Text Representation Research Articles

Category Mapping of Emergency Supplies Classification Standard Based on BERT-TextCNN

In recent years, the escalation in emergency occurrences has underscored the pressing need for expedient responses in delivering essential supplies. Efficient integration and precise allocation of emergency resources under joint government–enterprise stockpiling models are pivotal for enhancing emergency response effectiveness and minimizing economic repercussions. However, current research predominantly focuses on contract coordination and cost-sharing within these joint reserve modes, overlooking significant discrepancies in emergency supply classification standards between government and enterprise sectors, as well as the asymmetry in cross-sectoral and cross-regional supply information. This oversight critically impedes the timeliness and accuracy of emergency supply responses. In practice, manual judgment has been used to match the same materials under differing classification standards between government and enterprise reserves. Still, this approach is inefficient and prone to high error rates. To mitigate these challenges, this study proposes a methodology leveraging the BERT pre-trained language model and TextCNN neural network to establish a robust mapping relationship between these classification criteria. The approach involves abstracting textual representations of both taxonomical classes, generating comparable sentence vectors via average pooling, and calculating cosine similarity scores to facilitate precise classification mapping. Illustrated with China’s Classification and Coding of Emergency Supplies standards and Global Product Classification standards, empirical validation on annotated data demonstrates the BERT-TextCNN model’s exceptional accuracy of 98.22%, surpassing other neural network methodologies such as BERT-CNN, BERT-RNN, BERT-BiLSTM, etc. This underscores the potential of advanced neural network techniques in enhancing emergency supply management across diverse sectors and regions.

An End-to-End Scene Text Recognition for Bilingual Text

Text localization and recognition from natural scene images has gained a lot of attention recently due to its crucial role in various applications, such as autonomous driving and intelligent navigation. However, two significant gaps exist in this area: (1) prior research has primarily focused on recognizing English text, whereas Arabic text has been underrepresented, and (2) most prior research has adopted separate approaches for scene text localization and recognition, as opposed to one integrated framework. To address these gaps, we propose a novel bilingual end-to-end approach that localizes and recognizes both Arabic and English text within a single natural scene image. Specifically, our approach utilizes pre-trained CNN models (ResNet and EfficientNetV2) with kernel representation for localization text and RNN models (LSTM and BiLSTM) with an attention mechanism for text recognition. In addition, the AraElectra Arabic language model was incorporated to enhance Arabic text recognition. Experimental results on the EvArest, ICDAR2017, and ICDAR2019 datasets demonstrated that our model not only achieves superior performance in recognizing horizontally oriented text but also in recognizing multi-oriented and curved Arabic and English text in natural scene images.

Rethinking Multilingual Scene Text Spotting: A Novel Benchmark and a Character-Level Feature Based Approach

End-to-end multilingual scene text spotting aims to integrate scene text detection and recognition into a unified framework. Actually, the accuracy of text recognition largely depends on the accuracy of text detection. Due to the lackage of benchmarks with adequate and high-quality character-level annotations for multilingual scene text spotting, most of the existing methods train on the benchmarks only with word-level annotations. However, the performance of multilingual scene text spotting are not that satisfied training on the existing benchmarks, especially for those images with special layout or words out of vocabulary. In this paper, we proposed a simple YOLO-like baseline named CMSTR for character-level multilingual scene text spotting simultaneously and efficiently. Technically, for each text instance, we represent the character sequence as ordered points and model them with learnable explicit point queries. After passing a single decoder, the point queries have encoded requisite text semantics and locations, thus can be further decoded to the center line, boundary, script, and confidence of text via very simple prediction heads in parallel. Furthermore, we show the surprisingly good extensibility of our method, in terms of character class, language type, and task. On the one hand, DeepSolo not only performs well in English scenes but also masters the Chinese transcription with complex font structure and a thousand-level character classes. On the other hand, based on the extensibility of DeepSolo, we launch DeepSolo++ for multilingual text spotting, making a further step to let Transformer decoder with explicit points solo for multilingual text detection, recognition, and script identification all at once.

ETBHD‐HMF: A Hierarchical Multimodal Fusion Architecture for Enhanced Text‐Based Hair Design

AbstractText‐based hair design (TBHD) represents an innovative approach that utilizes text instructions for crafting hairstyle and colour, renowned for its flexibility and scalability. However, enhancing TBHD algorithms to improve generation quality and editing accuracy remains a current research difficulty. One important reason is that existing models fall short in alignment and fusion designs. Therefore, we propose a new layered multimodal fusion network called ETBHD‐HMF, which decouples the input image and hair text information into layered hair colour and hairstyle representations. Within this network, the channel enhancement separation (CES) module is proposed to enhance important signals and suppress noise for text representation obtained from CLIP, thus improving generation quality. Based on this, we develop the weighted mapping fusion (WMF) sub‐networks for hair colour and hairstyle. This sub‐network applies the mapper operations to input image and text representations, acquiring joint information. The WMF then selectively merges image representation and joint information from various style layers using weighted operations, ultimately achieving fine‐grained hairstyle designs. Additionally, to enhance editing accuracy and quality, we design a modality alignment loss to refine and optimize the information transmission and integration of the network. The experimental results of applying the network to the CelebA‐HQ dataset demonstrate that our proposed model exhibits superior overall performance in terms of generation quality, visual realism, and editing accuracy. ETBHD‐HMF (27.8 PSNR, 0.864 IDS) outperformed HairCLIP (26.9 PSNR, 0.828 IDS), with a 3% higher PSNR and a 4% higher IDS.

Software bug prediction using graph neural networks and graph-based text representations

While open-source communities keep growing at an impressive pace, the corresponding platforms where software engineers share their work, deliberate about software-related issues, point out undetected bugs and suggest potential features that could be incorporated, usually suffer from information overload. This paper proposes a hybrid approach that builds on the strengths of word embeddings techniques, graph-based representation of textual data and graph neural networks for the prediction of software bugs. Existing approaches aim to improve each of these components individually, thus neglecting structural or semantical underlying information. On the contrary, the approach presented in this paper aims to leverage both types of information by representing each text as a graph and utilizing Graph Attention Networks (GATs). The evaluation carried out in this paper demonstrates an overall improvement for a variety of metrics (i.e., accuracy, precision and recall), in comparison with a plethora of graph-based machine learning models. The experiments took place using four datasets of short-text GitHub and Jira issues, which are publicly available on the kaggle.com and zenodo.com platforms respectively.

Directional Texture Editing for 3D Models

AbstractTexture editing is a crucial task in 3D modelling that allows users to automatically manipulate the surface materials of 3D models. However, the inherent complexity of 3D models and the ambiguous text description lead to the challenge of this task. To tackle this challenge, we propose ITEM3D, a Texture Editing Model designed for automatic 3D object editing according to the text Instructions. Leveraging the diffusion models and the differentiable rendering, ITEM3D takes the rendered images as the bridge between text and 3D representation and further optimizes the disentangled texture and environment map. Previous methods adopted the absolute editing direction, namely score distillation sampling (SDS) as the optimization objective, which unfortunately results in noisy appearances and text inconsistencies. To solve the problem caused by the ambiguous text, we introduce a relative editing direction, an optimization objective defined by the noise difference between the source and target texts, to release the semantic ambiguity between the texts and images. Additionally, we gradually adjust the direction during optimization to further address the unexpected deviation in the texture domain. Qualitative and quantitative experiments show that our ITEM3D outperforms the state‐of‐the‐art methods on various 3D objects. We also perform text‐guided relighting to show explicit control over lighting. Our project page: https://shengqiliu1.github.io/ITEM3D/.

VTPL: Visual and text prompt learning for visual-language models

Visual-language (V-L) models have achieved remarkable success in learning combined visual–textual representations from large web datasets. Prompt learning, as a solution for downstream tasks, can address the forgetting of knowledge associated with fine-tuning. However, current methods focus on a single modality and fail to fully use multimodal information. This paper aims to address these limitations by proposing a novel approach called visual and text prompt learning (VTPL) to train the model and enhance both visual and text prompts. Visual prompts align visual features with text features, whereas text prompts enrich the semantic information of the text. Additionally, this paper introduces a poly-1 information noise contrastive estimation (InfoNCE) loss and a center loss to increase the interclass distance and decrease the intraclass distance. Experiments on 11 image datasets show that VTPL outperforms state-of-the-art methods, achieving 1.61%, 1.63%, 1.99%, 2.42%, and 2.87% performance boosts over CoOp for 1, 2, 4, 8, and 16 shots, respectively.

Semantic-Constraint Matching for transformer-based weakly supervised object localization

Weakly supervised object localization (WSOL) strives to localize objects with only image-level supervision. WSOL often faces challenges such as incomplete localization due to classifier bias and over-localization in real scenes where objects and backgrounds are strongly associated or structurally similar. While the latest Transformer-based methods effectively enhance localization performance by leveraging long-range feature dependencies, they may inadvertently amplify divergent background activation and remain susceptible to classification bias. To this end, we proposed a novel Semantic-Constraint Matching (SeCM) plug-in module tailored for transformer-based approaches. In detail, a local patch shuffle strategy is first introduced to disentangle partial contextual linkages, thereby creating image pairs. Then a semantic matching module extracts co-object knowledge from the primal-shuffled image pairs, drives the network to identify the association of foreground with semantic label to suppress divergent activation. Moreover, to alleviate incomplete localization and prevent excessive suppression of activation, we propose leveraging multi-modal class-specific textual representations to guide object localization by complementing intra-class priori diverse knowledge. Extensive experimental results conducted on CUB-200-2011 and ILSVRC datasets show that our method can achieve the new state-of-the-art performance.

Combining Multiple Text Representations for Improved Automatic Evaluation of Indonesian Essay Answers

multiple-choices, regarding students’ learning achievement. When the number of students in a class is huge, however, examinations using essay questions become hard to conduct and take a long evaluation time. Automatic essay evaluation has, therefore, become a potential approach in this situation. Various methods have been proposed, however, optimal solutions for such evaluation in the Indonesian language are less known. Furthermore, with the rapid development of machine learning approaches, in particular deep learning approaches, the investigation of such optimal solutions becomes more necessary. Method: To address the aforementioned issue, this study proposed the investigation of text representation approaches for optimal automatic evaluation of Indonesian essay answers. The investigation compared pre-trained word embedding methods such as Word2vec, GloVe, FastText, and RoBERTa, as well as compared text encoding methods such as long short-term memories (LSTMs) and transformers. LSTMs are able to capture temporal semantics by employing state variables, while transformers are able to capture long-term dependency between parts of their input sequences. Additionally, we investigated classification-based and similarity-based training to build text encoders. We expected that these training approaches allowed encoders to extract different views of information. We compared classification results produced by different text encoders and combinations of text encoders. Result: We evaluated various text representation approaches using the UKARA dataset. Our experiments showed that the FastText word embedding method outperformed the Word2vec, GloVe, and RoBERTa methods. The FastText method achieved an F1-score of 75.43% on validation sets, while the Word2vec, GloVe, and RoBERTa methods achieved F1-scores of 69.56%, 74.53%, and 72.87%, respectively. In addition, the experiments showed that combinations of text encoders outperformed individual encoders. The combination of the LSTM encoder, the transformer encoder, and the TF-IDF encoder obtained an F1-score of 77.22% in the best case, which is better than the best F1-scores of the individual LSTM encoders (75.35%), the best combination of transformer encoders (71.49%), and the individual TF-IDF encoder (76.69%). We observed that LSTM encoders produced better performance when they were built using classification-based training. Meanwhile, the transformer encoders obtained better performance when built using similarity-based training. Novelty: The novelty proposed in this research is the optimal combination of text encoders specifically constructed for the evaluation of essay answers in the Indonesian language. Our experiments showed that the combination of three encoders - namely the LSTM encoder built using classification-based training, the transformer encoder built using classification-based and similarity-based training, and the TF-IDF encoder - obtained the best classification performance.

The Research on Intelligent News Advertisement Recommendation Algorithm Based on Prompt Learning in End-to-End Large Language Model Architecture

With the explosive growth of information on the internet, users are increasingly facing the problem of information overload, making precise news and ad recommendations an important area of research. While traditional recommendation algorithms can meet user needs to some extent, they still have limitations in dealing with complex and changing user behaviors and dynamic content environments. This paper addresses the shortcomings of existing news and ad recommendation systems by proposing an intelligent recommendation algorithm based on an end-to-end large language model architecture. Firstly, we utilize the BERT model as the foundation, leveraging its powerful text representation capabilities to achieve deep semantic understanding of news and ad content, thereby capturing more detailed content features. Secondly, we apply prompt learning to fine-tune the BERT model, designing specific prompts for the model to better understand the implicit needs and preferences of users. Finally, we integrate these steps into an end-to-end architecture, enabling the model to achieve automated learning and optimization throughout the entire process from input to output, thus improving the precision and efficiency of recommendations. Experimental results demonstrate that the proposed method significantly outperforms traditional methods in the task of news and ad recommendation, not only enhancing the accuracy and relevance of recommendations but also effectively improving the model's interpretability and flexibility. This research explores new possibilities for the application of large language models in recommendation systems.

Multimodal fusion network for ICU patient outcome prediction

Over the past decades, massive Electronic Health Records (EHRs) have been accumulated in Intensive Care Unit (ICU) and many other healthcare scenarios. The rich and comprehensive information recorded presents an exceptional opportunity for patient outcome predictions. Nevertheless, due to the diversity of data modalities, EHRs exhibit a heterogeneous characteristic, raising a difficulty to organically leverage information from various modalities. It is an urgent need to capture the underlying correlations among different modalities. In this paper, we propose a novel framework named Multimodal Fusion Network (MFNet) for ICU patient outcome prediction. First, we incorporate multiple modality-specific encoders to learn different modality representations. Notably, a graph guided encoder is designed to capture underlying global relationships among medical codes, and a text encoder with pre-fine-tuning strategy is adopted to extract appropriate text representations. Second, we propose to pairwise merge multimodal representations with a tailored hierarchical fusion mechanism. The experiments conducted on the eICU-CRD dataset validate that MFNet achieves superior performance on mortality prediction and Length of Stay (LoS) prediction compared with various representative and state-of-the-art baselines. Moreover, comprehensive ablation study demonstrates the effectiveness of each component of MFNet.

PEPT: Expert Finding Meets Personalized Pre-training

Finding experts is essential in Community Question Answering (CQA) platforms as it enables the effective routing of questions to potential users who can provide relevant answers. The key is to personalized learning expert representations based on their historical answered questions, and accurately matching them with target questions. Recently, the application of Pre-trained Language Models (PLMs) have gained significant attraction due to their impressive capability to comprehend textual data, and are widespread used across various domains. There have been some preliminary works exploring the usability of PLMs in expert finding, such as pre-training expert or question representations. However, these models usually learn pure text representations of experts from histories, disregarding personalized and fine-grained expert modeling. For alleviating this, we present a personalized pre-training and fine-tuning paradigm, which could effectively learn expert interest and expertise simultaneously. Specifically, in our pre-training framework, we integrate historical answered questions of one expert with one target question, and regard it as a candidate aware expert-level input unit. Then, we fuse expert IDs into the pre-training for guiding the model to model personalized expert representations, which can help capture the unique characteristics and expertise of each individual expert. Additionally, in our pre-training task, we design: 1) a question-level masked language model task to learn the relatedness between histories, enabling the modeling of question-level expert interest; 2) a vote-oriented task to capture question-level expert expertise by predicting the vote score the expert would receive. Through our pre-training framework and tasks, our approach could holistically learn expert representations including interests and expertise. Our method has been extensively evaluated on six real-world CQA datasets, and the experimental results consistently demonstrate the superiority of our approach over competitive baseline methods. 1

Contrastive Learning with Transformer Initialization and Clustering Prior for Text Representation

Acquiring labeled data for learning sentence embeddings in Natural Language Processing poses challenges due to limited availability and high costs. In order to tackle this issue, we introduce a novel method called Contrastive Learning with Transformer Initialization and Clustering Prior for Text Representation (CLTC). Our method utilizes Pre-Layernorm Transformers without warm-up, stabilizing the training process while also increasing the final performance. We employ Contrastive Learning (CL) with dropout-based augmentation to enhance sentence embeddings. Additionally, we integrate prior knowledge into the contrastive learning framework within an efficient clustering strategy. When evaluated on the SentEval task, our approach showcases a competitive performance when compared to state-of-the-art approaches in the contrastive learning domain. Our method offers stability, improved embeddings, and the utilization of prior knowledge for enhanced unsupervised representation learning in Natural Language Processing applications.

FrameSum: Leveraging Framing Theory and Deep Learning for Enhanced News Text Summarization

Framing theory is a widely accepted theoretical framework in the field of news communication studies, frequently employed to analyze the content of news reports. This paper innovatively introduces framing theory into the text summarization task and proposes a news text summarization method based on framing theory to address the global context of rapidly increasing speed and scale of information dissemination. Traditional text summarization methods often overlook the implicit deep-level semantic content and situational frames in news texts, and the method proposed in this paper aims to fill this gap. Our deep learning-based news frame identification module can automatically identify frame elements in the text and predict the dominant frame of the text. The frame-aware summarization generation model (FrameSum) can incorporate the identified frame feature into the text representation and attention mechanism, ensuring that the generated summary focuses on the core content of the news report while maintaining high information coverage, readability, and objectivity. Through empirical studies on the standard CNN/Daily Mail dataset, we found that this method performs significantly better in improving summary quality and maintaining the accuracy of news facts.

Frame-level nonverbal feature enhancement based sentiment analysis

Multimodal Sentiment Analysis (MSA) comprehensively utilizing data from multiple modalities to obtain more accurate sentiment attribute, has important applications in other fields, such as social media analysis, user experience evaluation and medical health, etc. It is worth noting that previous studies have paid little attention to the inconsistency of the initial representation granularity between verbal (textual) and nonverbal (acoustic and visual) modalities. As a result, the imbalanced emotional information between them complicates the interaction process, and ultimately affects the model’s performance. To solve this problem, this paper proposes a Frame-level Nonverbal feature Enhancement Network (FNENet) to improve performance on MSA by reducing the gap and integrating asynchronous affective information between modalities. Specifically, Vector Quantization (VQ) is applied to nonverbal modalities to reduce the granularity differences and improve the performance of the model. Additionally, nonverbal information is integrated through the Sequence Fusion mechanism (SF) into a pre-trained language model to enhance the textual representation, which benefits the word-level semantic expression according to the asynchronous affective cues preserved in unaligned frame-level nonverbal features. Extensive experiments on three benchmark datasets demonstrate that FNENet significantly outperforms baseline methods. It indicates that our model has potential application on MSA.

Multi-dimensional feature fusion-based expert recommendation in community question answering

PurposeCommunity question answering (CQA) platforms play a significant role in knowledge dissemination and information retrieval. Expert recommendation can assist users by helping them find valuable answers efficiently. Existing works mainly use content and user behavioural features for expert recommendation, and fail to effectively leverage the correlation across multi-dimensional features.Design/methodology/approachTo address the above issue, this work proposes a multi-dimensional feature fusion-based method for expert recommendation, aiming to integrate features of question–answerer pairs from three dimensions, including network features, content features and user behaviour features. Specifically, network features are extracted by first learning user and tag representations using network representation learning methods and then calculating questioner–answerer similarities and answerer–tag similarities. Secondly, content features are extracted from textual contents of questions and answerer generated contents using text representation models. Thirdly, user behaviour features are extracted from user actions observed in CQA platforms, such as following and likes. Finally, given a question–answerer pair, the three dimensional features are fused and used to predict the probability of the candidate expert answering the given question.FindingsThe proposed method is evaluated on a data set collected from a publicly available CQA platform. Results show that the proposed method is effective compared with baseline methods. Ablation study shows that network features is the most important dimensional features among all three dimensional features.Practical implicationsThis work identifies three dimensional features for expert recommendation in CQA platforms and conducts a comprehensive investigation into the importance of features for the performance of expert recommendation. The results suggest that network features are the most important features among three-dimensional features, which indicates that the performance of expert recommendation in CQA platforms is likely to get improved by further mining network features using advanced techniques, such as graph neural networks. One broader implication is that it is always important to include multi-dimensional features for expert recommendation and conduct systematic investigation to identify the most important features for finding directions for improvement.Originality/valueThis work proposes three-dimensional features given that existing works mostly focus on one or two-dimensional features and demonstrate the effectiveness of the newly proposed features.

Video text tracking with transformer-based local search

Video text tracking is a highly significant branch in the field of multi-object tracking (MOT), aiming to detect all text instances in video frames and construct trajectories for each text. Conventional text tracking methods typically follow the Tracking-By-Detection (TbD) paradigm which involves two separate steps of detection and association. In complex scenarios, abundant detection omission of text instances is caused by exposure, occlusion, motion blur, and the like, leading to text trajectory breakpoints in the TbD methods. To this end, video text tracking with transformer-based local search (LSTrack) is proposed in this paper. We utilize historical trajectory information to estimate the approximate search areas where omitted texts are located. In this case, our local search tracker leverages the text images in historical trajectories as references to directly recall text instances in the search areas. In addition, our tracking framework can combine the explicit semantic information output from an OCR model to obtain the semantic version (SV) of LSTrack, which for the first time uses the text edit distance as the distance measurement in matching stages to achieve better text tracking results. Ultimately, our method achieves an advanced performance on several public benchmarks. In particular, LSTrack (SV) achieves state-of-the-art performance on the Minetto benchmark.

DPGS: Cross-cooperation guided dynamic points generation for scene text spotting

End-to-end text spotting aims to combine scene text detection and recognition into a unified framework. Dealing with the relationship between the two sub-tasks plays a pivotal role in designing effective spotters. While polygon or segmentation-based methods eliminate heuristic post-processing, they still face challenges such as background noise and high computational burden. In this study, we introduce DPGS, a coarse-to-fine learning framework that lets Dynamic Points Generation for text Spotting. DPGS simultaneously learns character representations for both detection and recognition tasks. Specifically, for each text instance, we represent the character sequence as ordered points and model them with learnable point queries. This approach progressively selects appropriate key points covering character and leverages group attention to associate similar information from different positions, improving detection accuracy. After passing through a single decoder, the point queries encode text semantics and locations, facilitating decoding to central line, boundary, script, and confidence of text through simple prediction heads. Additionally, we introduce an adaptive cooperative criterion to combine more useful feature knowledge, enhancing training efficiency. Extensive experiments show the superiority of our DPGS when handling scene text detection and recognition tasks. Compared to their respective top-1 methods, DPGS has significantly improved the average recognition accuracy by 3.7%, 1.9%, and 0.7% on the Total-Text, ICDAR15, and CTW1500 datasets, respectively.

Joint coordinate attention mechanism and instance normalization for COVID online comments text classification.

The majority of extant methodologies for text classification prioritize the extraction of feature representations from texts with high degrees of distinction, a process that may result in computational inefficiencies. To address this limitation, the current study proposes a novel approach by directly leveraging label information to construct text representations. This integration aims to optimize the use of label data alongside textual content. The methodology initiated with separate pre-processing of texts and labels, followed by encoding through a projection layer. This research then utilized a conventional self-attention model enhanced by instance normalization (IN) and Gaussian Error Linear Unit (GELU) functions to assess emotional valences in review texts. An advanced self-attention mechanism was further developed to enable the efficient integration of text and label information. In the final stage, an adaptive label encoder was employed to extract relevant label information from the combined text-label data efficiently. Empirical evaluations demonstrate that the proposed model achieves a significant improvement in classification performance, outperforming existing methodologies. This enhancement is quantitatively evidenced by its superior micro-F1 score, indicating the efficacy of integrating label information into text classification processes. This suggests that the model not only addresses computational inefficiencies but also enhances the accuracy of text classification.

ProcessCarbonAgent: A large language models-empowered autonomous agent for decision-making in manufacturing carbon emission management

Knowledge-intensive production represents a primary trend in industrial manufacturing, which heavily relies on the production logs of large-scale, historically similar orders for enhancing production efficiency and process quality. These logs are essential for predicting resource allocation and identifying bottlenecks in throughput. As a result, root cause analysis of the production process state is crucial for supporting decision-making in these settings. However, current methodologies heavily depend on expert knowledge, making the analysis time-consuming and inefficient for large-scale, multivariable processes. Although the development of large language models and autonomous agents presents a potential solution, these models are limited in their direct interaction with event logs due to inadequate data representation, token constraints, and insufficient accuracy. Therefore, enabling the interactive capabilities of large language models to overcome these specific limitations in process event data and industrial domain illusions poses a significant challenge. To address these issues, this paper introduces the ProcessCarbonAgent framework, an autonomous agent empowered by large language models, designed to enhance decision-making within industrial processes. Initially, a process data agent combines predefined semantic text representation methods with process template prompting strategies to improve interaction capabilities. Subsequently, an intention agent utilizing self-information and large language models is developed to address context length limitations by identifying and eliminating redundancies. Finally, a two-stage confidence estimation method is implemented to refine the precision of decision-making assistance, thereby improving the accuracy of decisions supported by large language models. Experiments with textile industry carbon emission data reveal that the assisted decision-making scores employing a compression ratio of 0.5, closely align with scores from manually labeled evaluations, with a 98% overlap across scoring intervals. Moreover, in contrast to relying solely on the original evaluation method, the two-stage confidence estimation method has led to a 20% increase in accuracy performance. The ProcessCarbonAgent achieved scores of 16.64, 55.13, 26.32, and 34.17 on METEOR, BERTScore, NUBIA, and BLEURT, respectively. The results demonstrate that the ProcessCarbonAgent framework significantly enhances the decision-making process for high-carbon emission states in industrial production, providing technical support for the low-carbon transformation and intelligent upgrading of these processes.