Method and system for intelligent analysis and text optimization of foreign financial discourse

Abstract

The application provides a kind of intelligent analysis and text optimization method and system for foreign financial discourse, belongs to the cross field of information technology and natural language processing, including: collecting English financial news corpus for preprocessing and warehousing, forming English financial news database;Train language feature labeling model based on machine learning to label language features, identify the language features of English text;Build a financial discourse analysis framework, with dimension score and language feature frequency as the basis for financial discourse analysis framework analysis;Train the foreign financial discourse vertical large model, analyze the English financial news text based on the financial discourse analysis framework and propose optimization suggestions.The application realizes accurate labeling, language feature analysis and high-quality text polishing of English financial text, effectively improves the readability, accuracy and influence of financial text among English native speakers.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of information technology and natural language processing, and specifically relates to an intelligent analysis and text optimization method and system for foreign financial discourse. Background Technology

[0002] Existing English text analysis and optimization methods or systems focus only on simple English part-of-speech or grammatical analysis. When applied to international financial discourse, most have technical shortcomings, such as:

[0003] (1) It is impossible to accurately evaluate the expression of English financial discourse through the co-occurrence analysis of linguistic features under the financial domain.

[0004] (2) It is impossible to accurately locate and systematically analyze the stance expressed in English financial discourse.

[0005] (3) It is impossible to obtain professional suggestions for optimizing English financial discourse and relevant example sentences in the real domain through intelligent interaction with the self-trained large model.

[0006] In conclusion, existing English text analysis and optimization methods or systems are insufficient for accurately annotating, analyzing language features, and polishing high-quality English financial texts, making it difficult to improve the readability, accuracy, and influence of financial texts among native English speakers. Summary of the Invention

[0007] The purpose of this invention is to provide an intelligent analysis and text optimization method and system for international financial discourse. By combining machine learning technology and statistical analysis methods, an evaluation system for English financial discourse is constructed to achieve the analysis, evaluation and improvement of financial texts.

[0008] To achieve the above objectives, the technical solution of the present invention is as follows:

[0009] An intelligent analysis and text optimization method for international financial discourse includes:

[0010] S1. Core Data Collection and Cleaning: Collect English financial news corpora from foreign media and domestic and foreign propaganda media, preprocess them and store them in the database to form an English financial news database.

[0011] S2. Training the language feature annotation model: Based on machine learning methods, three rounds of model training are conducted. The first round of model training identifies language features including parts of speech, simple grammar, and lexical stance. The second round of model training identifies language features including complex grammar. The third round of model training is used for iterative improvement, and finally, an English text language feature annotation model is obtained.

[0012] S3. Construct a financial discourse analysis framework: Calculate the 1,000-word frequency of language features, obtain several dimensions through factor analysis, and use the standard scores of each dimension and the standardized frequency of language features as the basis for the analysis of the financial discourse framework to calculate the Euclidean distance between the text to be analyzed.

[0013] S4. Training a large-scale model for international financial discourse: Using an open-source large model as a base, fine-tuning it to enable it to master the knowledge from steps S1-S3, and obtain an intelligent interactive large-scale model for international financial discourse, providing optimization suggestions for improving the quality of English financial texts.

[0014] Furthermore, the preprocessing method described in step S1 includes:

[0015] S101. Deduplication: Use regular expressions to filter out English financial news corpora that may be duplicated, and perform deduplication based on the similarity calculation of text content;

[0016] S102. Data Filtering: Use regular expressions to find text containing consecutive uppercase letters, determine if there are problems with non-standard formatting or low content quality, and whether it needs to be removed.

[0017] Furthermore, in step S1, after preprocessing and before data entry, data standardization and cleaning are required, including: adjusting the title format; extracting proper nouns to form a proper noun dictionary; standardizing the capitalization, punctuation, common abbreviations, and initial capitalization of the main text; cleaning the text to remove irrelevant content including illegal characters or URL links, and restoring the word form of the cleaned title and main text.

[0018] Furthermore, the data entry in step S1 includes: segmenting the text content to separate the title, date, source, watermark information, and body text; constructing a database with eight fields: ID, title, cleaned title, watermark, source, date, body text, and cleaned body text; writing the segmented content into the corresponding fields, with each data entry having a unique ID, thus forming an English financial news database; and performing multiple clustering operations on the cleaned text and titles in the English financial news database by fine-tuning the LDA topic clustering model to form several first-, second-, and third-level topics.

[0019] Furthermore, the first round of training of the language feature annotation model in step S2 includes:

[0020] S201. Apply a part-of-speech tagging tool to perform "one-time tagging" on the English text content, tagging the basic part-of-speech label for each word;

[0021] S202. Based on the aforementioned basic part-of-speech tags and combined with English grammar rules, programmatically perform "secondary tagging" on the relevant words to tag grammatical feature tags;

[0022] S203. Based on the position words in English and the corresponding clause grammar rules, programmatically determine whether the word expresses a position. If it does, mark the position word with a "three-fold label".

[0023] S204. Divide the dataset formed after three annotations into training set A, validation set of A, and test set B;

[0024] S205. Using RoBERTa-Large as the base model, the training set A is labeled and trained. The validation set of A is used to check the learning effect of each learning round. The model performance is tested through the test set B to obtain the DDU Tagger 1.0 language feature annotation model.

[0025] Furthermore, step S205 specifically includes: using RobertaTokenizerFast to split the sentence into WordPiece subwords, and using the first subword of each word as a token to map to the corresponding label of the word; then assigning the annotation training task to RobertaForTokenClassification, and using cross-entropy loss to allow the model to learn how to predict rule labels on a large-scale corpus, and using gradient accumulation, mixed precision, and learning rate warmup during the training process to adapt to the memory requirements of RoBERTaLarge; finally, the DDU Tagger 1.0 language feature annotation model is obtained through training.

[0026] Furthermore, the second round of training of the language feature annotation model in step S2 includes:

[0027] S211. The complex language features in the English text that cannot be accurately labeled at present are autonomously encoded for the labeling of complex grammatical features.

[0028] S212. Convert different electronic versions of English grammar dictionaries into searchable text and import them into the database; through indexing and classification, form different batches of original example sentences; use DDU Tagger 1.0 to pre-annotate the different batches of original example sentences, reserving items to be annotated for complex language features; after all batches of the same language feature are annotated, they are taken as a subset, and this process is repeated to generate subsets C1, C2, C3, ..., Cn, where n is the number of language feature types. All subsets together form a complete dataset C, and 10% is reserved as a test set D;

[0029] S213. Using the DDU Tagger 1.0 language feature annotation model as the initial weights, train it using dataset C to obtain the DDU Tagger 2.0 language feature annotation model.

[0030] Furthermore, the third round of training for the language feature annotation model in step S2 includes:

[0031] The DDU Tagger 2.0 language feature annotation model was used to re-annotate the test set B used in the first round of training. The correct and incorrect labels were counted to verify the accuracy of the model annotation. The high error rate labels were iteratively improved. Then, all erroneous instances and newly added labels were merged into the final training set E. Based on this, the DDU Tagger 2.0 language feature annotation model was retrained with the initial weights until the macro average F1 was ≥0.92, the recall of each new label was ≥0.80, and the overall accuracy was ≥98%, thus obtaining the final language feature annotation model DDU Tagger 3.0.

[0032] Furthermore, step S3 specifically includes:

[0033] S301. Use the language feature annotation model to re-annotate the English text in the English financial news database to obtain annotation results containing all language features;

[0034] S302. Calculate the 1000-word frequency of all language features in each English text data. The calculation method is: Language feature 1000-word frequency = (Actual frequency of language features / Total number of text tokens) * 1000;

[0035] S303. Import the 1000-word frequency of all language features into the social science statistical software SPSS for factor analysis, delete language features with loading values ​​below 0.35, and form several feature dimensions. Each dimension contains several language features with obvious co-occurrence trends.

[0036] S304. Statistically analyze the distribution of language features co-occurring in key sentences under the same dimension, and determine the specific classification and naming of each dimension by observing the distribution.

[0037] S305. A dimension naming validation dataset is formed by sampling, and language features are assigned to corresponding dimensions for quantitative testing and qualitative evaluation. The quantitative testing includes calculating the variance explained by the feature loadings and the Cronbach's alpha coefficient for each dimension on the dimension naming validation dataset, and testing the correlation coefficients between dimensions. Subsequently, the dimension naming validation dataset is clustered according to the projected dimension scores to evaluate the effectiveness of the dimension scores in distinguishing text differences. The qualitative evaluation includes reviewing the list of high-load features corresponding to each dimension, selecting typical sentences, and judging whether these sentences match the semantic category implied by the dimension naming. If most sentences match the naming, the naming is confirmed to be reasonable. If there is a significant mismatch, the feature assignments or names need to be readjusted.

[0038] S306. Using the 1,000-word frequency of the language features as the standardized frequency, and calculating the standard score for each dimension, the calculation formula is: Dimension score = (sum of 1,000-word frequencies of all positive features under this dimension - sum of 1,000-word frequencies of all negative features under this dimension); Based on the standardized frequency of each language feature and the standard score for each dimension, an English financial discourse evaluation standard is formed, which is used to compare with the text to be analyzed and form the analysis results.

[0039] In another aspect, this invention proposes an intelligent analysis and text optimization system for international financial discourse, comprising:

[0040] The core data collection and cleaning layer is sequentially connected, the language feature annotation model training layer is connected, the financial discourse analysis framework module is connected, and the large-scale model of external financial discourse is connected.

[0041] Core data collection and cleaning layer: Collect English financial news corpora from foreign media and domestic and foreign propaganda media, preprocess them and store them in the database to form an English financial news database;

[0042] Language feature annotation model training layer: Based on machine learning methods, three rounds of model training are carried out. The first round of model training identifies language features including parts of speech, simple grammar, and lexical stance. The second round of model training identifies language features including complex grammar. The third round of model training is used for iterative improvement, and finally, the English text language feature annotation model is obtained.

[0043] Financial discourse analysis framework module: Calculates the 1,000-word frequency of language features, obtains several dimensions through factor analysis, and uses the standard scores of each dimension and the standardized frequency of language features as the basis for the analysis of financial discourse, which is used to calculate the Euclidean distance between the text to be analyzed.

[0044] Large-scale model for international financial discourse: This model uses an open-source large-scale model as a foundation for fine-tuning, enabling it to master the knowledge of the core data collection and cleaning layer, the language feature annotation model training layer, and the financial discourse analysis framework module. This results in an intelligent interactive large-scale model for international financial discourse, providing optimization suggestions for improving the quality of English financial texts.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] 1. The English language feature annotation method, discourse analysis framework, and large-scale vertical model for financial discourse analysis constructed in this invention can achieve accurate annotation, language feature analysis, and high-quality text polishing of English financial texts.

[0047] 2. This invention can provide full-process support for editing English texts related to foreign financial news: covering multiple stages such as data cleaning, extraction of language features such as parts of speech and syntax, identification of themes, viewpoints and positions, as well as text parsing, evaluation and improvement.

[0048] 3. This invention effectively enhances the readability, accuracy, and influence of financial news among native English-speaking readers. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the method flow of Embodiment 1 of the present invention;

[0050] Figure 2 This is a schematic diagram of the system architecture of Embodiment 2 of the present invention. Detailed Implementation

[0051] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0052] The present invention will now be described in detail with reference to specific embodiments and accompanying drawings.

[0053] Example 1:

[0054] This embodiment applies the intelligent analysis and text optimization method for foreign financial discourse of the present invention. Many steps are automated by programming with programming languages. In this embodiment, Python is preferred because it uses Python's programming functions. However, it is not limited to Python. Using other programming languages ​​to achieve the same function is also within the protection scope of the present invention.

[0055] like Figure 1 As shown, the intelligent analysis and text optimization methods for international financial discourse specifically include:

[0056] I. Core Data Collection and Cleaning.

[0057] 1.1 Data Collection: Collect English financial news data from mainstream British and American media and major Chinese foreign propaganda media, including headlines, text, dates, sources, authors, etc.

[0058] 1.2 Data Preprocessing: Systematically preprocess the raw text, specifically including:

[0059] 1.2.1 Remove duplicate text.

[0060] (1) Extract the title prefix and group it:

[0061] The prefix is ​​extracted from the first line of the file title using the Python regular expression PREFIX_SEPARATOR; the files are grouped according to the extracted prefix, and the number of files and unique titles in each group is counted; the purpose is to quickly narrow down the comparison range, group files with potentially related content into one category, and avoid the inefficiency of comparing all files pairwise.

[0062] Only groups with more than one file are retained as candidate deduplication targets. Further filtering is applied to groups with "no subclasses" (all files within the group have identical titles) and a file count ≤ SIMILARITY_GROUP_MAX_SIZE (default 10). Only these groups are used for subsequent content comparison. Retaining groups with more than one file is to focus only on objects with a high probability of duplication; individual files do not require deduplication, and direct filtering reduces the processing load. Filtering "no subclasses" groups is because files with completely identical titles have a very high probability of content duplication; prioritizing these groups significantly improves the deduplication hit rate. Limiting the file count to ≤ 10 balances deduplication efficiency and coverage; comparing too many files (e.g., more than 10) increases computational cost, while 10 or fewer covers most repetitive scenarios while ensuring processing speed.

[0063] (2) Parallel similarity matching with two rules:

[0064] Calculate the standard similarity ratio of the document content. If it is greater than or equal to the set threshold of 0.8, it is considered similar and marked as "Standard". When the standard similarity ratio is not met, the first N characters of the longer content and the shorter content (N = the length of the shorter content) are compared to calculate the subset similarity ratio. If the subset similarity ratio is greater than or equal to the set threshold of 0.9, it is considered similar and marked as "Subset". Delete the matched text with a similarity ratio higher than the threshold and keep only one copy.

[0065] 1.2.2 Data filtering: Remove texts of low quality or irrelevant topics.

[0066] (1) Using the subcategories of news titles found in the previous step, keywords can be used as filtering conditions. For example, news with a keyword XX can be found using the startwith statement. Then, by sampling the news content, it can be determined whether news under a certain subcategory needs to be removed in batches.

[0067] (2) Use the regular expressions VALID_WORD_REGEX and MULTI_WORD_PATTERN to find the consecutive uppercase parts in the title and body of the article. These articles may have problems with non-standard format or low content quality. Manually judge whether they need to be removed.

[0068] 1.3 Data Cleaning:

[0069] This includes: adjusting title formatting; extracting proper nouns to create a proper noun dictionary; using a Python-written text formatting program to standardize the text content for capitalization, punctuation, common abbreviations, and initial capitalization; applying Python's NLTK natural language processing toolkit and Python regular expressions to clean the text, removing illegal characters, URL links, and other irrelevant content; and using the Spacy natural language processing library to perform lemmatization on the cleaned titles and text.

[0070] 1.4 Data import:

[0071] This includes: segmenting text content using regular expressions to separate it into several parts: title, date, source, watermark information, and body; constructing a PostgreSQL database with eight fields: ID, title, cleaned title, watermark, source, date, body, and cleaned body; writing the segmented content into the corresponding fields, with each data entry having a unique ID, forming an English financial news database α; and performing multiple clustering operations on the cleaned text and titles in the English financial news database α by fine-tuning the LDA topic clustering model to form several first-, second-, and third-level topics.

[0072] II. Training the language feature annotation model.

[0073] This embodiment uses RoBERTa-Large (the largest parameter version of the RoBERTa model) as the base model for labeled training.

[0074] RoBERTa is a pre-trained model based on BERT, developed by Facebook AI Research. Building upon BERT, it significantly improves model performance through optimization of the pre-training process. By pre-training on a large-scale corpus, RoBERTa learns rich linguistic knowledge and semantic representations, enabling it to accurately understand the meaning of text and capture grammatical, semantic, and contextual information. Because it has accumulated a large amount of general-purpose linguistic knowledge during pre-training, even with only a small amount of task-specific labeled data, the model can effectively learn and adjust during fine-tuning, resulting in relatively accurate annotations.

[0075] 2.1 First round of training:

[0076] The first round of training mainly focuses on the annotation of language features, including parts of speech, simple grammar, and lexical stance.

[0077] 2.1.1 The Stanford POS Tagger tool was used to perform a "first-time tagging" of the British and American financial news texts in the database, obtaining basic part-of-speech tags for each word. For example, the sentence "I will be a teacher." would be tagged with "I_PRP will_MD be_VB a_teacher_NN." where PRP represents a personal pronoun; MD represents a modal verb; VB represents the base form of a verb; and NN represents a countable or uncountable common noun. Some language features can be identified through these basic part-of-speech tags.

[0078] 2.1.2 Combining the basic part-of-speech tags and English grammar knowledge mentioned above, a rule-based "secondary annotation" program was written using Python to annotate the text. This primarily targets grammatical annotations with obvious features that can be converted into rule-based judgments with high accuracy, resulting in grammatical feature labels. Taking the sentence "I will be a teacher." as an example, the annotated version is: "I_FP will_MOD be_COP a_DT teacher_NN.", where FP represents the first-person pronoun; MOD represents a modal verb (different from the modal verb label MD in the primary annotation), which belongs to the Dependency Relation system and describes the grammatical relationship between the modal verb and the following verb; COP represents a copula; DT represents a determiner; and NN represents a singular noun. These grammatical feature labels can identify some grammatical language features.

[0079] The features are obvious, and the syntax can be converted into rule-based judgments with high accuracy. It can be understood as easily identifiable English grammar rules, and its rule judgment can be implemented programmatically. For example, COP is a linking verb, and its judgment method is implemented as follows:

[0080] (1) Define a set of candidate copulas, including all forms of the verb "be" and verbs such as seem, appear, feel, look, become; and define a set of adjective exceptions, including participles commonly used as adjectives, such as tired, interested, exciting, boring, etc.

[0081] (2) Check the position of the copula candidate word in the sentence (index j). If there is a word labeled VBN or VBG in at most 2 positions to its right (no other punctuation marks except quotation marks), and the stem of the VBN or VBG exists in the adjective exception set, then mark j as COP.

[0082] (3) If VBN or VBG appearing in at most two positions to the right of the candidate word at position j does not meet the conditions of rule 2, then further determine whether the word segment is VBG and whether the right side of it is immediately adjacent to a punctuation mark other than quotation marks. If so, mark j as COP.

[0083] (4) Define a prep set. If there are no VBN or VBG within two positions to the right of the candidate word at position j, but the part of speech of the next right position is a noun (starting with N), an adjective (starting with JJ), or a preposition (a word in the prep set), then mark j as COP.

[0084] (5) If a copula candidate does not meet any of the above rules, the word is not marked as COP.

[0085] In the example above, the features are obvious, and the syntax can be converted into rule-based judgments with high accuracy. These features are all labeled in the "secondary labeling". For complex grammatical features that are not obvious and are not easy to label accurately, the labeling is carried out in the subsequent second round of training.

[0086] 2.1.3 Based on the basic part-of-speech tags of primary annotation and the grammatical tags of secondary annotation, combined with Stanford PosTagger and English grammar knowledge, a rule-based "tertiary annotation" program was written in Python to annotate the text and obtain position tags.

[0087] In English, stances are expressed through specific stance vocabulary, broadly categorized into stance verbs, stance nouns, and stance adjectives. However, the mere presence of a word does not automatically indicate a stance; grammatical rules must be considered. For instance, a stance verb / adjective / noun + to / that clause is a form of stance expression, and corresponding rules can be established for its identification. Words that conform to these rules and are listed in the stance vocabulary can be considered to express a certain linguistic stance. The specific type of stance expressed is determined by its classification within the stance vocabulary.

[0088] The position vocabulary in this embodiment is a collection of existing position vocabulary in English and its synonyms and near-synonyms, and is classified by part of speech and type of complement clause.

[0089] Taking the sentence "I will be a teacher" as an example, the annotated version is: "I_FP will_MOD [PRSMD]be_COP a_DT teacher_NN." The added PRSMD is a position label, representing a modal verb expressing a predictive / willful stance. These position labels can be used to identify some linguistic features related to linguistic stance.

[0090] After three rounds of annotation, a standard JSON-formatted English financial news annotation dataset is obtained. This dataset includes sentences, vocabulary labels for each round of annotation, timestamps, sentence indexes, news dates, news sources, and types. The entire dataset is then divided into three parts: 80% is the training set A, 10% is the validation set of A, and 10% is the test set B. The validation set is used to evaluate the learning performance of the model in each training round, facilitating parameter tuning.

[0091] 2.1.4 Use training set A as the data source for training the RoBERTaLarge model, allowing the model to learn from the labeled results.

[0092] The high-performance tokenizer RobertaTokenizerFast, based on the RoBERTa model, breaks down sentences into WordPiece subwords. The first subword of each word is used as a token representing that word, mapped to its corresponding label, while subsequent subwords are ignored (-100). RobertaTokenizerFast uses a subword segmentation method, and WordPiece is a specific algorithm that implements this subword segmentation. Its core idea is to balance vocabulary size with the ability to handle unknown words. Unlike simply segmenting by spaces, WordPiece learns from a large corpus, breaking down text into meaningful subword units. For complex words like "unbelievably," it doesn't treat them as unrecognizable unknown words, but rather breaks them down into more common and meaningful segments like "un," "believe," and "##ably." Prefixes like "##" indicate that the subword segment is a subsequent part of a word, rather than the beginning of an independent word, which helps the model understand the internal structure of words. Because a complete word may be broken down into multiple subwords, it is necessary to map the label to the first subword and ignore subsequent subwords when performing sequence labeling tasks. This is to ensure that each original word is evaluated only once, mainly to avoid performance issues caused by repeated calculations.

[0093] The annotation training task was then assigned to the sequence labeling model RobertaForTokenClassification, based on the RoBERTa model. Cross-entropy loss was used to teach the model how to predict rule-based labels on a large corpus. In single-label tasks, cross-entropy loss simply maps the true label of each token to an integer ID and calculates it on the model's output logits (the raw, unnormalized scores generated by the model in the last layer) using nn.CrossEntropyLoss (the single-label classification loss function). Setting ignore_index=-100 automatically ignored padding (padding is a technique in NLP tasks to unify sentences / sequences of different lengths to the same length; in actual training, the model can only process tensor batches of fixed length at a time, so "placeholders"—PAD tokens—are added after each shorter sentence to make the entire batch have a consistent shape). For multi-label tasks, each token's label needs to be encoded as a 0 / 1 multi-hot vector, using nn.BCEWithLogitsLoss (a multi-label classification loss function). Similarly, padding is filtered out using a mask (a mask is used to mark the padding positions in the sentence and excludes these positions in necessary calculation steps, such as loss, attention, word vector aggregation, etc.). During training, the loss can be divided by the gradient accumulation steps, combined with mixed precision and warmup scheduling (warmup means that in the early stage of training, a very small learning rate is used to allow the model to gradually adapt to the gradient; then the learning rate is gradually increased to the target value, and then decayed according to a preset strategy) to ensure optimal memory usage and learning rate changes.

[0094] During training, gradient accumulation, mixed precision, and learning rate warmup are employed to accommodate the memory requirements of RoBERTaLarge. When training the model, batches are first split into several sub-batches (gradient accumulation). Each sub-batch is first autocast() to enable mixed precision. After obtaining logits from the forward pass, the loss is calculated using CrossEntropyLoss (single-label) or BCEWithLogitsLoss (multi-label). Then, the loss is divided by the accumulated steps and backpropagated. After the accumulated steps are complete, optimizer.step() (the optimizer is a gradient updater that converts gradient information into actual parameter changes, allowing the model to gradually approach the optimal solution) is called, and the gradients are reset to zero. Simultaneously, learning rate warmup is used. A linear or cosine scheduler is used to gradually increase the learning rate from 0 to the target value in the first 10% of the total steps, and then decays it according to a preset strategy in the remaining steps. This saves memory, improves training speed, and stabilizes the model in the early stages of training.

[0095] Ultimately, the fine-tuned model is able to generate annotation results consistent with expectations for sentences in the test set, realizing the transfer of existing knowledge to the model.

[0096] 2.1.5 The model performance was tested using test set B. A complete inference was performed using the fine-tuned RoBERTaLarge to obtain the predicted label for each token. The predicted label was then compared with the labeled dataset labels. The overall accuracy, macro / micro average F1, and precision and recall for each label were calculated to assess the model's transfer effect on language features. For labels with low recognition accuracy, the model was trained again on a new dataset until the model's labeling accuracy was close to that of the training set (≈98%). The first round of training for this language feature labeling model was then considered complete, and it was tentatively named DDU Tagger 1.0.

[0097] 2.2 Second round of training:

[0098] The first round of training mainly focuses on the annotation of language features containing complex grammar.

[0099] 2.2.1 Complex language features that could not be accurately labeled in the first round of training are autonomously encoded. Autonomous encoding refers to finding language features and assigning them specific labels. Because some language features containing complex grammar cannot be identified by writing rules, it is necessary to first obtain a dataset through manual annotation, and then feed the dataset into the model for training, ultimately obtaining a model-predictive-based labeler rather than a rule-based labeler. The significance of encoding is that in subsequent manual annotation, specific language features are uniformly labeled as complex grammatical feature labels, facilitating the model's learning of these labels in the second round of training.

[0100] For example, the relative pronouns (RP) that / which / who / that introduce a restrictive relative clause (RAC) modifying the subject (S) of the main clause and acting as the subject (S) of the subordinate clause can be encoded as [SS-RPRAC]. In the example sentence, this encoding is used as: "He that [SS-RPRAC] would eat the fruit must climb the tree."

[0101] 2.2.2 Obtain the dataset required for the second round of training.

[0102] First, the electronic versions of different English grammar dictionaries are converted into searchable text using OCR (Optical Character Recognition) and imported into a PostgreSQL database for quick retrieval of example sentences. Through retrieval and classification, different batches of original example sentences are formed.

[0103] Next, DDU Tagger 1.0 is used to pre-annotate all the original example sentences from different batches. Labels are generated under each existing rule-based token. Language features that cannot be accurately labeled are labeled as 0, reserved as unlabeled items. These can be manually cross-labeled by filling in the corresponding complex grammatical feature labels according to the self-encoding in step 2.2.1. The manual cross-labeling refers to the fact that during the manual labeling process, there may be situations where the identification of a certain complex grammatical feature label is inaccurate. Therefore, multiple people (at least two) need to cross-label the sentences in the same batch. If the labeling results are consistent, the dataset is submitted; otherwise, the administrator needs to perform the final labeling.

[0104] Labeling is performed for a specific language feature. Once all batches are labeled, they become a subset of the dataset. This process is repeated to generate C1, C2, C3...Cn, where n is the number of language feature types. All subsets together form the complete dataset C, with 10% reserved as the test set D.

[0105] 2.2.3 Train the complete dataset C using DDU Tagger 1.0 (using gradient accumulation, mixed precision, and learning rate warmup), evaluate macro / micro F1, precision, and recall, and perform error analysis on low-performing labels:

[0106] The model is considered qualified when the macro average F1 score is ≥0.92 and the recall of all new labels is ≥0.80. The model is evaluated on the test set D. If the overall accuracy is ≥98% and the F1 score of new labels is significantly improved, the second round of training of the model is completed and it is temporarily named DDU Tagger 2.0.

[0107] 2.3 Third round of training:

[0108] 2.3.1 Use DDU Tagger 2.0 to re-annotate the test set B used in the first round of training, count the correct and incorrect labels, and verify the model's labeling accuracy;

[0109] 2.3.2 For labels with high error rates, the model errors are checked, and an error label report is generated. Detailed error analysis is conducted based on dimensions such as error rate, confusion, rarity, and semantic ambiguity. Based on the analysis results, incremental collection, data augmentation, class weight adjustment, label smoothing, and rule supplementation are used to iteratively improve the high error rate labels. Subsequently, all erroneous instances and newly added labels are merged into the final training set E. On this basis, fine-grained learning rate, gradient accumulation, and mixed precision training are performed using DDU Tagger 2.0 as the initial weights until the macro-average F1 ≥ 0.92, the recall rate of each new label ≥ 0.80, and the overall accuracy ≥ 98%. The third round of model training is completed, resulting in a complete language feature annotation model, named DDU Tagger 3.0.

[0110] III. Forming a framework for analyzing financial discourse.

[0111] 3.1 Data Re-annotation. The English financial news database α was re-annotated using DDU Tagger 3.0 to obtain annotation results containing all language features.

[0112] 3.2 Standardized Frequency Calculation. Calculate the 1000-word frequency of all language features in each data point in the English financial news database. The calculation method is: Language Feature 1000-word Frequency = (Actual Frequency of Language Feature / Total Number of Tokens) * 1000.

[0113] 3.3 Factor Analysis Dimensionality Reduction. The 1000-word frequency of all linguistic features was imported into the SPSS statistical software package for factor analysis. Linguistic features with loadings below 0.35 were deleted, resulting in several feature dimensions. Each dimension contains several linguistic features with a clear co-occurrence trend. This clear co-occurrence trend refers to an observable, verifiable, and unambiguous co-occurrence trend; it is not a vague, accidental association, but a strong association that can be confirmed through specific manifestations.

[0114] 3.4 Dimension Naming. An English financial news database retrieval platform, tentatively named DDURetrieval 1.0, was developed using Python. This platform visualizes the database and allows the use of DDU Tagger 3.0 for word segmentation and tagging of articles. It automatically calculates the 1000-word frequency and dimensional scores of language features, batch retrieves key sentences based on language features, and statistically analyzes the distribution of co-occurrence of language features within the same dimension in key sentences. By observing this distribution, the specific classification and naming of each dimension were ultimately determined.

[0115] 3.5. Validation and Evaluation. A 10% sample of the English financial news database α was used to form a dimension-named validation dataset β. The annotation results of this dataset were summarized, and the language features were assigned to the corresponding dimensions for quantitative validation and qualitative evaluation.

[0116] 3.5.1 Quantitative Tests: On the β dataset, the variance explained and Cronbach's alpha of the loading features for each factor (dimension) are calculated, and the correlation coefficients between dimensions are tested. Subsequently, the text in the β dataset is clustered according to the projected dimensional scores to evaluate the effectiveness of the dimensional scores in distinguishing text differences. If the explained value is ≥50%, α ≥0.70, dimensional correlation ≤0.40, and the silhouette coefficient of the cluster is ≥0.3, then the dimensional structure is considered statistically reliable.

[0117] 3.5.2 Qualitative Assessment. Review the list of high-load features for each dimension and select typical sentences that appear in it. It is necessary to determine whether these sentences match the semantic category implied by the dimension name; if most sentences match the name, the name can be confirmed as reasonable; if there is a significant mismatch, the feature attribution or naming needs to be readjusted.

[0118] 3.6. Establishing Evaluation Criteria. The 1,000-word frequency of language features in the English financial news database α is used as the standardized frequency of these language features. Standardized scores for each dimension are calculated using the formula: Dimension Score = (Sum of 1,000-word frequencies of all positive features in this dimension - Sum of 1,000-word frequencies of all negative features in this dimension). Evaluation criteria for English financial discourse are established based on the standardized frequencies of each language feature and the standardized scores for each dimension. For any financial news text, DDU Tagger 3.0 is used to complete annotation, calculate the 1,000-word frequency of each language feature, automatically classify the language features into the predefined dimensions, calculate the score of the text in each dimension, and compare the feature frequencies and dimension scores with the standards to form the analysis results.

[0119] IV. Training a large-scale model for international financial discourse.

[0120] 4.1 Base Model:

[0121] This embodiment uses an open-source large-scale model as the base model for fine-tuning. GPT-OSS-120B (a high-parameter version of the GPT-OSS model) is used, but it is not limited to this; other similar large-scale models can also be used. All project code and documentation knowledge (the annotation part is still done by DDU Tagger 3.0) are segmented into short sentences or formulas. A unique token is generated using the GPT-OSS tokenizer, and 200 special IDs are reserved at the end of the vocabulary. Subsequently, a trainable Knowledge Embedding Table is added after the Transformer's embedding layer, assigning an ID to each piece of knowledge. A small MLP adapter is added next to the attention layer of each layer to specifically handle ID knowledge, enabling the model to ultimately master all the knowledge content of the research. This model, specifically trained for the English-language financial discourse vertical domain in this invention, is tentatively named INDEP v1, preparing for subsequent evaluation and question answering.

[0122] 4.2: Application Scenarios:

[0123] 4.2.1 Evaluation and improvement of English financial news discourse.

[0124] a. A user uploads an English financial news article. The INDEP v1 large-scale model for external financial discourse calls DDUTagger 3.0 to perform word segmentation and annotation on the English financial news article.

[0125] b.INDEP v1 calls the computation interface in DDU Retrieval 1.0 to automatically calculate the 1,000-word frequency and dimension score of language features;

[0126] c. Based on the news content and calculation results, automatically match the top 50 articles with high topic relevance in the English financial news database α to form a temporary dataset E, and calculate the linguistic features of this dataset, including the word frequency and dimensionality scores. As mentioned above, the English financial news database α has already undergone multiple clustering of the cleaned text and titles by fine-tuning the LDA topic clustering model during database construction, forming several first-, second-, and third-level topics, so topic relevance matching can be performed directly.

[0127] d. The INDEP v1.0 model calculates the similarity between the 1000-word frequency and dimensionality scores of the user-uploaded language features and the 1000-word frequency and dimensionality scores of the temporary dataset E. In this embodiment, the similarity is actually calculated by calculating the Euclidean distance between the two. This determines which language features the user can optimize, what the differences are in the dimensional features compared to the temporary dataset E, and searches for reference example sentences in the English financial news database α, presenting them to the user. INDEP v1 also calls the previously trained built-in knowledge base to reorganize the language, providing the user with improved sentences and explaining the differences in language features between the two, providing the user with a reference.

[0128] 4.2.2 Assist users in writing English financial press releases.

[0129] a. In DDU Retrieval 1.0, users select financial topics provided by the system. INDEP v1 presents all articles under that topic, sorted by relevance. Users can view the specific article's dimensional features, language features, stance features, main viewpoints, text summary, etc.

[0130] b. Users ask relevant questions to INDEP v1, enter prompt words, and INDEP v1 analyzes and responds.

[0131] This embodiment enables precise annotation, language feature analysis, and high-quality text polishing of English financial texts, providing full-process support for the editing of financial-related English texts for external communication, and effectively improving the readability, accuracy, and influence of financial texts among native English speakers.

[0132] Example 2:

[0133] This embodiment proposes an intelligent analysis and text optimization system for international financial discourse, such as... Figure 2 As shown, it includes:

[0134] Core data collection and cleaning layer: Collect English financial news corpora from foreign media and domestic and foreign propaganda media, preprocess them and store them in the database to form an English financial news database;

[0135] Language feature annotation model training layer: Based on machine learning methods, three rounds of model training are carried out. The first round of model training identifies simple language features including parts of speech, simple grammar, and language stance. The second round of model training identifies complex language features including complex grammar. The third round of model training is used for iterative improvement to obtain a language feature annotation model that identifies language features of English text.

[0136] Financial discourse analysis framework module: Calculates the 1,000-word frequency of language features, obtains several dimensions through factor analysis, and uses the standard scores of each dimension and the standardized frequency of language features as the basis for the analysis of financial discourse, which is used to calculate the Euclidean distance between the text to be analyzed.

[0137] Large-scale model for international financial discourse: An open-source large-scale model is selected as the base model and fine-tuned to enable it to call the language feature annotation model and perform analysis and calculation through the financial discourse analysis framework. Finally, an intelligent interactive large-scale model for international financial discourse is obtained, providing optimization suggestions for improving the quality of English financial texts.

[0138] The core data acquisition and cleaning layer includes:

[0139] Financial News Data Collection Module: Collects English financial news data from mainstream British and American media and major Chinese external propaganda media, including data such as titles, text, dates, sources, and authors.

[0140] Data preprocessing module: Performs data preprocessing, including deduplication: using regular expressions to filter English financial news corpora with the possibility of repetition, and performing deduplication based on text content similarity calculation; Data filtering: using regular expressions to find text with consecutive uppercase letters, judging whether there are problems with non-standard formatting or low content quality, and whether they need to be removed; Data standardization and cleaning, including: adjusting the title format; extracting proper nouns to form a proper noun dictionary; standardizing the capitalization, punctuation, common abbreviations, and capitalization of the first letter of sentences in the body text; cleaning the text, removing irrelevant content including illegal characters or URL links, and performing lemmatization on the cleaned title and body text.

[0141] English Financial News Database: The text content is segmented into title, date, source, watermark information, and body text; a database is constructed with eight fields: ID, title, cleaned title, watermark, source, date, body text, and cleaned body text; the segmented content is written into the corresponding fields, and each data entry has a unique ID, forming the English financial news database; by fine-tuning the LDA topic clustering model, the cleaned text and titles in the English financial news database are clustered multiple times to form several first-, second-, and third-level topics.

[0142] The language feature annotation model training layers include:

[0143] The first round of training involves: applying part-of-speech tagging tools to perform "primary tagging" on the English text content, labeling each word with its basic part-of-speech tag; based on these basic part-of-speech tags, and according to English grammar rules, programming is used to perform "secondary tagging" on the relevant words, labeling them with grammatical feature tags; based on position words and clauses following them in English, programming is used to determine whether a word expresses a position, and if so, "tertiary tagging" is performed on that word, labeling it with a position tag; the dataset formed after the three tagging processes is divided into training set A, validation set A, and test set B; RoBERTa-Large is used as the base model to train the annotation on training set A. The process involves using RobertaTokenizerFast to split sentences into WordPiece subwords, and using the first subword of each word as a token, mapping it to the corresponding tag of the word; then the annotation training task is assigned to RobertaForTokenClassification, using cross-entropy loss to allow the model to learn how to predict rule tags on a large-scale corpus. During training, gradient accumulation, mixed precision, and learning rate warmup are used to adapt to the memory requirements of RoBERTaLarge; finally, the DDU Tagger 1.0 language feature annotation model is obtained. The model performance was tested using test set B.

[0144] The second round of training involves: autonomously encoding complex language features in the English text that cannot be accurately labeled at present, for use in labeling complex grammatical features; converting different electronic versions of English grammar dictionaries into searchable text and importing them into the database; generating different batches of original example sentences through retrieval and classification; using DDU Tagger 1.0 to pre-label the different batches of original example sentences, reserving items to be labeled for complex grammatical features; after all batches of the same language feature are labeled, they are treated as a subset, and this process is repeated to generate subsets C1, C2, C3...Cn, where n is the number of language feature types. All subsets together form the complete dataset C, and 10% is reserved as the test set D; the DDU Tagger 1.0 language feature labeling model is used to train the complete dataset C to obtain the DDU Tagger 2.0 language feature labeling model.

[0145] The third round of training: The DDU Tagger 2.0 language feature annotation model was used to annotate the test set B used in the first round of training again, and the correct and incorrect labels were counted to verify the model's annotation accuracy. The high error rate labels were iteratively improved. Then, all subsets, erroneous instances and newly added labels were merged into the final training set E, and retrained on this set with DDU Tagger 2.0 as the initial weights until the macro average F1 ≥ 0.92, the recall of each new label ≥ 0.80 and the overall accuracy ≥ 98%, thus obtaining the language feature annotation model DDU Tagger 3.0.

[0146] The financial discourse analysis framework module includes:

[0147] Language feature annotation: The English text in the English financial news database was re-annotated using the DDU Tagger 3.0 language feature annotation model to obtain annotation results containing all language features;

[0148] Thousand-word frequency calculation: Calculate the thousand-word frequency of all language features in each English text data. The calculation method is: Language feature thousand-word frequency = (actual frequency of language feature / total number of tokens) * 1000;

[0149] Factor analysis dimensionality reduction: The 1000-word frequency of all language features was imported into the social science statistical software SPSS for factor analysis. Language features with loadings below 0.35 were deleted, forming several feature dimensions. Each dimension contains several language features with obvious co-occurrence trends. The obvious co-occurrence trends refer to co-occurrence trends that are observable, verifiable, and unambiguous. They are not vague or accidental associations, but strong associations that can be confirmed through specific manifestations.

[0150] Dimensional Feature Classification: This involves statistically analyzing the distribution of co-occurrence of linguistic features within the same dimension in key sentences. The statistical method involves using DDU Tagger 3.0 to segment and annotate the article, automatically calculating the 1000-word frequency and dimensional score of the linguistic features, and then batch retrieving key sentences using these features. By observing the co-occurrence distribution, the specific classification names for each dimension are determined. A dimension naming validation dataset is formed through sampling, and linguistic features are assigned to their corresponding dimensions for quantitative and qualitative evaluation. The quantitative evaluation includes calculating the variance explained by the feature loadings and the Cronbach's alpha coefficient for each dimension on the dimension naming validation dataset, and testing the correlation coefficients between dimensions. Subsequently, the dimension naming validation dataset is clustered according to the projected dimensional scores to evaluate the effectiveness of the dimensional scores in distinguishing text differences. The qualitative evaluation includes reviewing the list of high-load features for each dimension and selecting those that are... Typical sentences are selected to determine whether they conform to the semantic categories implied by the dimension naming. If most sentences match the naming, the naming is confirmed to be reasonable. If significant mismatches occur, the feature attribution or naming needs to be readjusted. The 1,000-word frequency of the language features is used as the standardized frequency, and the standard score of each dimension is calculated. The calculation formula is: Dimension score = (Sum of 1,000-word frequencies of all positive features under this dimension - Sum of 1,000-word frequencies of all negative features under this dimension). Based on the standardized frequency of each language feature and the standard score of each dimension, an evaluation standard for English financial discourse is formed for comparison and analysis results.

[0151] A comprehensive model of international financial discourse:

[0152] GPT-OSS-120B (a high-parameter version of the GPT-OSS model) was selected as the base model for fine-tuning. All project code and document knowledge (the annotation part was still done by DDU Tagger 3.0) were segmented into short sentences or formulas. The GPT-OSS tokenizer was used to generate exclusive tokens, and 200 special IDs were reserved at the end of the vocabulary. Subsequently, a trainable knowledge embedding table was added after the embedding layer of the Transformer, with each knowledge item corresponding to an ID. A small MLP adapter was added next to the attention layer to specifically handle ID knowledge, so that the model could finally master all the knowledge content of the research. As a large-scale model for the English foreign financial discourse vertical domain specifically trained for this invention, it is temporarily named INDEP v1, and is prepared for subsequent evaluation and question answering.

[0153] A database retrieval platform for English financial news, named DDURetrieval 1.0, was developed using Python. It provides a user interface, an intelligent analysis system based on the INDEP v1 large-scale model of foreign financial discourse, a web-based annotation tool that can call the DDU Tagger 3.0 language feature annotation model, and a database retrieval platform for searching English financial news databases.

[0154] A user uploads an English financial news article. The INDEP v1 large-scale model for the international financial discourse vertical domain calls DDUTagger 3.0 to segment and annotate the English financial news article; it calls the computation interface in DDU Retrieval 1.0 to automatically calculate the word frequency and dimensionality score of language features; based on the news content and calculation results, it automatically matches the top 50 articles with high topic relevance in the English financial news database α to form a temporary dataset E, and calculates the word frequency and dimensionality score of the language features in this dataset; it calculates the Euclidean distance between the word frequency and dimensionality score of the user's uploaded language features and the word frequency and dimensionality score of the temporary dataset E, determines which language features the user can optimize, what the differences are in the dimensional features compared to the dimensional features of the temporary dataset E, and searches for reference example sentences in the English financial news database α to present to the user. INDEP v1 also calls the previously trained built-in knowledge base to reorganize the language and provide the user with improved sentences, explaining the differences in language features between the two, providing the user with a reference.

[0155] The intelligent analysis and text optimization system for foreign financial discourse proposed in this embodiment can realize the intelligent analysis and text optimization method for foreign financial discourse described in Embodiment 1, and has the same beneficial technical effects as Embodiment 1.

[0156] The above-described embodiments are merely preferred embodiments of the present invention and are only used to help understand the method and core ideas of this application. The scope of protection of the present invention is not limited to the above embodiments, and all technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for intelligent analysis and text optimization of international financial discourse, characterized in that, include: S1. Core Data Collection and Cleaning: Collect English financial news corpora from foreign media and domestic and foreign propaganda media, preprocess them and store them in the database to form an English financial news database. S2. Training the language feature annotation model: Based on machine learning methods, three rounds of model training are conducted. The first round of model training identifies language features including parts of speech, simple grammar, and lexical stance. The second round of model training identifies language features including complex grammar. The third round of model training is used for iterative improvement, and finally, an English text language feature annotation model is obtained. S3. Constructing a framework for financial discourse analysis: Calculating the 1000-word frequency of language features, obtaining several dimensions through factor analysis, and using the scores of each dimension and the standardized frequency of language features as the analytical basis for the financial discourse analysis framework, used to calculate the Euclidean distance between the text and the text to be analyzed; including: S301. Use the language feature annotation model to re-annotate the English text in the English financial news database to obtain annotation results containing all language features; S302. Calculate the 1000-word frequency of all language features in each English text data. The calculation method is: Language feature 1000-word frequency = (Actual frequency of language features / Total number of text tokens) * 1000; S303. Import the 1000-word frequency of all language features into the social science statistical software SPSS for factor analysis, delete language features with loading values ​​below 0.35, and form several feature dimensions. Each dimension contains several language features with obvious co-occurrence trends. S304. Statistically analyze the distribution of language features co-occurring in key sentences under the same dimension, and determine the specific classification and naming of each dimension by observing the distribution. S305. A dimension naming validation dataset is formed by sampling, and language features are divided into corresponding dimensions for quantitative testing and qualitative evaluation. The quantitative testing includes calculating the variance explained by the feature loadings and the Cronbach's alpha coefficient for each dimension on the dimension naming validation dataset, and testing the correlation coefficients between dimensions. Then, the dimension naming validation dataset is clustered according to the projected dimension scores to evaluate the effectiveness of the dimension scores in distinguishing text differences. The qualitative evaluation includes reviewing the list of high-load features corresponding to each dimension, selecting typical sentences, and judging whether these sentences match the semantic category implied by the dimension naming. If the sentences match the naming, the naming is confirmed to be reasonable. If there is a mismatch, the feature assignment or naming needs to be readjusted. S306. Using the 1000-word frequency of the aforementioned language features as the standardized frequency, and calculating the scores for each dimension, the calculation formula is: Dimension score = (Sum of 1000-word frequencies of all positive features under this dimension - Sum of 1000-word frequencies of all negative features under this dimension); Based on the standardized frequency of each language feature and the scores for each dimension, an evaluation standard for English financial discourse is formed, which is used to compare with the text to be analyzed and form the analysis results; S4. Training a large-scale model for international financial discourse: Using an open-source large model as a base, fine-tuning it to enable it to master the knowledge from steps S1-S3, and obtain an intelligent interactive large-scale model for international financial discourse, providing optimization suggestions for improving the quality of English financial texts.

2. The intelligent analysis and text optimization method for international financial discourse according to claim 1, characterized in that, The preprocessing method described in step S1 includes: S101. Deduplication: Use regular expressions to filter out English financial news corpora that may be duplicated, and perform deduplication based on the similarity calculation of text content; S102. Data Filtering: Use regular expressions to find text containing consecutive uppercase letters, determine if there are problems with non-standard formatting or low content quality, and whether it needs to be removed.

3. The intelligent analysis and text optimization method for international financial discourse according to claim 1, characterized in that, In step S1, after preprocessing and before data entry, data standardization and cleaning are required, including: adjusting the title format; extracting proper nouns to form a proper noun dictionary; standardizing the capitalization, punctuation, common abbreviations, and capitalization of the first letter of sentences in the main text; cleaning the text to remove irrelevant content, including illegal characters or URL links, and restoring the word form of the cleaned title and main text.

4. The intelligent analysis and text optimization method for international financial discourse according to claim 1, characterized in that, The data entry in step S1 includes: segmenting the text content into title, date, source, watermark information, and body text; constructing a database with eight fields: ID, title, cleaned title, watermark, source, date, body text, and cleaned body text; writing the segmented content into the corresponding fields, with each data entry having a unique ID, thus forming an English financial news database; and performing multiple clustering operations on the cleaned text and titles in the English financial news database by fine-tuning the LDA topic clustering model to form several first-, second-, and third-level topics.

5. The intelligent analysis and text optimization method for international financial discourse according to claim 1, characterized in that, The first round of training of the language feature annotation model in step S2 includes: S201. Apply a part-of-speech tagging tool to perform "one-time tagging" on the English text content, tagging the basic part-of-speech label for each word; S202. Based on the aforementioned basic part-of-speech tags and combined with English grammar rules, programmatically perform "secondary tagging" on the relevant words to tag grammatical feature tags; S203. Based on the English position words and the corresponding clause grammar rules, programmatically determine whether the word expresses a position. If it does, mark the position word with a "three-fold label". S204. Divide the dataset formed after three annotations into training set A, validation set of A, and test set B; S205. Using RoBERTa-Large as the base model, the training set A is labeled and trained. The validation set of A is used to check the learning effect of each learning round. The model performance is tested through the test set B to obtain the DDU Tagger 1.0 language feature annotation model.

6. The intelligent analysis and text optimization method for international financial discourse according to claim 5, characterized in that, Step S205 specifically includes: using RobertaTokenizerFast to split the sentence into WordPiece subwords, and using the first subword of each word as a token to map to the corresponding label of the word; then assigning the annotation training task to RobertaForTokenClassification, and using cross-entropy loss to allow the model to learn how to predict rule labels on a large-scale corpus. During the training process, gradient accumulation, mixed precision, and learning rate warm-up are used to adapt to the memory requirements of RoBERTa-Large; finally, the DDU Tagger 1.0 language feature annotation model is obtained.

7. The intelligent analysis and text optimization method for international financial discourse according to claim 5, characterized in that, The second round of training of the language feature annotation model in step S2 includes: S211. The language features containing complex grammar in the English text that cannot be accurately labeled at present are autonomously encoded for the labeling of complex grammar feature tags. S212. Convert different electronic versions of English grammar dictionaries into searchable text and import them into the database; through indexing and classification, form different batches of original example sentences; use DDU Tagger 1.0 to pre-annotate the different batches of original example sentences, reserving items to be annotated for complex language features; after all batches of the same language feature are annotated, they are taken as a subset, and this process is repeated to generate subsets C1, C2, C3, ..., Cn, where n is the number of language feature types. All subsets together form a complete dataset C, and 10% is reserved as a test set D; S213. Using the DDU Tagger 1.0 language feature annotation model as the initial weights, train it using dataset C to obtain the DDU Tagger 2.0 language feature annotation model.

8. The intelligent analysis and text optimization method for international financial discourse according to claim 5, characterized in that, The third round of training for the language feature annotation model in step S2 includes: The DDU Tagger 2.0 language feature annotation model was used to re-annotate the test set B used in the first round of training. The correct and incorrect labels were counted to verify the accuracy of the model annotation. The high error rate labels were iteratively improved. Then, all erroneous instances and newly added labels were merged into the final training set E. Based on this, the DDU Tagger 2.0 language feature annotation model was retrained with the initial weights until the macro average F1 was ≥0.92, the recall of each new label was ≥0.80, and the overall accuracy was ≥98%, thus obtaining the final language feature annotation model DDU Tagger 3.

0.

9. An intelligent analysis and text optimization system for international financial discourse, characterized in that, include: The core data collection and cleaning layer is sequentially connected, the language feature annotation model training layer is connected, the financial discourse analysis framework module is connected, and the large-scale model of external financial discourse is connected. Core data collection and cleaning layer: Collect English financial news corpora from foreign media and domestic and foreign propaganda media, preprocess them and store them in the database to form an English financial news database; Language feature annotation model training layer: Based on machine learning methods, three rounds of model training are carried out. The first round of model training identifies language features including parts of speech, simple grammar, and lexical stance. The second round of model training identifies language features including complex grammar. The third round of model training is used for iterative improvement, and finally, the English text language feature annotation model is obtained. The financial discourse analysis framework module calculates the 1,000-word frequency of language features, obtains several dimensions through factor analysis, and uses the scores of each dimension and the standardized frequency of language features as the basis for the analysis of the financial discourse analysis framework to calculate the Euclidean distance between the text to be analyzed; including: using the language feature annotation model to re-annotate the English text in the English financial news database to obtain annotation results containing all language features; The 1,000-word frequency of all linguistic features in each English text dataset was calculated as follows: linguistic feature 1,000-word frequency = (actual frequency of linguistic feature / total number of text tokens) * 1000. The 1,000-word frequencies of all linguistic features were imported into SPSS (a social science statistical software) for factor analysis. Linguistic features with loadings below 0.35 were removed, resulting in several feature dimensions. Each dimension contained several linguistic features with a clear co-occurrence trend. The distribution of co-occurrence of linguistic features within the same dimension in key sentences was statistically analyzed. Based on the observation of the distribution, the specific classification and naming of each dimension were determined. A dimension naming validation dataset was formed by sampling, and linguistic features were assigned to their corresponding dimensions for quantitative testing and qualitative evaluation. The quantitative testing included calculating the variance explained by the loading features of each dimension on the dimension naming validation dataset and the Cronbach's alpha coefficients. The data is counted, and the correlation coefficients between dimensions are tested. Then, the dimension naming validation dataset is clustered according to the projected dimension scores to evaluate the effectiveness of the dimension scores in distinguishing text differences. The qualitative evaluation includes reviewing the list of high-load features corresponding to each dimension and selecting typical sentences to determine whether these sentences match the semantic category implied by the dimension naming. If the sentences match the naming, the naming is confirmed to be reasonable; if there is a mismatch, the feature attribution or naming needs to be readjusted. The 1000-word frequency of the language features is used as the standardized frequency, and the score of each dimension is calculated using the formula: Dimension score = (Sum of 1000-word frequencies of all positive features under this dimension - Sum of 1000-word frequencies of all negative features under this dimension). Based on the standardized frequency of each language feature and the score of each dimension, an evaluation standard for English financial discourse is formed, which is used to compare with the text to be analyzed and form the analysis results. Large-scale model for international financial discourse: This model uses an open-source large-scale model as a foundation for fine-tuning, enabling it to master the knowledge of the core data collection and cleaning layer, the language feature annotation model training layer, and the financial discourse analysis framework module. This results in an intelligent interactive large-scale model for international financial discourse, providing optimization suggestions for improving the quality of English financial texts.

Images