Method for training financial report analysis model, financial report analysis method, device and medium

Abstract

Embodiments of the present disclosure relate to a method for training a financial report analysis model, a financial report analysis method, a device and a medium, comprising: constructing a vocabulary containing sentiment words, entity words and bigrams and determining the type and weight attribute of each word in the vocabulary; receiving input text and performing word segmentation processing, annotating the type and weight attribute of the text after word segmentation processing based on the constructed vocabulary and predefined annotation rules; matching and combining the annotated texts based on predefined matching rules; calculating the sentiment score corresponding to a sentence composed of single words based on the scores of single words; repeatedly performing the above operations on financial report samples to obtain report text and corresponding sentiment scores; and performing model training on the obtained report text and corresponding sentiment scores based on a pre-trained BERT model to obtain a desired financial report analysis model.

Description

Technical Field

[0001] Embodiments of the present disclosure generally relate to the field of financial information processing, and more specifically to a method for training a financial report analysis model, a financial report analysis method, a computing device, and a computer-readable storage medium. Background Art

[0002] With the continuous development of the financial market, more and more investors have begun to pay attention to using text mining and sentiment analysis technology to obtain information on financial market trends, obtain emotional data of market participants from texts collected through news reports, social media, public opinion and other channels, and then provide decision-making support for investors.

[0003] Using traditional machine learning classification models to perform text sentiment analysis requires a huge amount of manually annotated data, that is, before machine learning, dedicated manual labelers are needed to classify the sentiment of a huge amount of text data (for example, into three categories: positive, negative, and neutral), and then use the machine learning model to learn and train these text data so that the model can learn to imitate humans in classifying text data sentiment. However, in the above process, there are several prominent problems.

[0004] 1. For machine learning models, especially modern deep learning models, the amount of training data required is often huge, which means that the workload of manual labeling required in this process is also huge, which requires a lot of manpower and time.

[0005] 2. For the sentiment classification of data, manual labeling has a strong subjective tendency and it is difficult to unify the classification standards. This can easily lead to large deviations in the data labeled by different labelers due to personal subjective differences and inconsistent standards.

[0006] 3. Currently, most sentiment analysis models and algorithms are implemented through text classification; for example, text is divided into two categories, positive and negative, or into three categories, positive, neutral and negative. Since sentiment analysis is inherently highly subjective, there is a lack of unified classification standards to clarify whether the target text should adopt the aforementioned two-category, three-category or four-category sentiment classification methods. If one of the classification methods is fixed, it will not be able to meet the diverse needs of different customers.

[0007] 4. General text sentiment analysis methods can only perform sentiment analysis on a sentence or an entire article. They lack multi-level and multi-angle sentiment analysis of paragraphs and sections of the text, and are not suitable for more complex application scenarios proposed by customers.

[0008] In summary, the shortcomings of traditional financial report sentiment analysis solutions are: large workload, lack of unified data sentiment classification standards, inflexible text classification methods, and single application scenarios of sentiment analysis. Summary of the invention

[0009] In response to the above problems, the present disclosure provides a method for training a financial report analysis model, a financial report analysis method based on the financial report analysis model obtained by the model training method, a computing device and a computer-readable storage medium, which can realize automatic annotation and sentiment scoring of text data in the financial field, while saving a large amount of manual annotation workload, eliminating the subjective bias generated when annotating sentiment data, performing three-dimensional sentiment analysis of financial reports from multiple angles such as sentences, paragraphs and chapters, and assigning corresponding sentiment scores to texts of each granularity to meet customers' more complex application scenarios.

[0010] According to a first aspect of the present disclosure, a method for training a financial report analysis model is provided, comprising: constructing a vocabulary containing sentiment words, entity words and bigrams and determining the type and weight attribute of each word in the vocabulary; receiving input text and performing word segmentation processing, annotating the type and weight attribute of the text after word segmentation processing based on the constructed vocabulary and predefined annotation rules; matching and combining the annotated texts based on predefined matching rules; calculating the sentiment score corresponding to the sentence composed of single words based on the score of the single word; repeatedly performing the above operations on the collected financial report samples to obtain the report text and the sentiment score corresponding to the report text as training samples for training the financial report analysis model; and training the financial report analysis model on the obtained report text and the sentiment score corresponding to the report text based on the pre-trained language representation BERT model to obtain the desired financial report analysis model.

[0011] According to a second aspect of the present disclosure, a financial report analysis method is provided, comprising: inputting a financial report into the financial report analysis model as described above to predict a sentiment score of the input financial report.

[0012] According to a third aspect of the present disclosure, a computing device is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the method of the first aspect of the present disclosure.

[0013] In a fourth aspect of the present disclosure, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to cause a computer to execute the method of the first aspect of the present disclosure.

[0014] In some embodiments, performing word segmentation processing on the input text includes: splitting the text into one or more sentences based on punctuation marks in the text; using a word segmentation algorithm to split the split sentences into one or more words; and marking the corresponding types and weight attributes of the words that match the sentiment words, entity words and bigrams in the vocabulary with preset marking symbols.

[0015] In some embodiments, matching and combining annotated text based on predefined matching rules includes: determining a fine-grained scale and a coarse-grained scale for the text based on a predefined granularity threshold; at the fine-grained scale, combining annotated entity words or adjacent entity words that meet predefined rules into entity combinations; at the fine-grained scale, combining annotated sentiment words or adjacent sentiment words that meet predefined rules into sentiment combinations; and at the coarse-grained scale, matching and combining entity combinations and sentiment combinations.

[0016] In some embodiments, matching and combining entity combinations and sentiment combinations includes: preferentially matching entity combinations and sentiment combinations in the same sentence, wherein the matching form of the entity combination and the sentiment combination takes precedence over the matching form of the sentiment combination and the entity combination; secondly matching entity combinations and sentiment combinations in different sentences, wherein only the matching form of the entity combination and the sentiment combination is retained, and the matching form of the sentiment combination and the entity combination is not retained; in response to no matching of the entity combination, retaining the sentiment combination as a separate word combination; and in response to no matching of the sentiment combination, performing a correction operation.

[0017] In some embodiments, calculating the sentiment score corresponding to a sentence composed of single words based on the scores of single words includes: for a word combination composed of words, multiplying the weights or absolute values ​​of the weights of all words in the word combination as the sentiment score of the word combination; and for a sentence composed of a word combination, adding the sentiment scores of all word combinations and bigrams in the sentence as the sentiment score of the sentence.

[0018] In some embodiments, the method for training a financial report analysis model further includes: processing the sentiment scores of the obtained sentences based on different text granularities, and applying the processed sentiment scores to the financial report analysis model training, thereby obtaining a financial report analysis model for predicting the sentiment scores of texts of corresponding granularity.

[0019] In some embodiments, obtaining a financial report analysis model for predicting sentiment scores of text of corresponding granularity includes: applying the sentiment scores of sentences to financial report analysis model training to obtain a financial report analysis model for predicting the sentiment scores of sentences; or synthesizing the sentiment scores of paragraphs based on the sentiment scores of sentences, applying the sentiment scores of paragraphs to financial report analysis model training to obtain a financial report analysis model for predicting the sentiment scores of paragraphs; or synthesizing the sentiment scores of chapters based on the sentiment scores of sentences, applying the sentiment scores of chapters to financial report analysis model training to obtain a financial report analysis model for predicting the sentiment scores of chapters.

[0020] In some embodiments, the two words in the bigram are in the same sentence, and there are no other entity words or sentiment words between the two words, and the entity combination and the sentiment combination are not matched and combined across any word in the bigram.

[0021] In some embodiments, the financial report analysis method further includes: processing the obtained predicted sentiment score of the financial report based on different text granularities.

[0022] In some embodiments, processing the predicted sentiment scores of financial reports obtained based on different text granularities includes: after predicting the sentiment score of a sentence in response to the financial report analysis model, synthesizing the predicted sentiment score of the paragraph or the chapter based on the predicted sentiment score of the sentence; or after predicting the sentiment score of the paragraph in response to the financial report analysis model, synthesizing the predicted sentiment score of the chapter based on the predicted sentiment score of the paragraph; or applying the financial report analysis model to directly predict the sentiment score of the chapter.

[0023] It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present disclosure, nor is it intended to limit the scope of the present disclosure. Other features of the present disclosure will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The above and other features, advantages and aspects of the embodiments of the present disclosure will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. In the accompanying drawings, the same or similar reference numerals represent the same or similar elements.

[0025] Figure 1 A schematic diagram of a system for implementing a financial report analysis method according to an embodiment of the present invention is shown.

[0026] Figure 2 A flowchart of a method 200 for training a financial report analysis model according to an embodiment of the present disclosure is shown.

[0027] Figure 3A flowchart of a financial report analysis method 300 according to an embodiment of the present disclosure is shown.

[0028] Figure 4 A block diagram of an electronic device 400 according to an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0029] The following is a description of exemplary embodiments of the present disclosure in conjunction with the accompanying drawings, including various details of the embodiments of the present disclosure to facilitate understanding, which should be considered as merely exemplary. Therefore, it should be recognized by those of ordinary skill in the art that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Similarly, for the sake of clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description.

[0030] As used herein, the term "including" and its variations mean open inclusion, i.e., "including but not limited to". Unless otherwise stated, the term "or" means "and / or". The term "based on" means "based at least in part on". The terms "an example embodiment" and "an embodiment" mean "at least one example embodiment". The term "another embodiment" means "at least one additional embodiment". The terms "first", "second", etc. may refer to different or the same objects. Other explicit and implicit definitions may also be included below.

[0031] As described above, using traditional machine learning classification models to perform text sentiment analysis requires a huge amount of manual data annotation, which inevitably consumes a lot of manpower and time. In addition, manual annotation has strong subjective bias and lacks unified text classification standards and sentiment analysis of texts of different granularities. Therefore, the existing financial report sentiment analysis methods cannot meet the increasingly complex and diverse application requirements put forward by customers.

[0032] In order to at least partially solve one or more of the above problems and other potential problems, the example embodiment of the present disclosure proposes a method for training a financial report analysis model. By establishing a basic vocabulary for annotating texts, predefining matching rules for annotated words in texts, and presetting a sentiment score calculation method from a single word to the entire text, the machine automatically annotates data with unified standards, and the computer automatically scores the sentiment of texts of various granularities such as sentences, paragraphs, and chapters, providing investors with objective and reliable sentiment data, and assisting investors in making correct investment decisions based on the sentiments expressed by financial analysts in the texts. Since the calculated sentiment scores are continuous, customers can perform customized sentiment classification based on the obtained sentiment scores, thereby making the solution of the present invention more widely applicable.

[0033] Figure 1FIG. 2 is a schematic diagram of a system for implementing a financial report analysis method according to an embodiment of the present invention. Figure 1 As shown in FIG, the system includes a computing device 110, a financial report data management device 130, and a network 140. The computing device 110 and the financial report data management device 130 can perform data exchange through the network 140 (eg, the Internet).

[0034] The financial report data management device 130 may store one or more financial report data. The financial report data management device 130 may also send the stored financial report data to the computing device 110. The financial report data management device may be, for example, but not limited to, an electronic computer, a network server, a storage calculator, etc. The financial report data management device 130 may collect financial report data from multiple sources, such as news reports, company annual reports, industry surveys, government information, etc., and may classify and manage the collected one or more financial report data.

[0035] Regarding the computing device 110, it is used, for example, to obtain financial report data from the financial report data management device 130; and further informatize the obtained financial report data. The computing device 110 can also generate multiple visual graphics to display the financial report data based on the informatized financial report data. By adjusting the visual graphics that display the financial report data, the screening and training of the financial report data can be achieved. The computing device 110 can receive text data from a user or the financial report data management device 130. By applying the method disclosed in the present invention, the computing device 110 can perform text sentiment analysis on research reports in the financial field.

[0036] The computing device 110 may have one or more processing units, including dedicated processing units such as GPU, FPGA and ASIC, and general processing units such as CPU. In addition, one or more virtual machines may also be running on each computing device 110. In some embodiments, the computing device 110 and the financial report data management device 130 may be integrated or separately arranged. In some embodiments, the computing device 110 includes, for example, a vocabulary library arrangement and maintenance unit 112, a text sentiment automatic scoring unit 114, a financial report analysis model training unit 116, and a financial report sentiment score prediction unit 118.

[0037] A word library collating and maintaining unit 112, wherein the word library collating and maintaining unit 112 is configured to construct a word library including sentiment words, entity words and bigrams and determine the type and weight attribute of each word in the word library;

[0038] The text sentiment automatic scoring unit 114 is configured to perform word segmentation processing on the input text, annotate the text type and weight attribute after word segmentation processing based on the constructed vocabulary and predefined annotation rules, match and combine the annotated text based on predefined matching rules, and calculate the sentiment score corresponding to the sentence composed of single words based on the scores of single words;

[0039] Financial report analysis model training 116, wherein the financial report analysis model training 116 is configured to repeatedly perform sentiment scoring operations on the collected financial report samples, obtain report texts and sentiment scores corresponding to the report texts, and perform financial report analysis model training on the obtained report texts and sentiment scores corresponding to the report texts based on a pre-trained language representation BERT model, thereby obtaining a desired financial report analysis model;

[0040] A financial report sentiment score prediction unit 118 is configured to predict a sentiment score of an input financial report based on a desired financial report analysis model.

[0041] Figure 2 1 is a flowchart of a method 200 for training a financial report analysis model according to an embodiment of the present disclosure. The method 200 may be performed as follows: Figure 1 The computing device 110 shown may also be executed in Figure 4 The method 200 is executed at the electronic device 400. It should be understood that the method 200 may further include additional blocks not shown and / or may omit the blocks shown, and the scope of the present disclosure is not limited in this respect.

[0042] In step 202 , the computing device 110 may construct a vocabulary including sentiment words, entity words, and bigrams and determine a type and a weight attribute of each word in the vocabulary.

[0043] The weight is a floating point number: a positive number represents positive emotion, a negative number represents negative emotion, 0 represents neutral, and the absolute value represents the strength of the emotion.

[0044] Taking into account the complexity of natural language, the present invention sets different types for words in the vocabulary, which can capture part of the complexity in natural language and improve the situation where the rules are usually too simple and rigid, resulting in misunderstanding of text emotions.

[0045] Sentiment words in the vocabulary are usually divided into three categories: sentiment 0, sentiment 1 and sentiment 2, and entity words are divided into three categories: entity 0, entity 1 and entity 2.

[0046] Sentiment 0 represents directional sentiment words, such as increase, decrease, etc., which are used to indicate the direction of entity change. The sentiment tendency changes with the change of the matched entity. Example: Net Profit | Increase, Revenue | Decrease, Cost | Increase; among which Net Profit Increase is a positive sentiment, while Revenue Decrease and Cost Increase are negative sentiments.

[0047] Sentiment 1 represents non-directional sentiment words, such as improvement, pessimism, etc. Its sentiment tendency is fixed and will not change with the change of the matched entity. Example: Revenue|Improvement, Cost|Improvement; no matter what the entity word in the word combination is, it always indicates a positive sentiment.

[0048] Emotion 2 represents modifying emotional words, such as greatly, slightly, not, and difficult to modify emotional words, indicating degree or negation, which can solve some situations where emotional reversal is caused by negative prefixes or suffixes. Examples: greatly|increase, slightly|decrease, not|decline, and difficult to|improve; among which greatly increase indicates increasing emotional intensity, slightly decrease indicates weakening emotional intensity, and not decline and difficult to improve indicate reversing emotional direction.

[0049] Some common sentiment words are shown in Table 1 below:

[0050]

[0051] Entity 0 represents regular entity words, such as net profit, cost, performance, sales volume, etc. For example: net profit|increase, cost|increase; because net profit increase expresses positive emotions, net profit is a positive entity word, and cost increase expresses negative emotions, so cost is a negative entity word.

[0052] Entity 1 represents abstract entity words, such as decline, growth rate, etc., which are often followed by regular entity words to form a new entity. Examples: net profit | decline | expansion, sales volume | growth rate | decline.

[0053] Entity 2 represents non-directional abstract entity words, such as risk, shortage, pressure, uncertainty, etc. When matched with sentiment words or other entities, the positive and negative signs of the combination are based on the signs of entity 2. Examples: profit|risk, downside|risk, bad debt|risk|increase, profit|pressure; these combinations all express negative emotions.

[0054] Some common entity words are shown in Table 2 below:

[0055]

[0056] A bigram is a phrase consisting of two fixed combinations of words. Generally speaking, if a commonly used word combination can express emotions but cannot be reflected as a matching combination of entity words and sentiment words, they can be included in the bigram word library. For example: (ranking, top), (international, first), (capacity, climbing), (risk, clearance), etc.

[0057] Some common bigrams are shown in Table 3 below:

[0058]

[0059] In step 204 , the computing device 110 may receive the input text and perform word segmentation processing, and annotate the text after word segmentation processing with types and weight attributes based on the constructed vocabulary and predefined annotation rules.

[0060] The word segmentation processing performed on the input text includes: splitting an article or text block into one or more sentences based on punctuation marks in the text, such as periods, question marks, exclamation marks, semicolons, colons, etc.; using a word segmentation algorithm (such as jieba) to split the split sentences into one or more words; and marking the corresponding types and weight attributes of the sentiment words, entity words and bigrams in the vocabulary with preset marking symbols.

[0061] When marking words, the symbol "|" is usually used to separate the words in the sentence and extract the words that need to be focused on. On this basis, if a word belongs to an entity word, sentiment word or bigram in the vocabulary, it is marked with a special mark symbol.

[0062] In one embodiment, the marking rules include but are not limited to:

[0063] 1. Enclose the entity word in curly brackets and indicate the type and weight of the entity word;

[0064] Example: {inventory:entity0,-1.0};

[0065] 2. Use angle brackets to enclose the sentiment words and indicate the type and weight of the sentiment words;

[0066] Example: <control:emotion1,0.0>;

[0067] 3. Enclose the bigram in parentheses and use "bigram 1" and "bigram 2" to represent the first and second words in the bigram, respectively. Note that the two words must be in the same sentence, that is, the text part separated by a comma or semicolon, to match;

[0068] Example: There is a bigram phrase (affected: bigram 1)|(weakened: bigram 2) in the vocabulary. For the participle "epidemic|affect|gradually|weakened|,|performance|stopped falling and rebounded", the form of "epidemic|[(affected: bigram 1)|gradually|(weakened: bigram 2)]|,|performance|stopped falling and rebounded" can be matched. However, for the participle "affected by|epidemic|,|overseas|expansion|momentum|weakened", since bigram 1 and bigram 2 are divided into two different short sentences, the matching form of "affected by|epidemic|(affected: bigram 1)|,|overseas|expansion|momentum|(weakened: bigram 2)" is not allowed;

[0069] 4. In addition, if there is an entity word or sentiment word between bigram 1 and bigram 2, this bigram cannot be matched;

[0070] Example: There is a bigram phrase [(impact: bigram 1) | (weaken: bigram 2)] in the vocabulary. For the segmentation "epidemic|impact|lead|inventory risk|rise|and|growth momentum|weaken", the bigram phrase match is invalid because bigram 1 and bigram 2 are separated by the entity word "growth momentum";

[0071] 5. If a word corresponds to two attributes in the vocabulary at the same time, it is a bisexual word. In this article, it is necessary to eliminate ambiguity. In this case, the symbol " / " is used to connect the two attributes;

[0072] Examples: Risk: physical / emotional;

[0073] 6. The tagging priority of bigrams is higher than that of entity words and sentiment words. That is, when a word is both bigram 1 or bigram 2 in a bigram and an entity word or sentiment word, the bigram is valid and the word is no longer matched as an entity word or sentiment word.

[0074] Example: entity + (sentiment / bigram 1) + bigram 2. Since the second word is both a sentiment word and bigram 1, the attributes of bigram 1 are retained and the attributes of the sentiment word are ignored. That is, the correct matching result of the bigram after eliminating ambiguity is: entity + bigram 1 + bigram 2.

[0075] In step 206 , the computing device 110 may match and combine the annotated texts based on predefined matching rules.

[0076] Matching and combining the annotated text based on predefined matching rules includes: determining a fine-grained scale and a coarse-grained scale for the text based on a predefined granularity threshold; combining annotated entity words or adjacent entity words that meet predefined rules into entity combinations at the fine-grained scale; combining annotated sentiment words or adjacent sentiment words that meet predefined rules into sentiment combinations at the fine-grained scale; and matching and combining entity combinations and sentiment combinations at the coarse-grained scale.

[0077] For easier identification, when words are matched and combined into a larger phrase, square brackets are used to enclose the combination from the first word to the last word;

[0078] Example: [<Control: Sentiment 1, 0.0>|{Inventory: Entity 0, -1.0}]|[(To reduce: Bigram 1)|(Risk: Bigram 2)]| and|[<Continue to increase: Sentiment 0, 3.0>|{R&D investment: Entity 0, 1.0}].

[0079] Matching rules at a fine-grained scale:

[0080] 1. Entity words are matched in the following priority order to form several entity combinations;

[0081] a) Entity 0 + Entity 1 / 2

[0082] The above two words must be in the same sentence and in a fixed order. If the order of the two words is reversed, the match fails. Example: For the word segmentation "net profit|growth rate|slowdown", the match can be "{net profit: entity 0, 2.0}|{growth rate: entity 1, 1.5}|<slowdown: sentiment 0, -1>";

[0083] b) Single entity 0 / 1 / 2

[0084] Example: For the word "net profit|improvement", the matching result is "{net profit: entity 0, 2.0}|<improvement: sentiment 0, 1.0>";

[0085] 2. The sentiment words are matched in order of priority to form several sentiment combinations;

[0086] a) Emotion 2 + Emotion 0 / 1

[0087] The above two words must be in the same sentence and in a fixed order. If the order of the two words is reversed, the match fails. Example: For the word "net profit|substantially|increase", the match can be "{net profit: entity 0, 2.0}|<substantially: sentiment 2, 2.0>|<increase: sentiment 0, 1.0>";

[0088] b) Single emotion 0 / 1

[0089] Example: For the word "net profit|increase", the matching result is "{net profit: entity 0, 2.0}|<increase: sentiment 0, 1.0>".

[0090] Matching rules on a coarse-grained scale:

[0091] For several entity combinations and sentiment combinations that have been successfully matched at a fine-grained scale, a higher-grained matching combination is performed in the following order of priority:

[0092] 1. Entity combination + sentiment combination, and the entity combination and sentiment combination are in the same short sentence:

[0093] a) [Entity 0 + Entity 1 / 2] + [Emotion 2 + Emotion 0 / 1]

[0094] Example: For the word "revenue|decline|gradually|reduce", the matching results are "[{revenue: entity 0, 1.5}|{decline: entity 1, -1.0}|<gradually: sentiment 2, 0.8>|<reduce: sentiment 0, -1.0>]";

[0095] b) [Entity 0 + Entity 1 / 2] + [Emotion 0 / 1]

[0096] Example: For the word "profit|growth rate|slowdown", the matching results are "[{profit: entity 0, 2.0}|{growth rate: entity 1, 1.5}|<slowdown: sentiment 0, -1.0>]";

[0097] c) [Entity 0 / 1 / 2] + [Emotion 2 + Emotion 0 / 1]

[0098] Example: For the word "pressure|continue|increase", the matching result is "[{pressure: entity 2, -1.0}|<continue: emotion 2, 2.0>|<increase: emotion 0, 1.0>]";

[0099] d) [Entity 0 / 1 / 2] + [Emotion 0 / 1]

[0100] Example: For the word "performance|improvement", the matching result is "[{performance: entity 0, 2.0}|<improvement: sentiment 0, 1.0>]".

[0101] 2. Sentiment combination + entity combination, and the entity combination and sentiment combination are in the same short sentence:

[0102] a) [Emotion 2 + Emotion 0 / 1] + [Entity 0 + Entity 1 / 2]

[0103] Example: For the word "gradually|expand|cost|decline", the matching results are "[<gradually:sentiment 2, 0.8>|<expand:sentiment 0, 3.0>|{cost:entity 0, -2.0}|{decline:entity 1, -1.0}]";

[0104] b) [Emotion 0 / 1] + [Entity 0 + Entity 1 / 2]

[0105] Example: For the word "reduce|profit|growth rate", the matching results are "[<reduce:sentiment 0, -1.0>|{profit:entity 0, 2.0}|{growth rate:entity 1, 1.5}]";

[0106] c) [Emotion 2 + Emotion 0 / 1] + [Entity 0 / 1 / 2]

[0107] Example: For the word "gradually|reduce|cost", the matching results are "[<gradually: sentiment 2, 0.8>|<reduce: sentiment 0, -1.0>|{cost: entity 0, -2.0}]";

[0108] d) [Emotion 0 / 1] + [Entity 0 / 1 / 2]

[0109] Example: For the word "increase|profit", the matching result is "[<increase:sentiment0,1.0>|{profit:entity0,2.0}]".

[0110] 3. Entity combination + sentiment combination, and the entity combination and sentiment combination are in different short sentences:

[0111] a) [Entity 0 + Entity 1 / 2] + [Emotion 2 + Emotion 0 / 1]

[0112] Example: For the word "net profit|increase|is|50%|,|continues|to rise", the matching results are "[{net profit: entity 0, 3.0}|{increase: entity 0, 3.0}|is|50%|,|<continues: sentiment 2, 2.0>| <rise: sentiment 0, 1.0>]";

[0113] b) [Entity 0 + Entity 1 / 2] + [Emotion 0 / 1]

[0114] Example: For the word "Revenue|Increase|is|50%|,|significantly improved", the matching result is "[{Revenue: entity 0, 3.0}|{Increase: entity 0, 3.0}|is|50%|,|<significantly improved: sentiment 1, 1.0>]";

[0115] c) [Entity 0 / 1 / 2] + [Emotion 2 + Emotion 0 / 1]

[0116] Example: For the word "net profit|is|100|billion yuan|,|significantly|increased", the matching result is "[{net profit: entity 0, 3.0}|is|100|billion yuan|,|<significantly: sentiment 2, 2.0> <increased: sentiment 0, 1.0>]";

[0117] d) [Entity 0 / 1 / 2] + [Emotion 0 / 1]

[0118] Example: For the word "net profit|is|100|billion yuan|,|increased|20%", the matching result is "[{net profit: entity 0, 3.0}|is|100|billion yuan|,|<increased: sentiment 0, 1.0>]|20%".

[0119] 4. When there is no entity combination that can be matched, the sentiment combination itself can be retained as a word combination:

[0120] a) [Emotion 2] + [Emotion 0 / 1]

[0121] Example: For the word "slightly|decline", the matching result is "[<slightly: sentiment 2, 0.5>|<decline: sentiment 0, -1>]";

[0122] b) [Emotion 0 / 1]

[0123] Example: For the word "bad recession", you can directly match "[<bad recession: sentiment 1, -3>]".

[0124] When there is no matching sentiment combination, a correction operation is performed on the entity combination.

[0125] Example: Detect an entity combination (entity 0 + entity 1) that is not matched to a sentiment combination, and determine whether entity 1 in the entity combination is also a sentiment word 0 / 1. If so, the entity combination can be modified to an entity + sentiment combination (entity 0 + sentiment 0 / 1).

[0126] It should also be noted that when the entity combination and sentiment combination are in different short sentences, the matching of sentiment combination + entity combination is invalid. In addition, the entity combination and sentiment combination cannot be matched across any word in the bigram.

[0127] Based on the above matching rules, the present invention also shows the following examples with more complex combinations:

[0128] Example 1: [emotion combination] + [entity combination + emotion combination]

[0129] For the participle "steady|development|,|at home and abroad|space|broad", the matching results are "[<steady: emotion 2, 1.2>|<development: emotion 1, 1.0>] |,|at home and abroad|[{space: entity 1, 1.0}|<broad: emotion 1, 20>]";

[0130] Example 2: [emotion combination] + [entity combination + emotion combination] + entity (invalid)

[0131] For the word "there is an expectation of |overcapacity|, |inventory|increase|", "there is an expectation (invalid entity) of |[<overcapacity: entity 1, -3.0>]|, |[{inventory: entity 0, -1.0}|<increase: sentiment 0, 1.0>]|" can be matched;

[0132] Example 3: [emotion combination] + [emotion combination + entity combination]

[0133] For the participle "Benefiting from the depreciation of the RMB since Q2, 2022, it is expected to improve the profitability of the company", the matching results are "[<Benefit: Sentiment 1, 1.0>] |from the depreciation of the RMB since Q2, 2022, |[<Expected to: Sentiment 2, 0.5>|<Improvement: Sentiment 1, 1.0>|company|{profitability: entity 0, 2.0}]".

[0134] In addition, for a bisexual word that is both a sentiment word and an entity word, when selecting at the coarse-grained level based on the priority during matching, we should also try to match this word to eliminate the ambiguity of bisexual words. Specifically:

[0135] a) The word of both sexes comes first

[0136] Example: Expansion has two attributes: entity 0 and emotion 1

[0137] For the word "expand production|not up to expectations", that is, (sentiment / entity) + sentiment, we can match "[{expand production: entity 0, 1.0}|<not up to expectations: sentiment 1, -3.0>]", which is entity + sentiment after eliminating ambiguity;

[0138] For the word "intended|expand|60,000 tons|capacity", that is, (sentiment / entity)+entity, we can match "intended|[<expand: sentiment 1, 1.0>|60,000 tons|{capacity: entity 0, 1.0}]", which is sentiment+entity after eliminating ambiguity;

[0139] b) The word of both sexes comes last

[0140] Example: Shortage has two attributes: Entity 2 and Sentiment 1

[0141] For the word "overlay|raw materials|shortage", that is, entity+(emotion / entity), we can match "overlay|[{raw materials: entity 0, -1.5}|<shortage: emotion 1, -1.0>]", which is entity+emotion after eliminating ambiguity;

[0142] For the word segmentation "gradually|relieve|chip|shortage|problem", that is, sentiment+(sentiment / entity), the matching result is "[<gradually: sentiment 2, 0.8>|<relieve: sentiment 1, 0.0>|{shortage: entity 2, -1.0}|problem]", which is sentiment+entity after eliminating ambiguity.

[0143] By following the above rules, ambiguity can be avoided and the annotation results of words in the text can be ensured to be true and in line with expectations.

[0144] In step 208 , the computing device 110 may calculate a sentiment score corresponding to the sentence composed of the single words based on the scores of the single words.

[0145] Calculating the sentiment score corresponding to a sentence composed of single words based on the scores of single words includes: multiplying the weights or absolute values ​​of the weights of all words in the single word combination as the sentiment score of the word combination; adding the sentiment scores of all word combinations and bigrams in a single sentence as the sentiment score of the sentence; and normalizing sentence weights.

[0146] Words without brackets indicate invalid matches / combinations and are not included in the above calculation process. In order to calculate the sentiment score more accurately, the following rules based on word attributes need to be followed:

[0147] 1. If the sentiment words in the word combination contain non-directional sentiment words, such as (sentiment 1) or (sentiment 2 + sentiment 1), the entity words are multiplied by the absolute value of the weight without taking the positive or negative sign; the positive or negative direction of the entire word combination is determined by the non-directional sentiment words;

[0148] 2. Entity 2, i.e. non-directional abstract entity words, follows special calculation rules in the following two cases:

[0149] a) [emotion combination + entity 2]: The symbol of this complete combination is directly taken from the symbol of entity 2, and the other calculation logic remains unchanged;

[0150] b) [Entity 0 + Entity 2]: The sign of this entity combination is directly taken from the sign of entity 2, and the other calculation logic remains unchanged.

[0151] The final normalization method is to divide by the number of words. Compared with the methods of dividing by the number of characters or the number of combinations, dividing by the number of words is the most effective. This is because usually, a long sentence and a short sentence both contain the same entity / emotion combination, and the rest are objective descriptions. In terms of semantics and emotional feelings, the emotion of the short sentence is stronger.

[0152] In step 210 , the computing device 110 repeatedly performs the above operations on the collected financial report samples to obtain the report text and the sentiment score corresponding to the report text as training samples for training the financial report analysis model.

[0153] In step 212, based on the pre-trained language representation BERT model, a financial report analysis model is trained on the acquired report text and the sentiment score corresponding to the report text to obtain a desired financial report analysis model.

[0154] Before starting model training, since the core of the algorithm of the present invention is to score the sentiment of sentences and texts have different granularities, the actual application scenario and text features are used.

[0155] The specific situations include the following:

[0156] 1. Directly apply the sentiment score of the sentence to the financial report analysis model training to obtain the financial report analysis model for predicting the sentiment score of the sentence;

[0157] 2. The sentiment scores of the paragraphs are synthesized based on the sentiment scores of the sentences, and the sentiment scores of the paragraphs are applied to the training of the financial report analysis model to obtain the financial report analysis model for predicting the sentiment scores of the paragraphs;

[0158] 3. The sentiment score of the passage is synthesized based on the sentiment scores of the sentences, and the sentiment score of the passage is applied to the training of the financial report analysis model to obtain the financial report analysis model for predicting the sentiment score of the passage.

[0159] In the process of synthesizing the sentiment scores of the paragraphs or chapters based on the sentiment scores of the sentences, the sentiment scores of the sentences are usually synthesized by weighted average. The method of setting the sentence weights includes but is not limited to:

[0160] a) From the perspective of sentiment score

[0161] In the field of financial reporting, the number of negative sentiment statements is relatively small, but they are also more valuable and have a greater role in investment reference. Therefore, it is possible to consider increasing the weight of the scores of negative sentiment statements. This method has achieved good results in quantitative backtesting;

[0162] b) From the perspective of text structure

[0163] In texts in any field, especially longer texts, each part may contain information of different values. Therefore, different weights can be assigned to sentences or paragraphs in different parts according to their importance and reference value, and then weighted average is performed to generate the total sentiment score of the text. This can usually improve the scoring effect of text sentiment and better predict the overall mood of the passage.

[0164] Methods for determining the weight of each part include: judging according to business logic, adjusting the weight according to the experimental results of quantitative backtesting, etc. Set the highest weight to the title of an article, the second highest weight to the summary, and the third is the weight of the full text. Content that is generally considered to be completely irrelevant to emotions, such as the statement page at the end of the report, can be set to 0. Different weights can also be given to core ideas and events in the report summary text. These weighting methods all work well in quantitative backtesting.

[0165] By using the BERT pre-trained model to continue training on the text sentiment label dataset of financial report sentences, paragraphs, and chapters, the text features are learned, and the sentiment score labels are regressed to learn how to predict the emotions expressed by financial report analysts.

[0166] In one embodiment, the process of training a financial report analysis model based on a pre-trained language representation BERT model includes:

[0167] 1. Divide the labeled text emotion files into training set and verification set: Divide the labeled text emotion files into training set and verification set in a ratio of 7:3;

[0168] 2. Input the training set data into the model, calculate the loss function, and update the model;

[0169] 3. Test the model effect and save the model: Based on the validation set, you can use multiple indicators to test the model effect and save the model to the model library;

[0170] 4. Adjust model parameters: adjust the learning rate of the model in different ranges (0.1, 0.001, 0.0001, 0.00001), etc., select different pre-trained financial report analysis models for training, and compare the effects of using different pre-trained models;

[0171] 5. Repeat the above process in a loop and retain the model with the best performance in each indicator in the test.

[0172] The present invention uses financial report text and corresponding scores as training data, utilizes the pre-trained deep learning model BERT, and splices a fully connected layer and an L2 MSE loss layer after the standard BERT model, and fine-tunes the BERT model on the training data, so that the BERT model can automatically perform sentiment scoring on sentences, paragraphs or entire articles in the financial report text.

[0173] Figure 3 1 is a flowchart of a financial report analysis method 300 according to an embodiment of the present disclosure. The method 300 may be performed as follows: Figure 1 The computing device 110 shown may also be executed in Figure 4 The method 300 is executed at the electronic device 400 shown. It should be understood that the method 300 may further include additional blocks not shown and / or may omit the blocks shown, and the scope of the present disclosure is not limited in this respect.

[0174] At step 302 , computing device 110 receives an input financial report.

[0175] At step 304 , computing device 110 calls the trained financial report analysis model.

[0176] At step 306 , computing device 110 predicts a sentiment score for the financial report based on the financial report analysis model.

[0177] Since the financial report analysis model is related to the text granularity, according to the actual application scenario and text characteristics, the user can process the obtained predicted sentiment scores based on different text granularities. The processing logic is related to and similar to the above-mentioned logic of processing the sentiment scores of different text granularities used for training the financial report analysis model.

[0178] The specific situations include the following:

[0179] 1. Obtaining a financial report analysis model for predicting the sentiment score of a sentence, and in response to the financial report analysis model predicting the sentiment score of the sentence, synthesizing the predicted sentiment score of the sentence to obtain a predicted sentiment score of the paragraph or a predicted sentiment score of the chapter;

[0180] 2. Obtaining a financial report analysis model for predicting the sentiment score of a paragraph, and in response to the financial report analysis model predicting the sentiment score of the paragraph, synthesizing the predicted sentiment score of the chapter based on the predicted sentiment score of the paragraph;

[0181] 3. Obtain a financial report analysis model for predicting the sentiment score of an article, and apply the financial report analysis model to directly predict the sentiment score of the article.

[0182] In addition, whether it is the sentiment score calculated based on the score of a single word or the sentiment score predicted based on the financial report analysis model, it can be standardized, but the specific method of standardization needs to be combined with the needs of the specific application scenario, such as narrowing the range to the interval of [-1, 1].

[0183] In scenarios where score standardization is required, you can choose to standardize when automatically scoring text sentiment, and then use the standardized scores to train the financial report analysis model. You can also choose to directly use the original text sentiment scores to train the financial report analysis model, and finally just need to standardize the predicted sentiment scores obtained using the model.

[0184] By applying the financial report analysis method disclosed in the present invention, computer automatic labeling and scoring can be realized, which saves a lot of manual labeling workload compared with the traditional manual labeling method. Moreover, the labeling algorithm used in the present invention is objective and implemented by computer, with unified standards, eliminating the subjective deviation of sentiment labeling, and the sentiment scores given to the text are continuous. Customers can perform customized sentiment classification based on the obtained sentiment scores, without having to face the problem of inconsistency between the number / standard of product classification and customer needs, thereby making the applicability of the present invention solution more extensive.

[0185] The present invention also proposes a score synthesis algorithm that can be applied to the needs of text sentiment scoring at different granularities, and can perform three-dimensional sentiment scoring on texts of various granularities such as sentences, paragraphs and chapters in financial reports, which can meet the multi-faceted and multi-scenario usage needs of various types of users, provide investors with processed sentiment data, and assist investors in making correct investment decisions based on the emotions expressed by financial analysts in the text.

[0186] Figure 4 4 shows a schematic block diagram of an example electronic device 400 that can be used to implement an embodiment of the present disclosure. Figure 1 The computing device 110 shown may be implemented by an electronic device 400. Figure 4 As shown, the electronic device 400 includes a central processing unit (CPU) 401, which can perform various appropriate actions and processes according to computer program instructions stored in a read-only memory (ROM) 402 or computer program instructions loaded from a storage unit 408 to a random access memory (RAM) 403. In the random access memory 403, various programs and data required for the operation of the electronic device 400 can also be stored. The central processing unit 401, the read-only memory 402, and the random access memory 403 are connected to each other via a bus 404. An input / output (I / O) interface 405 is also connected to the bus 404.

[0187] Multiple components in the electronic device 400 are connected to the input / output interface 405, including: an input unit 406, such as a keyboard, a mouse, a microphone, etc.; an output unit 407, such as various types of displays, speakers, etc.; a storage unit 408, such as a disk, an optical disk, etc.; and a communication unit 409, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 409 allows the device 400 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0188] The various processes and processing described above, such as methods 200 and 300, may be performed by the central processing unit 401. For example, in some embodiments, methods 200 and 300 may be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as a storage unit 408. In some embodiments, part or all of the computer program may be loaded and / or installed on the device 400 via the read-only memory 402 and / or the communication unit 409. When the computer program is loaded into the random access memory 403 and executed by the central processing unit 401, one or more actions of the methods 200 and 300 described above may be performed.

[0189] The present disclosure relates to methods, apparatuses, systems, electronic devices, computer-readable storage media and / or computer program products. The computer program products may include computer-readable program instructions for executing various aspects of the present disclosure.

[0190] A computer-readable storage medium may be a tangible device that can hold and store instructions used by an instruction execution device. A computer-readable storage medium may be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the above. More specific examples (a non-exhaustive list) of computer-readable storage media include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disk read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, such as a punch card or a raised structure in a groove on which instructions are stored, and any suitable combination of the above. The computer-readable storage medium used herein is not to be interpreted as a transient signal itself, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagating through a waveguide or other transmission medium (e.g., a light pulse through a fiber optic cable), or an electrical signal transmitted through a wire.

[0191] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to each computing / processing device, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network may include copper transmission cables, optical fiber transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge computing devices. The network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in a computer-readable storage medium in each computing / processing device.

[0192] The computer program instructions for performing the operations of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages, such as Smalltalk, C++, etc., and conventional procedural programming languages, such as "C" language or similar programming languages. The computer-readable program instructions may be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer, partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., using an Internet service provider to connect through the Internet). In some embodiments, by using the state information of the computer-readable program instructions to personalize an electronic circuit, such as a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA), the electronic circuit may execute the computer-readable program instructions, thereby implementing various aspects of the present disclosure.

[0193] Various aspects of the present disclosure are described herein with reference to flowcharts and / or block diagrams of methods, devices (systems) and computer program products according to embodiments of the present disclosure. It should be understood that each box in the flowchart and / or block diagram and the combination of boxes in the flowchart and / or block diagram can be implemented by computer-readable program instructions.

[0194] These computer-readable program instructions can be provided to a processing unit of a general-purpose computer, a special-purpose computer, or other programmable data processing device, thereby producing a machine, so that when these instructions are executed by the processing unit of the computer or other programmable data processing device, a device that implements the functions / actions specified in one or more boxes in the flowchart and / or block diagram is generated. These computer-readable program instructions can also be stored in a computer-readable storage medium, and these instructions cause the computer, programmable data processing device, and / or other equipment to work in a specific manner, so that the computer-readable medium storing the instructions includes a manufactured product, which includes instructions for implementing various aspects of the functions / actions specified in one or more boxes in the flowchart and / or block diagram.

[0195] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device so that a series of operating steps are performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to implement the functions / actions specified in one or more boxes in the flowchart and / or block diagram.

[0196] The flow chart and block diagram in the accompanying drawings show the possible architecture, function and operation of the system, method and computer program product according to multiple embodiments of the present disclosure. In this regard, each square box in the flow chart or block diagram can represent a part of a module, program segment or instruction, and a part of the module, program segment or instruction includes one or more executable instructions for realizing the specified logical function. In some alternative implementations, the function marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two continuous square boxes can actually be executed substantially in parallel, and they can sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs the specified function or action, or can be implemented with a combination of special hardware and computer instructions.

[0197] The embodiments of the present disclosure have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles of the embodiments, practical applications, or technical improvements in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. A method for training a financial report analysis model, characterized in that: include: Constructing a vocabulary containing sentiment words, entity words and bigrams and determining the type and weight attribute of each word in the vocabulary, wherein the sentiment words include directional sentiment words, non-directional sentiment words and modified sentiment words, and the entity words include regular entity words, abstract entity words and non-directional abstract entity words; Receive input text and perform word segmentation processing, and annotate the type and weight attributes of the directional sentiment words, non-directional sentiment words, modified sentiment words, regular entity words, abstract entity words, non-directional abstract entity words and bigrams after word segmentation processing based on the constructed vocabulary and predefined annotation rules; Matching and combining the annotated texts based on predefined matching rules, including determining a fine-grained scale and a coarse-grained scale for the text based on a predefined granularity threshold, combining annotated entity words or adjacent entity words that meet the predefined rules into entity combinations on the fine-grained scale, combining annotated sentiment words or adjacent sentiment words that meet the predefined rules into sentiment combinations on the fine-grained scale, and matching and combining entity combinations and sentiment combinations on the coarse-grained scale; Calculate the sentiment score of the sentence composed of single words based on the scores of single words; Repeat the above operation for the collected financial report samples to obtain the report text and the sentiment score corresponding to the report text as training samples for the financial report analysis model training; as well as Based on the pre-trained language representation BERT model, the financial report analysis model is trained on the acquired report text and the sentiment score corresponding to the report text to obtain the expected financial report analysis model.

2. The method according to claim 1, characterized in that The word segmentation process for the input text includes: Split the text into one or more sentences based on the punctuation marks in the text; Using a word segmentation algorithm, split the split sentence into one or more words; and The words that match the sentiment words, entity words and bigrams in the vocabulary are respectively marked with corresponding types and weight attributes using preset marking symbols.

3. The method according to claim 2, characterized in that The matching combinations of entity combinations and sentiment combinations include: Prioritize matching entity combinations and sentiment combinations in the same sentence, where the matching form of entity combinations and sentiment combinations takes precedence over the matching form of sentiment combinations and entity combinations; Secondly, the entity combinations and sentiment combinations in different sentences are matched, where only the matching form of the entity combination and the sentiment combination is retained, and the matching form of the sentiment combination and the entity combination is not retained; In response to no matching of the entity combination, retaining the sentiment combination as a word combination alone; and In response to not matching the sentiment combination, a corrective action is performed.

4. The method according to claim 3, characterized in that The sentiment scores of sentences composed of single words are calculated based on the scores of single words. For a word combination of words, multiply the weights or absolute values ​​of the weights of all words in the word combination as the sentiment score of the word combination; and For a sentence composed of word combinations, the sentiment scores of all word combinations and bigrams in the sentence are added together as the sentiment score of the sentence.

5. The method according to claim 4, characterized in that The method further comprises: The sentiment scores of the obtained sentences are processed based on different text granularities, and the processed sentiment scores are applied to the financial report analysis model training, so as to obtain the financial report analysis model for predicting the sentiment scores of texts with corresponding granularities.

6. The method according to claim 5, characterized in that The financial report analysis model for obtaining sentiment scores for corresponding granular text includes: Applying the sentiment score of the sentence to the training of the financial report analysis model to obtain the financial report analysis model for predicting the sentiment score of the sentence; or Synthesizing the sentiment scores of the sentences to obtain the sentiment scores of the paragraphs, applying the sentiment scores of the paragraphs to the training of the financial report analysis model, and obtaining the financial report analysis model for predicting the sentiment scores of the paragraphs; or The sentiment score of the passage is obtained by synthesizing the sentiment scores based on the sentences, and the sentiment score of the passage is applied to the training of the financial report analysis model to obtain the financial report analysis model for predicting the sentiment score of the passage.

7. The method according to any one of claims 1 to 6, characterized in that The two words in the bigram are in the same sentence, and there are no other entity words or sentiment words between the two words, and the entity combination and the sentiment combination do not cross any word in the bigram for matching and combination.

8. A financial report analysis method, characterized in that: include: The financial report is input into the financial report analysis model obtained based on any one of the methods of claims 1-7 to predict the sentiment score of the input financial report.

9. The method according to claim 8, characterized in that The method further comprises: The predicted sentiment scores of the obtained financial reports are processed based on different text granularities.

10. The method according to claim 9, characterized in that The predicted sentiment scores of the obtained financial reports based on different text granularities are processed as follows: In response to the sentiment score of the sentence predicted by the financial report analysis model, a predicted sentiment score of the paragraph or a predicted sentiment score of the chapter is synthesized based on the predicted sentiment score of the sentence; or In response to the financial report analysis model predicting the sentiment score of the paragraph, the predicted sentiment score of the chapter is synthesized based on the predicted sentiment scores of the paragraph; or Apply financial report analysis models to directly predict the sentiment scores of passages.

11. A computing device, characterized in that: include: at least one processor; as well as a memory communicatively coupled to the at least one processor; The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method according to any one of claims 1 to 10.

12. A non-transitory computer-readable storage medium storing computer instructions, characterized in that: The computer instructions are used to cause the computer to execute the method according to any one of claims 1 to 10.

Images