A method and device for extracting course concepts of a MOOC

Abstract

This invention provides a method and apparatus for extracting course concepts for MOOCs. First, it fully utilizes the knowledge of a pre-trained language model to learn the domain distribution features of each concept in the vocabulary. Then, it uses the domain distribution features of each concept in the vocabulary to assist in the automatic annotation of course concepts in video subtitle text, removing domain-irrelevant noise. Next, based on the automatically annotated video subtitle text, it constructs a MOOC course concept annotation model to minimize concept noise and missing concepts in the BIO annotation of video subtitle text. Finally, it applies the MOOC course concept annotation model to the MOOC course concept extraction task, thereby efficiently and accurately extracting high-quality concepts from the course.

Description

Technical Field

[0001] This invention relates to the field of text data processing technology, and in particular to a method and apparatus for extracting course concepts for MOOCs. Background Technology

[0002] MOOCs aim to provide the best courses to a diverse range of learners worldwide via the internet. Compared to traditional classroom formats, online courses have brought revolutionary changes to education: text, video, audio, and images enrich educational content; the internet breaks down the time and space limitations of teaching; big data recommendations change the form of teaching; and knowledge services are influencing the future blueprint of education. MOOC concept extraction can help teachers organize knowledge content and build course frameworks, and it can also help students master important knowledge points and create their own mind maps, providing very practical educational guidance. However, with the continuous evolution and rapid changes in online education, how to extract high-quality course concepts from massive amounts of online courses has become a new challenge.

[0003] Currently, MOOC concept extraction methods can be broadly categorized into two types. The first is part-of-speech (POS) based method. This approach incorporates lexical information to aid entity boundary identification and controls the weight contribution of words. It integrates lexical information into character representations more simply and efficiently while minimizing model complexity. Furthermore, it utilizes attention mechanisms to introduce global contextual information, ultimately improving concept extraction. However, this method requires a large amount of manually labeled training data. Manually labeling course concepts introduces numerous errors, and the results vary from person to person. Given the diverse and rapidly growing nature of MOOCs, with significant differences in domains between courses, obtaining high-quality manually labeled results is impractical. This presents a significant challenge to concept extraction in MOOCs from a data perspective. The second method is text entity extraction. This approach uses vocabulary matching to first identify noisy entities, then iteratively refines and denoises the entity results to obtain reliable entity results. It avoids the need for labeled samples, saving the trouble of manual labeling. However, this method suffers from two problems: noise caused by incorrect labeling and missing labels due to limited vocabulary coverage. The second method only improves the results, removes noise, and improves the extraction quality by progressively improving the results based on rule mining. However, it has poor transferability in the MOOC scenario because the vocabulary of MOOCs has a wide variety of concepts, with many differences and low-quality concepts, so the vocabulary matching and annotation will have a lot of noise. At the same time, the limitation of vocabulary size and the high growth rate of courses will lead to serious problems with missing annotations.

[0004] Therefore, there is an urgent need to provide a MOOC course concept extraction technology that can overcome the above problems. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method and apparatus for extracting course concepts for MOOCs. First, it fully utilizes the knowledge of a pre-trained language model to learn the domain distribution characteristics of each concept in the vocabulary. Then, it uses the domain distribution characteristics of each concept in the vocabulary to assist in the automated annotation of course concepts from video subtitle text, removing domain-irrelevant noise. Next, based on the automatically annotated video subtitle text, it constructs a MOOC course concept annotation model to minimize concept noise and missing concepts in the BIO annotation of the video subtitle text. Finally, it applies the MOOC course concept annotation model to the MOOC course concept extraction task, thereby efficiently and accurately extracting high-quality concepts from the course.

[0006] In a first aspect, the present invention provides a method for extracting course concepts from MOOCs, the method comprising:

[0007] The probability distribution of each concept in the vocabulary across different subject areas is studied; wherein, the vocabulary consists of high-quality concepts from different subject areas.

[0008] Based on the aforementioned probability distribution, BIO annotation is performed on the video subtitle text of each course in the MOOC corpus.

[0009] Self-supervised learning is performed on each of the video subtitle texts and its BIO annotations to obtain a MOOC course concept annotation model;

[0010] The MOOC course concept annotation model is used to extract course concepts from the video subtitle text of the target course.

[0011] Secondly, the present invention provides a course concept extraction device for MOOCs, the device comprising:

[0012] The learning module is used to learn the probability distribution of each concept in the vocabulary across different subject areas; wherein, the vocabulary consists of high-quality concepts from different subject areas;

[0013] The BIO annotation module is used to perform BIO annotation on the video subtitle text of each course in the MOOC corpus based on the distribution probability.

[0014] The MOOC course concept annotation model construction module is used to construct a MOOC course concept annotation model based on each video subtitle text and its BIO annotation, which minimizes concept noise and concept missingness in the BIO annotation of the video subtitle text.

[0015] The course concept extraction module is used to extract course concepts from the video subtitle text of the target course using the MOOC course concept annotation model.

[0016] This invention provides a method and apparatus for extracting course concepts for MOOCs. First, it fully utilizes the knowledge of a pre-trained language model to learn the domain distribution features of each concept in the vocabulary. Then, it uses the domain distribution features of each concept in the vocabulary to assist in the automatic annotation of course concepts in video subtitle text, removing domain-irrelevant noise. Next, based on the automatically annotated video subtitle text, it constructs a MOOC course concept annotation model to minimize concept noise and missing concepts in the BIO annotation of video subtitle text. Finally, it applies the MOOC course concept annotation model to the MOOC course concept extraction task, thereby efficiently and accurately extracting high-quality concepts from the course. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the course concept extraction method for MOOCs provided by the present invention.

[0019] Figure 2 This is a schematic diagram of the MOOC course concept extraction device provided by the present invention;

[0020] Figure 3 This is a schematic diagram of the structure of the electronic device provided by the present invention;

[0021] Figure label:

[0022] 310: Processor; 320: Communication interface; 330: Memory; 340: Communication bus. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0024] The following is combined Figures 1-3 This invention describes a method and apparatus for extracting course concepts for MOOCs.

[0025] Explanation of the abbreviations and key terms of this invention:

[0026] Course concept extraction: Concept extraction aims to extract semantic concepts or terms of interest to users from unstructured text data, focusing on specific domains. For example, from a data structures textbook, the sentence "A stack is a last-in, first-out linear list that requires deletion and insertion operations only at the tail of the list" can be used to extract computer science concepts of interest to the reader, such as "stack" and "linear list." Focusing on the education field, course concept extraction refers to extracting high-quality, course-related concepts from educational resources such as textbooks. For instance, in the sentence above, since it comes from a computer science textbook's introduction to stacks, we want to obtain the computer science concept of "stack," rather than irrelevant domain concepts such as "requirement," "list," or "operation."

[0027] BIO annotation: A method for performing BIO annotation on sequence words. B is the starting word position of the concept, I is the other word positions of the concept, and O is the annotation result of non-concept words.

[0028] Pre-trained language models: A type of language model whose parameters have been initialized on another task. For example, GPT2 completed the pre-training task of autoregressive encoding and initialized its parameters before being applied to downstream tasks.

[0029] Hint Templates: Given a downstream task and a pre-trained language model, we construct a template to transform the input format into the original task format of the pre-trained language model, thus prompting the language model to output the answer in the original pre-trained task format. For example, given a sentiment classification task and a GPT2 model, and a known instance input of "The weather is nice today," we need to construct a hint template like "The weather is nice today, and I'm in a good mood," so that GPT2 can predict the next word and perform sentiment classification based on whether the predicted word is good or bad.

[0030] In previous course concept extraction tasks, the video subtitle text in MOOCs was automatically annotated using a vocabulary containing course concepts to obtain course-related concepts. This method not only saves the cost of manual annotation and enables automated annotation of course concepts under the conditions of rapid MOOC growth, but also helps online education platforms continuously provide high-quality knowledge services, which has important educational application significance. However, it also brings two problems: first, concept noise caused by incorrect annotation, and second, concept missing due to the limited coverage of the vocabulary. For example, the correct course concept extraction result for the MOOC teaching text "the promoting effect of auxin on the elongation growth of roots and stems" should be: "auxin", "root", and "stem". However, the annotation result based on the vocabulary is: "growth", "root", "growth", and "effect". The noise annotation here is the false positive redundant annotation, such as the annotation of "growth" and "effect". Missing annotations manifest in two forms: one is incomplete annotation, such as in the example above, "auxin" is only annotated as "growth". Incomplete annotation usually manifests as missing prefixes and suffixes, such as missing prefixes, for example, the concept "no free lunch theorem" is only annotated as "theorem". Second, blank spaces are marked. For example, the concept of "stem" was not extracted in the example above.

[0031] Furthermore, given the rapid increase in MOOC courses and the diverse, varied, and often low-quality concepts they cover, existing efforts to overcome annotation noise and missing annotations have not yielded the desired results. For example, mainstream approaches to address annotation noise augmentation involve statistical, empirical, and model-based methods, but these require some degree of human intervention. Algorithms also employ noise-tolerant loss functions or transfer supervised learning, reinforcement learning, and bagging algorithms to denoising scenarios. Essentially, these all involve iteratively removing low-quality annotations and selecting high-quality ones to train the model. However, these algorithms only minimize the negative impact of noise and do not fully utilize the annotation data itself, often failing with noisy MOOC data. Finally, in terms of model design, pre-trained language models have become the preferred choice in recent years, but they are prone to overfitting to erroneous data early on during fine-tuning.

[0032] To address the missing concept problem, mainstream research introduces external knowledge through template-based approaches. Many works mine candidate concepts based on part-of-speech templates or construct larger vocabularies based on linguistic patterns to improve concept coverage. While this works well in specific domain scenarios and with stable terminology sets, it is not suitable for MOOC scenarios characterized by high growth and multidisciplinary integration. Some works also incorporate rich syntactic knowledge from pre-trained language models to improve recall. However, in the context of course scenarios, the concepts in the humanities are sparse and the semantic spaces vary greatly, significantly reducing the model's ability to handle missing concepts. Instead, it introduces excessive noise, greatly lowering accuracy, making it counterproductive.

[0033] Therefore, there is an urgent need to provide a MOOC course concept extraction technology that can overcome annotation noise and missing annotations.

[0034] In view of this, the present invention provides a method for extracting course concepts for MOOCs, such as... Figure 1 As shown, the method includes:

[0035] S11. Learn the probability distribution of each concept in the vocabulary across different disciplines; wherein, the vocabulary consists of high-quality concepts from different disciplines.

[0036] S12. Based on the aforementioned probability distribution, perform BIO annotation on the video subtitle text of each course in the MOOC corpus;

[0037] S13. Based on each of the video subtitle texts and its BIO annotations, construct a MOOC course concept annotation model to minimize concept noise and concept missingness in the BIO annotations of the video subtitle texts;

[0038] S14. Using the MOOC course concept annotation model, extract course concepts from the video subtitle text of the target course.

[0039] This invention provides a method for extracting course concepts from MOOCs. First, it fully utilizes the knowledge of a pre-trained language model to learn the domain distribution features of each concept in the vocabulary. Then, it uses the domain distribution features of each concept in the vocabulary to assist in the automatic annotation of course concepts in video subtitle text, removing domain-irrelevant noise. Next, based on the automatically annotated video subtitle text, it constructs a MOOC course concept annotation model to minimize concept noise and missing concepts in the BIO annotation of video subtitle text. Finally, it applies the MOOC course concept annotation model to the MOOC course concept extraction task, thereby efficiently and accurately extracting high-quality concepts from the course.

[0040] Specifically, the method includes two modes. In the first mode, S11 includes:

[0041] Step 1: Manually label the subject area of ​​each concept in the thesaurus; here, the single subject area labeling of each concept is manually labeled.

[0042] Step 2: Connect the linear classification layer, the single-class labeling prediction layer, and the optimizer sequentially after the BERT pre-trained language model to obtain the BERT concept classification model;

[0043] Step 3: Using the BERT pre-trained language model and the linear classification layer, process each concept in the vocabulary to obtain the predicted distribution probability value of each concept in the vocabulary in different subject areas;

[0044] Step 4: In the single-class labeling prediction layer, based on the predicted distribution probability values ​​of each concept in the vocabulary across different subject areas, generate the subject area prediction result for each concept in the vocabulary;

[0045] Step 5: In the optimizer, the cross-entropy between the subject domain prediction result and the subject domain label of each concept in the vocabulary is used as the training loss of each concept in the vocabulary. The training loss of all concepts in the vocabulary is used to optimize the BERT concept classification model.

[0046] Step 6: Repeat steps 3 to 5 until the BERT concept classification model converges, and regard the predicted distribution probability of each concept in the vocabulary in different subject areas when the BERT concept classification model converges as the learned distribution probability of each concept in the vocabulary in different subject areas.

[0047] This section utilizes a vocabulary-based classification task to fine-tune the BERT pre-trained language model, specifically for a vocabulary T = {c} i} i＝1,...,Q Each concept c i The expected output is concept c in the vocabulary. i The invention presents multi-classification results across various academic disciplines. To fully utilize the knowledge of the BERT pre-trained language model, a linear classification layer, a single-class labeling prediction layer, and an optimizer are subsequently added to the BERT pre-trained language model to generate an initial BERT concept classification model. The BERT concept classification model is then trained using the vocabulary's single-class labeling classification results until it converges.

[0048] Preferably, the BERT concept classification model is trained using the single-label classification results of the vocabulary concept until it converges, including:

[0049] For any concept in the dictionary

[0050] The concept is input into the BERT pre-trained language model for processing to obtain the encoded features of the concept;

[0051] Linear classification layers are used to process encoded features in order to predict the probability distribution of the concept in different disciplines.

[0052] In the single-class labeling prediction layer, the softmax function is used to activate the probability distribution, and the argmax function is used to select the subject area with the highest probability as the single subject area learning result of the concept.

[0053] Using Adam as the optimizer, the following loss function is used in Adam to calculate the learning loss of the concept, and the parameters of the BERT concept classification model are optimized using the learning loss of all concepts in the vocabulary.

[0054] Repeat the above steps until the BERT concept classification model converges.

[0055] The loss function formula is:

[0056]

[0057] Where C represents the number of subject area categories, p i q is the true value. i These are predicted values.

[0058] Ultimately, the BERT concept classification model possesses domain recognition capabilities. Simultaneously, the distribution probability of each concept in the linear classification layer across different subject domains is derived when the BERT concept classification model converges. The individual classification results of the vocabulary concepts themselves are enhanced by the pre-trained language model to represent the distribution across all subject domains. In this way, errors such as mislabeling, omissions, and outdated information in manual domain classification of the vocabulary can be compensated for.

[0059] S12 includes:

[0060] S12-A: Perform part-of-speech tagging and regular expression matching on the video subtitle text of each course to merge non-noun tags of 0 and adjacent nouns with the same part of speech into an untagged noun phrase, generating the initial tagged text of the video subtitle text of each course;

[0061] S12-B: Perform BI annotation on unannotated noun phrases in the initial annotation text that meet the first condition, and perform O annotation on unannotated noun phrases that do not meet the first condition, so as to achieve BIO annotation for the video subtitle text of each course.

[0062] The first condition is that the unlabeled noun phrase is a concept in the vocabulary, and the intersection of the K subject areas with the highest probability of the unlabeled noun phrase in different subject areas with the subject area label of each course is not empty.

[0063] Automated annotation related to the course domain is usually done by directly matching the longest match between the concepts in the vocabulary and the video subtitle text. However, when the course concepts are complex and there are many domains, a lot of noise from irrelevant domains will be introduced. If the results generated by S11 are used, the noise from irrelevant domains can be removed, and the purpose of noise reduction can be achieved to a certain extent.

[0064] The automated annotation related to the course domain of this invention is as follows:

[0065] Traverse the MOOC corpus for courses C that require automated tagging i ={S i F i}, where S iThis is the video subtitle text for the course, F i This refers to the subject area to which the course belongs, such as the F subject area for a course "Discrete Mathematics". i This may include the disciplines of computer science and mathematics.

[0066] For S i After word segmentation, obvious errors are removed by part-of-speech tagging and regular expression matching, and missing concepts are identified to obtain S. i Preliminary marking; S i After word segmentation, phrases of different parts of speech are connected. If they meet the requirements of regular expressions, they are considered to be expandable concepts. For example, since "cardiovascular disease" and "disease" are both in the vocabulary, they can be expanded to "cardiovascular disease," which solves the problem of incomplete annotation to some extent. At the same time, only nouns and noun phrases are retained to remove many obvious errors during the annotation process. For example, "recorded in classics and historical books" annotates the concept of "classics and history," but here "classics" means "through" and is a preposition, while "historical books" is a noun and does not mean "classics, history, philosophy, and literature."

[0067] Traverse S i The initial marking, removing those that do not belong to F i Irrelevant noise.

[0068] Among them, for a noun / noun phrase pc i Only when the domain labeling results overlap with the course classification itself should BI be labeled; otherwise, it is noise unrelated to the course domain and should be labeled O.

[0069] This invention combines part-of-speech tagging and regular expression matching to retain only nouns / noun phrases. i Then PC i top K The classification results are applied to the BIO automated annotation of the course video subtitle text to filter high-quality concepts related to the course and improve the accuracy of the annotation process.

[0070] S13 includes:

[0071] S13-A: Train the first student model and the second student model; wherein, the first student model / second student model is obtained by performing self-supervised learning on each video subtitle text and its BIO annotation for a certain number of iterations based on the BERT concept classification model;

[0072] S13-B: Generate a first teacher model that is consistent with the first student model, and generate a second teacher model that is consistent with the second student model;

[0073] S13-C: Perform joint optimization on the first student model, the second student model, the first teacher model, and the second teacher model to remove concept labeling noise and mine missing concept labels, and select the one with the best performance after optimization as the MOOC course concept labeling model.

[0074] After preparing the BIO-annotated data (i.e., video subtitle text and its BIO annotations), a suitable model and robust algorithm are needed to further address the issues of blank annotations, redundant annotations, and incomplete annotations. In S11, the BERT concept classification model obtained by fine-tuning the BERT pre-trained model using a classification task can perceive the course domain distribution of each word. In this step, the present invention uses the prepared BIO-annotated data to further fine-tune the BERT concept classification model, generating a first student model and a second student model. Specifically, the BERT concept classification model uses BERT as a feature encoder, which enhances the model's perception of the concept domain. The input is the token id obtained after the word segmentation encoder, which then enters the BERT pre-trained model to extract rich text features to obtain an output vector of [batch_size*max_seq_len*emb_size]. The output vector is passed through Bi-LSTM to extract the features required for concept entity recognition, resulting in a vector of [batch_size*max_seq_len*(2*hidden_size)]. Finally, it enters the CRF layer for decoding to calculate the optimal annotation sequence. Although cross-entropy and CRF are essentially the same at the word level, CRF exhibits significantly higher recall and precision at the concept word level due to the inclusion of constraints on annotation transition probabilities. Furthermore, this invention employs teacher-student self-supervised training, constructing a first teacher model-first student model and a second teacher model-second student model. These four models are optimized from the perspectives of removing concept annotation noise and uncovering missing concept annotations, ultimately selecting the optimal one as the concept annotation model for MOOC courses.

[0075] S14 includes:

[0076] Input the video subtitle text of the target course into the MOOC course concept annotation model to obtain the BIO annotation of the video subtitle text of the target course;

[0077] Extract continuous BI fields from the BIO annotation of the video subtitle text of the target course as a course concept to achieve course concept extraction of the target course.

[0078] In the second mode, S11 includes:

[0079] S11-I: Construct a prompt template for each concept in the vocabulary; wherein, the i-th concept c in the vocabularyI The prompt template represents the [MASK] used to fill in the subject area, including c i ;

[0080] S11-II: For each subject area covered by the vocabulary and each concept in the vocabulary, the subject area is filled into the prompt template constructed for the concept to obtain the first prompt template;

[0081] S11-III: Under the prompt of the first prompt template, use the BERT pre-trained language model to predict the probability of the concept in the subject area; integrate to obtain the distribution probability of each concept in the vocabulary in different subject areas.

[0082] Considering that in reality, the vocabulary may not contain any domain-related auxiliary information and there is no training data to fine-tune the BERT pre-trained language model, this invention constructs a prompt template to learn the classification results of concepts in the vocabulary under different disciplines.

[0083] For example: For vocabulary T = {c i} i＝1，...，Q Each concept c i :

[0084] First, according to the prompt template x′=f prompt (ci), where f prompt (·) Functions are templates that characterize [MASK] used to fill in subject areas, including [concept] used to fill in concepts, for example: f prompt (·) means "[MASK] includes [concept]", when a concept c comes from the vocabulary. I When "identifying", use c I Fill in the corresponding "[concept]" slot to obtain information about concept c. I One prompt is "[MASK] includes identification." While human prompt templates may align with human thought processes, this isn't always the case for machine learning models. In practice, human prompt templates are often unstable, with different templates producing significantly different results, or even minor changes of just a few words causing huge fluctuations. Therefore, this invention employs the Hearst model to optimize the human prompt template. The Hearst model is based on language behavior patterns extracted from a large corpus. Prompt models inspired by the Hearst model are better adapted to machine learning models.

[0085] Then, fill in [MASK] according to the subject areas covered by the vocabulary to obtain f. fill (x′,f j ), where f jThe set of subject areas F can be further subdivided into finer-grained categories according to the "Catalogue of Disciplines and Majors for Granting Doctoral and Master's Degrees and Training Graduate Students" promulgated by the Ministry of Education of the People's Republic of China in 1997. fill A function is based on x′ and uses f j This is obtained after filling the [MASK] slot.

[0086] Then, the concept c is obtained using the BERT pre-trained language model. i The probability distribution across F subject areas: p j (x′)=LM(f fill (x′,f j );θ), p j (x′) Concept c i In f j The probability distribution across subject areas is given by LM (a BERT pre-trained language model) and θ (the parameters of the BERT pre-trained language model). Considering the varying word segmentation lengths across different subject areas, this invention uses GLM to predict the probability distribution across F subject areas. For example, in the above example, because the concept of "identification" appears more frequently in the "chemistry" field, when f... j When corresponding to the subject of "chemistry," the expected probability is higher than that of other unrelated fields, thus activating the necessary knowledge from the BERT pre-trained language model to obtain auxiliary information. This invention learns the distribution probability of each concept in the vocabulary across different subject areas. The vocabulary, through the BERT pre-trained language model, perceives subject-related knowledge, thereby helping to reduce noise in the subsequent automated annotation process.

[0087] S12 includes:

[0088] The steps are the same as in step S12 of the first mode, and will not be repeated here.

[0089] S13 includes:

[0090] S13-I: Embed the distribution probability of each concept in the vocabulary across different subject areas into the BERT feature encoding layer of the BERT pre-trained language model to obtain a first BERT pre-trained language model that can perceive the course domain distribution of each word segmentation;

[0091] The feature encoding layer of the first BERT pre-trained language model is represented by the following formula:

[0092]

[0093] Where, x j As a word segment, E j For the feature encoding layer of the first BERT pre-trained language model, the word segmentation x jEncoded features, For the i-th concept c in the vocabulary i The distribution probability in different subject areas, Encoder() is the encoding operation of the feature encoding layer of the BERT pre-trained language model, and W is the learning matrix for learning the feature mapping relationship between the subject area dimension of the vocabulary and the encoding dimension of Encoder().

[0094] S13-II: Train the first student model and the second student model; wherein, the first student model / second student model is obtained by performing self-supervised learning on each video subtitle text and its BIO annotation for a certain number of iterations based on the first BERT pre-trained language model;

[0095] S13-III: Generate a first teacher model that is consistent with the first student model, and generate a second teacher model that is consistent with the second student model;

[0096] S13-IV: Perform joint optimization on the first student model, the second student model, the first teacher model, and the second teacher model to remove concept annotation noise and mine missing concept annotations, and select the one with the best performance after optimization as the MOOC course concept annotation model;

[0097] This invention reconstructs the BERT encoding layer in the BERT pre-trained language model by introducing concept domain distribution features, thereby enhancing the model's perception of the concept domain; specifically, for concepts... The word segmentation results can be used to learn the domain distribution of the language through a BERT pre-trained language model: Then each x f (t≤f≤t+|c i |) will learn through matrix Mapped to the BERT embedding dimension, it is then concatenated with the word-level embeddings of BERT to obtain new model feature inputs.

[0098] That is, for any word x j , encoding x j Feature E after j for:

[0099]

[0100] Specifically, if x j When there is no corresponding concept label, use 0 to fill in its domain distribution U.

[0101] Then, using the prepared BIO labeled data, the reconstructed BERT pre-trained language model is fine-tuned to generate the first student model and the second student model. In addition, similar to step S13 in one mode, this step also adopts teacher-student self-supervised training to construct the first teacher model-first student model and the second teacher model-second student model. These four models are optimized from the perspective of removing concept labeling noise and mining missing concept labels, and then the best one is selected as the concept labeling model for MOOC courses.

[0102] S14 includes:

[0103] The steps are the same as those in the first mode, and will not be repeated here.

[0104] Furthermore, the methods used in S13-C of the first mode and S13-IV of the second mode are consistent, specifically including:

[0105] The first student model, the second student model, the first teacher model, and the second teacher model are jointly optimized according to the first, second, third, and fourth preset rules until all four converge.

[0106] The first preset rule is: update the parameters in the first student model / second student model to the first teacher model / second teacher model according to the preset decay rate;

[0107] The second preset rule is: every certain number of iterations, the BIO annotations generated by the first teacher model and the second teacher model for each video subtitle text are interchanged; the BIO annotations generated by the first teacher model and the second teacher model for each video subtitle text are composed of the BIO annotations generated by the first teacher model and the second teacher model for each word segmentation in each video subtitle text and their confidence scores.

[0108] The third preset rule is: in each iteration, based on the BIO annotation, consistency detection and high confidence detection generated by the first teacher model / second teacher model for each video subtitle text, determine the pseudo BIO annotation of each video subtitle text in the first student model / second student model;

[0109] The fourth preset rule is: in each iteration, based on the pseudo-BIO annotation of each video subtitle text in the first student model / second student model, determine the first loss for removing concept annotation noise and the second loss for mining missing concept annotations, and use the first loss and the second loss to optimize the parameters of the first student model / second student model.

[0110] Furthermore, the third preset rule involves BIO annotation, consistency detection, and high-confidence detection generated for each of the video subtitle texts based on the first teacher model / second teacher model.

[0111] Determine the pseudo-BIO annotation for each of the video subtitle texts in the first student model / second student model, including:

[0112] Determine whether the BIO annotations generated by the first teacher model / second teacher model for each word segmentation in each of the video subtitle texts simultaneously meet the consistency requirement and the high confidence requirement, and record the word segments corresponding to the BIO annotations that do not simultaneously meet the consistency requirement and the high confidence requirement in the first word segmentation set;

[0113] Based on the BIO annotations generated by the first teacher model / second teacher model for each word in the video subtitle text that does not belong to the first word segmentation set, and the pseudo-BIO annotations generated by the first student model / second student model for each word in the video subtitle text that belongs to the second word segmentation set in the previous iteration, pseudo-BIO annotations for each video subtitle text in the first student model / second student model are generated.

[0114] If the BIO annotation generated by the first teacher model / second teacher model for each word segmentation in each video subtitle text is consistent with its previous iteration, then the consistency requirement is met.

[0115] If the confidence level of the BIO annotation generated by the first teacher model / second teacher model for each word segmentation in each of the video subtitle texts is greater than a preset threshold, then the high confidence level requirement is met.

[0116] The fourth preset rule determines a first loss for removing concept annotation noise based on the pseudo-BIO annotation of each video subtitle text in the first student model / second student model, including:

[0117] Based on the pseudo-BIO annotation of each video subtitle text in the first student model / second student model, and the BIO annotation generated by the first student model / second student model for each video subtitle text, a first loss for removing concept annotation noise is determined;

[0118] Based on the pseudo-BIO annotations of each video subtitle text in the first student model / second student model, a second loss for mining missing concept annotations is determined, including:

[0119] The parts marked "B" and "1" in the pseudo-BIO annotation of each video subtitle text in the first student model / second student model are mapped to 1, and the parts marked "O" are mapped to 0, so as to generate each video subtitle text carrying 0-1 annotations.

[0120] Each video subtitle text with 0-1 labels is split into a positive sample consisting of words labeled 1 and a negative sample consisting of words labeled 0, to generate a positive sample set and a negative sample set.

[0121] Using the positive sample set and the negative sample set, a second loss is determined for mining missing concept labels.

[0122] This invention is based on a teacher-student model and designs a self-supervised learning framework aimed at denoising and reducing missing labels. Then, a collaborative learning approach for teacher models is further introduced. Single teacher-student self-supervised training can tolerate a certain degree of noise and reduce missing labels to some extent. However, the interaction between teacher models further enhances the utilization of data and can further enhance the model's denoising ability and ability to reduce missing labels.

[0123] Specifically, the teacher-student model was initially proposed in knowledge distillation. Teacher models are often more complex but possess superior performance and generalization ability. The teacher model is used to guide the training of simpler student models, enabling them to achieve performance comparable to the teacher model with fewer parameters and simpler computations. This invention applies the teacher-student model to self-supervised learning, where the teacher model's role is to provide pseudo-labels for the student model's text annotation training.

[0124] This invention uses f(·;θ) respectively t ) and f(·;θ s Let f represent the teacher model and the student model. n，c (·;·) represents the probability that the nth word in a sentence belongs to the nth subject area, so for the automated annotation results Where X m =[x m，1 , ..., x m，N ] represents the N segmentations of the first sentence, D m =[d m，1 , ..., d m，N The [] represents the corresponding BIO automated annotation result. During the student model construction process, the student model will be minimized within [the specified range]. We use cross-entropy loss to learn model parameters:

[0125]

[0126] in, The cross-entropy loss is the BIO annotation generated by the student model for the nth word in the mth video subtitle text and the BIO annotation result corresponding to the nth word in the mth video subtitle text.

[0127] The teacher model is initially initialized by copying the parameters of the student model. During the self-supervised process, it is updated at each step using an exponential moving average, thereby improving the stability and robustness of the teacher model. Here, α is the decay rate, controlling the speed at which the teacher model updates its parameters from the student model.

[0128] θ t ←αθ t +(1-α)θ s

[0129] The key to self-supervised training lies in how the student model selects suitable data from the pseudo-labels generated by the teacher model. This invention summarizes the purpose of teacher-student self-supervised learning into two aspects: First, the requirement for consistency in labeling. Since the model should produce similar outputs for similar inputs, the selection process encourages retaining data with consistent model predictions. Second, the requirement for high confidence in the labeling. Initially, for more difficult data, the model's predictions tend to fluctuate, leading to errors. Therefore, a threshold is set to filter out data from which the model has a higher level of confidence. In summary, consistency and high confidence define the data that the student model filters from the pseudo-labels generated by the teacher model as follows:

[0130]

[0131] In self-supervised teacher-student self-supervised learning, although noise can be removed by selecting consistent and high-confidence predictions to avoid impacting model performance, the labeled data is not fully utilized. If the ability to detect errors can be improved to the ability to correct them, the corrected data can be used as high-quality training data in the next iteration. Therefore, this invention enhances overall data application by using two self-supervised teacher-learning models. Specifically, this invention randomly assigns different label masks to the student models of the two teacher models during initial initialization, allowing each student model to have its own strengths. After a certain number of outer loop cycles, the pseudo-labels generated by the two teacher models are provided to the corresponding student models to correct errors and fully utilize the data.

[0132] Therefore, the self-supervised learning algorithm process is as follows:

[0133] A: First, iterate a certain number of steps to pre-train two student models: θ s1 ,θ s2

[0134] B: Copy the parameters of each student model and initialize the teacher model: θ t1 ←θ s1 ,θ t2 ←θ s2

[0135] C: During the teacher-student model self-supervised learning phase, a batch of training data is generated.

[0136] C1: Teachers generate false annotations

[0137] C2: Selection based on consistency and high confidence requirements

[0138] C3: Update θ after backpropagation of the student model s1 ,θ s2 ;

[0139] C4: The teacher model updates θ using an exponential moving average. t1 ,θ t2 ;

[0140] C5: At a specific number of iterations, the pseudo-labels generated by swapping the two teacher models further enhance the model's ability to utilize the data.

[0141]

[0142]

[0143] D: Ultimately, in the four models θ s1 ,θ s2 ,θ t1 ,θ t2 The one with the best performance was selected as the MOOC course concept annotation model.

[0144] Although the self-supervised collaborative learning mentioned above already has good noise reduction capabilities, recall is of extra concern in real-world scenarios. Therefore, this invention optimizes the cross-entropy loss function of the student model and adds the loss of positive sample unlabeled learning, because it can consistently and unbiasedly evaluate the task loss, just like supervised learning, thereby uncovering more missing concepts.

[0145] Because unlabeled learning of positive samples can only handle binary classification, this invention maps "O" to 0 and "B" and "I" to 1, thus enabling the automated labeling results D to be processed. m =[d m,1 ,...,d m,N The dataset is divided into a set of positive samples with concept labels. and unlabeled sample set Then based on the positive sample set and unlabeled sample set To improve the model recall, a positive sample unlabeled loss (i.e., the second loss) is designed.

[0146] In summary, this invention proposes a framework for course concept annotation in MOOCs. Simply matching based on a simple vocabulary often leads to missing and noisy annotations. Furthermore, the numerous domains, complex concepts, and rapid growth of MOOCs exacerbate these problems. Therefore, the framework designed in this invention addresses the challenges of noise and missing concepts in three steps. First, it fully utilizes the knowledge of a pre-trained language model to learn the domain distribution characteristics of concepts in the vocabulary. Second, it uses these domain distribution characteristics to assist in the automated annotation of course concepts in video subtitle text, removing domain-irrelevant noise. Third, based on the automatically annotated video subtitle text, it constructs a MOOC course concept annotation model that minimizes concept noise and missing concepts in BIO annotation of video subtitle text, building upon the introduction of the domain distribution characteristics of concepts from the vocabulary.

[0147] On the Chinese dataset based on XuetangX, other methods suffer from low accuracy and recall due to overfitting to noisy data or their inability to handle the highly open nature of MOOCs. In contrast, this invention embeds concept domain classification features into the model, enabling it to perceive more domain information and significantly improving the accuracy of model annotation. This results in a substantial improvement over baseline solutions in unsupervised scenarios. The designed self-supervised learning algorithm selects high-quality data that meets consistency and high confidence requirements for training during iterations, thereby greatly improving the accuracy of annotation prediction and achieving a 7% improvement in F1 score on top of a strong baseline. Considering the model's high recall requirements, a positive sample unlabeled learning loss term is added to extract more missing concepts from negative examples. The entire course concept annotation framework specifically addresses the noise and missing concepts problems in automated annotation, helping MOOCs efficiently and accurately extract high-quality course-related concepts automatically.

[0148] Secondly, a course concept extraction device for MOOCs provided by the present invention will be described. The course concept extraction device for MOOCs described below can be referred to in correspondence with the course concept extraction method for MOOCs described above. Figure 2 This example illustrates the structure of a MOOC course concept extraction device, such as... Figure 2 As shown, the device includes:

[0149] Learning module 21 is used to learn the probability distribution of each concept in the vocabulary across different subject areas; wherein, the vocabulary consists of high-quality concepts from different subject areas;

[0150] BIO annotation module 22 is used to perform BIO annotation on the video subtitle text of each course in the MOOC corpus based on the distribution probability.

[0151] MOOC course concept annotation model construction module 23 is used to construct a MOOC course concept annotation model based on each video subtitle text and its BIO annotation, which minimizes concept noise and concept missingness in the BIO annotation of the video subtitle text.

[0152] The course concept extraction module 24 is used to extract course concepts from the video subtitle text of the target course using the MOOC course concept annotation model.

[0153] This invention provides a MOOC course concept extraction device. First, it fully utilizes the knowledge of a pre-trained language model to learn the domain distribution features of each concept in the vocabulary. Then, it uses the domain distribution features of each concept in the vocabulary to assist in the automatic annotation of course concepts in video subtitle text, removing domain-irrelevant noise. Next, based on the automatically annotated video subtitle text, it constructs a MOOC course concept annotation model to minimize concept noise and missing concepts in the BIO annotation of video subtitle text. Finally, it applies the MOOC course concept annotation model to the MOOC course concept extraction task, thereby efficiently and accurately extracting high-quality concepts from the course.

[0154] Thirdly, Figure 3 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 3 As shown, the electronic device may include a processor 310, a communications interface 320, a memory 330, and a communication bus 340, wherein the processor 310, communications interface 320, and memory 330 communicate with each other via the communication bus 340. The processor 310 can call logical instructions in the memory 330 to execute a MOOC course concept extraction method. This method includes: learning the distribution probability of each concept in a vocabulary across different subject areas; wherein the vocabulary consists of high-quality concepts from different subject areas; performing BIO annotation on the video subtitle text of each course in the MOOC corpus based on the distribution probability; constructing a MOOC course concept annotation model based on each video subtitle text and its BIO annotation to minimize concept noise and missing concepts in the video subtitle text BIO annotation; and using the MOOC course concept annotation model to extract course concepts from the video subtitle text of the target course.

[0155] Furthermore, the logical instructions in the aforementioned memory 330 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0156] Fourthly, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute a MOOC course concept extraction method provided by the methods described above. The method includes: learning the distribution probability of each concept in a vocabulary across different subject areas; wherein the vocabulary consists of high-quality concepts from different subject areas; performing BIO annotation on the video subtitle text of each course in the MOOC corpus based on the distribution probability; constructing a MOOC course concept annotation model based on each video subtitle text and its BIO annotation to minimize concept noise and missing concepts in the BIO annotation of the video subtitle text; and using the MOOC course concept annotation model to extract course concepts from the video subtitle text of a target course.

[0157] Fifthly, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a method for extracting course concepts for MOOCs provided by the methods described above. This method includes: learning the distribution probability of each concept in a vocabulary across different subject areas; wherein the vocabulary consists of high-quality concepts from different subject areas; performing BIO annotation on the video subtitle text of each course in the MOOC corpus based on the distribution probability; constructing a MOOC course concept annotation model based on each video subtitle text and its BIO annotation to minimize concept noise and missing concepts in the BIO annotation of the video subtitle text; and using the MOOC course concept annotation model to extract course concepts from the video subtitle text of a target course.

[0158] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0159] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0160] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for extracting course concepts for MOOCs, characterized in that, The method includes: The probability distribution of each concept in the vocabulary across different subject areas is studied; wherein, the vocabulary consists of high-quality concepts from different subject areas. Based on the aforementioned probability distribution, BIO annotation is performed on the video subtitle text of each course in the MOOC corpus; wherein, BIO annotation is performed on sequential single words, B is the starting word position of the concept, I is the other word position of the concept, and O is the annotation result of non-concepts; Based on each video subtitle text and its BIO annotation, a MOOC course concept annotation model is constructed to minimize concept noise and concept missingness in the BIO annotation of the video subtitle text. Using the MOOC course concept annotation model, course concepts can be extracted from the video subtitle text of the target course. Each course includes video subtitle text and subject area annotation; based on the aforementioned probability distribution, BIO annotation is performed on the video subtitle text of each course in the MOOC corpus, including: For each course's video subtitle text, perform part-of-speech tagging and regular expression matching to merge non-noun tags of 0 and adjacent nouns with the same part of speech into an untagged noun phrase, generating the initial tagged text for each course's video subtitle text; Unlabeled noun phrases in the initial labeled text that meet the first condition are labeled with BI, and unlabeled noun phrases that do not meet the first condition are labeled with O, so as to achieve BIO labeling of the video subtitle text of each course. The first condition is that the unlabeled noun phrase is a concept in the vocabulary, and the intersection of the K subject areas with the highest probability of the unlabeled noun phrase in different subject areas with the subject area label of each course is not empty.

2. The method for extracting course concepts for MOOCs according to claim 1, characterized in that, Learn the probability distribution of each concept in the vocabulary across different subject areas, including: Step 1: Obtain the subject area label for each concept in the vocabulary; Step 2: Connect the linear classification layer, the single-class labeling prediction layer, and the optimizer sequentially after the BERT pre-trained language model to obtain the BERT concept classification model; Step 3: Using the BERT pre-trained language model and the linear classification layer, process each concept in the vocabulary to obtain the predicted distribution probability value of each concept in the vocabulary in different subject areas; Step 4: In the single-class labeling prediction layer, based on the predicted distribution probability values ​​of each concept in the vocabulary across different subject areas, generate the subject area prediction result for each concept in the vocabulary; Step 5: In the optimizer, the cross-entropy between the subject domain prediction result and the subject domain label of each concept in the vocabulary is used as the training loss of each concept in the vocabulary. The training loss of all concepts in the vocabulary is used to optimize the BERT concept classification model. Step 6: Repeat steps 3 to 5 until the BERT concept classification model converges, and regard the predicted distribution probability of each concept in the vocabulary in different subject areas when the BERT concept classification model converges as the learned distribution probability of each concept in the vocabulary in different subject areas.

3. The method for extracting course concepts for MOOCs according to claim 1, characterized in that, Learn the probability distribution of each concept in the vocabulary across different subject areas, including: For each concept in the vocabulary, construct a prompt template; wherein, the first concept in the vocabulary... One concept The prompt template represents the [MASK] used to fill in the subject area, including ; For each subject area covered by the vocabulary and for each concept in the vocabulary, the subject area is filled into a prompt template constructed for the concept to obtain a first prompt template; Following the prompts of the first prompt template, the probability of the concept in the subject area is predicted using a BERT pre-trained language model. The probability distribution of each concept in the vocabulary across different disciplines is obtained by integration.

4. The MOOC course concept extraction method according to claim 2, characterized in that, Based on each video subtitle text and its BIO annotation, a MOOC course concept annotation model is constructed to minimize concept noise and concept missingness in the BIO annotation of the video subtitle text, including: Train a first student model and a second student model; wherein the first student model / second student model is obtained by self-supervised learning of each video subtitle text and its BIO annotation by performing a certain number of iterations based on the BERT concept classification model; Generate a first teacher model that is consistent with the first student model, and generate a second teacher model that is consistent with the second student model; The first student model, the second student model, the first teacher model, and the second teacher model are jointly optimized to remove concept labeling noise and mine missing concept labels, and the one with the best performance after optimization is selected as the concept labeling model for MOOC courses.

5. The method for extracting course concepts for MOOCs according to claim 3, characterized in that, Based on each video subtitle text and its BIO annotation, a MOOC course concept annotation model is constructed to minimize concept noise and concept missingness in the BIO annotation of the video subtitle text, including: The distribution probability of each concept in the vocabulary across different subject areas is embedded into the BERT feature encoding layer of the BERT pre-trained language model to obtain a first BERT pre-trained language model that can perceive the course domain distribution of each word segment. Train a first student model and a second student model; wherein the first student model / second student model is obtained by performing self-supervised learning on each video subtitle text and its BIO annotation for a certain number of iterations based on the first BERT pre-trained language model; Generate a first teacher model that is consistent with the first student model, and generate a second teacher model that is consistent with the second student model; The first student model, the second student model, the first teacher model, and the second teacher model are jointly optimized to remove concept annotation noise and mine missing concept annotations, and the one with the best performance after optimization is selected as the concept annotation model for MOOC courses. The feature encoding layer of the first BERT pre-trained language model is represented by the following formula: in, For a word segmentation, For word segmentation, the feature encoding layer of the first BERT pre-trained language model... Encoded features, For the word in the vocabulary One concept The probability distribution across different academic disciplines This refers to the encoding operation of the feature encoding layer of the BERT pre-trained language model. To the subject domain dimension of the aforementioned vocabulary A learning matrix for learning the feature mapping relationships between encoding dimensions.

6. The method for extracting course concepts for MOOCs according to claim 4 or 5, characterized in that, The joint optimization of the first student model, the second student model, the first teacher model, and the second teacher model regarding removing concept labeling noise and mining missing concept labels includes: The first student model, the second student model, the first teacher model, and the second teacher model are jointly optimized according to the first, second, third, and fourth preset rules until all four converge. The first preset rule is: update the parameters in the first student model / second student model to the first teacher model / second teacher model according to the preset decay rate; The second preset rule is: every certain number of iterations, the BIO annotations generated by the first teacher model and the second teacher model for each video subtitle text are interchanged; the BIO annotations generated by the first teacher model and the second teacher model for each video subtitle text are composed of the BIO annotations generated by the first teacher model and the second teacher model for each word segmentation in each video subtitle text and their confidence scores. The third preset rule is: in each iteration, based on the BIO annotation, consistency detection and high confidence detection generated by the first teacher model / second teacher model for each video subtitle text, determine the pseudo BIO annotation of each video subtitle text in the first student model / second student model; The fourth preset rule is: in each iteration, based on the pseudo-BIO annotation of each video subtitle text in the first student model / second student model, determine the first loss for removing concept annotation noise and the second loss for mining missing concept annotations, and use the first loss and the second loss to optimize the parameters of the first student model / second student model.

7. The method for extracting course concepts for MOOCs according to claim 6, characterized in that, Based on the BIO annotations, consistency checks, and high-confidence checks generated by the first teacher model / second teacher model for each of the video subtitle texts, the pseudo-BIO annotations for each of the video subtitle texts in the first student model / second student model are determined, including: Determine whether the BIO annotations generated by the first teacher model / second teacher model for each word segmentation in each of the video subtitle texts simultaneously meet the consistency requirement and the high confidence requirement, and record the word segments corresponding to the BIO annotations that do not simultaneously meet the consistency requirement and the high confidence requirement in the first word segmentation set; Based on the BIO annotations generated by the first teacher model / second teacher model for each word in the video subtitle text that does not belong to the first word segmentation set, and the pseudo-BIO annotations generated by the first student model / second student model for each word in the video subtitle text that belongs to the second word segmentation set in the previous iteration, pseudo-BIO annotations for each video subtitle text in the first student model / second student model are generated. If the BIO annotation generated by the first teacher model / second teacher model for each word segmentation in each video subtitle text is consistent with its previous iteration, then the consistency requirement is met. If the confidence level of the BIO annotation generated by the first teacher model / second teacher model for each word segmentation in each of the video subtitle texts is greater than a preset threshold, then the high confidence level requirement is met.

8. The method for extracting course concepts for MOOCs according to claim 7, characterized in that, Based on the pseudo-BIO annotations of each video subtitle text in the first student model / second student model, a first loss for removing concept annotation noise is determined, including: Based on the pseudo-BIO annotation of each video subtitle text in the first student model / second student model, and the BIO annotation generated by the first student model / second student model for each video subtitle text, a first loss for removing concept annotation noise is determined; Based on the pseudo-BIO annotations of each video subtitle text in the first student model / second student model, a second loss for mining missing concept annotations is determined, including: The parts marked "B" and "1" in the pseudo-BIO annotation of each video subtitle text in the first student model / second student model are mapped to 1, and the parts marked "O" are mapped to 0, so as to generate each video subtitle text carrying 0-1 annotations. Each video subtitle text with 0-1 labels is split into a positive sample consisting of words labeled 1 and a negative sample consisting of words labeled 0, to generate a positive sample set and a negative sample set. Using the positive sample set and the negative sample set, a second loss is determined for mining missing concept labels.

9. A MOOC course concept extraction device, characterized in that, The device includes: The learning module is used to learn the probability distribution of each concept in the vocabulary across different subject areas; wherein, the vocabulary consists of high-quality concepts from different subject areas; The BIO annotation module is used to perform BIO annotation on the video subtitle text of each course in the MOOC corpus based on the aforementioned probability distribution. BIO annotation is performed on sequential single characters, where B is the starting character position of the concept, I is the other character positions of the concept, and O is the annotation result of non-concepts. The MOOC course concept annotation model construction module is used to construct a MOOC course concept annotation model based on each video subtitle text and its BIO annotation, which minimizes concept noise and concept missingness in the BIO annotation of the video subtitle text. The course concept extraction module is used to extract course concepts from the video subtitle text of the target course using the MOOC course concept annotation model. Each course includes video subtitle text and subject area annotation; based on the aforementioned probability distribution, BIO annotation is performed on the video subtitle text of each course in the MOOC corpus, including: For each course's video subtitle text, perform part-of-speech tagging and regular expression matching to merge non-noun tags of 0 and adjacent nouns with the same part of speech into an untagged noun phrase, generating the initial tagged text for each course's video subtitle text; Unlabeled noun phrases in the initial labeled text that meet the first condition are labeled with BI, and unlabeled noun phrases that do not meet the first condition are labeled with O, so as to achieve BIO labeling of the video subtitle text of each course. The first condition is that the unlabeled noun phrase is a concept in the vocabulary, and the intersection of the K subject areas with the highest probability of the unlabeled noun phrase in different subject areas with the subject area label of each course is not empty.

Images