Search

The task of emotional support dialogue requires providing supportive responses based on a thorough understanding of the user’s psychological state, with the aim of alleviating their emotional distress. Most existing studies employ end-to-end generation methods, where small pre-trained language models are fine-tuned to adapt to the emotional support task. However, these methods lack a fine-grained understanding of the user’s psychological state, resulting in insufficient empathy, and the model decision process is opaque, resulting in poor interpretability. To address these issues, inspired by the excellent reasoning capabilities of current large language models, this paper proposes an emotional support dialogue reasoning framework based on large language models called CoES (chain-of-emotional-support). This framework transforms the end-to-end generation problem into a step-by-step reasoning problem, breaking down the complex task of emotional support into simpler subtasks to be solved sequentially. The framework comprises three reasoning chains: the emotional reasoning chain, the strategy reasoning chain, and the response generation chain, which are used for the fine-grained exploration of the user’s psychological state, the selection of emotional support strategies, and the generation and optimization of responses, respectively. Additionally, this paper designs various external knowledge augmentation strategies to improve the reasoning effectiveness of the large model in the psychological state exploration and support strategy selection processes. Both manual and automatic evaluation results on the ESConv dataset demonstrate that the proposed reasoning method achieves advanced performance in terms of the interpretability of emotional support and the quality of content generation.

Knowledge Graph-to-Text is a new task in the field of knowledge graph,which aims to transform knowledge graph into readable text describing these knowledge.With the deepening of research in recent years,the generation technology of Graph-to-Text has been applied to the fields of product review generation,recommendation explanation generation,paper abstract generation and so on.The translation model in the existing methods adopts the method of first-plan-then-realization,which fails to dynamically adjust the planning according to the generated text and does not track the static content planning,resulting in incohe-rent semantics before and after the text.In order to improve the semantic coherence of generated text,a Graph-to-Text model based on dynamic memory and two-layer reconstruction enhancement is proposed in this paper.Through three stages of static content planning,dynamic content planning and two-layer reconstruction mechanism,this model makes up for the structural difference between knowledge graph and text,focusing on the content of each triple while generating text.Compared with exis-ting generation models,this model not only compensates for the structural differences between knowledge graphs and texts,but also improves the ability to locate key entities,resulting in stronger factual consistency and semantics in the generated texts.In this paper,experiments are conducted on the WebNLG dataset.The results show that,compared with the current exis-ting models in the task of Graph-to-Text,the proposed model generates more accurate content planning.The logic between the sentences of the generated text is more reasonable and the correlation is stronger.The proposed model outperforms existing methods on me-trics such as BLEU,METEOR,ROUGE,CHRF++,etc.

Object detection is one of the most popular directions in computer vision,which is widely used in military,medical and other important fields.However,most object detection models can only recognize visible objects.There are often covered(invisible) target objects in pictures in daily life.It is difficult for existing object detection models to show ideal detection performance for covered objects in pictures.Therefore,this paper proposes a novel class reasoning model towards covered area in given image based on informed knowledge graph reasoning and multi-agent collaboration(IMG-KGR-MAC).Specifically,first,IMG-KGR-MAC constructs a global prior knowledge graph according to the visible objects of all pictures in a given picture library and the positional relationship between them.At the same time,according to the objects contained in the pictures themselves and their positional relationships,picture knowledge graphs are established for each picture respectively.The covered objects information in each picture is not included in the global prior knowledge graph and the picture's own knowledge graph.Second,deep deterministic policy gradient(DDPG) deep reinforcement learning idea is adopted to build two cooperative agents.Agent 1 selects the “category label” that is most suitable for the covered object from the global prior knowledge graph according to the semantic information of the current picture,and adds it to the knowledge graph of the given picture as a new entity node.Agent 2 further selects 〈entities,relationships〉 from the global prior knowledge graph according to the newly added entities of agent 1,and expands the graph structure associated with the new entity nodes.Third,agent 1 and agent 2 share the task environment and communicate the reward value,and cooperate with each other to carry out forward and reverse reasoning according to the principles of ‘picture covered target(entity) → associated graph structure' and ‘associated graph structure → picture covered object(entity)',so as to effectively estimate the most likely category label of the covered object of a given picture.Experimental results show that,compared with the existing related methods,the proposed IMG-KGR-MAC model can learn the semantic relationship between the covered picture of a given picture and the global prior knowledge graph,effectively overcome the shortcomings of the existing models that it is difficult to detect the covered object,and has good reasoning ability for the covered object.It has more than 20% improvement in many indicators such as MR(mean rank) and mAP(mean average precision).

Automatic text summarization aims to compress a given document, which can efficiently reflect the main idea of the source document with a short summary. At present, abstractive summarization method has become a research hotspot in the field of text summarization because it can paraphrase the source document with flexible and abundant vocabulary. However, existing abstractive summarization model reorganizes original words and adds new words when generating summary. That’s why it can easily cause the inconsistency and low readability. In addition, the traditional supervised learning based on labeled data requires high cost to improve the coherence of summary sentences, which limits the practical application. Therefore, this paper proposes an abstractive text summarization model with coherence reinforcement and no ground truth dependency (ATS_CG). On the one hand, based on the embdding of the source document, the model generates extractive label to describe the filtering process of the key information. And then, the filtered sentence embeddings are decoded by the decoder. On the other hand, based on the original word probability distribution output by the decoder, two types of summarization are generated according to “probability selection” and “Softmax-greedy selection”. And then, the model will compute the overall rewards of the two types of summarization from the two aspects of coherence and content. Next, the model will learn to filter key sentences and decode them through the self-critical policy gradient, so as to generate abstractive summarizaion with high coherence and quality. Experiments show that ATS_CG is superior to the existing text summarization methods in terms of evaluation scores on the whole, even without any ground truth. At the same time, abstractive summarization generated by ATS_CG is also better than the existing methods in coherence, relevance, redundancy, novelty and perplexity.

In recent years, with the rapid development of the Internet, people frequently post comments about a specific object on the Internet. Mastering the critical information from the crowdsourced comments promptly is crucial to the decision-making and service adjustment, with non-negligible application value. Therefore, it is imperative to devote effort to the research on crowdsourced comment integration problem. The goal of the crowdsourced comment integration is to integrate different users comments on the target object into a shorter integrated document by a given compression rate, so as to form a comparatively matched description of the target object according to the public cognition. To solve such problem, a two-phase crowdsourced comment integration method based on reward prediction and policy gradient is proposed. The proposed method does not rely on any man-made ground truth, only requiring the crowdsourced comments. Then, an agent, guided by the experience or reward, will extract key sentence from the crowdsourced comments to generate the integrated comment. Specifically, in the first phase, measuring the content quality of the integrated comment by the relevance and redundancy of sentences, taking the content quality as reward, the long-term reward from selecting a current sentence to the end of the whole comment integration process will be predicted by Q-value, based on which the agent is guided to learn an optimal sentence selection policy. Then, in the second phase, taking the sentiment intensity of the integrated comment as reward, the sentence selection policy learnt in the first phase will be further adjusted by policy gradient, so that the integrated comment generated by the agent can highlight the sentiment intensity from an objective perspective and reflect users attitude more obviously, at the same time, maintaining the content quality. According to the experimental results, compared with the other existing methods, the proposed method can achieve the best overall performance in terms of the content quality as well as the sentiment intensity of the integrated comment, and the time consumed for generation is still controlled at an acceptable level.

It is difficult to accurately identify the winter wheat acreage in the Jianghuai region of China, and the fusion of high-resolution images and medium-resolution image data can improve the image quality and facilitate the identification and acreage extraction of winter wheat. Therefore, the objective of this study is to improve the accuracy of China's medium-spatial resolution image data (environment and disaster monitoring and forecasting satellite data, HJ-1/CCD) in extracting the large area of winter wheat planted. The fusion and object-oriented classification of the 30 m × 30 m HJ-1/CCD multispectral image and 2 m × 2 m GF-1 panchromatic image (GF-1/PMS) of winter wheat at the jointing stage in the study area were studied. The GF-1/PMS panchromatic images were resampled at 8 m, 16 m and 24 m to produce panchromatic images with four spatial resolutions, including 2 m. They were fused with HJ-1/CCD multispectral images by Gram Schmidt (GS). The quality of the fused images was evaluated to pick adequate scale images for the field pattern of winter wheat cultivation in the study area. The HJ-1/CCD multispectral image was resampled to obtain an image with the same scale as the suitable scale fused image. In the two images, the training samples SFI (samples of fused image) and SRI (samples of resampled image) containing spectral and texture information were selected. The fused image (FI) and resampled image (RI) were used for winter wheat acreage extraction using an object-oriented classification method. The results indicated that the fusion effect of 16 m × 16 m fused image was better than 2 m × 2 m, 8 m × 8 m and 24 m × 24 m fused images, with mean, standard deviation, average gradient and correlation coefficient values of 161.15, 83.01, 4.55 and 0.97, respectively. After object-oriented classification, the overall accuracy of SFI for the classification of resampled image RI16m was 92.22%, and the Kappa coefficient was 0.90. The overall accuracy of SFI for the classification of fused image FI16m was 94.44%, and the Kappa coefficient was 0.93. The overall accuracy of SRI for the classification of resampled image RI16m was 84.44%, and the Kappa coefficient was 0.80. The classification effect of SFI for the fused image FI16m was the best, indicating that the object-oriented classification method combined with the fused image and the extraction samples of the fused image (SFI) could extract the winter wheat planting area with precision. In addition, the object-oriented classification method combining resampled images and the extraction samples of fused images (SFI) could extract the winter wheat planting area more effectively. These results indicated that the combination of medium spatial resolution HJ-1/CCD images and high spatial resolution GF-1 satellite images could effectively extract the planting area information of winter wheat in large regions. [ABSTRACT FROM AUTHOR]