An archive full-text intelligent retrieval method based on semantic analysis

Abstract

The application provides a file full-text intelligent retrieval method based on semantic analysis, comprising: obtaining industry background text to which the initial term group belongs, determining the field distribution density of the initial term group by counting the co-occurrence frequency of the initial term group in the industry background text; filtering out terms whose core meanings have not substantially changed under the support of high field distribution density by analyzing the field distribution density of the initial term group and the expression change of the terms in cross-era archives, to form a fixed term group; and compressing the era semantic dispersion degree of each term in the fixed term group according to the field distribution density of the term in proportion, and obtaining the meaning anchoring strength of the fixed term group in a single industry archive system to which the archive retrieval text belongs.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to an intelligent full-text retrieval method for archives based on semantic analysis. Background Technology

[0002] Full-text intelligent retrieval of archives is the core hub for uncovering the value of historical documents. Accurately understanding and retrieving textual content spanning long periods has become crucial for utilizing archival resources. Conventionally, the greater the time span of the same technical term in archives, the more significant its semantic shift across different eras. This increased span directly renders the criteria for determining term similarity matching in cross-era searches ineffective. For example, in railway construction archives, early archives described "roadbed" as an earthen embankment structure mainly constructed by manual ramming, while modern archives describe "roadbed" as an engineered foundation system that has been mechanically compacted in layers and equipped with systematic drainage and waterproofing facilities. The two differ significantly in their construction methods and structural descriptions. Similarly, the term "rail sleeper" in the archives referred to wooden sleepers in the early archives, while modern archives include prestressed concrete sleepers and synthetic material sleepers. The names and specifications of materials have changed over time. Although these terms have changed significantly in terms of specific technical implementation, material composition, or process details over the years, their functional positioning and core referents in the context of railway engineering have remained consistent. "Roadbed" has always referred to the foundation structure under the track that bears the load transfer and stability function, and "rail sleeper" has always referred to the supporting components laid laterally on the track bed to fix the rail spacing and distribute wheel loads. Although these terms differ in their surface descriptions, they have not fundamentally shifted at the professional semantic level. If conventional benchmarks that rely solely on surface vocabulary or general semantic models are used, the functional anchoring effect within the aforementioned industry context may be ignored, leading to misclassification of substantially related archives as lowly relevant and causing retrieval omissions. Existing mechanisms fail to recognize this semantic constraint effect brought about by industry background, mechanically assuming that the passage of time inevitably leads to significant changes in meaning, resulting in matching confusion. Therefore, in cross-era archive retrieval, accurately assessing the industry distribution density of the archives to which terms belong, identifying term groups that maintain functional and referential consistency despite changes in surface expression, and then dynamically adjusting the similarity matching criteria, becomes a key issue in improving the accuracy of intelligent full-text archive retrieval. Summary of the Invention

[0003] This invention provides a semantic analysis-based intelligent full-text retrieval method for archives, mainly including:

[0004] At least one initial term group is extracted from archival retrieval text spanning multiple eras, and the first and last occurrence times of each term in the archives are recorded to obtain the era span corresponding to the initial term group in the archival retrieval text. Based on the era span, the context of the initial term group in each era subset is vectorized, and the cosine distance between the same term in different era vectors is used to characterize the semantic dispersion of the term in the time dimension. The era semantic dispersion of the initial term group is then summarized. The industry background text to which the initial term group belongs is obtained, and the frequency of co-occurrence of the initial term group in the industry background text is statistically analyzed to determine... The domain distribution density of the initial term group is determined; by analyzing the domain distribution density of the initial term group and the changes in the expression of the terms in cross-era archives, a fixed term group is selected; the temporal semantic dispersion of the fixed term group is compressed proportionally according to the domain distribution density of each term, and the meaning anchoring strength of the fixed term group under the single industry archive system to which the archive retrieval text belongs is obtained; the pre-established initial similarity matching interval is adjusted according to the meaning anchoring strength, and cross-era semantic matching of the fixed term group is performed in the full text of the archive using the adjusted initial similarity matching interval, and document fragments that meet the matching conditions are returned to form the target archive retrieval results.

[0005] Furthermore, the step of extracting at least one initial term group from archival retrieval text spanning multiple eras includes: dividing the full text into several era subsets according to the year the archives were filed; extracting professional terms in each era subset using named entity recognition through a conditional random field; and merging professional terms pointing to the same functional object in different era subsets into the same term entry based on the consistency of the functional objects referred to by the professional terms, thereby forming the initial term group.

[0006] Furthermore, obtaining the epoch span corresponding to the initial term group in the archive retrieval text includes: for each term in the initial term group, traversing the occurrence positions of the term in each epoch subset, recording the archiving year corresponding to the first occurrence position as the start time point, recording the archiving year corresponding to the last occurrence position as the end time point, taking the earliest value among all term start time points as the group start point, and the latest value among the end time points as the group end point, and determining the epoch span by the difference between the group end point and the group start point.

[0007] Furthermore, the step of vectorizing the context of the initial term group in each decade subset, and using the cosine distance between the vectors of the same term in different decades to characterize the semantic dispersion of the term in the time dimension, includes: for each term in the initial term group, extracting a sequence of context words within a preset window length centered on the term in each decade subset, and using a word vector algorithm to independently vectorize the corpus of each decade subset to obtain the semantic vector corresponding to the term in each decade subset; calculating the cosine distance between any two terms paired in the semantic vectors of different decade subsets, and taking the average of the cosine distances of all terms in the initial term group to form the decade semantic dispersion.

[0008] Furthermore, obtaining the industry background text to which the initial term group belongs includes: compiling industry standards, engineering manuals, professional journals, and engineering reports published during the same period from the time span covered by the era into the industry background text; using the occurrence position of any term in the initial term group as an anchor point to extend a preset sentence length forward and backward as a co-occurrence window; any two terms in the initial term group form a term pair; counting the number of co-occurrences of each term pair within the co-occurrence window; and taking the average of the co-occurrence frequencies of all term pairs in the initial term group after normalization to determine the domain distribution density.

[0009] Furthermore, the step of analyzing the domain distribution density of the initial terminology group and the changes in the expression of terms in cross-generational archives to screen out fixed terminology groups includes: for each term in the initial terminology group, extracting the limiting descriptions of the form, structure, and process of the sentence segment containing the term from each decade subset, and sorting them by the year of archiving to form a cross-generational expression change sequence; extracting the functional positioning and object description fragments referred to by each decade subset in the cross-generational expression change sequence, and performing coverage comparison by keyword set; if the referent and function relationship of the keyword set remains consistent, the core meaning has not changed substantially; otherwise, the core meaning has changed substantially; terms whose domain distribution density is higher than a preset density threshold and whose core meaning has not changed substantially are included in the fixed terminology group.

[0010] Furthermore, the step of compressing the temporal semantic dispersion of a fixed term group proportionally based on the domain distribution density of each term, and summarizing the meaning anchoring strength of the fixed term group within the single industry archive system to which the archive retrieval text belongs, includes: for each term in the fixed term group, retrieving concurrently published revised versions of industry standards and revised versions of engineering manuals, extracting collocation entries from different versions to form a collocation change record; based on the temporal subset pairs corresponding to the collocation change record, selecting corresponding components from the cosine distance set of temporal semantic dispersion and recording them as variable collocation offsets; subtracting the variable collocation offset from the temporal semantic dispersion to obtain a stable anchoring residual; dividing the domain distribution density of each term by the highest domain distribution density to obtain a normalized weight; and determining the meaning anchoring strength based on the normalized weight and the stable anchoring residual.

[0011] Furthermore, the process of summarizing the meaning anchoring strength of the fixed term group within the single industry archive system to which the retrieved text belongs includes: for each term in the fixed term group, comparing the domain distribution density of the term with the median value of the domain distribution density within the fixed term group; if the value is higher than the median value, the corresponding component of the term in the chronological semantic dispersion is classified as a stable anchoring source; otherwise, it is classified as a variable collocation source; traversing the revised versions of industry standards and engineering manuals, extracting the collocation term change entries for each version, recording the cosine distance component of the contribution of the chronological subset corresponding to the collocation term change entry as an offset component, and subtracting the offset component from the variable collocation source to obtain the variable collocation residual value; summing the cosine distance component corresponding to the stable anchoring source with the variable collocation residual value to form a dispersion suppression term, using the difference between a preset benchmark value and the dispersion suppression term as the suppression contribution, dividing the domain distribution density of each term by the highest domain distribution density to obtain a normalized weight, and weighting and summing the suppression contribution based on the normalized weight to form the meaning anchoring strength.

[0012] Furthermore, adjusting the pre-established initial similarity matching interval according to the meaning anchoring strength includes: lowering the lower bound of the pre-established initial similarity matching interval, wherein the lower bound is positively correlated with the meaning anchoring strength, and the higher the meaning anchoring strength, the greater the lower bound. The adjusted initial similarity matching interval is obtained after the lower bound of the initial similarity matching interval is lowered.

[0013] Furthermore, the step of using the adjusted initial similarity matching interval to perform cross-era semantic matching on the fixed term group in the full text of the archive, returning document fragments that meet the matching conditions, and forming the target archive retrieval results includes: traversing the archive text with terms in the fixed term group as retrieval anchors, extracting continuous document fragments containing the terms, averaging the semantic vectors of each term in the fixed term group to obtain a group-level semantic vector, calculating the cosine similarity value between the semantic vector of the document fragment and the group-level semantic vector, and including document fragments whose values ​​fall within the adjusted initial similarity matching interval in the target archive retrieval results.

[0014] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0015] This invention discloses an intelligent full-text retrieval method for archives based on semantic analysis. Addressing the problem of decreased retrieval accuracy due to semantic drift of terms in cross-era archive retrieval, it identifies fixed term groups by constructing a two-dimensional analysis framework of temporal semantic dispersion and domain distribution density. First, this invention extracts the temporal span of terms in the archives and vectorizes the context of different eras. Cosine distance is calculated to characterize the degree of semantic distribution dispersion. Simultaneously, the co-occurrence frequency of terms in industry background text is statistically analyzed to determine domain distribution density, filtering out terms with stable core meanings supported by high domain density. For these fixed term groups, this invention identifies variable collocation components introduced by industry standard updates, compresses temporal semantic dispersion according to the domain distribution density ratio, and eliminates semantic shifts caused by changes in word collocation, forming the semantic anchoring strength of terms within a single industry archive system. This anchoring strength is used to dynamically adjust the similarity matching interval, achieving accurate semantic matching of cross-era archives, effectively solving the retrieval omission problem caused by changes in terminology, and significantly improving the recall and accuracy of archive retrieval. Attached Figure Description

[0016] Figure 1 This is a flowchart of a semantic analysis-based intelligent full-text retrieval method for archives according to the present invention.

[0017] Figure 2 This is a schematic diagram of a semantic analysis-based intelligent full-text retrieval method for archives according to the present invention.

[0018] Figure 3 This is another schematic diagram of a semantic analysis-based intelligent full-text retrieval method for archives according to the present invention. Detailed Implementation

[0019] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0020] like Figures 1-3 This embodiment of a semantic analysis-based intelligent full-text retrieval method for archives may specifically include:

[0021] Step S101: Extract at least one initial term group from the archival retrieval text spanning multiple eras, and record the first and last appearance times of each term in the archives, thereby obtaining the era span corresponding to the initial term group in the archival retrieval text.

[0022] For the archival retrieval text spanning multiple eras, the full text is divided into several era subsets based on the archival filing year. Within each era subset, named entity recognition is performed using a conditional random field (CRF) to extract professional terms related to railway engineering. Based on the consistency of the functional objects referred to by these terms in the railway engineering context, professional terms pointing to the same functional object in different era subsets are grouped into the same term entry, forming at least one initial term group. For each term in the initial term group, the occurrence positions of the term in the several era subsets are traversed. The archival year corresponding to the first occurrence position of the term in the archive is recorded as the start time point, and the archival year corresponding to the last occurrence position is recorded as the end time point. The earliest value among all the start times of all terms in the initial term group is taken as the group starting point, and the latest value among all the end times of all terms is taken as the group ending point. The difference between the group ending point and the group starting point determines the era span corresponding to the initial term group in the archival retrieval text.

[0023] In one implementation, the cross-era archival retrieval text originates from archives in the railway engineering field, encompassing various historical documents such as roadbed construction records, track laying technical reports, bridge and tunnel engineering archives, and locomotive and rolling stock operation documents. The archival retrieval text uses the year of archiving as the carrier of time attributes, with each document containing an identifiable archiving year field. Before processing the archival retrieval text, the original archival text undergoes conventional word segmentation and character normalization processing to obtain a standardized text set suitable for subsequent business processing.

[0024] Specifically, the chronological segmentation is based on the year of archiving, and the archive retrieval text is divided into several consecutive chronological subsets according to a preset chronological interval length.

[0025] Preferably, the length of the time interval is 10 or 20 years to ensure the relative stability of the archival context within each time interval subset.

[0026] It should be noted that the conditional random field is a probabilistic graphical tool for sequence labeling. The input is a sequence of characters or words after word segmentation, and the output is a sequence of entity labels corresponding to each position.

[0027] In one possible implementation, the label set of the Conditional Random Field (CRF) includes entity labels for railway engineering, such as roadbed, sleepers, rails, ballast, fasteners, and bridge / tunnel structures, as well as non-entity labels. Within each time-based subset, the archived text is fed into the CRF sentence by sentence. For each word position within a sentence, the entity label for that position is output by combining its literal features, part-of-speech features, and contextual features of adjacent positions, thereby extracting railway engineering terminology from the time-based subset.

[0028] For example, in the early epoch subset, the conditional random field can identify technical terms such as "rammed earth subgrade," "wooden sleeper," and "fishplate"; in the modern epoch subset, the conditional random field can identify technical terms such as "concrete subgrade," "prestressed concrete sleeper," and "elastic fastener." Furthermore, the consistency of the functional objects is based on functional anchoring within the railway engineering context.

[0029] Specifically, the specialized terms extracted from different epochal subsets are compared according to their functional roles in railway engineering. Terms referring to the foundation structure beneath the track that bears the load transfer and stabilization functions are grouped into the same term entry; terms referring to the support components laid laterally on the ballast bed to fix the rail spacing and distribute wheel loads are grouped into another term entry. In this way, "rammed earth subgrade" in the early epochal subset and "concrete subgrade" in the modern epochal subset are grouped into the same term entry, while "wooden sleeper" and "prestressed concrete sleeper" are grouped into another term entry, thus forming at least one initial term group.

[0030] It is understood that each initial term group contains term members spanning multiple epochal subsets, and each term member is associated with its respective epochal subset. For each term in the initial term group, the plurality of epochal subsets are traversed, and the occurrence position of the term in each epochal subset is recorded.

[0031] For example, for the term "roadbed" which is formed after merging, if it first appears in the 1950s subset and last appears in the 2020s subset, then the year of its first appearance is recorded as the starting time point, and the year of its last appearance is recorded as the ending time point.

[0032] In one embodiment, the initial term group includes three term entries: "roadbed", "sleeper" and "track bed", each corresponding to its own time point group. The three time point groups together constitute the time point group set of the initial term group.

[0033] Specifically, based on the set of time points, the minimum start time of all term entries within the group is taken, and the resulting archiving year is recorded as the group start point; the maximum end time of all term entries within the group is taken, and the resulting archiving year is recorded as the group end point. The year corresponding to the group end point is subtracted from the year corresponding to the group start point, and the resulting number of years is the chronological span of the initial term group in the document retrieval text.

[0034] Preferably, the obtained time span serves as the time dimension basis for subsequent determination of the meaning shift of terms, providing a quantitative characterization of the time range for semantic matching of railway engineering archives in cross-era retrieval.

[0035] Step S102: Vectorize the context of the initial term group in each decade subset according to the decade span, and use the cosine distance between the same term and different decade vectors to characterize the semantic dispersion of the term in the time dimension, and summarize to obtain the decade semantic dispersion of the initial term group.

[0036] Based on several decade subsets defined by the aforementioned decade span, for each term in the initial term group, a sequence of context words within a preset window length is extracted from each decade subset centered on the term, resulting in a set of context fragments corresponding to the term in each decade subset. For each set of context fragments, a word vector algorithm is used to independently vectorize the corpus of each decade subset, obtaining a semantic vector corresponding to the term in each decade subset. The value of the semantic vector is determined by the statistical association strength between the term and co-occurring words in the context fragment of the corresponding decade subset. For each term in the initial term group, the semantic vectors of the term in different decade subsets are paired according to the decade subset, and the cosine distance between each pair of semantic vectors is calculated. The cosine distance is used to characterize the degree of semantic dispersion of the term in the time dimension. The average cosine distance of all terms in the initial term group is taken, and the summation yields the decade semantic dispersion of the initial term group.

[0037] For railway engineering terms such as "roadbed," "sleeper," and "ballast" included in the initial terminology group, sentences containing these terms are retrieved from the archival text corresponding to each epoch subset. Using the word position of the term as the center, a preset window length is extended forward and backward, and a continuous sequence of words within the window is extracted as the context fragment of the term in that epoch subset. All context fragments of the term in that epoch subset are then compiled into a context fragment set. In one embodiment, the preset window length is 5 to 10 words.

[0038] For example, in the sentence containing the word "roadbed" in the subset from the 1950s, "the roadbed was constructed by manually ramming earthen embankments," five words are taken before and after the sentence to extract the corresponding context fragments.

[0039] It should be noted that the word vector algorithm adopts a co-occurrence word distribution modeling method based on context window, with the input being the set of context fragments and the output being the semantic vector corresponding to the term.

[0040] Specifically, a vocabulary list is independently constructed for each decade subset. The frequency of collocation between the terms in that decade subset and the co-occurring words within the window is statistically analyzed. The higher the collocation frequency, the stronger the association between the term and the corresponding co-occurring word in the context of that decade. The statistical association strength of the co-occurring words is reflected in the value of the frequency of the word appearing around the term after global word frequency normalization.

[0041] For example, in the early chronological subset, the co-occurring terms "roadbed" are strongly associated with "rammed earth," "earth embankment," and "artificial," while in the modern chronological subset, "roadbed" is strongly associated with "layered," "compacted," and "drainage." The semantic vector is a numerical sequence of fixed dimensions, where the value of each dimension corresponds to the statistical association strength between the term and a co-occurring term.

[0042] Preferably, the vector dimension is between 100 and 300.

[0043] It is understood that, for each term in the initial term group, a corresponding number of semantic vectors are obtained in several epoch subsets covered by the epoch span, forming a cross-epoch vector sequence for the term.

[0044] Specifically, for the cross-decade vector sequence, the semantic vectors are paired according to decade subsets. That is, the semantic vector corresponding to the term in decade subset A is paired with the semantic vector corresponding to the term in decade subset B. Decade subset A and decade subset B are any two different sets of decade subsets. For each pair of semantic vectors, the cosine of the angle between them is calculated, and then the cosine distance is obtained by subtracting the cosine value from 1. The larger the cosine distance, the more significant the semantic distribution deviation of the term between decade subset A and decade subset B.

[0045] For example, the cosine distance between the term "rail sleeper" and the subsets of the 1950s and 2020s is significantly greater than the cosine distance between two adjacent subsets. For all terms within the initial term group, the arithmetic mean of the cosine distances between all terms in each subset is calculated, and this mean is denoted as the chronological semantic dispersion of the initial term group.

[0046] In one embodiment, for the three terms "roadbed", "sleeper" and "track bed", the cosine distance set between each of the chronological subset pairs is obtained. The three sets are merged and the mean is calculated. The obtained value is the chronological semantic dispersion of the initial term group in the archive retrieval text. The chronological semantic dispersion characterizes the overall semantic shift of the initial term group in cross-chronological archives.

[0047] Step S103: Obtain the industry background text to which the initial term group belongs, and determine the domain distribution density of the initial term group by statistically analyzing the co-occurrence frequency of the initial term group in the industry background text.

[0048] For each term in the initial terminology group, industry standards, technical manuals, professional journals, and engineering reports published during the same period are retrieved from the time span covered by the archival retrieval text. These retrieved texts are then compiled into an industry background text set to which the initial terminology group belongs. For this industry background text set, each text is traversed. Using the position of any term from the initial terminology group within the text as an anchor point, a pre-defined segment length is extended forward and backward to create a co-occurrence window. The number of times each pair of terms in the initial terminology group co-occurs within this window is counted, resulting in a co-occurrence frequency record between each pair of terms in the initial terminology group. Based on this co-occurrence frequency record, the co-occurrence frequency of each term pair in the initial terminology group is normalized to the total number of occurrences of that term pair in the industry background text set. The resulting value is used as the co-occurrence frequency of the term pair. The average co-occurrence frequency of all term pairs in the initial terminology group is then calculated to determine the domain distribution density of the initial terminology group.

[0049] For railway engineering terms such as "roadbed", "sleeper" and "ballast" included in the initial terminology group, the time range covered by the stated era span is used as the filtering scope. Technical materials published in the same period are retrieved from the railway industry's document repository, and the obtained texts are compiled into an industry background text set to which the initial terminology group belongs.

[0050] Specifically, the industry background text set consists of four types of text. The industry standards include mandatory or recommended technical documents such as railway engineering construction quality acceptance standards and railway subgrade design specifications; the technical manuals include compilations such as railway engineering design manuals and track engineering construction manuals; the professional journals include academic journals published concurrently, such as the Journal of Railway Engineering and the Journal of Railway Science; and the engineering reports include survey and design reports and completion acceptance reports for each section of the railway line. In one implementation, for each text in the industry background text set, the text is split into sentence segments using periods, semicolons, question marks, exclamation marks, and consecutive line breaks. For excessively short segments with fewer than 8 characters after splitting, they are merged into the adjacent preceding sentence segment to avoid semantic fragmentation. The scale of the co-occurrence window is denoted as w, and its value ranges from 3 to 5 consecutive sentence segments. The specific value is determined based on the text type: for industry standards and technical manuals, where the clauses are concise and the context span is small, w=3 is used; for professional journal texts, where the argumentation chain is of medium length, w=4 is used; for engineering reports, where the narrative is more fully developed and the related information spans a larger range, w=5 is used. When there are significant differences in paragraph length within the same text, linear interpolation can be performed within this range based on the average number of characters per sentence segment, followed by rounding to the nearest integer.

[0051] Specifically, the process iterates through the industry background text set, locates the position of any term in the initial terminology group within the text, and expands forward and backward by N segments, centered on the sentence segment where the term appears. Here, N = ⌊L / 2⌋, where L is the preset segment length and ⌊⌋ represents rounding down. The text segment covered by the co-occurrence window is extracted. For example, when the preset segment length is 5, N = 2 (expanding forward by 2 segments and backward by 2 segments); when the preset segment length is 3, N = 1 (expanding forward by 1 segment and backward by 1 segment). For each text segment, the occurrence of the remaining terms in the initial terminology group within the text segment is retrieved one by one. If they appear, it is recorded as a co-occurrence of the two terms within the co-occurrence window.

[0052] It should be noted that the co-occurrence frequency records are based on term pairs as the basic unit.

[0053] For example, for the term pair consisting of "roadbed" and "track bed", after traversing all texts in the industry background text set, the total number of times the term pair co-occurs in all co-occurrence windows is obtained; for the term pair consisting of "sleeper" and "track bed", the corresponding total number of co-occurrences is obtained in the same way.

[0054] Specifically, for each term pair in the initial term group, the total number of co-occurrences of the term pair in the industry background text set is used as the numerator, and the smaller of the individual occurrences of each term in the term pair in the industry background text set is used as the denominator. The result of dividing the two is the co-occurrence frequency of the term pair, denoted as f. For example, if the total number of co-occurrences of a term pair is 20, and the individual occurrences of the two terms are 50 and 80 respectively, then f = 20 / min(50,80) = 0.4. Since the number of co-occurrences will not exceed the occurrence of any single term, the value of f must fall between 0 and 1. The larger the value of f, the higher the degree of matching between the term pairs in the industry background text set. When f is greater than or equal to 0.6, it can be determined as a closely matched term pair.

[0055] For example, the co-occurrence frequency of "roadbed" and "track bed" is significantly higher than the co-occurrence frequency of "roadbed" and unrelated words.

[0056] It is understood that the initial term group contains several term pairs, and the co-occurrence frequency of each term pair is obtained in the manner described above. Further, the arithmetic mean of the co-occurrence frequencies of all term pairs within the initial term group is taken, and this average is denoted as the domain distribution density of the initial term group. The domain distribution density characterizes the overall tightness of the interrelationship among the terms in the initial term group within the railway engineering context represented by the industry background text set.

[0057] In one embodiment, for the term pair set consisting of the three term entries "roadbed", "sleeper" and "ballast", the co-occurrence frequency of each term pair is obtained. The mean of the co-occurrence frequency is calculated, and the obtained value is the domain distribution density of the initial term group in the railway engineering industry background text. The domain distribution density is used as a quantitative indicator to measure the breadth and density of the initial term group within the single industry to which the document retrieval text belongs.

[0058] Step S104: By analyzing the domain distribution density of the initial term group and the changes in the expression of the terms in the archives across different eras, terms whose core meanings have not changed substantially under the support of high domain distribution density are selected to form a fixed term group.

[0059] For each term in the initial terminology group, limiting descriptions related to its morphology, structural composition, and construction technology are extracted from the sentence segments containing the term in the archival retrieval text of each era. These limiting descriptions for the same term obtained from different era subsets are sorted by archiving year, forming a cross-era description change sequence for the term. This cross-era description change sequence is used to present the differences in the term's surface description. For this cross-era description change sequence, the functional positioning and target description fragments referred to in each era subset are extracted. The functional positioning and target description fragments from different era subsets are compared by keyword set coverage. If the keyword set of the functional positioning and target description fragments maintains a consistent referent and functional relationship within the cross-era description change sequence, the core meaning of the term is determined to have not undergone substantial change; if the keyword set shows a change in referent or functional relationship, the core meaning of the term is determined to have undergone substantial change. Based on the comparison between the domain distribution density of the initial term group and a preset density threshold, when both the domain distribution density is higher than the preset density threshold and the core meaning of the term has not changed substantially, the terms in the initial term group that meet the two conditions are aggregated and the aggregated terms are assigned to a fixed term group.

[0060] For railway engineering terms such as "roadbed", "sleeper", "ballast", "rail" and "fastener" included in the initial term group, the archival retrieval texts of each era are traversed to locate the sentence segment in each era subset of the archival text, and descriptive phrases that define and characterize the term are extracted from the sentence segment.

[0061] Specifically, the limiting description consists of three types of descriptive phrases. Morphological descriptive phrases describe the external dimensions and shape features of the entity corresponding to the term; structural composition descriptive phrases describe the material composition and internal components of the entity corresponding to the term; and construction process descriptive phrases describe the forming method and operational procedures of the entity corresponding to the term. Within each epoch subset, for each term, the three types of descriptive phrases are extracted one by one to form a set of limiting descriptions for that term within that epoch subset. The sets of limiting descriptions for the terms within each epoch subset are sequentially concatenated according to the archiving year from earliest to latest, and the resulting sequence is denoted as the cross-epoch description change sequence of the term.

[0062] For example, regarding the term "rail sleeper," the limiting description set under the 1950s subset includes descriptive phrases such as "rectangular cross-section wooden component," "oil-impregnated anti-corrosion treatment of sleepers," and "axial laying"; the limiting description set under the 2020s subset includes descriptive phrases such as "prestressed concrete component," "embedded steel reinforcement cage," and "casting in a mold." Further, for the limiting description set corresponding to each decade subset in the cross-decade description change sequence, descriptive phrases referring to the role of the term in railway engineering are selected from the limiting description set and denoted as functional positioning description fragments; descriptive phrases referring to the objects in which the term acts in railway engineering are selected and denoted as object description fragments. The functional positioning description fragments and the object description fragments are extracted from the limiting description set based on semantic roles.

[0063] It should be noted that the construction of the keyword set is based on word segmentation and stop word removal.

[0064] Specifically, for the functional positioning description fragment and the object description fragment under each decade subset, the words are segmented, meaningless auxiliary words and prepositions are removed, and meaningful content words are retained. The set of retained content words is denoted as the functional keyword set of the term under that decade subset. The coverage comparison is based on the ratio of the intersection to the union of sets. For the functional keyword sets under any two decade subsets in the cross-decade expression change sequence, the ratio of the number of intersection elements to the number of union elements is calculated. The obtained ratio is denoted as the functional consistency degree of the term under the two decade subsets. If the functional consistency degree is higher than a preset consistency threshold, it is determined that the referent and function relationship of the term under the two decade subsets are consistent; if the functional consistency degree is lower than the preset consistency threshold, it is determined that the referent and function relationship of the term under the two decade subsets have changed.

[0065] Preferably, the consistency threshold is between 0.6 and 0.8.

[0066] In one embodiment, for the term "roadbed", the set of functional keywords under the 1950s subset includes substantive words such as "bearing", "track", "foundation" and "stability", and the set of functional keywords under the 2020s subset includes substantive words such as "bearing", "track", "foundation", "stability" and "load transfer". If the functional consistency between the two is higher than a preset consistency threshold, it is determined that the referent and function of "roadbed" in the cross-era archives remain consistent, and the core meaning has not changed substantially.

[0067] It is understood that, for each term in the initial term group, the functional keyword sets under any two epochal subsets in the cross-epochal expression change sequence are compared pairwise. If all pairwise comparison results determine that the referent and the function relationship are consistent, then the core meaning of the term has not changed substantially; otherwise, it is determined that a substantial change has occurred.

[0068] Specifically, the comparison of the distribution density in the field is based on a preset density threshold.

[0069] Preferably, the density threshold is taken as the median value of the co-occurrence frequencies of commonly used term pairs in the industry background text set. When the domain distribution density of the initial term group is higher than the density threshold, it is determined that the initial term group has high distribution density support in the industry background text. Further, the terms in the initial term group are screened one by one. If the domain distribution density of the initial term group is higher than the density threshold, and the core meaning of the term has not changed substantially, and both conditions are met, then the term is grouped into a pending category; all pending categories that meet the two conditions are merged and classified into a fixed term group.

[0070] In one embodiment, for the five terms "subgrade", "sleeper", "ballast", "rail", and "fastener", the domain distribution density of the initial term group is higher than the density threshold. The core meaning of the four terms "subgrade", "sleeper", "ballast" and "rail" has not changed substantially. The four terms together constitute the fixed term group. The fixed term group reflects a set of railway engineering terms whose core meaning remains consistent across eras under the support of high domain distribution density.

[0071] Step S105: For the fixed term group, compress its temporal semantic dispersion proportionally according to the domain distribution density of each term, and summarize to obtain the meaning anchoring strength of the fixed term group under the single industry archive system to which the archive retrieval text belongs.

[0072] For each term in the fixed terminology group, the revised versions of industry standards and technical manuals published concurrently are retrieved from the industry background text set. Collocations appearing in different versions are extracted, and these collocations are arranged chronologically by version, resulting in a record of collocation changes during version updates. This record depicts the trajectory of collocation changes introduced by standard revisions and manual updates. Based on the decade subsets corresponding to the change points of each collocation in the collocation change record, the cosine distance components between these decade subsets are selected from the set of cosine distances included in the terminology's decade semantic dispersion. This cosine distance is denoted as the variable collocation offset. The remainder obtained by subtracting the variable collocation offset from the decade semantic dispersion is denoted as the stable anchoring residual. Divide the domain distribution density of each term in the fixed term group by the highest domain distribution density in the fixed term group to obtain the normalized weight of each term. Use the normalized weight to perform a weighted summation on the stable anchoring residual amount, and the resulting value is determined as the meaning anchoring strength of the fixed term group under the single industry archive system to which the archive retrieval text belongs.

[0073] For each term in the fixed terminology group, industry standards that have been revised multiple times and technical manuals that have been revised multiple times within the time span are selected from the industry background text set, and the text corresponding to each revision version of the standard and each revision version of the manual is extracted according to the release time.

[0074] Specifically, for each term, the sentence segments in which the term appears are retrieved in the revised version of the industry standard and the revised version of the technical manual. The limiting words and modifiers co-occurring within three words before and after the term in each sentence segment are extracted. The set of limiting words and modifiers is denoted as the collocation set of the term in that version. The collocation sets of the term in each version are listed sequentially according to the version release date from earliest to latest, and the collocation sets of adjacent versions are compared. During the comparison, terms that exist in the old version but not in the new version are first identified and marked as disappeared collocations, and terms that exist in the new version but not in the old version are marked as new collocations. Semantic matching is performed between disappeared and new collocations, and the cosine similarity of the word vectors is calculated. When the similarity is greater than or equal to 0.7 and the two terms modify the same position in the sentence, the term pair is marked as a replacement collocation, and the pair is removed from both the disappeared and new collocation sets. The recorded new collocation entries, disappearing collocation entries, and replaced collocation entries are records of the collocation changes of the terms during the version update process.

[0075] For example, regarding the term "subgrade," the set of collocations for the term in the 1980s revision of the Railway Subgrade Design Code includes "geotechnical filler," "roller compaction," and "manual compaction"; while the set of collocations for the term in the 2010s revision includes "layered filling," "vibratory compaction," and "drainage ditch." There are multiple alternative collocations between the two. In one implementation, each collocation replacement point corresponds to a version release year. The version release year is compared with the aforementioned determined subset of years to locate the two adjacent subsets of years to which the version release year falls. These two adjacent subsets constitute a subset pair, which serves as the mapping interval of the collocation replacement point in the semantic dispersion of years.

[0076] It should be noted that the stripping process of the variable pairing offset is based on the aforementioned set of cosine distances.

[0077] Specifically, the cosine distances obtained by pairing the terms between different epoch subsets have been compiled into a cosine distance set, where each cosine distance corresponds to an epoch subset pair. For the epoch subset pairs mapped to the time points of each collocation change in the collocation change record, the cosine distances corresponding to the epoch subset pairs are selected from the cosine distance set. The selected cosine distances are the collocation fluctuation components introduced by the specification revision and manual update. The collocation fluctuation components are accumulated, and the resulting value is recorded as the variable collocation offset of the term. The variable collocation offset is subtracted from the epoch semantic dispersion corresponding to the term, and the remaining amount is recorded as the stable anchoring residual of the term. The stable anchoring residual characterizes the true degree of cross-epoch semantic shift of the term after removing the disturbances to the collocation relationship caused by the specification revision and manual update.

[0078] In one embodiment, for the term "roadbed," the semantic dispersion of the era is 0.42, and the cumulative fluctuation components corresponding to the standard revision and manual update are 0.18, so the stable anchoring residual of the term is 0.24. Further, for each term in the fixed term group, its domain distribution density is obtained. The domain distribution densities of all terms in the fixed term group are extracted, and the maximum value obtained is recorded as the highest domain distribution density. The domain distribution density of each term is divided by the highest domain distribution density, and the quotient is recorded as the normalized weight of that term.

[0079] Specifically, the normalization weights range from 0 to 1. A larger normalization weight indicates a higher degree of coherence between the term and the industry context text. The normalization weights serve as weighting coefficients for the stable anchoring residual.

[0080] In one possible implementation, for each term within the fixed term group, the stable anchoring residual is multiplied by the normalized weight, and the resulting product is recorded as the weighted stable anchoring term for that term; the weighted stable anchoring terms of all terms within the fixed term group are summed and normalized using the sum of the normalized weights, and the resulting value is the meaning anchoring strength of the fixed term group within the single industry archive system to which the archive retrieval text belongs.

[0081] In one embodiment, for the four fixed terminology group members "subgrade", "sleeper", "ballast", and "rail", their respective stable anchoring residuals and normalized weights are obtained. The values ​​obtained by the weighted aggregation method are the meaning anchoring strength of the fixed terminology group. The meaning anchoring strength reflects the degree of stable support of the core meaning of the fixed terminology group across eras under the single industry archive system of railway engineering.

[0082] For each term in the fixed term group, stable anchoring sources and variable collocation sources are divided according to the density of term distribution in the term domain. The offset components corresponding to the variable collocation sources introduced by industry standard updates and technical manual revisions are identified. The contribution of stable anchoring sources to suppressing the semantic dispersion of time after removing offset components is analyzed. The meaning anchoring strength of the fixed term group in the archival system is evaluated and formed.

[0083] For each term in the fixed term group, the domain distribution density of the term is compared with the median value of the domain distribution density within the fixed term group. If the domain distribution density of the term is higher than the median value, the corresponding component of the term in the chronological semantic dispersion is classified as a stable anchoring source; if the domain distribution density of the term is lower than or equal to the median value, the corresponding component of the term in the chronological semantic dispersion is classified as a variable collocation source. For the variable collocation sources, the updated versions of the industry standard and the revised versions of the technical manual are traversed, and the collocation term change entries of the term in each version are extracted. The cosine distance component contributed by the chronological subset pair corresponding to the collocation term change entry in the chronological semantic dispersion is recorded as the offset component. The offset component is subtracted from the variable collocation sources to obtain the variable collocation residual value of the term. For each term in the fixed term group, the cosine distance component corresponding to the stable anchoring source is summed with the variable collocation residual value to form a dispersion suppression term. The difference between the preset benchmark value and the dispersion suppression term is used as the suppression contribution of the term. Based on the normalized weight of the distribution density of each term domain in the fixed term group, the suppression contribution is weighted and summed. The obtained value is used to evaluate and form the meaning anchoring strength of the fixed term group in the archival system.

[0084] For railway engineering terms such as "roadbed", "sleeper", "ballast bed" and "rail" included in the fixed term group, the source attributes of each component in the cosine distance set are divided according to the distribution density of the field.

[0085] Specifically, the median value of the domain distribution density is obtained by arranging the domain distribution densities of all terms within the fixed term group in ascending order and taking the value at the center of the arrangement as the median value. For each term, the domain distribution density of the term is compared with the median value. If the domain distribution density of the term is higher than the median value, it is determined that the term has a strong degree of collocation in the industry background text, and all cosine distance components corresponding to the term in the calculation of the epochal semantic dispersion of all epochal subset pairs are classified as stable anchoring sources. If the domain distribution density of the term is lower than or equal to the median value, all cosine distance components corresponding to the term in the calculation of the epochal semantic dispersion of all epochal subset pairs are classified as variable collocation sources.

[0086] For example, for the four terms "roadbed," "sleeper," "ballast," and "rail," corresponding domain distribution density values ​​are obtained. The domain distribution density of "roadbed" and "sleeper" is higher than the median value, and they are classified as stable anchoring sources; the domain distribution density of "ballast" and "rail" is lower than or equal to the median value, and they are classified as variable collocation sources. Further, for terms classified as variable collocation sources, the updated versions of the industry standards and the revised versions of the technical manuals in the industry background text set are traversed. Collocation entries for each term are extracted in the updated and revised versions, and the differences in collocation entries between adjacent versions are compared. New, disappeared, and replaced collocation entries are recorded as collocation change entries. In one implementation, the process of identifying the offset component is as follows: Each collocation term change entry corresponds to a version release year. The version release year is compared with the previously determined period subset division to locate the period subset pair consisting of two adjacent period subsets to which the version release year falls. The cosine distance component corresponding to the period subset pair is retrieved from the previously obtained cosine distance set. The retrieved cosine distance component is the offset component contributed by the collocation term change entry. The offset components corresponding to all collocation term change entries of the term are accumulated, and the accumulated value is recorded as the total offset component of the term. The total offset component is subtracted from the overall value of the cosine distance component corresponding to the variable collocation source, and the remaining value is recorded as the variable collocation residual value of the term.

[0087] In one embodiment, for the term "track bed", the cosine distance component corresponding to the variable matching source is generally taken as 0.36, the total offset component is 0.14, and the variable matching residual value obtained by subtracting the two is 0.22.

[0088] Specifically, for each term in the fixed term group, the cosine distance component corresponding to the stable anchoring source is summed with the variable collocation residual value, and the resulting value is denoted as the dispersion suppression term for that term. The smaller the dispersion suppression term, the smaller the degree of cross-decade semantic shift of the term after removing the influence of specification revisions and manual updates.

[0089] It should be noted that the preset benchmark value is taken as the maximum value among the semantic dispersion of the dates of each term in the fixed term group.

[0090] Specifically, for each term, the difference between the preset benchmark value and the dispersion suppression term is recorded as the suppression contribution of that term. The larger the suppression contribution, the more significant the suppression effect of the term on the semantic dispersion of dates.

[0091] In one possible implementation, the aforementioned normalized weights are used, whereby the normalized weights are the quotient of the distribution density of each term domain and the highest domain distribution density within the fixed term group. For each term in the fixed term group, the suppression contribution is multiplied by the normalized weight to obtain a weighted suppression term for that term; the weighted suppression terms for all terms in the fixed term group are accumulated, and the accumulated value is used to evaluate and form the semantic anchoring strength of the fixed term group within the archival system to which the archival retrieval text belongs.

[0092] In one embodiment, for the four fixed terminology group members "roadbed", "sleeper", "ballast", and "rail", their respective dispersion suppression terms, suppression contribution, and normalization weights are obtained. The values ​​obtained by weighted summation are the meaning anchoring strength of the fixed terminology group under the railway engineering archive system. The meaning anchoring strength characterizes the degree of stable support of the core meaning of the fixed terminology group across generations after removing the disturbances of standard revisions and manual updates.

[0093] Step S106: Adjust the pre-established initial similarity matching interval according to the anchoring strength of the meaning, and use the adjusted initial similarity matching interval to perform cross-era semantic matching on the fixed term group in the full text of the archive, return the document fragments that meet the matching conditions, and form the target archive retrieval results.

[0094] Based on the anchoring strength of meaning, the lower bound of the pre-established initial similarity matching interval is adjusted downwards. The adjustment magnitude corresponds positively to the anchoring strength of meaning; the higher the anchoring strength, the greater the adjustment magnitude. After adjustment, the lower bound of the initial interval becomes the adjusted initial similarity matching interval. Based on the adjusted initial similarity matching interval, for the document retrieval text, the document text is traversed one by one using terms in the fixed terminology group as retrieval anchors. Continuous document segments containing the terms are extracted, and the mean of the semantic vectors of each term in the fixed terminology group is calculated to obtain the group-level semantic vector. The semantic similarity value between the semantic vector of the document segment and the group-level semantic vector is calculated using cosine similarity. By comparing the semantic similarity value with the adjusted initial similarity matching interval, if the semantic similarity value falls within the adjusted initial similarity matching interval, the document segment is determined to meet the matching condition, and the document segment is included in the target document retrieval result set. The target document retrieval result set obtained after traversal forms the target document retrieval result.

[0095] For cross-era semantic matching of railway engineering archives, a pre-established initial similarity matching interval is a pair of value ranges. The value range consists of a lower bound and an upper bound. The upper bound is 1, and the lower bound is a preset value between 0.65 and 0.75. The semantic similarity falling within the value range is determined as a match hit.

[0096] Specifically, the implementation of the downsizing process involves using the meaning anchoring strength as the basis for the downsizing coefficient, which is the product of the meaning anchoring strength and a preset downsizing upper limit. For the lower bound of the initial interval, the downsizing coefficient is subtracted from the lower bound, and the difference is used as the adjusted lower bound. A higher meaning anchoring strength indicates stronger stability of the core meaning of the fixed term group across different eras within the railway engineering archives system to which the retrieval text belongs; the corresponding downsizing coefficient is larger, and the downsizing of the lower bound is greater. The upper bound of the initial interval remains unchanged, and together with the adjusted lower bound, forms the adjusted similarity matching initial interval.

[0097] In one embodiment, for the cross-era matching of railway subgrade and sleeper archives, the meaning anchoring strength is set to 0.82, the upper limit of the downward adjustment is set to 0.15, resulting in a downward adjustment coefficient of 0.123. The lower limit of the initial interval is lowered from 0.7 to 0.577, and the adjusted initial similarity matching interval is 0.577 to 1. Further, for each archival text in the archival retrieval text, terms such as "subgrade," "sleeper," "ballast," and "rail" from the fixed terminology group are used as retrieval anchors. The positions of these terms in the archival text are located by sentence segment. A preset sentence segment length is extended forward and backward from the sentence segment where the position is located, extracting the continuous document fragment containing the term in the archival text.

[0098] It should be noted that the group-level semantic vector is formed as follows: Using the semantic vectors of each term in each decade subset, the arithmetic mean of the semantic vectors of all terms within the fixed term group in all decade subsets is calculated for each dimension. The resulting sequence of equal-dimensional values ​​is denoted as the group-level semantic vector. For the extracted continuous document fragments, the document fragments are vectorized using a word vectorization algorithm to obtain the semantic vectors corresponding to the document fragments.

[0099] Specifically, the semantic vector corresponding to the document fragment and the group-level semantic vector are multiplied by their corresponding dimensions and then summed. The sum is then normalized using the product of the magnitudes of the two semantic vectors. The resulting value is the semantic similarity value between the document fragment and the fixed term group. The semantic similarity value falls between 0 and 1; a higher value indicates a higher degree of semantic fit between the document fragment and the fixed term group.

[0100] In one possible implementation, the semantic similarity value is compared with the adjusted initial similarity matching interval. If the semantic similarity value falls within the adjusted initial similarity matching interval, the document fragment is determined to meet the matching condition, and the document fragment, along with the archiving year and archive number of its respective archive, is included in the target archive retrieval result set. If the semantic similarity value does not fall within the adjusted initial similarity matching interval, the document fragment is determined not to meet the matching condition and is removed.

[0101] In one embodiment, for cross-generational archive retrieval of railway engineering documents, the fixed terminology group covers the four terms "roadbed," "sleeper," "ballast," and "rail." After traversing all the archive texts in the retrieval text, the target archive retrieval result set contains several archival document fragments from the 1950s to the 2020s. The target archive retrieval result set forms the target archive retrieval result, which realizes the retrieval and recall of core meaning-consistent fragments of railway engineering archives across different eras.

[0102] It should also be noted that the various specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the present invention will not describe the various possible combinations separately. Furthermore, various different embodiments of the present invention can also be arbitrarily combined, as long as they do not violate the spirit of the present invention, and should also be regarded as the content disclosed by the present invention.

Claims

1. A semantic analysis-based intelligent full-text retrieval method for archives, characterized in that, The method includes: extracting at least one initial term group from archival retrieval text spanning multiple eras, recording the first and last occurrence times of each term in the archives, and obtaining the era span corresponding to the initial term group in the archival retrieval text; vectorizing the context of the initial term group in each era subset according to the era span, using the cosine distance between the same term in different era vectors to characterize the semantic dispersion of the term in the time dimension, and summarizing to obtain the era semantic dispersion of the initial term group; obtaining the industry background text to which the initial term group belongs, and statistically analyzing the common occurrences of the initial term group in the industry background text. Frequency is used to determine the domain distribution density of the initial term group; by analyzing the domain distribution density of the initial term group and the changes in the expression of terms in cross-generational archives, fixed term groups are selected; based on the domain distribution density of each term, the temporal semantic dispersion of the fixed term groups is compressed proportionally, and the meaning anchoring strength of the fixed term groups under the single industry archive system to which the archive retrieval text belongs is obtained; the pre-established initial similarity matching interval is adjusted according to the meaning anchoring strength, and cross-generational semantic matching of the fixed term groups is performed in the full text of the archive using the adjusted initial similarity matching interval, returning document fragments that meet the matching conditions, forming the target archive retrieval results.

2. The method for intelligent full-text retrieval of archives based on semantic analysis according to claim 1, characterized in that, The step of extracting at least one initial term group from archival retrieval text spanning multiple eras includes: dividing the full text into several era subsets according to the year the archives were filed; extracting professional terms in each era subset using named entity recognition through a conditional random field; and merging professional terms pointing to the same functional object in different era subsets into the same term entry based on the consistency of the functional objects referred to by the professional terms, thus forming the initial term group.

3. The method for intelligent full-text retrieval of archives based on semantic analysis according to claim 1, characterized in that, The step of obtaining the epoch span corresponding to the initial term group in the archive retrieval text includes: for each term in the initial term group, traversing the occurrence positions of the term in each epoch subset, recording the archiving year corresponding to the first occurrence position as the start time point, recording the archiving year corresponding to the last occurrence position as the end time point, taking the earliest value among all term start time points as the group start point, and the latest value among the end time points as the group end point, and determining the epoch span by the difference between the group end point and the group start point.

4. The intelligent full-text retrieval method for archives based on semantic analysis according to claim 1, characterized in that, The step of vectorizing the context of the initial term group in each decade subset, and using the cosine distance between the vectors of the same term in different decades to characterize the semantic dispersion of the term in the time dimension, includes: for each term in the initial term group, extracting a sequence of context words within a preset window length centered on the term in each decade subset, and using a word vector algorithm to independently vectorize the corpus of each decade subset to obtain the semantic vector corresponding to the term in each decade subset; calculating the cosine distance between any two terms paired in the semantic vectors of different decade subsets, and taking the average of the cosine distances of all terms in the initial term group to form the decade semantic dispersion.

5. The method for intelligent full-text retrieval of archives based on semantic analysis according to claim 1, characterized in that, The process of obtaining the industry background text to which the initial term group belongs includes: compiling industry standards, engineering manuals, professional journals, and engineering reports published during the same period from the time span covered by the terminology; using the occurrence position of any term in the initial term group as an anchor point to extend a preset sentence length forward and backward as a co-occurrence window; forming a term pair from any two terms in the initial term group; counting the number of co-occurrences of each term pair within the co-occurrence window; and taking the average of the co-occurrence frequencies of all term pairs in the initial term group after normalization to determine the domain distribution density.

6. The method for intelligent full-text retrieval of archives based on semantic analysis according to claim 1, characterized in that, The process involves analyzing the domain distribution density of the initial terminology group and the changes in terminology expression across archival periods to select a fixed terminology group. This includes: for each term in the initial terminology group, extracting the limiting descriptions of morphology, structure, and technology from the sentence segments containing the term in each decade subset, and sorting them by archival year to form a cross-generational expression change sequence; extracting the functional positioning and object description fragments referred to by each decade subset in the cross-generational expression change sequence, and performing a coverage comparison by keyword set. If the referent and function relationship of the keyword set remains consistent, the core meaning has not changed substantially; otherwise, the core meaning has changed substantially. Terms whose domain distribution density is higher than a preset density threshold and whose core meaning has not changed substantially are included in the fixed terminology group.

7. The intelligent full-text retrieval method for archives based on semantic analysis according to claim 1, characterized in that, The method of compressing the temporal semantic dispersion of a fixed term group proportionally based on the domain distribution density of each term, and summarizing the meaning anchoring strength of the fixed term group within the single industry archive system to which the archive retrieval text belongs, includes: for each term in the fixed term group, retrieving concurrently published revised versions of industry standards and revised versions of engineering manuals, extracting collocation entries from different versions to form a collocation change record; based on the temporal subset pairs corresponding to the collocation change record, selecting corresponding components from the cosine distance set of temporal semantic dispersion and recording them as variable collocation offsets; subtracting the variable collocation offsets from the temporal semantic dispersion to obtain a stable anchoring residual; dividing the domain distribution density of each term by the highest domain distribution density to obtain a normalized weight; and determining the meaning anchoring strength based on the normalized weights and the stable anchoring residual.

8. The intelligent full-text retrieval method for archives based on semantic analysis according to claim 1, characterized in that, The process of summarizing the meaning anchoring strength of the fixed term group within the single industry archive system to which the retrieved text belongs includes: for each term in the fixed term group, comparing the domain distribution density of the term with the median value of the domain distribution density within the fixed term group; if the density is higher than the median value, the corresponding component of the term in the chronological semantic dispersion is classified as a stable anchoring source; otherwise, it is classified as a variable collocation source; traversing the revised versions of industry standards and engineering manuals, extracting the collocation term change entries for each version, recording the cosine distance component of the contribution of the chronological subset corresponding to the collocation term change entry as an offset component, and subtracting the offset component from the variable collocation source to obtain the variable collocation residual value; summing the cosine distance component corresponding to the stable anchoring source with the variable collocation residual value to form a dispersion suppression term, using the difference between a preset benchmark value and the dispersion suppression term as the suppression contribution, dividing the domain distribution density of each term by the highest domain distribution density to obtain a normalized weight, and weighting and summing the suppression contribution based on the normalized weight to form the meaning anchoring strength.

9. The method for intelligent full-text retrieval of archives based on semantic analysis according to claim 1, characterized in that, The step of adjusting the pre-established initial similarity matching interval according to the meaning anchoring strength includes: lowering the lower bound of the pre-established initial similarity matching interval, wherein the lower bound is positively correlated with the meaning anchoring strength, and the higher the meaning anchoring strength, the greater the lower bound. The adjusted initial similarity matching interval is obtained after the lower bound of the initial similarity matching interval is lowered.

10. The intelligent full-text retrieval method for archives based on semantic analysis according to claim 1, characterized in that, The process of utilizing the adjusted initial similarity matching interval to perform cross-era semantic matching on the fixed term group in the full text of the archive, returning document fragments that meet the matching conditions, and forming the target archive retrieval results includes: traversing the archive text using terms in the fixed term group as retrieval anchors, extracting continuous document fragments containing the terms, averaging the semantic vectors of each term in the fixed term group to obtain a group-level semantic vector, calculating the cosine similarity value between the semantic vector of the document fragment and the group-level semantic vector, and including document fragments whose similarity value falls within the adjusted initial similarity matching interval in the target archive retrieval results.

Images