Mutual knowledge transfer method for cross-domain subjective recognition

Abstract

The invention discloses a reciprocal knowledge transfer method for cross-domain subjective recognition, and relates to the technical field of knowledge transfer. The method comprises the steps of constructing a heterogeneity teacher model, constructing a multi-task student model and introducing a reciprocity teaching controller RTC, a static fusion strategy in a traditional method is abandoned through the heterogeneity teacher model, and meanwhile, a confidence evaluation network CAN is configured through the multi-task student model; in the training process, the reliability of guidance of a teacher for each sample is learned and evaluated, and a dynamic weight is given to distillation loss according to the reliability, so that students can selectively and critically absorb knowledge, potential noise or uncertainty guidance of the teacher is effectively filtered, the learning efficiency and robustness are improved, and the learning efficiency and robustness are improved. And meanwhile, by introducing a reciprocal teaching controller RTC, a reciprocal learning closed loop that the students feed back to the teachers is constructed, a two-way information flow between teaching and learning is established, and the accuracy and adaptability of knowledge transmission are greatly improved.

Description

Technical Field

[0001] This invention belongs to the field of knowledge transfer technology, and in particular relates to a reciprocal knowledge transfer method for cross-domain subjectivity identification. Background Technology

[0002] Aspect-Based Sentiment Analysis (ABSA), as a core technology of fine-grained sentiment computing, aims to extract users' sentiment towards specific product attributes or service dimensions from unstructured text, providing accurate decision support for intelligent recommendation, quality control, and market insights. However, existing ABSA systems generally suffer from an underestimated but crucial technical deficiency:

[0003] The lack of effective ability to discern the subjectivity of text means that a significant proportion of sentences in real user review data only contain objective factual descriptions, such as "This smartphone is equipped with a Snapdragon 888 processor" or "The restaurant is located in the downtown business district." These are essentially neutral information statements that do not carry any emotional signals that can be mined. If the sentiment analysis system cannot accurately identify and filter such objective content, it will lead to the introduction of a large amount of noisy data, thereby weakening the model's accuracy in discerning texts that truly contain emotional information.

[0004] Building a high-performance detection system capable of stable operation across multiple domains still faces the following technical challenges:

[0005] 1. Domain distribution differences and model generalization bottlenecks: Different application domains exhibit significant heterogeneity in vocabulary distribution, syntactic structure, expression conventions and semantic features. This cross-domain semantic drift phenomenon leads to severe performance degradation when models trained in a single domain are directly transferred to the target domain, which greatly limits the practical deployment value of the model.

[0006] 2. Task asymmetry dilemma and scarcity of annotation resources: The ideal solution to the domain adaptation problem is to provide sufficient subjective annotation data for each target domain. However, subjective annotation is highly subjective and complex, requiring annotators to have profound linguistic literacy, which leads to high annotation costs. Meanwhile, precise annotations specifically for subjective annotation are extremely scarce. This imbalance in resource allocation forms a typical "task asymmetry" dilemma. How to achieve effective cross-domain knowledge transfer under such constraints has become a key technical bottleneck that urgently needs to be overcome.

[0007] 3. Inherent limitations of the traditional one-way knowledge transfer paradigm: Most existing knowledge distillation and domain adaptation methods follow a one-way "teacher-student" architecture, that is, a pre-trained teacher model transfers knowledge unilaterally to the target student model. This static teaching paradigm has two fundamental flaws:

[0008] First, any single teacher model inevitably carries certain inductive biases and knowledge gaps, which may convey biased or erroneous information to students.

[0009] Secondly, rigid teaching strategies cannot adapt to the dynamic changes in the learning state of student models during the training process, and may even impair the final performance of students due to negative transfer effects.

[0010] An ideal knowledge transfer system should support two-way interaction and dynamic adjustment between teachers and students to achieve truly reciprocal learning. To address this issue, we designed a reciprocal knowledge transfer method for cross-domain subjectivity identification. Summary of the Invention

[0011] To address the aforementioned technical problems, this invention provides a reciprocal knowledge transfer method for cross-domain subjectivity identification, comprising the following steps:

[0012] S1. Construct a cross-domain data system, collect source domain data and target domain data, and perform text preprocessing;

[0013] S2. Construct a heterogeneous teacher model, a team of experts consisting of multiple classifiers with different inductive biases, responsible for generating high-quality subjective pseudo-labels for unlabeled target domain data;

[0014] S3. Construct a multi-task student model consisting of a shared encoder from RoBERTa and two independent classification heads, while learning subjectivity from teacher guidance and emotional polarity from the true labels of the target domain.

[0015] S4. Introduce the Reciprocal Teaching Controller (RTC), a GRU network that acts as a meta-learner. After each training cycle, it "observes" the students' macro-learning status and dynamically adjusts the teacher's internal teaching strategies to better adapt to the students' learning needs.

[0016] The entire learning process unfolds iteratively in epoch i, with each epoch following a three-stage process:

[0017] (1) Guidance: The teacher provides data for the target domain. Generate subjective soft labels t (i) ;

[0018] (2) Learning: Students are trained on mixed data while optimizing subjective distillation loss and polarity classification loss;

[0019] (3) Feedback: At the end of the cycle, the students' overall learning outcomes are quantified as feedback signals f. (i) This information is then transmitted to the RTC for updating the teacher's strategy before the start of the next cycle.

[0020] The source domain data contains a sufficient amount of labeled subjective samples, such as product reviews, while the target domain data includes low-labeled or unlabeled subjective samples, such as film reviews.

[0021] The measures taken in S1 to perform text preprocessing include the following steps:

[0022] S11. Text cleaning operation: remove special symbols, stop words and meaningless short sentences, wherein the length of the short sentences is ≤2 characters;

[0023] S12. Word segmentation and semantic standardization processing: domain-adaptive annotation is performed on synonyms and ambiguous sentences. For example, synonyms such as "easy to use" and "easy to use" are subjective in the product domain and objective in the news domain.

[0024] The heterogeneous teacher model integrates judgments from multiple heterogeneous sources. It consists of a dictionary-rule-based classifier, a pre-trained language model-based classifier, and an implication-based classifier. Its outputs are aggregated using DS evidence theory under a dynamically weighted scheme. The dictionary-rule-based classifier, the pre-trained language model-based classifier, and the implication-based classifier are only used on source domain data. Pre-training is performed on the target domain to extract expert knowledge about subjectivity, and the aggregated output is used for the target domain. Generate pseudo tags.

[0025] The dictionary-rule-based classifier, the pre-trained language model-based classifier, and the entailment-based classifier are three independent classifiers.

[0026] The dictionary-rule-based classifier:

[0027] Using the MPQA subjective dictionary, we identify explicit subjective expressions in a given text and apply a rule set adapted from Wiebe et al. to analyze their contextual usage, while also considering phenomena such as negation and intensifiers. This outputs a probability distribution P covering the {subjective, objective} categories for each aspect of the text. lex This provides stable, high-precision anchor points for heterogeneous teacher models;

[0028] The classifier based on the pre-trained language model:

[0029] For a sentence-aspect pair (x, a), construct the input sequence "[CLS]x[SEP]a[SEP]", input the sequence into the encoder, and select the hidden state h of [CLS]. c As a text feature representation, the subjective and objective classification probabilities P are generated after processing by a two-layer feedforward network. plm :

[0030] Pplm =softmax(W2(ReLU(W1·h)) c +b1))+b2)

[0031] Where W1, b1, W2, and b2 are all learnable parameters;

[0032] The classifier based on implication reasoning:

[0033] For each sentence-aspect pair (s, a), construct two opposing hypotheses:

[0034] h subj "The evaluation of 'a' is subjective."

[0035] h obj "The evaluation of A is objective";

[0036] Taking the original sentence s as a premise, we pair it with the two hypotheses respectively to form two input sequences:

[0037] [CLS]s[SEP]h subj [SEP] and [CLS]s[SEP]h obj [SEP];

[0038] These two sequences are fed into a pre-trained NLI model to obtain the probabilities p(entail|s,h) of the "implication" relationship. subj ) and p(entail|s,h obj These two probabilities, after normalization, constitute the final output P of the classifier. nli =p subj p obj .

[0039] Evidence convergence for the DS theory:

[0040] In DS theory, the probability output of each classifier is transformed into a basic probability assignment function, or quality function m(·), which assigns confidence to all possible subsets of the identification frame Θ = Subj, Obj, including the uncertainty term m(Subj, Obj);

[0041] Introduce a set of dynamic weights To adjust the influence of each evidence source in the i-th training cycle;

[0042] First, the quality function m of each classifier k Through weight Offer discounts;

[0043] Then, using Dempster's combination rules Iterative fusion of the discounted evidence:

[0044]

[0045] Finally, the aggregated quality function is transformed back into a probability distribution, forming the final soft label t that the teacher provides to the student in that period. (i) .

[0046] The multi-task student model in S3 includes the following:

[0047] a. Domain-adaptive shared representation learning

[0048] A domain adversarial training mechanism is introduced to enable the shared encoder to learn domain-invariant feature representations. This mechanism is achieved by introducing a domain discriminator and a gradient inversion layer. The domain discriminator is a multilayer perceptron attached to the output layer of the shared encoder.

[0049] GRL acts as an identity mapping during forward propagation, and during backpropagation, it passes the gradient from the discriminator multiplied by a negative constant to the encoder, along with the corresponding domain adversarial loss. The overall optimization objective incorporated into the student model;

[0050] b. Selective knowledge assimilation mechanism

[0051] Design a confidence evaluation network (CAN), whereby CAN acts as a lightweight module to generate confidence weights w for each target domain sample. conf ∈(0,1), the confidence weight directly adjusts the intensity of the knowledge distillation loss, thereby achieving selective knowledge assimilation:

[0052] f conf =[t;P subj ;|tP subj |;H(s subj )]

[0053] The input to the CAN is a carefully constructed concatenated feature vector f. conf This comprehensively depicts the learning context of the current sample;

[0054] Where t represents the teacher's soft label, and P subj For students' own subjective predictions, |ts subj The degree of current disagreement between teachers and students is quantified, with H(·) representing Shannon entropy. Through a two-layer MLP structure, the CAN maps the feature vector to a confidence scalar w. conf ;

[0055] The CAN uses source domain data with subjective, truthful labels. Training was conducted on [the subject].

[0056] Build the proxy task in the source domain:

[0057] When the teacher's prediction t matches the true label y s When consistent, w conf The target value is 1, otherwise it is 0;

[0058] c. Unify optimization goals

[0059] Multi-task student model and CAN module via source domain With the target domain The end-to-end training of mixed batches is optimized, and the overall optimization objective consists of source domain loss, target domain loss, and domain adversarial loss.

[0060] For source domain samples The loss function includes a subjective classification loss supervised by real labels and a binary cross-entropy loss from CAN training:

[0061]

[0062] Among them, P subj Output the student's subjective classification header, l CE Let λ represent the cross-entropy function. c To balance the hyperparameters;

[0063] For target domain samples loss function Fusion from true polarity label y t The supervisory signal and the distillation signal of the teacher pseudo-label t:

[0064]

[0065] The first term is the standard polarity classification loss, where P pola The output is a polarity classification header for students, and the second term is the subjective knowledge distillation loss, using KL divergence D. KL Measuring the difference in predicted distributions among teachers and students;

[0066] The loss term is generated by the confidence weight w using CAN. conf Modulation:

[0067] When CAN determines that the teacher's guidance is unreliable, i.e., w conf When the number of students approaches zero, they primarily rely on self-directed learning; conversely, when the number of students approaches zero, they fully absorb the knowledge taught by the teacher.

[0068] Finally, the total loss function for the student model is:

[0069]

[0070] Where, λ adv Hyperparameters are used to balance various losses.

[0071] The reciprocal teaching controller in S4 includes the following:

[0072] ① Definition of student feedback signals

[0073] We designed the Knowledge Assimilation Index (KAI) to measure learning quality and the Cognitive Effort (CE) to quantify learning costs.

[0074] After each training epoch i, the multi-task student model updates its parameters. For the target dataset An evaluation is performed to generate a structured feedback vector F. (i) The structured feedback vector provides a macro-level overview of the learning experience, reflecting both the quality of knowledge assimilation and the degree of cognitive engagement.

[0075] F (i) =[KAI (i) CE (i) ]

[0076] The knowledge assimilation index (KAI) assesses the quality of students' posterior knowledge state; therefore, KAI is defined as the product of students' average confidence score (C) and teacher-student consensus score (S).

[0077] KAI (i) =C (i) ·S (i)

[0078] in:

[0079]

[0080] Where sim(·,·) is the similarity function between two distributions, and 1-JSD(·||·) is used as its implementation, where JSD represents the Jensen-Shannon divergence, and a high KAI value indicates that students have completed the learning with high confidence and fidelity;

[0081] Cognitive effort (CE) measures the "cost" of learning by quantifying the magnitude of belief changes during the learning process. It is defined as the average KL divergence from teacher guidance to the student's final belief state.

[0082]

[0083] ② Teacher strategy adaptation based on Gated Recurrent Unit (GRU)

[0084] Before the start of the i-th training cycle, the RTC receives the feedback vector f calculated at the end of the previous cycle. (i-1) The GRU unit utilizes the new feedback f (i-1) To update its internal hidden state, from h (i-2) Updated to h (i-1):

[0085] h (i-1) =GRU(f (i-1) h (i-2) )

[0086] Among them, h (i-1) It is a compressed representation of the student's entire learning journey so far;

[0087] The updated hidden state is fed into a linear layer and activated by the Softmax function to generate aggregate weights w that guide the teaching throughout the i-th cycle. (i) :

[0088] w (i) =Softmax((W h ·h (i-1) +b h ) / τ)

[0089] Where τ is the temperature parameter.

[0090] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0091] This invention constructs a heterogeneous teacher model composed of three different paradigms: language rules, deep contextual understanding, and logical inference. It abandons the static fusion strategy in traditional methods and introduces Dempster-Shafer evidence theory. The fusion weights are dynamically generated by a meta-learning controller, so that the teacher is no longer a fixed source of knowledge, but an adaptive teaching expert who can adjust his own strategy according to the teaching effect.

[0092] At the same time, a reciprocal learning loop of students feeding back to teachers was constructed, and a two-way information flow between "teaching" and "learning" was established. After each training cycle, the macro-learning status of students was quantified by two indicators: the knowledge assimilation index (KAI) and cognitive effort (CE). These feedback signals were transmitted to the teaching controller, which directly affected the teaching strategy in the next round. This mechanism of students giving feedback to teachers is the embodiment of the "reciprocal" nature.

[0093] Meanwhile, a student learning mechanism with critical thinking was designed, equipped with a confidence evaluation network (CAN). During training, the network learns to evaluate the reliability of the teacher's guidance for each sample and assigns dynamic weights to the distillation loss accordingly. This enables students to selectively and critically absorb knowledge, effectively filtering out potential noise or uncertainty in the teacher's guidance, thereby improving the efficiency and robustness of learning.

[0094] By introducing the Teaching Strategy Controller (RTC) as a meta-learner to perceive student status and adjust teacher behavior, and working in conjunction with the Confidence Assessment Network (CAN) to empower students with critical learning abilities, a deep coupling and co-evolution of "teaching" and "learning" is achieved. This mutually beneficial mechanism not only improves the accuracy and adaptability of knowledge transfer, but also provides a new theoretical perspective and technological paradigm for cross-disciplinary learning.

[0095] The reciprocal knowledge transfer MRKT method proposed in this invention achieves significant superiority over current mainstream knowledge distillation and domain adaptation methods in all aspects by constructing a two-way interactive learning system. This fully demonstrates the superiority of the reciprocal knowledge transfer method for cross-domain subjectivity identification in solving cross-domain challenges with asymmetric tasks.

[0096] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0097] Figure 1 This is a framework diagram of the reciprocal teacher-student learning model in this invention;

[0098] Figure 2 This is a flowchart of a reciprocal knowledge transfer method for cross-domain subjectivity identification according to the present invention. Detailed Implementation

[0099] like Figures 1-2 As shown, a specific embodiment of the present invention is a reciprocal knowledge transfer method for cross-domain subjectivity identification, comprising the following steps:

[0100] S1. Construct a cross-domain data system, collect source domain data and target domain data, and perform text preprocessing;

[0101] S2. Construct a heterogeneous teacher model, a team of experts consisting of multiple classifiers with different inductive biases, responsible for generating high-quality subjective pseudo-labels for unlabeled target domain data;

[0102] S3. Construct a multi-task student model consisting of a shared encoder from RoBERTa and two independent classification heads, while learning subjectivity from teacher guidance and emotional polarity from the true labels of the target domain.

[0103] S4. Introduce the Reciprocal Teaching Controller (RTC), a GRU network that acts as a meta-learner. After each training cycle, it "observes" the students' macro-learning status and dynamically adjusts the teacher's internal teaching strategies to better adapt to the students' learning needs.

[0104] The entire learning process unfolds iteratively in epoch i, with each epoch following a three-stage process:

[0105] ⑴. Guidance: The teacher generates subjective soft labels t for the target domain data ; (i) ;

[0106] ⑵. Learning: The student trains on the mixed data while optimizing both the subjective distillation loss and the polarity classification loss;

[0107] ⑶. Feedback: After the end of the cycle, the overall learning effectiveness of the student is quantified as a feedback signal f (i) , and is transmitted to the RTC to update the teacher's strategy before the start of the next cycle.

[0108] The source domain data contains a sufficient amount of labeled subjective samples, such as product reviews, and the target domain data includes low-labeled or unlabeled subjective samples, such as movie reviews.

[0109] The measures taken for the execution of text preprocessing in S1 include the following steps:

[0110] S11. Text cleaning operation to remove special symbols, stop words, and meaningless short sentences with a short sentence length ≤ 2 characters;

[0111] S12. Word segmentation and semantic standardization processing, domain adaptation annotation for synonyms and ambiguous sentences. Synonyms such as "easy to use" and "user-friendly", and ambiguous sentences such as "This price is very worth it", which are subjective in the product domain and objective in the news domain;

[0112] By cleaning and standardizing the obtained source domain data and target domain data, it effectively avoids the influence of sentences in different domains and contexts in the data on the subjective judgment in the subsequent learning process, and effectively solves the drawback that significant heterogeneity in vocabulary distribution, syntactic structure, expression convention, and semantic features in different application domains affects subjective detection.

[0113] To generate reliable and robust supervision signals, the heterogeneous teacher model integrates judgments from multiple heterogeneous sources;

[0114] Such as Figure 1 as shown in (a), the heterogeneous teacher model consists of a lexicon-rule-based classifier, a pre-trained language model-based classifier, and an entailment reasoning-based classifier. Its output is aggregated through the DS evidence theory under a dynamic weighting scheme. The lexicon-rule-based classifier, the pre-trained language model-based classifier, and the entailment reasoning-based classifier are only pre-trained on the source domain data to extract their expert knowledge about subjectivity, and the aggregated output is used to generate pseudo-labels for the target domain .

[0115] The dictionary-rule-based classifier, the pre-trained language model-based classifier, and the entailment-based classifier are three independent classifiers.

[0116] Dictionary-rule-based classifiers:

[0117] The word typicality classifier provides basic judgments based on mature linguistic knowledge, adopts the MPQA subjective dictionary, identifies explicit subjective expressions in a given text, and applies a rule set adapted from Wiebe et al. to analyze its contextual usage, while also considering phenomena such as negation and intensification words. It outputs a probability distribution P covering {subjective, objective} categories for each aspect of the text. lex This provides stable, high-precision anchor points for heterogeneous teacher models;

[0118] Classifiers based on pre-trained language models:

[0119] For a sentence-aspect pair (x, a), construct the input sequence "[CLS]x[SEP]a[SEP]", input the sequence into the encoder, and select the hidden state h of [CLS]. c As a text feature representation, the subjective and objective classification probabilities P are generated after processing by a two-layer feedforward network. plm :

[0120] P plm =softmax(W2(ReLU(W1·h)) c +b1))+b2)

[0121] Where W1, b1, W2, and b2 are all learnable parameters;

[0122] Classifiers based on implication reasoning:

[0123] For each sentence-aspect pair (s, a), construct two opposing hypotheses:

[0124] h subj "The evaluation of 'a' is subjective."

[0125] h obj "The evaluation of A is objective";

[0126] Taking the original sentence s as a premise, we pair it with the two hypotheses respectively to form two input sequences:

[0127] [CLS]s[SEP]h subj [SEP] and [CLS]s[SEP]j obj [SEP];

[0128] These two sequences are fed into a pre-trained NLI model to obtain the probabilities p(entail|s,h) of the "implication" relationship. subj) and p(entail|s,h obj These two probabilities, after normalization, constitute the final output P of the classifier. nli =p subj p obj .

[0129] Evidence convergence for the DS theory:

[0130] Faced with the independent probability distributions P generated by the three classifiers mentioned above lex P plm P nli Simple weighted averaging ignores the conflicts and uncertainties between different sources of evidence. DS evidence theory allows confidence to be explicitly assigned to uncertainty itself, so that the influence of sources of evidence with high uncertainty is naturally weakened when aggregated. This ability to model the 'known unknown' makes the aggregation results more robust.

[0131] In DS theory, the probability output of each classifier is transformed into a basic probability assignment function, or quality function m(·), which assigns confidence to all possible subsets of the identification frame Θ = Subj, Obj, including the uncertainty term m(Subj, Obj);

[0132] Introduce a set of dynamic weights To adjust the influence of each evidence source in the i-th training cycle;

[0133] First, the quality function m of each classifier k Through weight Offer discounts;

[0134] Then, using Dempster's combination rules Iterative fusion of the discounted evidence:

[0135]

[0136] Finally, the aggregated quality function is transformed back into a probability distribution, forming the final soft label t that the teacher provides to the student in that period. (i) where the weight w (i) It is not static, but dynamically generated by the reciprocal teaching controller at the beginning of each cycle based on the student's historical learning status.

[0137] like Figure 1 As shown in (b), the multi-task student model in S3 includes the following:

[0138] a. Domain-adaptive shared representation learning

[0139] A domain adversarial training mechanism is introduced to enable the shared encoder to learn domain-invariant feature representations. This mechanism is achieved by introducing a domain discriminator and a gradient inversion layer. The domain discriminator is a multilayer perceptron attached to the output layer of the shared encoder.

[0140] In GRL, the gradient from the discriminator is multiplied by a negative constant and passed to the encoder during backpropagation. Through this adversarial optimization, the encoder's parameter update direction tends to "deceive" the discriminator, thus being motivated to generate a general representation that is effective for the main task but has ambiguous domain information. The corresponding domain adversarial loss is... The overall optimization objective incorporated into the student model;

[0141] b. Selective knowledge assimilation mechanism

[0142] Given the subjective pseudo-labels t generated by the heterogeneous teacher model (i) Inevitably containing noise, if the student model accepts all instruction without selection, it will lead to the internalization of erroneous knowledge. Therefore, a confidence evaluation network (CAN) is designed. CAN, as a lightweight module, generates confidence weights w for each target domain sample. conf ∈(0,1), the confidence weight directly adjusts the intensity of the knowledge distillation loss, achieving selective knowledge assimilation:

[0143] f conf =[t;P subj ;|tP subj |;H(s subj )]

[0144] The input to CAN is a carefully constructed concatenated feature vector f. conf This comprehensively depicts the learning context of the current sample;

[0145] Where t represents the teacher's soft label, and P subj For students' own subjective predictions, |ts subj |Quantify the current degree of disagreement between teachers and students, H(·) represents the Shannon entropy, and through a two-layer MLP structure, CAN maps the feature vector to a confidence scalar w. conf ;

[0146] CAN in source domain data with subjective truth labels Training was conducted on [the subject].

[0147] Build the proxy task in the source domain:

[0148] When the teacher's prediction t matches the true label y s When consistent, w confThe target value is 1 if the target value is 1 otherwise 0. CAN learns to distinguish between patterns of "reliable performance" and "potential errors" of teachers and generalizes this ability to the target domain.

[0149] c. Unify optimization goals

[0150] Multi-task student model and CAN module via source domain With the target domain The end-to-end training of mixed batches is optimized, and the overall optimization objective consists of source domain loss, target domain loss, and domain adversarial loss.

[0151] For source domain samples The loss function includes a subjective classification loss supervised by real labels and a binary cross-entropy loss from CAN training:

[0152]

[0153] Where, P subj Output the subjective classification header for students;

[0154] l CE Represents the cross-entropy function;

[0155] λ c To balance the hyperparameters;

[0156] For target domain samples loss function Fusion from true polarity label y t The supervisory signal and the distillation signal of the teacher pseudo-label t:

[0157]

[0158] The first term is the standard polarity classification loss, where P pola The output is a polarity classification header for students, and the second term is the subjective knowledge distillation loss, using KL divergence D. KL Measuring the difference in predicted distributions among teachers and students;

[0159] The loss term uses the confidence weights w generated by CAN. conf Modulation:

[0160] When CAN determines that the teacher's guidance is unreliable, i.e., w conf When the number of students approaches zero, they primarily rely on self-directed learning; conversely, when the number of students approaches zero, they fully absorb the knowledge taught by the teacher.

[0161] Finally, the total loss function for the student model is:

[0162]

[0163] Where, λ advHyperparameters to balance various losses;

[0164] This carefully designed unified optimization paradigm enables multi-task student models to learn prudently from teachers while addressing domain differences, laying a solid theoretical and practical foundation for achieving accurate, fine-grained sentiment analysis.

[0165] The reciprocal teaching controller in S4 includes the following:

[0166] like Figure 1 As shown in (c), the Reciprocal Teaching Controller (RTC) constructs a key feedback loop, which upgrades this framework from simple knowledge distillation to a truly interactive learning system. The Reciprocal Teaching Controller can dynamically adjust teacher strategies based on a quantitative understanding of students' learning status.

[0167] ① Definition of student feedback signals

[0168] To enable RTC to make informed decisions, feedback signals need to characterize students' learning status from two orthogonal dimensions: the 'quality' of learning and the 'cost' of learning. Quality refers to the degree to which students internalize knowledge, while cost reflects the cognitive effort students put into achieving the current state. To this end, the Knowledge Assimilation Index (KAI) was designed to measure learning quality, and the Cognitive Effort (CE) was designed to quantify learning cost.

[0169] After each training epoch i, the multi-task student model updates its parameters. For the target dataset An evaluation is performed to generate a structured feedback vector F. (i) Structured feedback vectors provide a macro-level overview of the learning experience, reflecting both the quality of knowledge assimilation and the degree of cognitive engagement.

[0170] F (i) =[KAI (i) CE (i) ]

[0171] The knowledge assimilation index (KAI) assesses the quality of students' posterior knowledge state. Successful knowledge transfer should enable students to achieve both internal and external consistency in their knowledge state. Therefore, KAI is defined as the product of the student's average confidence score (C) and the teacher-student consensus score (S).

[0172] KAI (i) =C (i) ·S (i)

[0173] in:

[0174]

[0175] Where sim(·,·) is the similarity function between two distributions, and 1-JSD(·||·) is used as its implementation, where JSD represents the Jensen-Shannon divergence, and a high KAI value indicates that students have completed the learning with high confidence and fidelity;

[0176] Cognitive effort (CE) measures the "cost" of learning by quantifying the magnitude of belief changes during the learning process. It is defined as the average KL divergence from teacher guidance to the student's final belief state.

[0177]

[0178] A high cognitive effort value indicates that the teacher's guidance is highly inspiring or "unexpected" for the student, forcing the student to make significant conceptual shifts. While this may indicate effective learning, a low KAI value may also suggest that the student is confused.

[0179] ② Teacher strategy adaptation based on Gated Recurrent Unit (GRU)

[0180] The core mission of RTC is to integrate students' historical learning feedback sequences. Transformed into the teacher aggregate weight w in the i-th training cycle (i) The system employs a gated recurrent unit (GRU) network as the core of the controller. The GRU can capture the temporal dependencies in the learning trajectory, so that the evolution of the teacher's strategy depends not only on the performance in the previous round, but on the entire learning history.

[0181] Before the start of the i-th training cycle, the RTC receives the feedback vector f calculated at the end of the previous cycle. (i-1) The GRU unit utilizes the new feedback f (i-1) To update its internal hidden state, from h (i-2) Updated to h (i-1) :

[0182] h (i-1) =GRU(f (i-1) h (i-2) )

[0183] Among them, h (i-1) It is a compressed representation of the student's entire learning journey so far;

[0184] The updated hidden state is fed into a linear layer and activated by the Softmax function to generate aggregate weights w that guide the teaching throughout the i-th cycle. (i) :

[0185] w (i) =Softmax((W h ·h (i-1) +b h ) / τ)

[0186] Where τ is the temperature parameter;

[0187] These newly generated weights w (i) Subsequently used by the teacher's DS evidence aggregator in all batches of the i-th cycle, this closed-loop mechanism—where the student's macro-learning state directly influences the teacher's micro-teaching strategy—is key to achieving intelligent and adaptive knowledge transfer, embodying the reciprocity of reciprocal knowledge transfer. This closed-loop mechanism enables the student's macro-learning state to directly influence the teacher's subsequent teaching strategies, reflecting the principle of interactive teaching, thereby promoting a more intelligent and adaptive knowledge transfer process.

[0188] The REINFORCE algorithm is used to optimize the parameters of RTC, with the goal of maximizing the expected cumulative reward. At the end of each period i, the calculated reward r is used to optimize the RTC parameters. i =KAI (i) -KAI (i-1) To update the parameters of the RTC, the direction of gradient update will result in more effective teacher weights.

[0189] Example:

[0190] Simulating real-world scenarios: Subjective labels can be obtained in a general domain (source domain), but the ultimate task is to perform sophisticated sentiment analysis in a specific domain (target domain) where only polarity labels exist;

[0191] Source domain: Subjective dataset used:

[0192] MPQA: This dataset, created by Wiebe et al., contains a large number of sentences extracted from news articles around the world. Each sentence is labeled as subjective or objective, using its v2.0 version and labeled with aspects using commercial annotation tools.

[0193] OpinionFinder: Released by the University of Pennsylvania, it is also one of the benchmark datasets for subjectivity analysis. It contains various news and commentary texts and provides rich annotations of opinion expressions, including sentence-level subjective and objective classifications. It also uses commercial annotation tools to annotate aspects.

[0194] Target domains: Benchmark datasets from three domains—SemEval-2014 Task 4, SemEval-2015 Task 12, emEval-2016 Task 5, and SentiHood—were used. ATSA and ACSA were treated as binary classification tasks. Neutral instances were ignored in all datasets, and aspect polarity was labeled as positive or negative. A high-quality manually labeled test set, FG-ABSA-Test, was constructed, with samples drawn from the official test sets of five widely used ABSA benchmarks to ensure domain diversity.

[0195] These datasets include: SemEval-2014 Task 4 (Restaurant and Laptop domains), SemEval-2015 Task 12 (Restaurant domains), SemEval-2016 Task 5 (Restaurant domains), and SentiHood (Community domains);

[0196] To build a sufficiently large annotation pool, 240 "sentence-aspect word" pairs were randomly selected from each dataset, totaling 1200 instances. The label definitions are shown in Table 1 below:

[0197] Table 1. Label Definitions for FG-ABSA-Test

[0198]

[0199] Three graduate students proficient in natural language processing were recruited as annotators and received standardized training.

[0200] To ensure the independence of the annotation process, the entire process was conducted in an isolated environment;

[0201] After the first round of annotation was completed, the inter-annotator agreement (IAA) was calculated. Using the Fleiss' Kappa index, κ = 0.87 was obtained. This value is usually interpreted as "near-perfect consistency", which strongly demonstrates the clarity of the annotation criteria and the high reliability of the results.

[0202] For instances with differing opinions, a majority voting strategy is used to determine the final label. If all three parties disagree, the sample is considered ambiguous and removed from the dataset. After filtering, the FG-ABSA-Test dataset containing 1064 high-quality samples was finally constructed. The label distribution of this dataset is shown in Table 2.

[0203] Table 2. Label distribution of the FG-ABSA-Test dataset

[0204] Fine-grained labels quantity Percentage (%) Subjective - Positive 352 32.8％ Subjective-Negative 298 27.7％ Objective - Positive 221 20.6％ Objective - Negative 203 18.9％

[0205] Use the default training and validation data split, and use accuracy and macro F1 metrics to measure performance;

[0206] It was compared with five diverse and representative benchmark models:

[0207] RoBERTa-finetune: A standard RoBERTa-base model that is pre-trained subjectively in the source domain and then fine-tuned only in the target domain for polarity classification. This base model does not produce fine labels and serves as a lower bound benchmark.

[0208] DANN: A leading domain adaptation method that introduces a domain discriminator to learn domain-insensitive features. This embodiment adapts it to transfer objective knowledge from the source domain to the target domain.

[0209] CDAN: An extension of DANN, which associates adversarial predictions with classifier features, providing more refined adaptability;

[0210] Vanilla KD: A classic knowledge distillation implementation scheme. The heterogeneous teacher model generates flexible subjective labels for the target domain, and the multi-task student model uses these labels as the distillation target in the KL divergence loss, while also using them for polarity classification loss.

[0211] TinyBERT: An advanced knowledge distillation method designed for Transformer models; this embodiment adapts it to the framework of this paper, distilling subjective teacher knowledge in addition to the prediction layer, as well as the attention layer and hidden layer, and imparting it to the student RoBERTa;

[0212] All RoBERTa-based models were initialized using the robotic-base version to ensure the reliability of the experimental data.

[0213] This embodiment uses the AdamW optimizer, with the learning rate linearly warmed up in the first 10% of training steps and then linearly decayed thereafter.

[0214] Unless otherwise specified, the batch size for all models is set to 32, and the maximum sequence length is 80.

[0215] For the MRKT framework in this embodiment, the learning rate of the RoBERTa encoder of the multi-task student model is 2e-5, and the hidden layer dimension of the teaching policy controller GRU is 128 with a learning rate of 1e-4.

[0216] The confidence evaluation network is a three-layer MLP with a learning rate of 1e-4 and a weight decay of 0.01.

[0217] For all benchmark models, this embodiment follows the recommended parameters in their original papers or is tuned on the validation set via grid search.

[0218] The reciprocal knowledge transfer method proposed in this embodiment effectively improves the performance of downstream sentiment polarity classification tasks.

[0219] Table 3 Comparison of Accuracy Rates for Fine Emotion Classification Tasks

[0220] Model Name RES14 RES15 RES16 LAP14 LAP15 LAP16 RoBERTa-finetune 92.74％ 90.63％ 93.25％ 86.89％ 91.69％ 93.71％ VanillaKD 92.34％ 89.75％ 91.76％ 86.84％ 90.68％ 90.94％ DANN 91.37％ 88.63％ 92.64％ 86.34％ 90.34％ 90.63％ CDAN 90.98％ 87.79％ 91.96％ 85.93％ 89.69％ 91.79％ TinyBERT 92.76％ 89.94％ 93.47％ 87.35％ 91.74％ 89.73％ MRKT(ours) 93.05％ 91.84％ 95.54％ 88.24％ 91.31％ 93.72％

[0221] Table 4 Comparison of Macro F1 Scores for Emotion Fine Classification Tasks

[0222] Model Name RES14 RES15 RES16 LAP14 LAP15 LAP16 RoBERTa-finetune 90.84％ 89.63％ 92.25％ 86.21％ 90.13％ 90.41％ VanillaKD 90.31％ 87.07％ 90.79％ 85.77％ 87.68％ 90.71％ DANN 90.12％ 88.31％ 89.34％ 85.34％ 90.31％ 89.51％ CDAN 89.48％ 87.56％ 90.02％ 85.53％ 89.69％ 90.19％ TinyBERT 91.63％ 88.04％ 92.12％ 86.05％ 89.71％ 89.52％ MRKT (the model in this paper) 91.93％ 90.44％ 93.86％ 88.13％ 90.07％ 90.51％

[0223] The results in Tables 3 and 4 show that:

[0224] Compared to the strongest baseline model TinyBERT, MRKT improves accuracy by an average of 1.45% and F1 score by 1.65% across all datasets.

[0225] It is worth noting that while static distillation methods such as Vanilla KD and TinyBERT are superior to simple fine-tuning benchmarks, their performance is still inferior to MRKT. The above data effectively demonstrates the superiority of dynamic teaching strategies.

[0226] MRKT can adjust the teaching focus according to the students' learning status, avoiding the static, one-size-fits-all knowledge transmission by teachers;

[0227] Furthermore, compared with domain-adaptive methods such as DANN / CDAN, the heterogeneous teacher model and the multi-task student model also show significant advantages, indicating that the application explicitly models and transmits the core knowledge of 'subjectivity and objectivity', which is more effective than implicit methods that learn domain-invariant features.

[0228] Table 5 shows the fine-grained four-class classification performance (Macro P / R / F1) on the manually labeled set FG-ABSA-Test.

[0229] Model Macro-P Macro-R MacroF1 RoBERTa-finetune 88.92％ 88.26％ 87.95％ VanillaKD 89.13％ 89.32％ 88.35％ TinyBERT-RoBERTa 90.61％ 90.35％ 89.43％ DANN 88.25％ 88.51％ 87.87％ CDAN 89.31％ 89.17％ 88.02％ MRKT(ours) 91.46％ 90.58％ 90.39％

[0230] To directly evaluate the model's ability to generate fine-grained labels, this embodiment conducted a four-class classification test on the manually labeled FG-ABSA-Test, and the results are shown in Table 5.

[0231] MRKT significantly outperformed other baseline models across all metrics, achieving a Macro F1-Score of 90.39%, nearly one percentage point higher than the second-best TinyBERT.

[0232] This directly proves that the reciprocal knowledge transfer method can not only improve the performance of downstream tasks, but also effectively decouple and integrate the subjective and objective knowledge of the source domain with the polar knowledge of the target domain, generating high-quality fine-grained sentiment tags.

[0233] Ablation Research

[0234] To analyze the MRKT model presented in this paper and verify the contribution of each key component, an in-depth ablation study was conducted. The following variations were designed in this embodiment:

[0235] MRKT w / o GRU: Using a static weighted and averaged strategy instead of a dynamic GRU controller to aggregate teacher evidence, which is equivalent to non-adaptive teaching;

[0236] MRKT w / o CAN: Remove the trust evaluation network and apply the subjective distillation loss uniformly to all samples, thereby eliminating filtering through discriminatory learning;

[0237] MRKT w / o KAI&CE: By removing macroscopic feedback signals, the GRU receives no input and therefore cannot adjust its strategy based on the learning history;

[0238] Table 6 Ablation Experiment Results

[0239] Model variants Macro-P Macro-R MacroF1 MRKT 91.46％ 90.58％ 90.39％ MRKT w / o GRU 90.39％ 90.02％ 89.85％ MRKTw / o CAN 90.43％ 89.87％ 89.79％ MRKT w / o KAI&CE 89.96％ 89.41％ 89.41％

[0240] Removing the GRU controller caused the Macro-F1 score to drop by 0.54 percentage points, indicating that while static teaching still has some effect, dynamically adjusting teaching strategies based on student feedback can bring significant performance gains, confirming the value of adaptive teaching.

[0241] Removing the CAN network caused the Macro-F1 score to drop by 0.60 percentage points, a drop even slightly greater than that caused by removing the GRU. This highlights the importance of enabling students to develop critical learning skills and filter out guidance that is highly inconsistent with their own understanding or that teachers lack confidence in. This is crucial for preventing the mistransmission of knowledge and improving learning efficiency.

[0242] Removing the feedback signal led to the most severe performance degradation, with the Macro-F1 dropping by 0.98 percentage points. This strongly demonstrates that 'reciprocity' is the cornerstone of the entire framework. Without student learning feedback, the teacher's dynamic adjustments become meaningless, the entire teaching and learning loop is broken, and the model degenerates into a poorly performing static distillation system.

[0243] Qualitative Analysis and Case Studies

[0244]

[0245]

[0246] In the above case, the sentence contains both factual descriptions and personalized assertions. The baseline model was likely misled by the latter, classifying the entire statement as subjective. However, MRKT correctly identified its objective basis. This indicates that our model successfully decoupled subjective and objective knowledge from emotional polarity knowledge during the distillation process, which is precisely the key role played by the CAN network.

[0247] During training, for samples containing strong objective facts, the teacher's guidance signals are more certain, and CAN will assign them greater learning weight, thereby strengthening the student model's ability to identify objective facts.

[0248] This embodiment presents a comprehensive experimental evaluation of the proposed reciprocal knowledge transfer method. The results fully demonstrate that the method outperforms a series of powerful basic models in three different target domains, and the overall performance comparison confirms the effectiveness of MRKT.

[0249] Furthermore, the ablation study validated the necessity of each innovative component, such as the adaptive GRU controller and the confidence evaluation network;

[0250] In summary, these empirical results strongly support the assertion that the meta-learning-based bidirectional knowledge transfer framework is an effective method for conducting refined and interpretable sentiment analysis.

[0251] This invention introduces mature languages ​​with different logics from different domains through a heterogeneous teacher model, forming a database that facilitates subjective judgment in multiple domains and contexts. It solves the problem that the significant heterogeneity in vocabulary distribution, syntactic structure, expression conventions and semantic features in different application domains leads to serious performance degradation when a single-domain trained model is directly transferred to the target domain, which greatly limits the actual deployment of the model and affects subjective judgment. Combined with the selective and critical learning of the multi-task student model, it greatly improves the efficiency and robustness of learning.

[0252] Simultaneously, a reciprocal teaching controller is introduced. Two indicators, the knowledge assimilation index (KAI) and cognitive effort (CE), are designed to quantify the students' macro-learning status. By feeding the signals of the knowledge assimilation index (KAI) and cognitive effort (CE) back to the teaching controller, a mechanism for students to provide feedback to teachers is generated, realizing the goal of reciprocal learning. This solves the problems of existing knowledge distillation and domain adaptation methods, which mostly follow a one-way "teacher-student" relationship and can only conduct one-sided knowledge transfer. This inevitably carries specific inductive biases and knowledge blind spots, and may convey biased or erroneous information to students. At the same time, the fixed teaching strategies cannot adapt to the dynamic learning status changes of student models during the training process, and may even damage the students' final performance due to negative transfer effects.

[0253] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0254] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A reciprocal knowledge transfer method for cross-domain subjectivity identification, characterized in that, Includes the following steps: S1. Construct a cross-domain data system, collect source domain data and target domain data, and perform text preprocessing; S2. Construct a heterogeneous teacher model, a team of experts consisting of multiple classifiers with different inductive biases, responsible for generating high-quality subjective pseudo-labels for unlabeled target domain data; S3. Construct a multi-task student model consisting of a shared encoder from RoBERTa and two independent classification heads, while learning subjectivity from teacher guidance and emotional polarity from the true labels of the target domain. S4. Introduce the Reciprocal Teaching Controller (RTC), a GRU network that acts as a meta-learner. After each training cycle, it "observes" the students' macro-learning status and dynamically adjusts the teacher's internal teaching strategies to better adapt to the students' learning needs. The entire learning process unfolds iteratively in the form of epochi cycles, with each cycle following a three-stage process: (1) Guidance: The teacher provides data for the target domain. Generate subjective soft labels t (i) ; (2) Learning: Students are trained on mixed data while optimizing subjective distillation loss and polarity classification loss; (3) Feedback: At the end of the cycle, the students' overall learning outcomes are quantified as a feedback signal f. (i) This information is then transmitted to the RTC for updating the teacher's strategy before the start of the next cycle.

2. The reciprocal knowledge transfer method for cross-domain subjectivity identification according to claim 1, characterized in that, The source domain data contains a sufficient amount of labeled subjective samples, such as product reviews, while the target domain data includes low-labeled or unlabeled subjective samples, such as film reviews.

3. The reciprocal knowledge transfer method for cross-domain subjectivity identification according to claim 1, characterized in that, The measures taken in S1 to perform text preprocessing include the following steps: S11. Text cleaning operation: remove special symbols, stop words and meaningless short sentences, wherein the length of the short sentences is ≤2 characters; S12. Word segmentation and semantic standardization processing: domain-adaptive annotation is performed on synonyms and ambiguous sentences. The synonyms are such as "easy to use" and "easy to use", and the ambiguous sentences are such as "this price is very worthwhile". In the product domain, it is subjective, and in the news domain, it is objective.

4. The reciprocal knowledge transfer method for cross-domain subjectivity identification according to claim 1, characterized in that, The heterogeneous teacher model integrates judgments from multiple heterogeneous sources. It consists of a dictionary-rule-based classifier, a pre-trained language model-based classifier, and an implication-based classifier. Its outputs are aggregated using DS evidence theory under a dynamically weighted scheme. The dictionary-rule-based classifier, the pre-trained language model-based classifier, and the implication-based classifier are only used on source domain data. Pre-training is performed on the target domain to extract expert knowledge about subjectivity, and the aggregated output is used for the target domain. Generate pseudo tags.

5. The reciprocal knowledge transfer method for cross-domain subjectivity identification according to claim 1, characterized in that, The dictionary-rule-based classifier, the pre-trained language model-based classifier, and the entailment-based classifier are three independent classifiers. The dictionary-rule-based classifier: Using the MPQA subjective dictionary, we identify explicit subjective expressions in a given text and apply a rule set adapted from Wiebe et al. to analyze their contextual usage, while also considering phenomena such as negation and intensifiers. This outputs a probability distribution P covering the {subjective, objective} categories for each aspect of the text. lex This provides stable, high-precision anchor points for heterogeneous teacher models; The classifier based on the pre-trained language model: For a sentence-aspect pair (x, a), construct the input sequence "[CLS]x[SEP]a[SEP]", input the sequence into the encoder, and select the hidden state h of [CLS]. c As a text feature representation, the subjective and objective classification probabilities P are generated after processing by a two-layer feedforward network. plm : P plm =softmax(W2(ReLU(W1·h c +b1))+b2) Where W1, b1, W2, and b2 are all learnable parameters; The classifier based on implication reasoning: For each sentence-aspect pair (s, a), construct two opposing hypotheses: h subj "The evaluation of 'a' is subjective." h obj "The evaluation of A is objective"; Taking the original sentence s as a premise, we pair it with the two hypotheses respectively to form two input sequences: [CLS]s[SEP]h subj [SEP]and[CLS]s[SEP]h obj [SEP]; These two sequences are fed into a pre-trained NLI model to obtain the probabilities p(entail|s,h) of the "implication" relationship. subj ) and p(entail|s,h obj These two probabilities, after normalization, constitute the final output P of the classifier. nli =p subj p obj .

6. The reciprocal knowledge transfer method for cross-domain subjectivity identification according to claim 1, characterized in that, Evidence convergence for the DS theory: In DS theory, the probability output of each classifier is transformed into a basic probability assignment function, or quality function m(·), which assigns confidence to all possible subsets of the identification frame Θ = Subj, Obj, including the uncertainty term m(Subj, Obj); Introduce a set of dynamic weights To adjust the influence of each evidence source in the i-th training cycle; First, the quality function m of each classifier k Through weight Offer discounts; Then, using Dempster's combination rules Iterative fusion of the discounted evidence: Finally, the aggregated quality function is transformed back into a probability distribution, forming the final soft label t that the teacher provides to the student in that period. (i) .

7. The reciprocal knowledge transfer method for cross-domain subjectivity identification according to claim 1, characterized in that, The multi-task student model in S3 includes the following: a. Domain-adaptive shared representation learning A domain adversarial training mechanism is introduced to enable the shared encoder to learn domain-invariant feature representations. This mechanism is achieved by introducing a domain discriminator and a gradient inversion layer. The domain discriminator is a multilayer perceptron attached to the output layer of the shared encoder. GRL acts as an identity mapping during forward propagation, and during backpropagation, it passes the gradient from the discriminator multiplied by a negative constant to the encoder, along with the corresponding domain adversarial loss. The overall optimization objective incorporated into the student model; b. Selective knowledge assimilation mechanism Design a confidence evaluation network (CAN), whereby CAN acts as a lightweight module to generate confidence weights w for each target domain sample. conf ∈(0,1), the confidence weight directly adjusts the intensity of the knowledge distillation loss, thereby achieving selective knowledge assimilation: f conf =[t;P subj ;|t-P subj |;H(s subj )] The input to the CAN is a carefully constructed concatenated feature vector f. conf This comprehensively depicts the learning context of the current sample; Where t represents the teacher's soft label, and P subj For students' own subjective predictions, |ts subj The degree of current disagreement between teachers and students is quantified, with H(·) representing Shannon entropy. Through a two-layer MLP structure, the CAN maps the feature vector to a confidence scalar w. conf ; The CAN uses source domain data with subjective, truthful labels. Training was conducted on [the subject]. Build the proxy task in the source domain: When the teacher's prediction t matches the true label y s When consistent, w conf The target value is 1, otherwise it is 0; c. Unify optimization goals Multi-task student model and CAN module via source domain With the target domain The end-to-end training of mixed batches is optimized, and the overall optimization objective consists of source domain loss, target domain loss, and domain adversarial loss. For source domain samples The loss function includes a subjective classification loss supervised by real labels and a binary cross-entropy loss from CAN training: Where, P subj Output the student's subjective classification header, l CE Let λ represent the cross-entropy function. c To balance the hyperparameters; For target domain samples loss function Fusion from true polarity label y t The supervisory signal and the distillation signal of the teacher pseudo-label t: The first term is the standard polarity classification loss, where P pola The output is a polarity classification header for students, and the second term is the subjective knowledge distillation loss, using KL divergence D. KL Measuring the difference in predicted distributions among teachers and students; The loss term is generated by the confidence weight w using CAN. conf Modulation: When CAN determines that the teacher's guidance is unreliable, i.e., w conf When the number of students approaches zero, they primarily rely on self-directed learning; conversely, when the number of students approaches zero, they fully absorb the knowledge taught by the teacher. Finally, the total loss function for the student model is: Where, λ adv Hyperparameters are used to balance various losses.

8. A reciprocal knowledge transfer method for cross-domain subjectivity identification according to claim 1, characterized in that, The reciprocal teaching controller in S4 includes the following: ① Definition of student feedback signals We designed the Knowledge Assimilation Index (KAI) to measure learning quality and the Cognitive Effort (CE) to quantify learning costs. After each training session of Epochi, the multi-task student model updates its parameters. For the target dataset An evaluation is performed to generate a structured feedback vector F. (i) The structured feedback vector provides a macro-level overview of the learning experience, reflecting both the quality of knowledge assimilation and the degree of cognitive engagement. FAVORITE (i) =[YOURSELF (i) ,SAY (i) ] The knowledge assimilation index (KAI) assesses the quality of students' posterior knowledge state; therefore, KAI is defined as the product of students' average confidence score (C) and teacher-student consensus score (S). KAI (i) =C (i) ·S (i) in: Where sim(·,·) is the similarity function between two distributions, and 1-JSD(·||·) is used as its implementation, where JSD represents the Jensen-Shannon divergence, and a high KAI value indicates that students have completed the learning with high confidence and fidelity; Cognitive effort (CE) measures the "cost" of learning by quantifying the magnitude of belief changes during the learning process. It is defined as the average KL divergence from teacher guidance to the student's final belief state. ② Teacher strategy adaptation based on Gated Recurrent Unit (GRU) Before the start of the i-th training cycle, the RTC receives the feedback vector f calculated at the end of the previous cycle. (i-1) The GRU unit utilizes the new feedback f (i-1) To update its internal hidden state, from h (i-2) Updated to h (i-1) : h (i-1) =GRU(f (i-1) ,h (i-2) ) Among them, h (i-1) It is a compressed representation of the student's entire learning journey so far; The updated hidden state is fed into a linear layer and activated by the Softmax function to generate aggregate weights w that guide the teaching throughout the i-th cycle. (i) : w (i) =Softmax((W h ·h (i-1) +b h ) / t) Where τ is the temperature parameter.

Images