas.

[0059] In this way, a second filtering threshold is adaptively set based on semantic similarity. Target extended words are selected from the union of the first and second extended word sets. Then, the retrieval weight corresponding to the target extended words is determined by combining semantic similarity and hierarchical distance. The retrieval formula is constructed using the target extended words based on the difference in retrieval weight. This not only eliminates invalid extended words with large semantic deviations but also distinguishes the matching priority of different extended words, ensuring that the retrieval formula fits the user's academic semantic boundaries. This expands the retrieval coverage while improving the relevance accuracy of candidate articles.

[0060] S105: Analyze the multi-round historical dialogue text to determine the user's target academic intent for academic retrieval, and filter the multiple candidate articles according to the target academic intent to obtain the target article.

[0061] In this step, the multi-round historical dialogue text can be parsed, and features such as contextual dialogue, content association logic, and academic retrieval scenario can be combined to comprehensively determine the user's true and complete target academic intent. Then, based on the target academic intent, the multiple candidate articles can be screened to determine the target articles that highly match the user's intent, thereby achieving accurate convergence and personalized adaptation of retrieval results.

[0062] In some possible implementations, parsing the multi-turn historical dialogue text to determine the user's target academic intent for academic retrieval includes: The multi-round historical dialogue text is parsed to determine multiple candidate academic intentions of the user in conducting academic searches; For each candidate academic intent, based on the mapping relationship between academic intent and representative words, the representative word coverage and representative word density of the multi-turn historical dialogue text relative to the candidate academic intent are determined; Based on the representative word coverage and the representative word density, the confidence level of the candidate academic intent is determined; Candidate academic intents with a confidence level greater than a confidence threshold are identified as the target academic intents for the user to conduct academic searches.

[0063] In the above steps, the multi-turn historical dialogue text can be parsed to filter out multiple candidate academic intentions for the user's academic retrieval from a number of preset academic intentions. For example, multiple candidate academic intentions can be determined by calling large language models, intelligent agents, etc. Then, the confidence of these multiple candidate academic intentions can be dynamically verified in subsequent steps. The number of candidate academic intentions can be determined according to actual needs and is not limited here; for example, it can be three.

[0064] Specifically, the types, tags, and core requirements of multiple pre-set academic intentions can be represented by the following table 1: Table 1 For each candidate academic intent, a representative word corresponding to the candidate academic intent is determined based on the mapping relationship between academic intent and representative words. Specifically, the mapping relationship between academic intent and representative words can be shown in Table 2 below: Table 2 Based on the representative words corresponding to the candidate academic intent and the multi-round historical dialogue text, determine the total number of representative words of the candidate academic intent, the number of representative words of the candidate academic intent appearing in the multi-round historical dialogue text, the total frequency of the representative words of the candidate academic intent appearing in the multi-round historical dialogue text, and the total number of words included in the multi-round historical dialogue text.

[0065] Based on the number of representative words appearing in the multi-round historical dialogue text corresponding to the candidate academic intent and the total number of representative words corresponding to the candidate academic intent, the representative word coverage rate of the multi-round historical dialogue text relative to the candidate academic intent is determined. Specifically, the ratio of the number of representative words appearing in the multi-round historical dialogue text corresponding to the candidate academic intent to the total number of representative words corresponding to the candidate academic intent can be determined as the representative word coverage rate of the multi-round historical dialogue text relative to the candidate academic intent. Specifically, the representative word coverage rate can be determined by the following formula (6): (6) in, Indicates the first The coverage rate of representative words for each candidate academic intent. Indicates the first The number of representative words corresponding to each candidate academic intent that appear in the multi-round historical dialogue text. Indicates the first The total number of representative words corresponding to each candidate academic intent.

[0066] Based on the total frequency of the representative words corresponding to the candidate academic intent appearing in the multi-round historical dialogue text and the total number of words included in the multi-round historical dialogue text, the representative word density of the multi-round historical dialogue text relative to the candidate academic intent is determined. Specifically, the ratio of the total frequency of the representative words corresponding to the candidate academic intent appearing in the multi-round historical dialogue text to the total number of words included in the multi-round historical dialogue text can be determined as the representative word density of the multi-round historical dialogue text relative to the candidate academic intent. Specifically, the representative word density can be determined by the following formula (7): (7) in, Indicates the first The density of representative words for each candidate academic intent. Indicates the first The total frequency of the representative words corresponding to the candidate academic intentions in the multi-round historical dialogue text. This indicates the total number of words in the multi-turn historical dialogue text.

[0067] Based on the representative word coverage and the representative word density, the confidence level of the candidate academic intent is determined. Specifically, the confidence level can be determined by the following formula (8): (8) in, Indicates the first Confidence level of each candidate's academic intent This represents the weight corresponding to the coverage of representative words. Indicates the first The coverage rate of representative words for each candidate academic intent. This represents the weight corresponding to the word density. Indicates the first The density of representative words for each candidate academic intent. This represents the maximum value among the representative word densities corresponding to all candidate academic intentions, used for normalization.

[0068] The sum of the weight corresponding to the representative word coverage rate and the weight corresponding to the representative word density is 1. The specific values ​​of the weight corresponding to the representative word coverage rate and the weight corresponding to the representative word density are determined based on the confidence level requirements and are not specifically limited here. For example, the weight corresponding to the representative word coverage rate is 0.7 and the weight corresponding to the representative word density is 0.3.

[0069] The average confidence score of all candidate academic intentions is determined as the confidence threshold. The confidence score of each candidate academic intention is compared with the confidence threshold, and the candidate academic intentions with a confidence score greater than the confidence threshold are determined as the target academic intentions of the user for academic retrieval.

[0070] Optionally, if the confidence levels of all candidate academic intentions are less than or equal to the confidence threshold, the candidate academic intention with the highest confidence level is determined as the target academic intention.

[0071] In this way, by determining the coverage and density of representative words corresponding to candidate academic intentions, the confidence level of each candidate academic intention is determined from the dual dimensions of vocabulary coverage and content density, thereby screening out the target academic intention. This effectively avoids the one-sidedness of single semantic judgment, can accurately identify users' deep and continuous academic search needs, and effectively improves the accuracy of determining the target academic intention.

[0072] In some possible implementations, the target academic intent carries a confidence level; the step of filtering the plurality of candidate articles according to the target academic intent to obtain the target article includes: For each candidate article, a ranking factor score is determined for the candidate article under multiple preset dimensions, and the ranking factor scores under multiple preset dimensions are normalized to obtain a normalized ranking factor score under multiple preset dimensions; the multiple preset dimensions include the number of article citations, publication year, author authority, article accessibility and journal influence. For each of the target academic intentions, a baseline weight for the target academic intention under multiple preset dimensions is determined based on the mapping relationship between academic intention and weight. Based on the confidence level of the target academic intent, the baseline weights under multiple preset dimensions are adjusted and normalized to obtain the adjusted weights of the target academic intent under multiple preset dimensions. Based on the normalized ranking factor scores under the multiple preset dimensions and the adjusted weights under the multiple preset dimensions, the ranking score corresponding to the candidate article is determined. The candidate articles with the highest ranking scores (first preset number) are selected as the target articles.

[0073] In the above steps, for each candidate article, a ranking factor score is determined under multiple preset dimensions. These preset dimensions include article citation count, publication year, author authority, article accessibility, and journal impact. Here, the ranking factor score based on article citation count can be determined based on the total number of times the candidate article is cited by other academic literature, and this score represents the candidate article's influence and recognition in the academic field. The ranking factor score based on publication year can be determined based on the official publication date of the candidate article, and this score distinguishes between newer and older candidate articles to meet timeliness requirements. The ranking factor score based on author authority can be determined based on the H-index of the first author of the candidate article. The H-index is used to assess the quantity and level of researchers' academic output, and this ranking factor score assesses the author's academic influence in that research direction. The ranking factor score under article accessibility can be determined based on whether the candidate article is an open access resource, whether the full text is available for free, and whether it supports previewing core content. This ranking factor score prioritizes ensuring users can easily access the full text of the article. The ranking factor score under journal influence can be determined based on the latest Journal Citation Reports (JCR) impact factor or the influence score corresponding to the Chinese Academy of Sciences' regional classification of the candidate article. This ranking factor score represents the academic standing and influence of the journal in the relevant discipline.

[0074] Specifically, the ranking factor scores of candidate articles across multiple preset dimensions can be represented as shown in Table 3 below: Table 3 For each candidate article, the mean and standard deviation of the ranking factor scores under multiple preset dimensions are determined. Based on the mean and standard deviation of the ranking factor scores under multiple preset dimensions, the scores are normalized by Z-score and then mapped to the [0,1] interval by the Sigmoid function to obtain the normalized ranking factor scores under multiple preset dimensions, thereby eliminating the scale difference between factors of different dimensions.

[0075] For each target academic intent, a baseline weight is determined across multiple preset dimensions based on the mapping relationship between academic intent and weight. Specifically, the mapping relationship between academic intent and weight can be shown in Table 4 below: Table 4 For each target academic intent, the baseline weights under multiple preset dimensions are adjusted based on the confidence level of the target academic intent to obtain the adjusted baseline weights of the target academic intent under multiple preset dimensions. Optionally, the baseline weights under each preset dimension can be adjusted separately based on the confidence level of the target academic intent to obtain the adjusted baseline weights of the target academic intent under each preset dimension. Alternatively, to improve data processing efficiency, the baseline weights of the target academic intent under multiple preset dimensions can be concatenated into a baseline weight vector, and the baseline weight vector can be adjusted based on the confidence level of the target academic intent to obtain the adjusted baseline weight vector. This allows for simultaneous adjustment of the baseline weights under multiple preset dimensions. It can be understood that the adjusted baseline weight vector is formed by concatenating the adjusted baseline weights of the target academic intent under multiple preset dimensions.

[0076] When adjusting based on the confidence level of the target academic intent, the higher the confidence level, the larger the adjusted baseline weight value corresponding to the ranking factor score of the target academic intent, thereby achieving gradient adjustment of the baseline weight of each target academic intent when there are multiple target academic intents.

[0077] Specifically, the adjusted benchmark weight vector can be determined by the following formula (9): (9) in, Indicates the first The adjusted baseline weight vector for each target academic intention Indicates the first A baseline weight vector for each target academic intention. Indicates the first Confidence level of the academic intent of the target, This indicates an adjustment of the step size.

[0078] The specific value of the adjustment step size is determined based on the needs of the benchmark weight adjustment, and is not specifically limited here. For example, the adjustment step size is 0.4.

[0079] The adjusted baseline weights of the target academic intent across multiple preset dimensions are normalized so that the sum of the weights across the multiple preset dimensions is 1, thus obtaining the adjusted weights of the target academic intent across multiple preset dimensions. Alternatively, each adjusted baseline weight in the adjusted baseline weight vector can be normalized to obtain the adjusted weight vector. It can be understood that the adjusted weight vector is formed by concatenating the adjusted weights of the target academic intent across multiple preset dimensions.

[0080] Based on the normalized ranking factor scores and adjusted weights under the multiple preset dimensions, the ranking scores corresponding to the candidate articles are determined. The first preset number of candidate articles with the highest ranking scores are selected as the target articles. For example, the candidate articles can be arranged in descending order of their ranking scores, and the first preset number of candidate articles in this descending order can be selected as the target articles. The specific value of the first preset number is determined based on the user's settings; if the user does not set a value, a default value, such as 20, can be used.

[0081] Traditional academic search relies on fixed ranking weights, which cannot adapt to the diverse academic intentions of users. However, in this embodiment, ranking factor scores of candidate articles across multiple preset dimensions are determined and normalized. Adjusted weights are obtained by dynamically adjusting and normalizing the baseline weights based on the confidence level of the target academic intention. A ranking score is then determined based on the normalized ranking factor scores and adjusted weights. Candidate articles are re-ranked according to these scores to filter out the target articles desired by the user. This overcomes the limitations of fixed weights and rigid ranking rules in traditional academic search. The influence ratio of each dimension can adaptively adjust according to the confidence level of the target academic intention, aligning with users' differentiated and in-depth academic search preferences, effectively improving the rationality, personalization, and accuracy of target article selection.

[0082] In some possible implementations, determining the ranking score corresponding to the candidate article based on the normalized ranking factor scores under the multiple preset dimensions and the adjusted weights under the multiple preset dimensions includes: When the target academic intent is singular, the normalized ranking factor scores under the multiple preset dimensions are weighted and summed based on the adjusted weights under the multiple preset dimensions to obtain the ranking score corresponding to the candidate article. When there are multiple target academic intentions, target weights under multiple preset dimensions are determined based on the confidence level of each target academic intention and the adjusted weights under multiple preset dimensions. The normalized ranking factor scores under the multiple preset dimensions are then weighted and summed based on the target weights under the multiple preset dimensions to obtain the ranking score corresponding to the candidate article.

[0083] Here, when there is only one target academic intent, the adjusted weights under the multiple preset dimensions are directly used. Alternatively, given the adjusted weight vector, the adjusted weights under the multiple preset dimensions can be extracted from the adjusted weight vector. Based on the adjusted weights under the multiple preset dimensions, the normalized ranking factor scores under the multiple preset dimensions are weighted and summed to obtain the ranking score corresponding to the candidate article.

[0084] When there are multiple target academic intentions, the adjusted weights of the multiple target academic intentions under multiple preset dimensions are weighted and fused according to confidence normalization to determine the target weights under multiple preset dimensions. Alternatively, the adjusted weight vector can be weighted and fused according to confidence normalization to obtain a target weight vector, and the target weights under the multiple preset dimensions can be extracted from the target weight vector. The sum of the target weights under the multiple preset dimensions is still 1.

[0085] Specifically, the target weight vector can be determined by the following formula (10): (10) in, Represents the target weight vector. Indicates the first Confidence level of the academic intent of the target, This represents the total number of target academic intentions. Indicates the first The percentage of confidence in the academic intent of each target. Indicates the first Adjusted weight vector for each target academic intention.

[0086] The normalized ranking factor scores under the multiple preset dimensions are weighted and summed based on the target weights under the multiple preset dimensions to obtain the ranking score corresponding to the candidate article.

[0087] Specifically, the ranking score can be determined by the following formula (11): (11) in, Indicates the sorted score. This represents the normalized ranking factor score based on the number of article citations. This represents the normalized ordinal factor score for the year of publication. This represents the normalized ranking factor score under author authority. This represents the normalized ranking factor score under journal influence. This represents the normalized ranking factor score based on article accessibility; assuming a single target academic intent, This indicates the adjusted weight under the article citation count. Indicates the adjusted weight under the publication year. This indicates an adjustment of weights based on the author's authority. This indicates the adjusted weight under the journal's influence. This indicates the adjusted weights based on article accessibility; when there are multiple target academic intentions, This represents the target weight based on the number of article citations. This indicates the target weight under the publication year. This indicates the target weight under the author's authority. This indicates the target weight under the journal's influence. This represents the target weight under the condition of article accessibility.

[0088] In this way, ranking scores are determined differently for single-objective and multi-objective academic intentions. In the case of a single-objective academic intention, the normalized ranking factor scores are weighted and summed directly using the adjusted weights under each dimension. In the case of a multi-objective academic intention, the target weight is obtained by normalizing and weighting the confidence levels of each target academic intention, and then the normalized ranking factor scores are weighted and summed. This effectively achieves a reasonable fusion and configuration of dimension weights under multiple academic intentions, effectively reducing the problems of traditional academic retrieval that cannot adapt to scenarios where multiple academic intentions coexist, lack of weight fusion, and rigid ranking rules. It can meet the complex and multi-level academic retrieval needs of users, take into account the emphasis of different retrieval intentions, and make the retrieval results more in line with the user's real retrieval preferences, thereby improving the matching degree of article selection.

[0089] To more clearly illustrate the article selection process, see [link / reference]. Figure 3 This is a schematic diagram illustrating an article filtering process as an exemplary embodiment of this application. Figure 3 As shown, the process involves acquiring multi-round historical dialogue texts, detecting topic drift, and attenuating the attenuation coefficient when it occurs. Dialogue weights are determined based on the attenuation coefficients and rounds. Word frequencies are then weighted and accumulated according to these dialogue weights to generate a globally weighted word frequency vector. Target words are then selected based on this globally weighted word frequency vector. Candidate articles are retrieved based on the target words. The academic intent is analyzed from the multi-round historical dialogue texts, and multiple candidate articles are selected according to the target academic intent to obtain the target article. Specific steps are described in the aforementioned embodiment and will not be repeated here.

[0090] To better understand the article selection process, an example is provided below. The process involves obtaining six rounds of historical dialogue text from the user's academic search using a large language model. In the first round of historical dialogue, a user asked: "I've recently been thinking about the boundaries of our traditional organic hydrocarbon generation theory. Do you have any knowledge of 'inorganic grafting hydrocarbon generation'? I mean the view that deep inorganic gas sources (such as those from mantle degassing) and sedimentary organic matter can 'graft' or mix under specific geological conditions, jointly participating in the hydrocarbon generation process. This seems to partially explain some anomalous discoveries in ultra-deep oil and gas reservoirs." The Big Language Model replied: "This is a very cutting-edge and challenging interdisciplinary direction. The 'inorganic grafting hydrocarbon generation theory' you mentioned attempts to bridge the gap between classical organic hydrocarbon generation theory and inorganic hydrocarbon generation under extreme conditions (such as Fischer-Tropsch synthesis). It is usually associated with regions of strong crust-mantle interaction, such as active fault zones and areas near deep faults, where deep fluids (rich in H2, CO2, He, etc.) may upwell, undergoing physicochemical reactions with organic-rich sedimentary layers, or providing abiotic substrates for microbial activity, thereby 'grafting' or 'catalyzing' hydrocarbon generation. Are you focusing on specific geochemical indicators for this 'mixed origin' mechanism?"

[0091] In the second round of historical dialogue, a user asked: "Yes, the discrimination index is a very difficult problem. But what interests me more is the generation mechanism itself. Besides the deep fluids you mentioned, I've read some preliminary research on the role of 'plasma excitation' in simulating extreme conditions in the deep crust—could high-energy particles disrupt the chemical equilibrium of the CHO system and promote hydrocarbon synthesis via non-biological pathways? What do you think of the theoretical and experimental progress in this direction?" The large language model replied: "Your focus is very insightful. Research on simulating geological processes using 'plasma excitation' or 'discharge plasma' does indeed provide a new perspective for exploring non-biological hydrocarbons (especially methane and light hydrocarbons). Recent experiments in high-pressure physics and deep Earth materials science have shown that under extreme temperature and pressure conditions simulating the mantle transition zone or subduction zone environment, plasma-state or high-energy particle flows can promote the direct conversion of CO2 and H2O into hydrocarbons. This provides new physical mechanism support for the 'inorganic origin of hydrocarbon generation,' but most current results are still concentrated in laboratory simulations and theoretical calculations. Scaling them up, geologically verifying them, and linking them to actual oil and gas reservoirs are the main challenges at present."

[0092] In the third round of historical dialogue, a user asked: "Empirical verification is key. Our traditional organic hydrocarbon generation model faces extreme challenges in terms of temperature and pressure at depths of over 10,000 meters. The discovery of the 'deep subsurface microbiome' introduces new variables. Is it possible that below the traditional 'death line,' there still chemosynthetic microorganisms driven by deep inorganic fluids (such as H2) exist, metabolizing and producing hydrocarbons themselves, or modifying inorganic hydrocarbons from the depths? Could this be considered another form of 'grafting'?" The Big Language Model replied: "You've connected two cutting-edge fields! That's very insightful. The latest research trends are indeed exploring the connection between the 'deep biosphere' and..." The coupling of "abiotic processes" is also relevant. In some mid-ocean ridges and serpentinized environments, archaea (methanogens) that synthesize methane using H2 and CO2 have been confirmed. Therefore, theoretically, similar ecological niches exist in serpentinized or water-rock reaction zones caused by deep intracontinental faults. These microorganisms may act as "biocatalysts" between "inorganic" fluids and "organic" hydrocarbons, achieving a "bio-grafting" from inorganic substrates to biological organic products. Currently, identifying this hybrid "biological-abiotic" origin through the combined use of specific biomarkers (such as lipids) and isotopes (such as C and H), along with geological tectonic background analysis, is an active research hotspot internationally.

[0093] In the fourth round of historical dialogue, a user stated: "Yes, biomarkers and isotope geochemistry remain our core tools. However, this places unprecedented demands on the acquisition and analytical precision of ultra-deep samples. We've recently been discussing the planning and interpretation schemes for 10,000-meter-level scientific exploration wells in the Tarim and Sichuan Basins, and we're troubled by these theoretical controversies. Simply applying the organic matter thermal degradation model doesn't explain many phenomena; a complete shift to inorganic origins lacks a systematic and widely accepted chain of evidence. The debate in this field seems to be moving from an 'either / or' to a 'multiple and complex' approach." The large language model replied: "Your summary is not..." This is a very insightful observation. The concept of "multi-element, complex genesis" is becoming a new paradigm for research on the origins of ultra-deep hydrocarbons. This is no longer a simple debate between organic and inorganic processes, but rather a focus on how, within the context of crust-mantle interaction, deep inorganic processes (fluid processes, high-energy physics), potential deep biological processes, and residual sedimentary organic matter, through energy and material exchange, collectively construct hydrocarbon reservoirs over long geological periods. Research has shifted its focus to modeling specific geological scenarios (such as the thermo-fluid-chemical-biological coupling processes in deep fault zones), more refined experimental simulations (multi-factor coupling), and high-resolution geochemical fingerprinting techniques.

[0094] In the fifth round of historical dialogue, a user stated: "Yes, modeling and experimentation need to keep up. I feel this field is at a critical juncture of theoretical breakthroughs and technology-driven advancements. Traditional petroleum geology needs to engage more deeply with geophysics, geochemistry, and even microbiology and high-pressure physics. We can no longer be confined to static basin models." The Big Language Model replied: "Completely agree. This is a typical interdisciplinary 'deep earth' frontier. Judging from the latest international conferences and project updates, research trends are focusing on: 1) Multi-scale in-situ observation and experimental techniques: such as diamond cavity combined with spectral analysis for extreme condition simulation, and high-fidelity core sampling and in-situ microbial detection techniques for deep samples; 2) High-precision geochemical tracing systems: such as the application of non-traditional stable isotopes (e.g., Cl, Zn) in distinguishing fluid sources and reaction pathways; 3) Digital twins and process simulation: integrating deep fluid transport, chemical reaction kinetics, and microbial metabolic models into the framework of geological structural evolution for numerical simulation. These advancements are propelling us from 'what' to 'how'."

[0095] In the sixth round of historical dialogue, the user asked, "Could you help me find some relevant literature?" Here, in the sixth round of historical dialogue, the user requested to search for literature, but did not specify search terms or their academic search intent.

[0096] For each round of historical dialogue text except the first round, the cosine similarity between the TF-IDF vector corresponding to that round's historical dialogue text and the TF-IDF vector corresponding to the previous round's historical dialogue text is determined. sim(1,2)=0.72, sim(2,3)=0.68, sim(3,4)=0.37, sim(4,5)=0.55, sim(5,6)=0.31. If the topic coherence threshold is 0.4, then the cosine similarity sim(3,4) < 0.4 for rounds 3 to 4, and the cosine similarity sim(5,6) < 0.4 for rounds 5 to 6, indicating topic drift in the fourth and sixth rounds of historical dialogue. If a pre-set decay coefficient is used... Punishment factor Then the attenuation coefficients for rounds 1, 2, and 3 are all... The attenuation coefficients for rounds 4 and 5 are both... The attenuation coefficient in the 6th round .

[0097] The dialogue weights corresponding to each round of historical dialogue are determined as follows: , , , , , Based on the dialogue weight corresponding to each round of historical dialogue, the word frequencies of each word contained in the text of multiple rounds of historical dialogue are weighted and accumulated to generate a global weighted word frequency vector. The dialogue weight corresponding to the 6th round of historical dialogue is the highest, and the word frequencies of words appearing in the 6th round of historical dialogue receive the maximum weight.

[0098] Based on the globally weighted term frequency vector, the term frequency-inverse document frequency (TF-IDF) score of each word is determined. For the score sequence composed of each TF-IDF score, the mean of the score sequence is determined. Standard deviation of fractional sequences Thus, the first filtering threshold is determined. Words with TF-IDF scores greater than a first filtering threshold are used to form a candidate word set. An extracted word set is then obtained from multi-turn historical dialogue text by using a large language model. The intersection of the candidate word set and the extracted word set is determined, and the five words with the highest TF-IDF scores in the intersection are identified as target words. For example, target words include organic and inorganic grafting hydrocarbon generation theory, inorganic hydrocarbon generation genesis, organic hydrocarbon generation genesis, plasma excitation, deep subsurface microbiome, crust-mantle interaction, and ultra-deep exploration.

[0099] The target word is expanded using synonyms and abbreviations to obtain a first set of expanded words. The semantic similarity and hierarchical distance between each expanded word in the first set and the target word are determined. The target word is then expanded using parent and child categories to obtain a second set of expanded words. The semantic similarity and hierarchical distance between each expanded word in the second set and the target word are determined. Based on the semantic similarity of each target word, a semantic similarity sequence is constructed, and the mean of the semantic similarity sequence is determined. and standard deviation Thus, the second filtering threshold is determined. For each target term, the union of the first and second extended term sets is determined. From this union, extended terms with semantic similarity less than the second filtering threshold, extended terms with a character length less than 2, and extended terms not included in the academic lexicon are removed. The remaining extended terms are then identified as the target extended terms corresponding to the target term. The retrieval weight for each target extended term is determined. Based on the retrieval weight, a retrieval formula is constructed using the target extended terms. For example, retrieval formula 1 is organic-inorganic coupling hydrocarbon generation, crust-mantle interaction, deep carbon cycle, and mixed-origin hydrocarbons; retrieval formula 2 is inorganic hydrocarbon generation, serpentinization, Fischer-Tropsch synthesis, non-thermal plasma, and plasma-catalyzed hydrocarbon synthesis; and retrieval formula 3 is organic hydrocarbon generation, deep subsurface biosphere, microbial methane generation, and ultra-deep oil and gas exploration. The retrieved articles are then aggregated and deduplicated to obtain multiple candidate articles. The number of candidate articles is generally greater than 300.

[0100] The historical dialogue text from multiple rounds was analyzed to identify three candidate academic intentions for users conducting academic searches: latest research trends, comparative analysis of methods, and high-impact research results. Dynamic confidence checks were then performed on each candidate academic intention. [Further details regarding the candidate academic intentions are needed for a complete translation.] Figure 1 Latest research trends and candidate academic ideas Figure 1 Representative terms include "latest, cutting-edge, recent years, progress, hot topics, emerging, dynamics," and candidate academic meanings. Figure 1 The total number of corresponding representative words Candidate academic intention Figure 1 The number of representative words appearing in the multi-turn historical dialogue text. (Latest, cutting-edge, recent years), candidate academic ideas Figure 1 The coverage rate of representative words The total number of words included in the multi-round historical dialogue text Candidate academic intention Figure 1 The total frequency of the corresponding representative words in the multi-turn historical dialogue text Candidate academic intention Figure 1 The density of representative words Regarding the academic intentions of the candidates Figure 2 Comparative analysis of methods, candidate academic significance Figure 2 Representative terms include "comparison, comparison, advantages and disadvantages, selection, differences, evaluation, benchmark," and candidate academic meanings. Figure 2 The total number of corresponding representative words Candidate academic intention Figure 2 The number of representative words appearing in the multi-turn historical dialogue text. (Comparison, contrast, difference, evaluation), candidate academic intent Figure 2 The coverage rate of representative words Candidate academic intention Figure 2 The total frequency of the corresponding representative words in the multi-turn historical dialogue text Candidate academic intention Figure 2 The density of representative words Regarding the academic intentions of the candidates Figure 3 High-impact research findings, candidate academic achievements Figure 3 Representative terms include "highly cited, top journals, top conferences, authoritative, benchmark, and influential," and candidate academic terms. Figure 3 The total number of corresponding representative words Candidate academic intention Figure 3 The number of representative words appearing in the multi-turn historical dialogue text. (Top Journal), Candidate Academic Intent Figure 3 The coverage rate of representative words Candidate academic intention Figure 3 The total frequency of the corresponding representative words in the multi-turn historical dialogue text Candidate academic intention Figure 3 The density of representative words The maximum value among the representative word densities corresponding to all candidate academic intentions. Determine the confidence level of each candidate's academic intent: , , Confidence threshold The confidence scores of each candidate academic intent are compared with a confidence threshold. , , Determine the user's academic intent in conducting academic searches as the latest research trends (confidence level). Comparative analysis of methods (confidence level) ).

[0101] For each candidate article, ranking factor scores are determined across multiple predefined dimensions. These scores are then normalized to obtain normalized ranking factor scores. The predefined dimensions include article citation count, publication year, author authority, article accessibility, and journal impact. The target academic intent is then defined. Figure 1 (Latest research trend) Benchmark weight vectors under multiple preset dimensions Based on the target academic intention Figure 1 The confidence level is used to adjust the baseline weight vector, resulting in the adjusted baseline weight vector. The adjusted baseline weights in the adjusted baseline weight vector are normalized to obtain the adjusted weight vector. Targeting academic interests Figure 2 (Method Comparative Analysis), Objective Academic Purpose Figure 2 Benchmark weight vectors under multiple preset dimensions Based on the target academic intention Figure 2 The confidence level is used to adjust the baseline weight vector, resulting in the adjusted baseline weight vector. The adjusted baseline weights in the adjusted baseline weight vector are normalized to obtain... The adjusted weights of the two target academic intentions across multiple preset dimensions are determined by weighting and fusing them using confidence-normalized methods. .

[0102] The normalized ranking factor scores under multiple preset dimensions are weighted and summed based on the target weights under multiple preset dimensions to obtain the ranking scores of the candidate articles. The 10 candidate articles with the highest ranking scores are then selected as the target articles.

[0103] The article filtering method provided in this application can mine multi-round historical dialogue texts during a user's academic search process. It dynamically adjusts the attenuation coefficients corresponding to each round of historical dialogue through topic drift detection, determines the dialogue weights based on the attenuation coefficients and rounds, and dynamically generates a globally weighted word frequency vector by combining word frequency distribution. This allows words in recent and topically coherent dialogues to have higher dialogue weights, thereby accurately filtering out target words for retrieval and more accurately reflecting the user's academic search focus. Furthermore, it can parse the user's true and in-depth academic intent from multi-round historical dialogue texts, retrieve candidate articles based on target words, and perform secondary filtering according to the target academic intent to obtain the target article. This adapts to changes in multi-round dialogue context and continuous search needs, reduces reliance on the free output of large language models, and ensures that target articles match the user's real needs. It effectively reduces the problems of one-sided demand mining and low search result matching caused by single-round search filtering, improving the accuracy and demand relevance of search results.

[0104] Those skilled in the art will understand that, in the above-described method of the specific implementation, the order in which each step is written does not imply a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of each step should be determined by its function and possible internal logic.

[0105] Corresponding to the embodiments of the aforementioned article screening method, this application also provides embodiments of an article screening device.

[0106] Please see Figure 4 This is a schematic diagram illustrating an article filtering device according to an exemplary embodiment of this application. Figure 4 As shown in the illustration, the article filtering device 400 provided in this application embodiment includes: The dialogue acquisition module 401 is used to acquire the multi-turn historical dialogue text of the user during the academic retrieval process through the large language model; The drift judgment module 402 is used to determine whether the topic of the historical dialogue in each round other than the first round has drifted relative to the previous round of historical dialogue for each round of historical dialogue text. If the topic of the historical dialogue in any round has drifted relative to the previous round of historical dialogue, the attenuation coefficient corresponding to all rounds of historical dialogue before the current round of historical dialogue is attenuated. The word filtering module 403 is used to weight and accumulate the word frequencies of each word contained in the multi-round historical dialogue text according to the dialogue weight corresponding to each round of historical dialogue, generate a global weighted word frequency vector, and filter out multiple target words from each word based on the global weighted word frequency vector; the dialogue weight corresponding to each round of historical dialogue is determined based on the attenuation coefficient and round number corresponding to each round of historical dialogue. Article retrieval module 404 is used to retrieve multiple candidate articles based on the multiple target words; The intent mining module 405 is used to parse the multi-round historical dialogue text, determine the user's target academic intent for academic retrieval, and filter the multiple candidate articles according to the target academic intent to obtain the target article.

[0107] In one optional implementation, when the intent mining module 405 parses the multi-turn historical dialogue text to determine the user's target academic intent for academic retrieval, it is specifically used for: The multi-round historical dialogue text is parsed to determine multiple candidate academic intentions of the user in conducting academic searches; For each candidate academic intent, based on the mapping relationship between academic intent and representative words, the representative word coverage and representative word density of the multi-turn historical dialogue text relative to the candidate academic intent are determined; Based on the representative word coverage and the representative word density, the confidence level of the candidate academic intent is determined; Candidate academic intents with a confidence level greater than a confidence threshold are identified as the target academic intents for the user to conduct academic searches.

[0108] In one optional implementation, when the word filtering module 403 is used to filter multiple target words from the various words based on the globally weighted word frequency vector, it is specifically used for: Based on the globally weighted word frequency vector, the word frequency-inverse document frequency (TF-IDF) score of each word is determined; For a score sequence composed of the various TF-IDF scores, a first filtering threshold is determined based on the mean and standard deviation of the score sequence; Based on words whose TF-IDF scores are greater than the first filtering threshold, multiple target words are determined.

[0109] In one optional implementation, the target academic intent carries a confidence level; when the intent mining module 405 filters the plurality of candidate articles according to the target academic intent to obtain the target article, it is specifically used for: For each candidate article, a ranking factor score is determined for the candidate article under multiple preset dimensions, and the ranking factor scores under multiple preset dimensions are normalized to obtain a normalized ranking factor score under multiple preset dimensions; the multiple preset dimensions include the number of article citations, publication year, author authority, article accessibility and journal influence. For each of the target academic intentions, a baseline weight for the target academic intention under multiple preset dimensions is determined based on the mapping relationship between academic intention and weight. Based on the confidence level of the target academic intent, the baseline weights under multiple preset dimensions are adjusted and normalized to obtain the adjusted weights of the target academic intent under multiple preset dimensions. Based on the normalized ranking factor scores under the multiple preset dimensions and the adjusted weights under the multiple preset dimensions, the ranking score corresponding to the candidate article is determined. The candidate articles with the highest ranking scores (first preset number) are selected as the target articles.

[0110] In an optional implementation, when the intent mining module 405 determines the ranking score corresponding to the candidate article based on the normalized ranking factor scores under the multiple preset dimensions and the adjusted weights under the multiple preset dimensions, it is specifically used for: When the target academic intent is singular, the normalized ranking factor scores under the multiple preset dimensions are weighted and summed based on the adjusted weights under the multiple preset dimensions to obtain the ranking score corresponding to the candidate article. When there are multiple target academic intentions, target weights under multiple preset dimensions are determined based on the confidence level of each target academic intention and the adjusted weights under multiple preset dimensions. The normalized ranking factor scores under the multiple preset dimensions are then weighted and summed based on the target weights under the multiple preset dimensions to obtain the ranking score corresponding to the candidate article.

[0111] In one optional implementation, the article retrieval module 404 is specifically used for: For each target word, the target word is expanded with synonyms and abbreviations to obtain a first set of expanded words, and the semantic similarity and hierarchical distance between each expanded word in the first set of expanded words and the target word are determined; The target word is expanded by parent and child classes to obtain a second set of expanded words, and the semantic similarity and hierarchical distance between each expanded word in the second set of expanded words and the target word are determined. Based on each target word and the first extended word set, the second extended word set corresponding to the target word, the semantic similarity, and the hierarchical distance, a retrieval formula is constructed. A search using the search query yields multiple candidate articles.

[0112] In an optional implementation, when the article retrieval module 404 constructs a retrieval formula based on each of the target words and the first extended word set, the second extended word set corresponding to the target words, the semantic similarity, and the hierarchical distance, it is specifically used for: A second filtering threshold is determined based on the semantic similarity corresponding to each of the target words; For each target word, the union of the first extended word set and the second extended word set is determined, and the target extended words corresponding to the target word are selected from the union based on the second filtering threshold. Based on the semantic similarity and hierarchical distance corresponding to the target extended words, the retrieval weight corresponding to the target extended words is determined; Based on the retrieval weights, a retrieval formula is constructed using the target expanded terms.

[0113] The specific implementation process of the functions and roles of each module in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0114] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0115] Based on the same technical concept, this application also provides a computer device 500, referring to... Figure 5 The diagram shown is a schematic representation of the structure of a computer device according to an exemplary embodiment of this application, comprising: The processor 510, memory 520, and bus 530 are included. The memory 520 is used to store execution instructions and includes main memory 521 and external memory 522. The main memory 521, also known as internal memory, is used to temporarily store the operation data in the processor 510 and the data exchanged with external memory 522 such as hard disk. The processor 510 exchanges data with external memory 522 through main memory 521.

[0116] In this embodiment, the memory 520 is specifically used to store application code that executes the solution of this application, and its execution is controlled by the processor 510. That is, when the computer device 500 is running, the processor 510 communicates with the memory 520 through the bus 530, or the processor 510 communicates with the memory 520 through other means, so that the processor 510 executes the application code stored in the memory 520, and then performs the steps of the article filtering method described in any of the foregoing embodiments.

[0117] The memory 520 may be, but is not limited to, random access memory (RAM), read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), etc.

[0118] Processor 510 may be an integrated circuit chip with signal processing capabilities. The aforementioned processor can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor can be a microprocessor or any conventional processor.

[0119] It is understood that the structures illustrated in the embodiments of this application do not constitute a specific limitation on the computer device 500. In other embodiments of this application, the computer device 500 may include more or fewer components than illustrated, or combine some components, or split some components, or have different component arrangements. The illustrated components may be implemented in hardware, software, or a combination of software and hardware.

[0120] This disclosure also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the article filtering method described in the above-described method embodiments. The storage medium may be a volatile or non-volatile computer-readable storage medium.

[0121] This disclosure also provides a computer program product storing a computer program. When the computer program is run by a processor, it executes the steps of the article filtering method provided in any of the above embodiments of this disclosure. For details, please refer to the above method embodiments, which will not be repeated here.

[0122] The aforementioned computer program product can be implemented through hardware, software, or a combination thereof. In one optional embodiment, the computer program product is specifically embodied in a computer storage medium, which can be a volatile or non-volatile computer-readable storage medium. In another optional embodiment, the computer program product is specifically embodied in a software product, such as a software development kit (SDK), etc.

[0123] Furthermore, embodiments of the subject matter and functional operation described in this specification can be implemented in the following ways: digital electronic circuits, tangibly embodied computer software or firmware, computer hardware including the structures disclosed in this specification and their structural equivalents, or combinations thereof. Embodiments of the subject matter described in this specification can be implemented as one or more computer programs, i.e., one or more modules of computer program instructions encoded on a tangible, non-transitory program carrier for execution by a data processing apparatus or for controlling the operation of a data processing apparatus. Alternatively or additionally, program instructions may be encoded on artificially generated propagation signals, such as machine-generated electrical, optical, or electromagnetic signals, which are generated to encode information and transmit it to a suitable receiving device for execution by the data processing apparatus. The computer storage medium may be a machine-readable storage device, a machine-readable storage substrate, a random or serial access memory device, or combinations thereof.

[0124] The processing and logic flow described in this specification can be executed by one or more programmable computers that execute one or more computer programs to perform corresponding functions by operating on input data and generating output. The processing and logic flow can also be executed by dedicated logic circuitry—such as FPGAs (Field-Programmable Gate Arrays) or ASICs (Application-Specific Integrated Circuits), and the device can also be implemented as dedicated logic circuitry.

[0125] Suitable computers for executing computer programs include, for example, general-purpose and / or special-purpose microprocessors, or any other type of central processing unit. Typically, the central processing unit receives instructions and data from read-only memory and / or random access memory. The basic components of a computer include a central processing unit for implementing or executing instructions and one or more memory devices for storing instructions and data. Typically, a computer will also include one or more mass storage devices for storing data, such as disks, magneto-optical disks, or optical disks, or the computer will be operatively coupled to such mass storage devices to receive data from or transfer data to them, or both. However, a computer is not required to have such devices. Furthermore, a computer can be embedded in another device, such as a mobile phone, a personal digital assistant (PDA), a mobile audio or video player, a game console, a global positioning system (GPS) receiver, or a portable storage device such as a universal serial bus (USB) flash drive, to name a few.

[0126] Computer-readable media suitable for storing computer program instructions and data include all forms of non-volatile memory, media, and memory devices, such as semiconductor memory devices (e.g., EPROM, EEPROM, and flash memory devices), magnetic disks (e.g., internal hard disks or removable disks), magneto-optical disks, and CD-ROM and DVD-ROM disks. Processors and memory may be supplemented by or incorporated into dedicated logic circuitry.

[0127] While this specification contains numerous specific implementation details, these should not be construed as limiting the scope of any invention or the scope of the claims, but rather are primarily intended to describe features of specific embodiments of a particular invention. Certain features described in the various embodiments herein may also be implemented in combination in a single embodiment. Conversely, various features described in a single embodiment may also be implemented separately in various embodiments or in any suitable sub-combination. Furthermore, while features may function in certain combinations as described above and even initially claimed in this way, one or more features from a claimed combination may be removed from that combination in some cases, and a claimed combination may refer to a sub-combination or a variation thereof.

[0128] Similarly, although the operations are depicted in a specific order in the accompanying drawings, this should not be construed as requiring these operations to be performed in the specific order shown or sequentially, or requiring all illustrated operations to be performed to achieve the desired result. In some cases, multitasking and parallel processing may be advantageous. Furthermore, the separation of various system modules and components in the above embodiments should not be construed as requiring such separation in all embodiments, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged into multiple software products.

[0129] Thus, specific embodiments of the subject matter have been described. Other embodiments are within the scope of the appended claims. In some cases, the actions recited in the claims may be performed in a different order and still achieve the desired result. Furthermore, the processes depicted in the drawings are not necessarily shown in a specific order or sequence to achieve the desired result. In some implementations, multitasking and parallel processing may be advantageous.

[0130] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. An article screening method, characterized in that, The method includes: Acquire the historical dialogue text of users during academic retrieval using a large language model; For each round of historical dialogue text except the first round, determine whether the topic has shifted relative to the previous round of historical dialogue. If the topic has shifted relative to the previous round of historical dialogue, attenuation processing is applied to the attenuation coefficients of all rounds of historical dialogue before the current round of historical dialogue. According to the dialogue weight corresponding to each round of historical dialogue, the word frequencies of each word contained in the multi-round historical dialogue text are weighted and accumulated to generate a global weighted word frequency vector. Based on the global weighted word frequency vector, multiple target words are selected from each word. The dialogue weight corresponding to each round of historical dialogue is determined based on the decay coefficient and round number corresponding to each round of historical dialogue. Based on the aforementioned target terms, a search is performed to obtain multiple candidate articles; The multi-round historical dialogue text is parsed to determine the user's target academic intent for academic retrieval. The multiple candidate articles are then filtered according to the target academic intent to obtain the target article.

2. The method according to claim 1, characterized in that, The step of parsing the multi-round historical dialogue text to determine the user's target academic intent for academic retrieval includes: The multi-round historical dialogue text is parsed to determine multiple candidate academic intentions of the user in conducting academic searches; For each candidate academic intent, based on the mapping relationship between academic intent and representative words, the representative word coverage and representative word density of the multi-turn historical dialogue text relative to the candidate academic intent are determined; Based on the representative word coverage and the representative word density, the confidence level of the candidate academic intent is determined; Candidate academic intents with a confidence level greater than a confidence threshold are identified as the target academic intents for the user to conduct academic searches.

3. The method according to claim 1, characterized in that, The process of selecting multiple target words from the various words based on the globally weighted word frequency vector includes: Based on the globally weighted word frequency vector, the word frequency-inverse document frequency (TF-IDF) score of each word is determined; For a score sequence composed of the various TF-IDF scores, a first filtering threshold is determined based on the mean and standard deviation of the score sequence; Based on words whose TF-IDF scores are greater than the first filtering threshold, multiple target words are determined.

4. The method according to claim 1, characterized in that, The target academic intent carries a confidence level; the process of filtering the multiple candidate articles according to the target academic intent to obtain the target article includes: For each candidate article, a ranking factor score is determined for the candidate article under multiple preset dimensions, and the ranking factor scores under multiple preset dimensions are normalized to obtain a normalized ranking factor score under multiple preset dimensions; the multiple preset dimensions include the number of article citations, publication year, author authority, article accessibility and journal influence. For each of the target academic intentions, a baseline weight for the target academic intention under multiple preset dimensions is determined based on the mapping relationship between academic intention and weight. Based on the confidence level of the target academic intent, the baseline weights under multiple preset dimensions are adjusted and normalized to obtain the adjusted weights of the target academic intent under multiple preset dimensions. Based on the normalized ranking factor scores under the multiple preset dimensions and the adjusted weights under the multiple preset dimensions, the ranking score corresponding to the candidate article is determined. The candidate articles with the highest ranking scores (first preset number) are selected as the target articles.

5. The method according to claim 4, characterized in that, The process of determining the ranking score corresponding to the candidate article based on the normalized ranking factor scores under the multiple preset dimensions and the adjusted weights under the multiple preset dimensions includes: When the target academic intent is singular, the normalized ranking factor scores under the multiple preset dimensions are weighted and summed based on the adjusted weights under the multiple preset dimensions to obtain the ranking score corresponding to the candidate article. When there are multiple target academic intentions, target weights under multiple preset dimensions are determined based on the confidence level of each target academic intention and the adjusted weights under multiple preset dimensions. The normalized ranking factor scores under the multiple preset dimensions are then weighted and summed based on the target weights under the multiple preset dimensions to obtain the ranking score corresponding to the candidate article.

6. The method according to claim 1, characterized in that, The search based on the multiple target words yields multiple candidate articles, including: For each target word, the target word is expanded with synonyms and abbreviations to obtain a first set of expanded words, and the semantic similarity and hierarchical distance between each expanded word in the first set of expanded words and the target word are determined; The target word is expanded by parent and child classes to obtain a second set of expanded words, and the semantic similarity and hierarchical distance between each expanded word in the second set of expanded words and the target word are determined. Based on each target word and the first extended word set, the second extended word set corresponding to the target word, the semantic similarity, and the hierarchical distance, a retrieval formula is constructed. A search using the search query yields multiple candidate articles.

7. The method according to claim 6, characterized in that, The process of constructing a retrieval formula based on each target word, the first extended word set corresponding to the target word, the second extended word set, the semantic similarity, and the hierarchical distance includes: A second filtering threshold is determined based on the semantic similarity corresponding to each of the target words; For each target word, the union of the first extended word set and the second extended word set is determined, and the target extended words corresponding to the target word are selected from the union based on the second filtering threshold. Based on the semantic similarity and hierarchical distance corresponding to the target extended words, the retrieval weight corresponding to the target extended words is determined; Based on the retrieval weights, a retrieval formula is constructed using the target expanded terms.

8. An article filtering device, characterized in that, The device includes: The dialogue acquisition module is used to acquire the historical dialogue text of users during academic retrieval using a large language model; The drift detection module is used to determine whether the topic of each round of historical dialogue has drifted relative to the previous round of historical dialogue for each round of historical dialogue text except the first round. If topic drift occurs in any round of historical dialogue relative to the previous round of historical dialogue, the attenuation coefficient corresponding to all rounds of historical dialogue before this round of historical dialogue is attenuated. The word filtering module is used to weight and accumulate the word frequencies of each word in the multi-round historical dialogue text according to the dialogue weight corresponding to each round of historical dialogue, generate a global weighted word frequency vector, and filter out multiple target words from each word based on the global weighted word frequency vector; the dialogue weight corresponding to each round of historical dialogue is determined based on the decay coefficient and round number corresponding to each round of historical dialogue. The article retrieval module is used to retrieve multiple candidate articles based on the multiple target words; The intent mining module is used to parse the multi-turn historical dialogue text, determine the user's target academic intent for academic retrieval, and filter the multiple candidate articles according to the target academic intent to obtain the target article.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the article filtering method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the article filtering method according to any one of claims 1 to 7.

Images