A news recommendation method, terminal and storage medium

Abstract

The application relates to the technical field of Internet information processing, and discloses a news optimization method, a terminal and a storage medium. The method acquires a news set to be processed, extracts entities and events contained in each piece of news to form a set of information units, constructs a binary covering matrix between the news set and the set of information units, defines a covering function and a redundancy function based on the binary covering matrix, introduces a redundancy penalty factor, constructs a news optimization objective function, converts the news optimization objective function into a linear programming model, constructs a heuristic algorithm based on a Lagrange multiplier to solve the linear programming model, iteratively selects news by calculating marginal comprehensive increments of candidate news until a preset stop condition is met, and obtains an optimal news set. Under the constraint of a limited number of news, the application maximizes the coverage of key enterprise and event information units and simultaneously reduces the redundancy influence caused by the repeated reporting of the same information by multiple pieces of news.

Description

Technical Field

[0001] This invention relates to the field of Internet information processing technology, specifically a news selection method, terminal, and storage medium. Background Technology

[0002] With the explosive growth of internet information, massive amounts of news data have placed a huge reading burden on users. To sift through this vast amount of news, existing technologies primarily focus on news deduplication or text deduplication. For example, existing technologies include methods that use term weight clustering to determine document duplication; methods for real-time deduplication of news or event data from multiple channels; and general technologies that use multi-level fingerprint comparison ("file fingerprint + text fingerprint + paragraph fingerprint") to determine text content duplication. However, these existing technologies mainly focus on text similarity, duplicate documents, and event identification. In practical applications, it is often necessary to present as much key information (such as the companies involved, event types, etc.) as possible within a limited number of news items (i.e., limited page space or recommendation slots). While this can eliminate completely duplicate articles, it does not address the issue of "coverage optimization." Specifically, in actual news collections, different news reports may contain the same company or event type information. If selection is based solely on the information content of a single news item, it can easily lead to multiple selected news items repeatedly reporting the same information units, causing information redundancy and reducing the overall information gain efficiency under limited quantity constraints. For example, when only a limited number of news items can be displayed, if the selected news items all focus on reporting the same trending event while ignoring other important events, it is impossible to maximize the coverage of key information. Therefore, how to maximize the coverage of key information units while effectively reducing the redundancy caused by duplicate information within the constraint of a limited number of news items is a pressing technical problem that needs to be solved.

[0003] In addition, in recent years, some technologies have proposed using optimization models such as integer linear programming to perform extractive summarization or text selection schemes. Such methods usually extract the topic saliency of documents and use continuous values ​​such as cosine similarity between sentences to perform "fuzzy measurement" of sentence coherence and redundancy in order to extract high-quality text summaries. However, when facing the global selection and display scenarios of massive news, there are obvious limitations: (1) Existing summarization or text selection methods mostly present a "one-to-many" membership extraction logic (i.e., extracting multiple sentence summaries from a specific document), while the actual news selection scenario is often a "many-to-many" global coverage logic, which requires presenting as much key corporate or event information as possible under the constraint of a limited number of news items (i.e., limited page space or recommendation slots). (2) "Fuzzy measurement" based on semantic similarity is not only prone to causing a computing power explosion when facing large-scale news data, but also cannot accurately measure whether a specific institution or event has been missed. (3) Under the constraint of a limited number of items, existing optimization solutions (such as standard greedy strategies or static constraint models) tend to repeatedly select news items containing a large amount of hot information, while ignoring other equally important "long-tail events". Summary of the Invention

[0004] To address the technical problems existing in the prior art, this invention provides a news selection method, terminal, and storage medium, aiming to maximize the coverage of key enterprise and event information units under the constraint of a limited number of news items, while reducing the redundancy caused by multiple news items repeatedly reporting the same information.

[0005] To achieve the above objectives, the present invention provides the following technical solution: This invention discloses a method for selecting news sources, comprising the following steps: S1. Obtain the set of news to be processed, extract the entities and events contained in each news item, thereby forming a set of information units; S2. Construct a binary covering matrix between the news set and the information unit set, and use the binary covering matrix to transform the inclusion relationship between the news and the information unit into a discrete matching state; S3. Based on the binary coverage matrix, define a coverage function and a redundancy function. The coverage function is used to represent the number of different information units covered by the selected news subset, and the redundancy function is used to represent the degree to which an information unit is repeatedly covered by multiple news articles. A redundancy penalty factor is introduced to construct a news optimization objective function with the goal of maximizing information unit coverage and minimizing redundant information, and with the upper limit of the number of selected news articles as a constraint. S4. Introduce a binary variable representing the news selection status and the information unit coverage status, and a redundancy variable representing the redundancy degree of the information unit, and transform the news selection objective function into a linear programming model; S5. Construct a heuristic algorithm based on Lagrange multipliers to solve the linear programming model. Use Lagrange multipliers to construct a dynamic penalty mechanism for uncovered information units, and integrate the dynamic penalty mechanism into the marginal comprehensive increment of candidate news for comprehensive evaluation. Iterate to select the news with the largest marginal comprehensive increment until the preset stopping condition is met to obtain the optimal news set.

[0006] As a further improvement to the above scheme, step S2 specifically includes: Let the news collection be The information unit set is ;in, For the first News item , Total number of news items; For the first Each information unit , The total number of information units, i.e., from the news collection The sum of the number of entities and events extracted from all news articles; Define a binary covering matrix Each element in is:

[0007] In the formula, Represents a binary covering matrix The Line 1 Column elements.

[0008] As a further improvement to the above scheme, in step S3, the objective function for news selection is:

[0009] In the formula, As a selected subset of news, This represents the number of selected news items in the set. This is the preset maximum number of news items. and These represent taking the maximum value and taking the minimum value, respectively. This is a redundancy penalty factor.

[0010] As a further improvement to the above scheme, in step S4, the expression of the linear programming model is:

[0011] In the formula, For information units The corresponding redundancy variable, ; This is a binary variable representing whether to select news. ; It is a binary variable, representing an information unit. Whether it has been overwritten.

[0012] As a further improvement to the above scheme, step S5 specifically includes: S51. Introduce Lagrange multipliers to transform the linear programming model into Lagrange form:

[0013] In the formula, and For Lagrange multipliers, It is a penalty for violating the size of the set. Information unit Penalties for not being covered; S52. For each candidate news item Calculate its addition to the set The marginal comprehensive increment after :

[0014] In the formula, Indicates the selected subset The comprehensive value, Indicates the selected subset Covering function, Indicates the selected subset Repeating units in; This is the marginal coverage gain, i.e. Add to collection The resulting increase in information unit coverage; Indicates information redundancy, that is, if... Add to collection , and There is already a lot of repetitive information in the news. Represents a single information unit In the selected news collection The number of times it is mentioned repeatedly; Indicates will Add to collection Then the information unit The resulting gains Representation of information unit In the set The extent to which it has already been covered; S53. Repeat step S52 to iteratively select news. In each iteration, select the news with the largest marginal total increment and add it to the set. The Lagrange multipliers are updated, and iteration stops when a preset condition is met. The set after the iteration ends is... This is the optimal news collection.

[0015] As a further improvement to the above scheme, in step S53, the iteration stops when any of the following conditions are met: Condition 1: Set Number of selected news items Reaching the preset news quantity limit ; Condition 2: Coverage saturation ; Condition 3: The maximum boundary of the current candidate news's comprehensive increment Less than the set threshold ;in, for Remove from The remaining news collection.

[0016] As a further improvement to the above scheme, in step S1, before obtaining the news collection to be processed, multi-source reports, announcements, and regulatory documents are converted into a unified HTML or plain text format, and non-text content is removed, thereby establishing the corpus input for the news collection.

[0017] As a further improvement to the above solution, step S1, which involves extracting the entities and events contained in each news item, specifically includes: Constructing a joint coding layer: A pre-trained model is used to obtain the contextual representation of each word in the news article, expressed as follows:

[0018] In the formula, News Lexical elements in Contextual representation; news collection as , Total number of news items; , News Total number of lexical elements, ; (·) indicates a pre-trained model; Constructing a joint decoding layer: For each word, entity extraction and event extraction are performed based on the context representation, outputting entity category probability vectors and event category probability vectors respectively:

[0019]

[0020] In the formula, Indicates word elements The entity category probability vector; Indicates word elements The event category probability vector; , , and For model weights; The entity set of each news item is obtained by parsing the word sequence and the entity category probability vector of each word. The event set of each news item is obtained by parsing the word sequence and the event category probability vector of each word.

[0021] The present invention also discloses a computer terminal, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the steps of the news optimization method as described above.

[0022] The present invention also discloses a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the news selection method described above.

[0023] Compared with the prior art, the beneficial effects of the present invention are: 1. The news selection method disclosed in this invention effectively solves the problem of optimal selection of large-scale news sets under limited constraints. First, the method extracts entities and events to construct information units and a binary coverage matrix, upgrading traditional text deduplication to core semantic coverage. Second, based on this matrix, an objective function containing a coverage function and a redundancy function is constructed. By introducing binary variables and redundancy variables, the nonlinear logic, which is difficult to solve directly, is cleverly transformed into a standard linear programming model, providing a computable foundation for complex optimization problems. Finally, addressing the problem of combinatorial explosion in solving large-scale data, this invention constructs a heuristic algorithm based on Lagrange multipliers, iteratively selecting news by calculating the marginal comprehensive increment including coverage gain, redundancy penalty, and constraint penalty. Therefore, the overall solution can quickly find a news set close to the optimal solution under limited news quantity constraints, maximizing the coverage of key enterprise and event information units while significantly reducing the redundancy caused by multiple news reports repeatedly reporting the same information.

[0024] 2. Traditional greedy algorithms often tend to repeatedly select news items containing a large amount of trending information, resulting in insufficient coverage of long-tail information and extremely high redundancy. This invention introduces a redundancy penalty factor and Lagrange multipliers (especially a dynamic penalty mechanism for uncovered information), enabling the algorithm to effectively suppress repeated selection of already covered trending information during iteration, while maintaining a high sensitivity to scarce long-tail information units. This dynamic adjustment mechanism ensures that the final selected news collection has better content diversity and information completeness.

[0025] 3. In practical application scenarios (such as limited app recommendation slots or morning news pages), this invention can significantly reduce the number of news items to be selected while achieving the same information coverage goal. This means that the system can deliver more comprehensive and richer core information with fewer news articles, avoiding the occupation of limited display resources by highly homogenized and repetitive information, thereby effectively improving distribution efficiency and user experience. Attached Figure Description

[0026] Figure 1 This is a flowchart of the news selection method in Embodiment 1 of the present invention.

[0027] Figure 2 This is a comparison chart showing the change in coverage of the optimized algorithm and the baseline algorithm in Embodiment 1 of the present invention with the number of selections.

[0028] Figure 3 This is a comparison chart showing the redundancy of the optimized algorithm and the baseline algorithm in Embodiment 1 of the present invention as a function of the number of selections.

[0029] Figure 4 This is a schematic diagram of the computer terminal structure in Embodiment 2 of the present invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Example 1

[0032] Please see Figure 1 This embodiment provides a news selection method, which includes the following steps, namely S1 to S5.

[0033] S1. Obtain the set of news to be processed, extract the entities and events contained in each news item, thereby forming a set of information units.

[0034] In step S1, before obtaining the news collection to be processed, multi-source reports, announcements, and regulatory documents are converted into a unified HTML or plain text format, and non-text content is removed, thereby establishing the corpus input for the news collection.

[0035] Since multi-source reports and announcements often contain a large amount of useless information such as HTML tags, layout code, or advertising links, removing non-text content can effectively reduce semantic drift and noise interference in subsequent entity and event extraction from news articles. In this embodiment, "entity" typically refers to key companies, institutions, or important figures mentioned in the news (e.g., "a certain company," "China Securities Regulatory Commission"); "event" refers to the core actions, states, or macro-level situations reported in the news (e.g., "release of financial reports," "changes in senior management," "product launch," "interest rate adjustments"). Unifying and abstracting these discrete entities and events into information units helps to break down the differences in wording across different news texts, allowing subsequent models to move beyond simple textual similarity and directly focus on the core substantive content conveyed by the news.

[0036] In step S1, extracting the entities and events contained in each news item specifically includes: Constructing a joint coding layer: A pre-trained model is used to obtain the contextual representation of each word in the news article, expressed as follows:

[0037] In the formula, News Lexical elements in Contextual representation; news collection as , Total number of news items; , News Total number of lexical elements, ; (·) indicates a pre-trained model; Constructing a joint decoding layer: For each word, entity extraction and event extraction are performed based on the context representation, outputting entity category probability vectors and event category probability vectors respectively:

[0038]

[0039] In the formula, Indicates word elements The entity category probability vector; Indicates word elements The event category probability vector; , , and For model weights; The entity set of each news item is obtained by parsing the word sequence and the entity category probability vector of each word, and is represented as follows: , For News A collection of entities Let be the total number of entities in the set; the event set of each news item is obtained by parsing the word sequence and the event category probability vector of each word, denoted as . , For News A collection of events The total number of events in the set.

[0040] The combined extraction result for each news item is: .

[0041] Information unit set is .

[0042] The steps described above—obtaining deep semantic features through a pre-trained model (such as BERT) and simultaneously predicting the probabilities of lexical entities and events using a joint decoding layer—can efficiently and accurately reduce the dimensionality of originally unstructured natural language text and map it into a structured set of information unit features, providing quantified data input for subsequent combinatorial optimization.

[0043] S2. Construct a binary covering matrix between the news set and the information unit set, and use the binary covering matrix to transform the inclusion relationship between the news and the information unit into a discrete matching state.

[0044] Step S2 specifically includes: Let the news collection be The information unit set is ;in, For the first News item , Total number of news items; For the first Each information unit , The total number of information units, i.e., from the news collection The sum of the number of entities and events extracted from all news articles.

[0045] Define a binary covering matrix Each element in is:

[0046] In the formula, Represents a binary covering matrix The Line 1 Column elements.

[0047] To more intuitively understand binary covering matrices The construction process is illustrated with a specific example: Suppose the current news collection to be processed... Entity set event collection The overall vector order of the information unit set is: If the news The entity was mentioned and events Then the matrix The first row contains 1s, and the rest are 0s, so the vector representation is [1,0,0,1,0]. Similarly, vectors for all news items are constructed, resulting in the final matrix. The contents of Table 1 can be used to illustrate this: Table 1: Binary Covering Matrix Example

[0048] As can be seen from Table 1: news Mentioning physical entities and events ; news Mentioning physical entities , and events ; news Mentioning physical entities , ; news Mentioning physical entities and events , .

[0049] With constraint K=2, by choosing and Optimal coverage can be achieved.

[0050] S3. Based on the binary coverage matrix, define a coverage function and a redundancy function. The coverage function is used to represent the number of different information units covered by the selected news subset, and the redundancy function is used to represent the degree to which an information unit is repeatedly covered by multiple news articles. A redundancy penalty factor is introduced to construct a news optimization objective function with the goal of maximizing information unit coverage and minimizing redundant information, and with the upper limit of the number of selected news articles as a constraint.

[0051] In step S3, the objective function for news selection is:

[0052] In the formula, As a selected subset of news, This represents the number of selected news items in the set. This is the preset maximum number of news items. and These represent taking the maximum value and taking the minimum value, respectively. This is a redundancy penalty factor.

[0053] In real-world news distribution or display scenarios, if a greedy selection is performed directly based on the number of information units contained, it is highly likely that multiple lengthy news articles reporting on the same major trending event will be selected, resulting in highly homogenized information seen by users. Therefore, the objective function designed in this step has a clear physical meaning: The coverage function with min logic aims to reflect the principle of effective coverage, meaning that the information gain contribution of an information unit is only calculated once (maximum 1), regardless of how many times it is repeatedly reported by selected news items. The redundancy function with max logic aims to reflect the principle of redundancy cost, meaning that the first reporting of an information unit is an effective gain, but from the second report onwards, each additional repeated coverage is included in the redundancy score, thus penalizing the overall score. The redundancy penalty factor is a user-defined value used to flexibly adjust the system's tolerance for duplicate information. The larger the value, the more the model tends to select news with significant differences in content and diverse perspectives; conversely, the model may allow for a moderate repetition of some key and trending information.

[0054] S4. Introduce a binary variable representing the news selection status and the information unit coverage status, and a redundancy variable representing the redundancy degree of the information unit, and transform the news selection objective function into a linear programming model.

[0055] In step S4, the expression for the linear programming model is:

[0056] In the formula, For information units The corresponding redundancy variable, ; This is a binary variable representing whether to select news. ; It is a binary variable, representing an information unit. Whether it has been overwritten.

[0057] While the objective function constructed in step S3 logically aligns with business requirements, it contains non-linear min and max operations. This type of combinatorial optimization problem with non-linear terms (similar to a complex set covering problem) is NP-hard, making it extremely difficult for computers to solve directly when faced with real-world industrial scenarios involving thousands of news articles. Therefore, the core objective of step S4 is to linearize the non-linear logic. This is achieved by introducing variables representing specific business actions. That is, using binary variables Mapping the decision of whether or not to select news items using binary variables. Mapping status feedback on whether certain information has been presented to the user, and using continuous variables. By quantifying the specific number of repetitions, the originally complex nonlinear objective is equivalently transformed into a standard linear inequality constraint. This transformation retains the original core business principle of "maximizing coverage and minimizing redundancy," while enabling the model to adapt to the standard linear programming solution framework.

[0058] S5. Construct a heuristic algorithm based on Lagrange multipliers to solve the linear programming model. Use Lagrange multipliers to construct a dynamic penalty mechanism for uncovered information units, and integrate the dynamic penalty mechanism into the marginal comprehensive increment of candidate news for comprehensive evaluation. Iterate to select the news with the largest marginal comprehensive increment until the preset stopping condition is met to obtain the optimal news set.

[0059] Step S5 specifically includes S51 to S53.

[0060] S51. Introduce Lagrange multipliers to transform the linear programming model into Lagrange form:

[0061] In the formula, and For Lagrange multipliers, It is a penalty for violating the size of the set. Information unit Penalties for not being covered.

[0062] Since pure integer linear programming models may still encounter the problem of combinatorial explosion leading to excessively long solution times under extremely large-scale data, this step introduces the Lagrange relaxation technique. Specifically, multipliers This is equivalent to a dynamically adjusted "threshold control fee," which applies when the size of the selected set approaches the set upper limit K. The multiplier will increase accordingly, imposing harsher penalties on subsequent attempts to include news; This is equivalent to a bounty mechanism for scarce information; if a certain information unit... It has not been covered by the current set for a long time, and its corresponding The weight will increase, thus driving the algorithm to prioritize news containing this niche / long-tail information in subsequent iterations, ensuring comprehensive information coverage.

[0063] S52. For each candidate news item Calculate its addition to the set The marginal comprehensive increment after :

[0064] In the formula, Indicates the selected subset The comprehensive value, Indicates the selected subset Covering function, Indicates the selected subset Repeating units in; This is the marginal coverage gain, i.e. Add to collection The resulting increase in information unit coverage; Indicates information redundancy, that is, if... Add to collection , and There is already a lot of repetitive information in the news. Represents a single information unit In the selected news collection The number of times it is mentioned repeatedly; Indicates will Add to collection Then the information unit The resulting gains Representation of information unit In the set The extent to which it has already been covered.

[0065] This marginal incremental This is the core basis for news selection in this invention. It not only evaluates the absolute information coverage benefit brought by adding a new news item, but also simultaneously deducts the redundancy cost caused by content duplication and the constraint penalty caused by approaching the quantity limit in the underlying logic. It is an accurate measure of the true comprehensive value of candidate news.

[0066] S53. Repeat step S52 to iteratively select news. In each iteration, select the news with the largest marginal total increment and add it to the set. The Lagrange multipliers are updated, and iteration stops when a preset condition is met. The set after the iteration ends is... This is the optimal news set. The iteration stops when any of the following conditions are met: Condition 1: Set Number of selected news items Reaching the preset news quantity limit (This is a rigid constraint on the number of physical layouts or recommended positions.) Condition 2: Coverage saturation (This is the ideal business state, meaning that all extracted key information points have been presented to the user, and there is no need to add extra news to waste reading time.) Condition 3: The maximum boundary of the current candidate news's comprehensive increment Less than the set threshold ;in, for Remove from The remaining news set (this is the utility termination condition, indicating that the remaining news can provide very limited fresh information, and the redundancy penalty even exceeds the coverage benefit due to high repetition, resulting in a very low or negative overall increment. At this point, continuing to select news has lost its business value, and the algorithm adaptively stops).

[0067] To verify the effectiveness of the news selection method of the present invention, the following comparative experiment is also provided in this embodiment.

[0068] 1. Sources and Environment of Experimental Data

[0069] To verify the improvement in information coverage effect of the news selection method of this invention under the constraint of a limited number of news items, this experiment uses a synthetic news dataset for testing. The experimental data is generated as follows: There are a total of N=200 news items collected, and each news item corresponds to several information units (abstract representations such as entities and events), with a total of Q=100 information units; Information units are generated according to hot topics and long tail distribution, with 15% of information units being hot topics (multiple news items appearing repeatedly) and the rest being long tail information (a small number of news items appearing). Each news item randomly contains several hot topics and long tail topics, with an average of 10 hot topics and 1 long tail topic. The experiment was run in a Python 3.10 environment and relied on common libraries such as NumPy, Pandas, and Matplotlib. It can be executed on operating systems including Windows, macOS, and Linux.

[0070] 2. Comparison Algorithm

[0071] This experiment compares two types of news selection algorithms: (1) Baseline algorithm (greedy selection): Each time, select the news with the most newly added coverage information until the upper limit of the number of news K=50 is reached.

[0072] (2) Optimization algorithm (coverage increment - redundancy penalty): When selecting news each time, both the new coverage increment and the redundancy penalty are considered, that is, the selection is made to maximize the overall value. The biggest news is that the penalty coefficient is set to λ=1.5.

[0073] 3. Experimental evaluation indicators

[0074] To measure the optimal performance of news collections, the following metrics are introduced: Information coverage: The total number of different information units covered by the selected news collection; Redundancy: The total number of information units that are repeatedly covered in the selected news collection; Coverage curve: As the number of news items increases, the cumulative coverage and redundancy are statistically analyzed to assess the dynamic performance of the algorithm selection process.

[0075] The experimental results are as follows: Table 2: Comparison of Coverage and Redundancy under a Fixed Number of News Items

[0076] As shown in Table 2, when the same news item limit K=50 is reached, the coverage of the optimized algorithm is significantly higher than that of the baseline greedy algorithm. For example, the baseline method covers more hot-topic information but insufficient long-tail information; the optimized algorithm, while suppressing redundancy, can cover hot-topic and long-tail information units more evenly; the baseline greedy method is prone to multiple news items repeatedly covering the same information unit during the selection process, resulting in high redundancy; the optimized algorithm effectively reduces the number of repeated coverages by penalizing redundancy.

[0077] Table 3: Number of News Items Required to Achieve Specified Coverage Targets

[0078] As shown in Table 3, to achieve the same information coverage target, the baseline algorithm requires significantly more news items to be selected than the optimized algorithm of this invention. For example, when the coverage target is 40 information units, the baseline algorithm needs to select 21 news items, while the method of this invention only needs 8. When the coverage target is increased to 50 information units, the number of news items required by the baseline algorithm increases to 33, while the method of this invention still only needs 11. These results demonstrate that by introducing coverage increment and redundancy suppression mechanisms in the news selection process, this invention can achieve higher information coverage efficiency with fewer news items, thereby effectively improving the utilization rate of limited display resources.

[0079] Please see Figure 2It can be seen that, under the same number of selected news items, the coverage of the optimization algorithm is consistently higher than that of the baseline greedy algorithm, and the coverage advantage gradually expands as the number of news items increases. Especially in the stage where the number of news items is small, the optimization algorithm can cover more information units with fewer news items, indicating that this method can prioritize news items that contribute more to the overall information coverage during the news selection process, thereby significantly improving the information coverage efficiency under the constraint of a limited number of news items.

[0080] Please see Figure 3 The results show that as the number of selected news items increases, the redundancy of the baseline greedy algorithm accumulates rapidly, while the redundancy of the optimized algorithm increases at a significantly lower rate. This indicates that the optimized algorithm effectively suppresses the repeated selection of already covered information units during the news selection process, reducing the accumulation of information redundancy and thus preventing limited display resources from being occupied by duplicate information.

[0081] In summary, under the constraint of a limited number of news items, the preferred method of this invention can balance information coverage and redundancy control, achieving a better news information combination effect than the baseline greedy strategy, thus verifying the effectiveness of this invention in news selection and coverage optimization.

[0082] Example 2

[0083] This embodiment provides a computer terminal, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the news selection method as described in Embodiment 1.

[0084] like Figure 4 As shown, the computer terminal provided in this embodiment includes: at least one processor 101, and a memory 102 connected to at least one processor 101. This embodiment does not limit the specific connection medium between the processor 101 and the memory 102. Figure 4 The example shown is the connection between processor 101 and memory 102 via bus 100. Bus 100 is... Figure 4 The connections between other components are shown in bold lines and are for illustrative purposes only, not as limiting information. Bus 100 can be divided into address bus, data bus, control bus, etc., for ease of representation. Figure 4 The bus is represented by a single thick line, but this does not indicate that there is only one bus or one type of bus. Alternatively, the processor 101 may also be called a controller; there is no restriction on the name.

[0085] In this embodiment, the memory 102 stores instructions that can be executed by at least one processor 101. The at least one processor 101 can execute the aforementioned method by executing the instructions stored in the memory 102.

[0086] The processor 101 is the control center of the device. It can connect to various parts of the control device through various interfaces and lines. By running or executing instructions stored in memory 102 and calling data stored in memory 102, the processor can perform various functions and process data, thereby monitoring the device as a whole.

[0087] In one possible design, processor 101 may include one or more processing units. Processor 101 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user interface, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may also not be integrated into processor 101. In some embodiments, processor 101 and memory 102 may be implemented on the same chip; in some embodiments, they may also be implemented on separate chips.

[0088] Processor 101 can be a general-purpose processor, such as a central processing unit (CPU), digital signal processor, application-specific integrated circuit, field-programmable gate array or other programmable logic device, discrete gate or transistor logic device, or discrete hardware component, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in the embodiments. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the preferred method for news disclosing embodiment 1 can be directly manifested as being executed by a hardware processor, or executed by a combination of hardware and software modules in processor 101.

[0089] Memory 102, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules. Memory 102 may include at least one type of storage medium, such as flash memory, hard disk, multimedia card, card-type memory, random access memory (RAM), static random access memory (SRAM), programmable read-only memory (PROM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), magnetic storage, magnetic disk, optical disk, etc. Memory 102 can be any other medium capable of carrying or storing desired program code in the form of instructions or data structures that can be accessed by a computer, but is not limited thereto. In this embodiment, memory 102 can also be a circuit or any other device capable of implementing storage functions for storing program instructions and / or data.

[0090] By designing and programming the processor 101, the code corresponding to the news selection method described in the foregoing embodiments can be embedded into the chip, thereby enabling the chip to execute the code during operation. Figure 1 The steps of the news selection method are shown. How to design and program the processor 101 is a technique well-known to those skilled in the art and will not be described further here.

[0091] Example 3

[0092] This embodiment provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the news selection method as described in Embodiment 1.

[0093] The computer-readable storage medium may include flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the storage medium may be an internal storage unit of a computer device, such as the hard disk or memory of the computer device. In other embodiments, the storage medium may also be an external storage device of the computer device, such as a plug-in hard disk, smart memory card, secure digital card, flash memory card, etc., provided on the computer device. Of course, the storage medium may include both internal storage units and external storage devices of the computer device. In this embodiment, the memory is typically used to store the operating system and various application software installed on the computer device. In addition, the memory can also be used to temporarily store various types of data that have been output or will be output.

[0094] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for selecting news sources, characterized in that, Includes the following steps: S1. Obtain the set of news to be processed, extract the entities and events contained in each news item, thereby forming a set of information units; S2. Construct a binary covering matrix between the news set and the information unit set, and use the binary covering matrix to transform the inclusion relationship between the news and the information units into a discrete matching state; step S2 specifically includes: Let the news collection be The information unit set is ;in, For the first News item , Total number of news items; For the first Each information unit , The total number of information units, i.e., from the news collection The sum of the number of entities and events extracted from all news articles; Define a binary covering matrix Each element in is: In the formula, Represents a binary covering matrix The Line 1 Column elements; S3. Based on the binary coverage matrix, define a coverage function and a redundancy function. The coverage function is used to represent the number of different information units covered by the selected news subset, and the redundancy function is used to represent the degree to which an information unit is repeatedly covered by multiple news articles. A redundancy penalty factor is introduced to construct a news optimization objective function with the goal of maximizing information unit coverage and minimizing redundant information, and with the upper limit of the number of selected news articles as a constraint. S4. Introduce a binary variable representing the news selection status and the information unit coverage status, and a redundancy variable representing the redundancy degree of the information unit, and transform the news selection objective function into a linear programming model; S5. Construct a heuristic algorithm based on Lagrange multipliers to solve the linear programming model. Use Lagrange multipliers to construct a dynamic penalty mechanism for uncovered information units, and integrate the dynamic penalty mechanism into the marginal comprehensive increment of candidate news for comprehensive evaluation. Iterate to select the news with the largest marginal comprehensive increment until the preset stopping condition is met to obtain the optimal news set.

2. The news selection method according to claim 1, characterized in that, In step S3, the objective function for news selection is: In the formula, As a selected subset of news, This represents the number of selected news items in the set. This is the preset maximum number of news items. and These represent taking the maximum value and taking the minimum value, respectively. This is a redundancy penalty factor.

3. The news selection method according to claim 2, characterized in that, In step S4, the expression for the linear programming model is: In the formula, For information units The corresponding redundancy variable, ; This is a binary variable representing whether to select news. ; It is a binary variable, representing an information unit. Whether it has been overwritten.

4. The news selection method according to claim 3, characterized in that, Step S5 specifically includes: S51. Introduce Lagrange multipliers to transform the linear programming model into Lagrange form: In the formula, and For Lagrange multipliers, It is a penalty for violating the size of the set. Information unit Penalties for not being covered; S52. For each candidate news item Calculate its addition to the set The marginal comprehensive increment after : In the formula, Indicates the selected subset The comprehensive value, Indicates the selected subset Covering function, Indicates the selected subset Repeating units in; This is the marginal coverage gain, i.e. Add to collection The resulting increase in information unit coverage; Indicates information redundancy, that is, if... Add to collection , and There is already a lot of repetitive information in the news. Represents a single information unit In the selected news collection The number of times it is mentioned repeatedly; Indicates will Add to collection Then the information unit The resulting gains Representation of information unit In the set The extent to which it has already been covered; S53. Repeat step S52 to iteratively select news. In each iteration, select the news with the largest marginal total increment and add it to the set. The Lagrange multipliers are updated, and iteration stops when a preset condition is met. The set after the iteration ends is... This is the optimal news collection.

5. The news selection method according to claim 4, characterized in that, In step S53, the iteration stops when any of the following conditions are met: Condition 1: Set Number of selected news items Reaching the preset news quantity limit ; Condition 2: Coverage saturation ; Condition 3: The maximum marginal aggregate increment of the current candidate news. Less than the set threshold ;in, for Remove from The remaining news collection.

6. The news selection method according to claim 1, characterized in that, In step S1, before obtaining the news collection to be processed, multi-source reports, announcements, and regulatory documents are converted into a unified HTML or plain text format, and non-text content is removed, thereby establishing the corpus input for the news collection.

7. The news selection method according to claim 1, characterized in that, In step S1, extracting the entities and events contained in each news item specifically includes: Constructing a joint coding layer: A pre-trained model is used to obtain the contextual representation of each word in the news article, expressed as follows: In the formula, News Lexical elements in Contextual representation; news collection as , Total number of news items; , News Total number of lexical elements, ; (·) indicates a pre-trained model; Constructing a joint decoding layer: For each word, entity extraction and event extraction are performed based on the context representation, outputting entity category probability vectors and event category probability vectors respectively: In the formula, Indicates word elements The entity category probability vector; Indicates word elements The event category probability vector; , , and For model weights; The entity set of each news item is obtained by parsing the word sequence and the entity category probability vector of each word. The event set of each news item is obtained by parsing the word sequence and the event category probability vector of each word.

8. A computer terminal, 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 computer program, it implements the steps of the news selection method as described in any one of claims 1 to 7.

9. 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 news selection method as described in any one of claims 1 to 7.

Images