Language learning ability evaluation method and system based on data multi-modal fusion

Abstract

The invention relates to a language learning ability evaluation method and system based on data multi-modal fusion, and relates to the technical field of data processing. The method mainly comprises the steps of obtaining to-be-evaluated language text data; the basic readability of the language text data to be evaluated is verified; if the basic readability verification of the language text data passes, preprocessing the language text data to obtain a word segmentation result, an entity recognition result and a paragraph division result; performing information structure analysis according to the word segmentation result, the entity recognition result and the paragraph division result to obtain an information structure analysis result; determining text semantic depth and text credibility according to an information structure analysis result, a word segmentation result, an entity recognition result and a paragraph division result; and calculating an evaluation score of the language text data to be evaluated according to the text semantic depth, the text credibility and the information structure analysis result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and in particular to a language learning ability evaluation method and system based on data multi-modal fusion. BACKGROUND

[0002] With the deepening of the globalization process and the rapid development of the Internet, text data presents an explosive growth, but these data sources are different, which affects the utilization efficiency of information. Therefore, how to evaluate the quality of the real-time collected data and recommend high-quality text information to users is an important basic problem in the field of text intelligence under the background of big data.

[0003] Usually, a classification model is trained by using artificially labeled samples, and then the classification model is used to predict the quality of the text, so as to filter out high-quality text from the text library and recommend it to the user. SUMMARY

[0004] The present application aims to provide a language learning ability evaluation method and system based on data multi-modal fusion to solve the problems existing in the prior art. The technical problems to be solved by the present application are solved by the following technical solutions.

[0005] The embodiment of the present application provides a language learning ability evaluation method based on data multi-modal fusion, which comprises: Obtaining language text data to be evaluated; and verifying the basic readability of the language text data to be evaluated; If the basic readability of the language text data is verified, the language text data is preprocessed to obtain a word segmentation result, an entity recognition result and a paragraph division result; According to the word segmentation result, the entity recognition result and the paragraph division result, an information structure analysis result is obtained by information structure analysis; According to the information structure analysis result, the word segmentation result, the entity recognition result and the paragraph division result, the text semantic depth and the text credibility are determined; The evaluation score of the language text data to be evaluated is calculated by the text semantic depth, the text credibility and the information structure analysis result.

[0006] In an optional embodiment, the information structure analysis result includes information entropy density, and the information structure analysis result is obtained according to the word segmentation result, the entity recognition result and the paragraph division result, comprising: The number of non-repeated entities is determined according to the entity recognition result, and the word segmentation result is matched with a predefined verb library to determine the number of key action verbs; remove the stop words in the word segmentation result and count to obtain the total number of valid words; obtain the information entropy density by dividing the sum of the non-repeated entity number and the key action verb number by the total number of valid words, the information entropy density being used to quantify the concentration of substantive information elements in the information entropy density.

[0007] In an optional embodiment, the information structure analysis result includes a logical coherence coefficient, the information structure analysis result being obtained according to the word segmentation result, the entity recognition result, and the paragraph division result, and including: determining the total number of paragraphs according to the paragraph division result, and matching the word segmentation result with the connecting words in a logical connecting word library to determine the number of valid connecting words, the connecting words including causal connecting words, transitional connecting words, and progressive connecting words; obtaining the logical coherence coefficient by calculating the ratio of the number of valid connecting words to the total number of paragraphs, the logical coherence coefficient being used to measure the logical connection strength between paragraphs in the language text data to be evaluated.

[0008] In an optional embodiment, the information structure analysis result includes a theme concentration degree, the information structure analysis result being obtained according to the word segmentation result, the entity recognition result, and the paragraph division result, and including: determining the core theme word frequency according to the word segmentation result corresponding to the title and the first paragraph, and determining the secondary theme word frequency according to the word segmentation result corresponding to all paragraphs; calculating the theme concentration degree according to the core theme word frequency and the secondary theme word frequency, the theme concentration degree being used to evaluate the degree of concentration of the language text data to be evaluated around the core theme.

[0009] In an optional embodiment, determining the text semantic depth according to the information structure analysis result, the word segmentation result, the entity recognition result, and the paragraph division result includes: obtaining all sentences containing core theme words based on the paragraph division result and the structure analysis result; extracting triples according to the sentences containing core theme words and the corresponding word segmentation result, and counting the number of nodes participating in reasoning chains exceeding a preset number of layers and the total number of nodes in the triples; calculating the text semantic depth by calculating the ratio of the number of nodes participating in reasoning chains exceeding a preset number of layers to the total number of nodes.

[0010] In an optional embodiment, determining the text credibility according to the information structure analysis result, the word segmentation result, the entity recognition result, and the paragraph division result includes: matching the key words in the word segmentation result and the entity recognition result with key words of the evidence verb library to obtain an evidence verb; determining a paragraph where the evidence verb is located through the paragraph division result, and extracting an evidence element corresponding to the evidence verb according to the paragraph where the evidence verb is located, the evidence element including a source entity and a time expression; determining the text credibility through the evidence element.

[0011] In an optional embodiment, after the information structure analysis result is obtained according to the word segmentation result, the entity recognition result and the paragraph division result, the method further includes: calculating a sentiment quality coupling coefficient according to the word segmentation result, the entity recognition result and the paragraph division result; calculating a sentiment quality score according to the sentiment quality coupling coefficient and the information structure analysis result; the calculating of the evaluation score of the language text data to be evaluated according to the text semantic depth, the text credibility and the information structure analysis result includes: the calculating of the evaluation score of the language text data to be evaluated according to the text semantic depth, the text credibility, the sentiment quality score and the information structure analysis result.

[0012] In an optional embodiment, the calculating of the sentiment quality coupling coefficient according to the word segmentation result, the entity recognition result and the paragraph division result includes: determining, for each paragraph corresponding to the paragraph division result, the word segmentation result and the entity recognition result, a number of positive sentiment sentences and a number of negative sentiment sentences of the corresponding paragraph; calculating a number of positive sentiment sentences and a number of negative sentiment sentences of the whole text according to all the word segmentation results and the entity recognition results; calculating a global sentiment polarity intensity through the number of positive sentiment sentences, the number of negative sentiment sentences and the total number of sentences of the whole text; calculating a local sentiment polarity intensity according to the number of positive sentiment sentences, the number of negative sentiment sentences and the total number of sentences of each paragraph; calculating a sentiment fluctuation coefficient through the global sentiment polarity intensity and the local sentiment polarity intensity, and calculating a sentiment quality coupling coefficient according to the sentiment fluctuation coefficient and the global sentiment polarity intensity.

[0013] In an optional embodiment, the calculating of the sentiment quality score according to the sentiment quality coupling coefficient and the information structure analysis result includes: The content basic quality is calculated by mechanically weighting the information entropy density, the logical coherence coefficient and the theme concentration in the information structure analysis result; The emotional quality score is calculated by the content basic quality and the emotional quality coupling coefficient.

[0014] The embodiment of the application provides a language learning ability evaluation system based on data multi-modal fusion, and the system comprises: An acquisition module is configured to acquire language text data to be evaluated, and verify the basic readability of the language text data to be evaluated; A preprocessing module is configured to perform preprocessing on the language text data to obtain a word segmentation result, an entity recognition result and a paragraph division result if the basic readability of the language text data is verified. An analysis module is configured to perform information structure analysis according to the word segmentation result, the entity recognition result and the paragraph division result to obtain an information structure analysis result. A determination module is configured to determine a text semantic depth and a text credibility according to the information structure analysis result, the word segmentation result, the entity recognition result and the paragraph division result. A calculation module is configured to calculate an evaluation score of the language text data to be evaluated by the text semantic depth, the text credibility and the information structure analysis result.

[0015] The embodiment of the application has the following advantages: The embodiment of the application provides a language learning ability evaluation method and system based on data multi-modal fusion. First, language text data to be evaluated is acquired, and the basic readability of the language text data to be evaluated is verified. If the basic readability of the language text data is verified, the language text data is preprocessed to obtain a word segmentation result, an entity recognition result and a paragraph division result. Then, information structure analysis is performed according to the word segmentation result, the entity recognition result and the paragraph division result to obtain an information structure analysis result. A text semantic depth and a text credibility are determined according to the information structure analysis result, the word segmentation result, the entity recognition result and the paragraph division result. Finally, an evaluation score of the language text data to be evaluated is calculated by the text semantic depth, the text credibility and the information structure analysis result. The language text data to be evaluated is evaluated in multiple dimensions, i.e., the text is evaluated from the dimensions of text structure analysis, text semantic depth and text credibility. Therefore, the accuracy of text evaluation can be improved by the application. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 is a flowchart of a language learning ability evaluation method based on data multi-modal fusion provided by the embodiment of the application. Figure 2 is a flowchart of another language learning ability evaluation method based on data multi-modal fusion provided by an embodiment of the present application; Figure 3 is a structural schematic diagram of a language learning ability evaluation system based on data multi-modal fusion provided by an embodiment of the present application. DETAILED DESCRIPTION

[0017] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0018] Please refer to Figure 1 A language learning ability evaluation method based on data multi-modal fusion provided by an embodiment of the present application, which specifically comprises S101-S105: S101, obtaining language text data to be evaluated; and verifying the basic readability of the language text data to be evaluated.

[0019] It should be noted that before verifying the basic readability of the language text data to be evaluated, sensitive word detection needs to be performed on the language text data to be evaluated, such as determining whether a sensitive word appears by comparing a pre-defined rule violation word library, and if no sensitive word appears, the basic readability of the language text data to be evaluated is verified.

[0020] In the present embodiment, readability verification can be determined by calculating a syntactic integrity index and a punctuation symbol abuse rate. The syntactic integrity index = number of complete sentences / total number of sentences (deduct points for fragmented sentences), and the punctuation symbol abuse rate can be determined by the number of consecutive uses of punctuation symbols, such as more than 3 consecutive exclamation marks or question marks, which can be determined as punctuation symbol abuse.

[0021] S102, if the basic readability verification of the language text data passes, the language text data is preprocessed to obtain word segmentation results, entity recognition results and paragraph division results.

[0022] The preprocessing can be word segmentation and part-of-speech tagging, input text → word segmentation tool → output word sequence with part-of-speech tagging. For example, the word segmentation result corresponding to "convolutional neural network (CNN) significantly improves image recognition accuracy" is: [convolution (noun) / neural network (noun) / CNN (entity) / significantly (adverb) / improve (verb) / image recognition (term) / accuracy (noun)].

[0023] In the embodiment, the entity and term recognition can use the NER tool to recognize: personal name, place name, organization name, professional term, and then obtain the entity recognition result (such as ["CNN", "image recognition"]); paragraph and sentence division: dividing paragraphs according to line breaks, dividing sentences according to punctuation, and recording: paragraph number N para, sentence number N sent.

[0024] In S103, information structure analysis is performed according to the word segmentation result, the entity recognition result, and the paragraph division result to obtain an information structure analysis result.

[0025] In an optional embodiment provided in the application, the information structure analysis result includes information entropy density, and the information structure analysis is performed according to the word segmentation result, the entity recognition result, and the paragraph division result to obtain the information structure analysis result, including: determining the number of non-repeated entities according to the entity recognition result; matching the word segmentation result with a predefined verb library to determine the number of key action verbs; removing stop words in the word segmentation result and counting to obtain the total number of valid words; and dividing the sum of the number of non-repeated entities and the number of key action verbs by the total number of valid words to obtain the information entropy density, which is used to quantify the concentration of substantive information elements in the information entropy density.

[0026] The predefined verb library stores a plurality of verbs, such as "promote", "solve", "optimize", "prove", etc. The word segmentation result is matched with the predefined verb library to determine the corresponding defined verbs, and then the number of defined verbs is counted to obtain the number of key action verbs. Finally, the sum of the number of non-repeated entities and the number of key action verbs is divided by the total number of valid words to obtain the information entropy density. Since the numerator in the application carries the entity and verb of core information (high-value content), and the denominator is the total information carrying unit (to avoid dilution of the score by redundancy), the information entropy density in the embodiment can be used to quantify the concentration of substantive information elements in the information entropy density.

[0027] For example, the language text data to be evaluated is "CNN extracts features through convolution layers, greatly improves classification accuracy", the number of non-repeated entities is ["CNN", "convolution layer", "feature"], the number of non-repeated entities is N unique entity=3 obtained by counting non-repeated entities, the key verbs ["extract", "promote"] are obtained by matching the predefined verb library, and the number of key action verbs is N key verb=2. The total number of valid words (after removing stop words) is 9, and the information entropy density IED=(3+2) / 9≈0.56 is calculated.

[0028] In an optional embodiment provided in the present application, the information structure analysis result includes a logical coherence coefficient, and the information structure analysis is performed according to the word segmentation result, the entity recognition result and the paragraph division result to obtain an information structure analysis result, including: determining a total number of paragraphs according to the paragraph division result; and matching the word segmentation result with connectives in a logical connective library to determine a number of valid connectives; the connectives include: causal connectives, transitional connectives and progressive connectives; and the logical coherence coefficient is obtained by calculating a ratio of the number of valid connectives to the total number of paragraphs; the logical coherence coefficient is used to measure the logical connection strength between paragraphs in the language text data to be evaluated.

[0029] The causal connectives can be ["therefore", "so", "because", "therefore"], the transitional connectives can be ["however", "but", "although", "on the contrary"], and the progressive connectives can be ["furthermore", "and", "further", "simultaneously"], which are not limited in the embodiment.

[0030] For example, the language text data (two paragraphs) to be evaluated is "CNN can effectively extract features. Therefore, it performs well in image recognition. However, a large amount of data is required for training." The valid connectives obtained by matching with the connectives in the logical connective library are ["therefore", "however"] -> the number of valid connectives is N_valid_connective=2; the total number of paragraphs is 2, and the logical connection strength between paragraphs LCC=2 / 2=1.0 is calculated.

[0031] In an optional embodiment provided in the present application, the information structure analysis result includes a theme concentration degree, and the information structure analysis is performed according to the word segmentation result, the entity recognition result and the paragraph division result to obtain an information structure analysis result, including: determining a title and a first paragraph according to the paragraph division result, and determining a core theme word frequency in the word segmentation result corresponding to the title and the first paragraph; determining a secondary theme word frequency in the word segmentation result corresponding to all paragraphs; and calculating the theme concentration degree according to the core theme word frequency and the secondary theme word frequency, the theme concentration degree being used to evaluate the degree of concentration of the language text data to be evaluated around a core theme. The secondary theme word frequency is the highest frequency noun after excluding the core word.

[0032] The core theme word frequency is the highest frequency noun in the title / first paragraph. In the embodiment, the title (or the first three sentences of the first paragraph when there is no title) or the first paragraph noun is extracted first, then the word frequency is counted, and then the highest frequency noun is selected as the core theme word. The word frequency table is generated by counting all nouns, then the core theme word is excluded, 1-3 highest frequency nouns in the remaining nouns are selected as the secondary theme word, and the secondary theme word frequency is obtained by counting the number of secondary theme words.

[0033] For example, the title is "Application of CNN in medical image diagnosis", the text is "CNN extracts features through... (8 times of 'CNN'), deep learning technology... ('deep learning' 6 times)", the determined core theme word is "CNN", the statistical core theme word frequency is Freq_core_theme=8; the secondary theme word is "deep learning", the statistical secondary theme word frequency is Freq_secondary_theme=6, and the calculated theme concentration TC=8 / 6≈1.33.

[0034] In S104, the text semantic depth is determined according to the information structure analysis result, the word segmentation result, the entity recognition result, and the paragraph division result.

[0035] In an optional embodiment provided in the application, the determination of the text semantic depth according to the information structure analysis result, the word segmentation result, the entity recognition result, and the paragraph division result comprises: obtaining all sentences containing core theme words based on the paragraph division result and the structure analysis result; extracting triplets according to the sentences containing core theme words and the corresponding word segmentation results; and counting the number of nodes participating in reasoning chains exceeding a preset number of layers and the total number of nodes in the triplets; and calculating the ratio of the number of nodes participating in reasoning chains exceeding the preset number of layers to the total number of nodes to obtain the text semantic depth. The preset number of layers of reasoning chains can be 2 or 3, which can be set according to actual conditions.

[0036] The triplet is (subject, relation, object). For example, the obtained sentence containing core theme words is "CNN extracts features through convolution layer", and the triplet extracted from the corresponding word segmentation result of the sentence is ("CNN", "through", "convolution layer"), ("convolution layer", "extract", "feature"). Then all nodes in the triplet are traversed to identify the depth path of each node in the triplet. For example, the continuous reasoning chain starting from the core theme word is ≥3 steps, and the qualified path is: CNN→convolution layer→feature extraction→classification precision improvement (4 layers). If the number of nodes participating in the depth path is N_deep_node=7, and the total number of knowledge nodes is N_total_node=12, then the text semantic depth DPR=7 / 12≈0.583.

[0037] In an optional embodiment provided in the present application, the determining of the text credibility according to the information structure analysis result, the word segmentation result, the entity recognition result, and the paragraph division result comprises: matching keywords in the word segmentation result and the entity recognition result with keywords in an evidence verb library to obtain an evidence verb; determining a paragraph where the evidence verb is located through the paragraph division result, and extracting an evidence element corresponding to the evidence verb according to the paragraph where the evidence verb is located, the evidence element comprising a source entity and a time expression; and determining the text credibility through the evidence element. The keywords in the evidence verb library can be, but are not limited to, "prove", "indicate", "verify", "experimental results show", and "data confirm". The source entity can be determined through an organization and an author, and the time expression can be a year or a date.

[0038] Specifically, the text credibility can be calculated by the following formula: ECW = (Σ(W_type*W_time)) / N_evidence. In the formula, ECW is the text credibility, W_type is a weight of an evidence type determined according to a source entity, such as 1.0 for an academic paper, 0.8 for an organization report, 0.6 for a business case, and 0.2 for no source; W_time is a weight of timeliness determined according to a time expression, such as 1.0 for a current year minus a publication year being less than or equal to 1 year, 0.7 for less than or equal to 3 years, and 0.3 for more than 3 years; and N_evidence is a total number of evidence entries, the total number of evidence entries being at least 1 (if there is no evidence, ECW = 0).

[0039] In the source entity, ".edu" or an academic journal name can determine an academic paper as the evidence type; "research institute" or "association" can determine an organization report as the evidence type; and "company" or "enterprise" can determine a business case as the evidence type.

[0040] In S105, the evaluation score of the language text data to be evaluated is calculated according to the text semantic depth, the text credibility, and the information structure analysis result.

[0041] Specifically, the evaluation score of the language text data to be evaluated can be calculated by weighting. For example, the evaluation score of the language text data to be evaluated can be calculated by the following formula: Quality Score = (IED×W1 + LCC × W2 + TC × W3) × DPR Factor ×ECW_Factor Wherein, the Quality Score is the evaluation score of the language text data to be evaluated, the IED is the information entropy density, W1 is the weight value of the information entropy density, the LCC is the logical connection strength between paragraphs, W2 is the weight value of the logical connection strength between paragraphs, the TC is the theme concentration, W3 is the weight value of the theme concentration, the Depth_Factor is the text semantic depth factor, the DPR Factor is the text credibility factor, DPR Factor = 1 + 0.2 x DPR (upper limit 1.5), and ECW_Factor = min(ECW, 1.0).

[0042] Then the evaluation score of the language text data to be evaluated is output, which can be divided into the following three levels: high quality (≥0.8): indicating that the language text data to be evaluated is high information density + strong logic + deep evidence; qualified (0.5-0.8): indicating that the language text data to be evaluated meets the basic standard but lacks depth; to be optimized (<0.5): indicating that the language text data to be evaluated is information sparse or logical broken.

[0043] The language learning ability evaluation method based on data multi-modal fusion provided by the embodiment first acquires language text data to be evaluated; and checks the basic readability of the language text data to be evaluated; if the basic readability of the language text data passes the check, the language text data is preprocessed to obtain word segmentation results, entity recognition results and paragraph division results, then information structure analysis is performed according to the word segmentation results, the entity recognition results and the paragraph division results to obtain information structure analysis results; the text semantic depth and the text credibility are determined according to the information structure analysis results, the word segmentation results, the entity recognition results and the paragraph division results; finally, the evaluation score of the language text data to be evaluated is calculated through the text semantic depth, the text credibility and the information structure analysis results. The language text data to be evaluated is evaluated in multiple dimensions by the application, that is, the text is evaluated from the dimensions of text structure analysis, text semantic depth and text credibility, so that the accuracy of text evaluation can be improved by the application.

[0044] Please refer to Figure 2 Another language learning ability evaluation method based on data multi-modal fusion is provided for the embodiment of the application, and the method specifically includes S201-S205. S201, language text data to be evaluated is acquired; and the basic readability of the language text data to be evaluated is checked.

[0045] S202, if the basic readability of the language text data passes the check, the language text data is preprocessed to obtain word segmentation results, entity recognition results and paragraph division results.

[0046] S203, performing information structure analysis according to the word segmentation result, the entity recognition result, and the paragraph division result to obtain an information structure analysis result; and calculating a sentiment quality coupling coefficient according to the word segmentation result, the entity recognition result, and the paragraph division result.

[0047] It should be noted that the related description of steps S201-S203 in this embodiment can refer to the related description of steps S201-S203 in the first aspect of the present application. Figure 1 According to the description of the corresponding steps, the present embodiment will not be described here.

[0048] In an optional embodiment provided in the present application, the calculating of the sentiment quality coupling coefficient according to the word segmentation result, the entity recognition result, and the paragraph division result comprises: S2031, determining the number of positive sentiment sentences and the number of negative sentiment sentences of each paragraph corresponding to the paragraph division result, the word segmentation result, and the entity recognition result.

[0049] In the present embodiment, the number of positive sentiment sentences and the number of negative sentiment sentences of each paragraph are determined according to the paragraph division result, the word segmentation result, and the entity recognition result. Among them, the sentence containing positive words (such as "outstanding" / "breakthrough") and without negative modification is determined as a positive sentiment sentence; and the sentence containing negative words (such as "defect" / "failure") and without negative modification is determined as a negative sentiment sentence.

[0050] S2032, calculating the number of positive sentiment sentences and the number of negative sentiment sentences of the whole text according to all the word segmentation results and the entity recognition results.

[0051] S2033, calculating the global sentiment polarity intensity by the number of positive sentiment sentences, the number of negative sentiment sentences, and the total number of sentences of the whole text.

[0052] S2034, calculating the local sentiment polarity intensity according to the number of positive sentiment sentences, the number of negative sentiment sentences, and the total number of sentences of each paragraph.

[0053] S2035, calculating the sentiment fluctuation coefficient by the global sentiment polarity intensity and the local sentiment polarity intensity; and calculating the sentiment quality coupling coefficient according to the sentiment fluctuation coefficient and the global sentiment polarity intensity.

[0054] Specifically, the present embodiment calculates the sentiment quality coupling coefficient by the formula EQC=2 / (1+e^(k×|PS|))×EFC. Wherein, EQC is the sentiment quality coupling coefficient, PS is the global sentiment polarity intensity, EFC is the sentiment fluctuation coefficient, and k is the attenuation coefficient (k=4.0 for academic texts and k=1.5 for review texts).

[0055] S204, determining the text semantic depth and text credibility based on the information structure analysis results, word segmentation results, entity recognition results, and paragraph division results; and calculating the sentiment quality score based on the sentiment quality coupling coefficient and the information structure analysis results.

[0056] Specifically, the emotional quality score is calculated based on the emotional quality coupling coefficient and the information structure analysis results, including: mechanically weighted calculation of the information entropy density, logical coherence coefficient, and topic concentration in the information structure analysis results to obtain the basic content quality; and calculating the emotional quality score through the basic content quality and the emotional quality coupling coefficient.

[0057] Specifically, the emotional quality score can be calculated using the formula EQDA_Score = CQ_base×min(2, 1+EQC). EQDA_Score is the emotional quality score, CQ_base is the basic content quality, and EQC is the emotional quality coupling coefficient.

[0058] S205 , calculating an evaluation score of the language text data to be evaluated based on the text semantic depth, the text credibility, the sentiment quality score, and the information structure analysis result.

[0059] Specifically, this embodiment can perform weighted calculation on the text semantic depth, the text credibility, the sentiment quality score and the information structure analysis result to obtain the text semantic depth, the text credibility, the sentiment quality score and the information structure analysis result.

[0060] This embodiment provides a language learning ability assessment based on multimodal data fusion. By dynamically coupling sentiment stability analysis with content quality, it solves the problem of the separation between sentiment and quality in traditional methods. In tests in scenarios such as e-commerce reviews, academic papers, and news reports, it improves by 12-18% compared to a single model.

[0061] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0062] In one embodiment, a language learning ability assessment system based on multimodal data fusion is provided. Figure 3 As shown, the functional modules of the language learning ability assessment system based on multimodal data fusion are described in detail as follows: An acquisition module 31 is configured to acquire language text data to be evaluated and verify the basic readability of the language text data to be evaluated. The preprocessing module 32 is configured to perform preprocessing on the language text data to obtain a word segmentation result, an entity recognition result, and a paragraph division result if the basic readability check of the language text data passes. The analysis module 33 is configured to perform information structure analysis on the word segmentation result, the entity recognition result, and the paragraph division result to obtain an information structure analysis result. The determination module 34 is configured to determine a text semantic depth and a text credibility according to the information structure analysis result, the word segmentation result, the entity recognition result, and the paragraph division result. The calculation module 35 is configured to calculate an evaluation score of the language text data to be evaluated according to the text semantic depth, the text credibility, and the information structure analysis result.

[0063] In an optional embodiment, the information structure analysis result includes an information entropy density, and the analysis module 33 is specifically configured to: determine a number of non-repeated entities according to the entity recognition result, and determine a number of key action verbs by matching the word segmentation result with a predefined verb library; remove stop words in the word segmentation result and count to obtain a total number of valid words; obtain the information entropy density by dividing a sum of the number of non-repeated entities and the number of key action verbs by the total number of valid words, where the information entropy density is used to quantify a concentration of substantive information elements in the information entropy density.

[0064] In an optional embodiment, the information structure analysis result includes a logical coherence coefficient, and the analysis module 33 is specifically configured to: determine a total number of paragraphs according to the paragraph division result, and determine a number of valid conjunctions by matching the word segmentation result with conjunctions in a logical conjunction library, where the conjunctions include causal conjunctions, transitional conjunctions, and progressive conjunctions; obtain the logical coherence coefficient by calculating a ratio of the number of valid conjunctions to the total number of paragraphs, where the logical coherence coefficient is used to measure a logical connection strength between paragraphs in the language text data to be evaluated.

[0065] In an optional embodiment, the information structure analysis result includes a theme concentration degree, and the analysis module 33 is specifically configured to: determine a title and a first paragraph according to the paragraph division result, determine a core theme word frequency according to the word segmentation result corresponding to the title and the first paragraph, and determine a secondary theme word frequency according to the word segmentation result corresponding to all paragraphs; calculate the theme concentration degree according to the core theme word frequency and the secondary theme word frequency, where the theme concentration degree is used to evaluate a degree of concentration of the language text data to be evaluated around a core theme.

[0066] In an optional embodiment, the determining module 34 is specifically configured to: obtain all sentences containing core theme words based on the paragraph division result and the structure analysis result; extract a triple from the sentences containing core theme words and the corresponding word segmentation result; and count the number of nodes participating in a reasoning chain exceeding a preset number of layers and the total number of nodes in the triple; calculate a ratio of the number of nodes participating in the reasoning chain exceeding the preset number of layers and the total number of nodes to obtain the semantic depth of the text.

[0067] In an optional embodiment, the determining module 34 is specifically configured to: match keywords in the word segmentation result and the entity recognition result with keywords in a key verb library to obtain a key verb; determine a paragraph where the key verb is located through the paragraph division result, and extract a corresponding evidence element of the key verb according to the paragraph where the key verb is located, the evidence element including a source entity and a time expression; determine the text credibility through the evidence element.

[0068] In an optional embodiment, the calculating module 35 is further configured to calculate a sentiment quality coupling coefficient according to the word segmentation result, the entity recognition result, the paragraph division result, and calculate a sentiment quality score according to the sentiment quality coupling coefficient and the information structure analysis result. The calculating module 35 is specifically configured to calculate an evaluation score of the language text data to be evaluated through the text semantic depth, the text credibility, the sentiment quality score, and the information structure analysis result.

[0069] In an optional embodiment, the calculating module 35 is specifically configured to: determine, for each paragraph corresponding to the paragraph division result, the word segmentation result, and the entity recognition result, a number of positive sentiment sentences and a number of negative sentiment sentences of the corresponding paragraph; calculate a number of positive sentiment sentences and a number of negative sentiment sentences of the whole text according to all the word segmentation results and the entity recognition results; calculate a global sentiment polarity intensity through the number of positive sentiment sentences and the number of negative sentiment sentences of the whole text and a total number of sentences; calculate a local sentiment polarity intensity according to the number of positive sentiment sentences and the number of negative sentiment sentences of each paragraph and a total number of sentences of each paragraph; calculate a sentiment fluctuation coefficient through the global sentiment polarity intensity and the local sentiment polarity intensity, and calculate a sentiment quality coupling coefficient according to the sentiment fluctuation coefficient and the global sentiment polarity intensity.

[0070] In an optional embodiment, the computing module 35 is specifically configured to: The content basic quality is calculated by mechanically weighting the information entropy density, the logical coherence coefficient and the theme concentration in the information structure analysis result. The emotional quality score is calculated by coupling the content basic quality and the emotional quality coupling coefficient.

[0071] It should be noted that the above detailed description is exemplary and is intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs.

[0072] For specific limitations of the language learning ability evaluation system based on data multi-modal fusion, refer to the limitations of the language learning ability evaluation method based on data multi-modal fusion in the above, which will not be repeated here. Each module in the above device can be realized by software, hardware and their combination in whole or in part. The above modules can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so as to be called and executed by the processor to perform the operations corresponding to each module.

[0073] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above functional units and modules is exemplified, and in actual application, the above functions can be completed by different functional units and modules according to needs, that is, the internal structure of the system is divided into different functional units or modules to complete all or part of the functions described above.

[0074] The above embodiments are only used to illustrate the technical solutions of the present application, rather than limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A language learning ability assessment method based on multimodal data fusion, characterized in that: The method comprises: Acquiring language text data to be evaluated; and verifying the basic readability of the language text data to be evaluated; If the basic readability check of the language text data passes, the language text data is preprocessed to obtain word segmentation results, entity recognition results, and paragraph division results; Performing information structure analysis based on the word segmentation result, the entity recognition result, and the paragraph division result to obtain an information structure analysis result; Determine the text semantic depth and text credibility based on the information structure analysis result, the word segmentation result, the entity recognition result, and the paragraph division result; An evaluation score of the language text data to be evaluated is calculated based on the text semantic depth, the text credibility, and the information structure analysis result.

2. The method according to claim 1, characterized in that The information structure analysis result includes information entropy density, and the information structure analysis is performed based on the word segmentation result, the entity recognition result, and the paragraph division result to obtain the information structure analysis result, including: Determine the number of non-repeated entities according to the entity recognition result; and match the word segmentation result with a predefined verb library to determine the number of key action verbs; Remove stop words from the word segmentation results and calculate the total number of valid words; The information entropy density is obtained by dividing the sum of the number of non-repeated entities and the number of key action verbs by the total number of valid words. The information entropy density is used to quantify the concentration of substantial information elements in the information entropy density.

3. The method according to claim 1, characterized in that The information structure analysis result includes a logical coherence coefficient, and the information structure analysis is performed based on the word segmentation result, the entity recognition result, and the paragraph division result to obtain the information structure analysis result, including: Determine the total number of paragraphs based on the paragraph segmentation results; and match the word segmentation results with the connectives in the logical connective vocabulary to determine the number of valid connectives; the connectives include: causal connectives, transitional connectives, and progressive connectives; The logical coherence coefficient is obtained by calculating the ratio of the number of the effective connectives to the total number of the paragraphs; the logical coherence coefficient is used to measure the strength of the logical connection between paragraphs in the language text data to be evaluated.

4. The method according to claim 1, wherein The information structure analysis result includes topic concentration, and the information structure analysis result obtained by performing information structure analysis based on the word segmentation result, the entity recognition result, and the paragraph division result includes: Determine the title and the first paragraph according to the paragraph division results, and determine the core subject word frequency according to the word segmentation results corresponding to the title and the first paragraph; determine the sub-subject word frequency according to the word segmentation results corresponding to all paragraphs; The topic concentration is calculated according to the core topic word frequency and the sub-topic word frequency. The topic concentration is used to evaluate the degree of focus of the language text data to be evaluated around the core topic.

5. The method according to any one of claims 1 to 4, characterized in that The determining of the text semantic depth according to the information structure analysis result, the word segmentation result, the entity recognition result, and the paragraph division result includes: Obtaining all sentences containing core keywords based on the paragraph division results and the structural analysis results; Extract triples based on sentences containing core keywords and corresponding word segmentation results; and count the number of nodes in the triples that participate in reasoning chains exceeding a preset number of layers and the total number of nodes; The text semantic depth is obtained by calculating the ratio of the number of nodes in the reasoning chain exceeding the preset number of layers to the total number of nodes.

6. The method according to any one of claims 1 to 4, characterized in that The determining of text credibility according to the information structure analysis result, the word segmentation result, the entity recognition result, and the paragraph division result includes: Matching the keywords in the word segmentation result and the entity recognition result with the keywords in the empirical verb library to obtain empirical verbs; Determine the paragraph where the empirical verb is located based on the paragraph division result, and extract the evidence elements corresponding to the empirical verb according to the paragraph where the empirical verb is located, wherein the evidence elements include source entities and time expressions; The credibility of the text is determined by the evidence elements.

7. The method according to claim 5, characterized in that After performing information structure analysis based on the word segmentation result, the entity recognition result, and the paragraph division result to obtain an information structure analysis result, the method further includes: Calculate the sentiment quality coupling coefficient according to the word segmentation result, the entity recognition result, and the paragraph division result; Calculating an emotional quality score based on the emotional quality coupling coefficient and the information structure analysis result; The calculating the evaluation score of the language text data to be evaluated based on the text semantic depth, the text credibility, and the information structure analysis result includes: An evaluation score of the language text data to be evaluated is calculated based on the text semantic depth, the text credibility, the sentiment quality score, and the information structure analysis result.

8. The method according to claim 7, characterized in that The calculating of the sentiment quality coupling coefficient according to the word segmentation result, the entity recognition result, and the paragraph division result includes: Determine the number of positive sentiment sentences and negative sentiment sentences in each paragraph corresponding to the paragraph division result, word segmentation result, and entity recognition result; Calculate the number of positive emotion sentences and negative emotion sentences in the full text according to all the word segmentation results and the entity recognition results; The global sentiment polarity strength is calculated by the number of positive sentiment sentences, negative sentiment sentences and total number of sentences in the whole text; The local sentiment polarity strength is calculated based on the number of positive sentiment sentences, the number of negative sentiment sentences, and the total number of sentences in each paragraph; The emotion fluctuation coefficient is calculated by using the global emotion polarity strength and the local emotion polarity strength; and the emotion quality coupling coefficient is calculated according to the emotion fluctuation coefficient and the global emotion polarity strength.

9. The method according to claim 7, characterized in that The calculating of the emotion quality score according to the emotion quality coupling coefficient and the information structure analysis result includes: The basic quality of the content is obtained by mechanically weighting the information entropy density, logical coherence coefficient, and topic concentration in the information structure analysis results; The emotional quality score is calculated by using the content basic quality and the emotional quality coupling coefficient.

10. A language learning ability assessment system based on multimodal data fusion, characterized in that: The system comprises: An acquisition module, configured to acquire language text data to be evaluated; and verify the basic readability of the language text data to be evaluated; A preprocessing module, configured to preprocess the language text data to obtain word segmentation results, entity recognition results, and paragraph division results if the basic readability check of the language text data passes; An analysis module, configured to perform information structure analysis based on the word segmentation result, the entity recognition result, and the paragraph division result to obtain an information structure analysis result; A determination module, configured to determine text semantic depth and text credibility based on the information structure analysis result, the word segmentation result, the entity recognition result, and the paragraph division result; A calculation module is used to calculate the evaluation score of the language text data to be evaluated based on the text semantic depth, the text credibility and the information structure analysis result.

Images