Understanding a discourse requires tracking entities and the relations that hold between them. While Large Language Models (LLMs) perform well on relational reasoning, the mechanism by which they bind entities, relations, and attributes remains unclear. We study discourse-level relational binding and show that LLMs encode it via a Cell-based Binding Representation (CBR): a low-dimensional linear subspace in which each “cell” corresponds to an entity–relation index pair, and bound attributes are retrieved from the corresponding cell during inference. Using controlled multi-sentence data annotated with entity and relation indices, we identify the CBR subspace by decoding these indices from attribute-token activations with Partial Least Squares regression. Across domains and two model families, the indices are linearly decodable and form a grid-like geometry in the projected space. We further find that context-specific CBR representations are related by translation vectors in activation space, enabling cross-context transfer. Finally, activation patching shows that manipulating this subspace systematically changes relational predictions and that perturbing it disrupts performance, providing causal evidence that LLMs rely on CBR for relational binding.

Incremental Object Detection (IOD) is hampered by catastrophic forgetting from continuous learning as object categories evolve, particularly in real-world tasks. This paper focuses on tackling IOD challenges and adopting the lightweight You Only Look Once (YOLO) detector suitable for prevalent edge devices. To preserve prior knowledge, we propose a training freezing strategy for network layers that extract common features of all classes in new category learning, employing the Deep Deterministic Policy Gradient (DDPG) to guide the layer chosen process. Additionally, a novel Knowledge Distillation (KD) approach is presented. It introduces the focal global to address the discrepancy between teacher-student feature maps inherent in KD and the response distillation to offer reasoning information from the old detector. We also build a spatial and channel attention module to enhance the distillation performance further. Moreover, a Bi-fold sampling method is developed according to the balance in category probability to keep the capability of picking representative exemplars continuous. Experiments on MS COCO show that our method gets State-Of-The-Art results.

Entity tracking is essential for complex reasoning.To perform in-context entity tracking, language models (LMs) must bind an entity to its attribute (e.g., bind a container to its content) to recall attribute for a given entity.For example, given a context mentioning "The coffee is in Box Z, the stone is in Box M, the map is in Box H", to infer "Box Z contains the coffee" later, LMs must bind "Box Z" to "coffee".To explain the binding behaviour of LMs, Feng and Steinhardt (2023) introduce a Binding ID mechanism and state that LMs use a abstract concept called Binding ID (BI) to internally mark entity-attribute pairs.However, they have not captured the Ordering ID (OI) from entity activations that directly determines the binding behaviour.In this work, we provide a novel view of the BI mechanism by localizing OI and proving the causality between OI and binding behaviour.Specifically, by leveraging dimension reduction methods (e.g., PCA), we discover that there exists a low-rank subspace in the activations of LMs, that primarily encodes the order (i.e., OI) of entity and attribute.Moreover, we also discover the causal effect of OI on binding that when editing representations along the OI encoding direction, LMs tend to bind a given entity to other attributes accordingly.For example, by patching activations along the OI encoding direction we can make the LM to infer "Box Z contains the stone" and "Box Z contains the map".The code and datasets used in this paper are available at https://github.com/cl-tohoku/OI-Subspace.

Bi-encoder architectures for distantlysupervised relation extraction are designed to make use of the complementary information found in text and knowledge graphs (KG).However, current architectures suffer from two drawbacks.They either do not allow any sharing between the text encoder and the KG encoder at all, or, in case of models with KG-to-text attention, only share information in one direction.Here, we introduce cross-stitch bi-encoders, which allow full interaction between the text encoder and the KG encoder via a cross-stitch mechanism.The cross-stitch mechanism allows sharing and updating representations between the two encoders at any layer, with the amount of sharing being dynamically controlled via cross-attention-based gates.Experimental results on two relation extraction benchmarks from two different domains show that enabling full interaction between the two encoders yields strong improvements.

In recent years, the metaverse becomes a promising method to provide an intelligent teaching platform for the teaching-learning process. Existing teaching evaluation methods rely on the exam results and the educator’s teaching experience, which is hard to reflect the detailed teaching outcomes and each student’s learning situation. Therefore, this paper proposes a four-layer metaverse architecture to build students’ virtual entities of learning situations, containing a data acquisition layer, a technology layer, a model building layer, and an application layer. Furthermore, the virtual entities are constructed on the event logs recorded in the Learning Management System (LMS) and visualized as spatiotemporal digital profiles for students. In the spatial dimension, the profile can reflect a student’s learning situation in different aspects, including the completion of an assignment, the mastery of knowledge, the practical ability, etc. In the temporal dimension, it can reflect a student’s learning situation at different learning stages. Students’ practical abilities are obtained by the Machine Learning method GBDT (Gradient Boosting Decision Tree), and other dimensions of the profile are generated by the Knowledge Graph technology. With these profiles, educators can do teaching intervention, teaching evaluation, and personalized cultivation, providing a new path for intelligent teaching. We take the CG (Course Grading) platform and Data Structure and Algorithm course as examples to validate our model strategy. The experimental results show that the spatiotemporal digital profiles can better describe students’ learning situations, providing data support for teaching evaluation.

This paper explores how the Distantly Supervised Relation Extraction (DS-RE) can benefit from the use of a Universal Graph (UG), the combination of a Knowledge Graph (KG) and a large-scale text collection.A straightforward extension of a current state-of-the-art neural model for DS-RE with a UG may lead to degradation in performance.We first report that this degradation is associated with the difficulty in learning a UG and then propose two training strategies: (1) Path Type Adaptive Pretraining, which sequentially trains the model with different types of UG paths so as to prevent the reliance on a single type of UG path; and (2) Complexity Ranking Guided Attention mechanism, which restricts the attention span according to the complexity of a UG path so as to force the model to extract features not only from simple UG paths but also from complex ones.Experimental results on both biomedical and NYT10 datasets prove the robustness of our methods and achieve a new state-ofthe-art result on the NYT10 dataset.The code and datasets used in this paper are available at https://github.com/baodaiqin/ UGDSRE.

Flooded area extraction using remote sensing data plays a fundamental role on precisely evaluating flooded area and conducting disaster rescue and relief. The flooded area images may have irregular and fuzzy outline. The traditional methods are easily affected by difference image with low contrast obtained by multi-temporal remote sensing images. In order to overcome the above limitations, a new automatic and un-supervised method for extracting flooded area is presented in this paper. The emphasis of this study lies in creating high contrast difference image via weighted combing different features and applying Level Set Method (LSM) to extract flooded area without predefined information. LSM Chan-Vese (C-V) model is a better choice because it can handle topology changes to extract object with variable shapes from image. Besides, the proposed method modifies the initial curve of C-V model to speed up the iteration and improve extraction precision. This paper selects Landsat 8 OLI data set to validate the methodology this paper study. The proposed method provides more accurate and efficient extraction of flooded area extent when compared with Fuzzy C-Mean (FCM) algorithm.

Delay is one of the most important parameters of VoIP which can significantly impact the voice quality. When crossing an ADSL downlink, a VoIP application has to contend for the bandwidth with other applications, e.g., TCP. With the assumption that the DSLAM only has a FCFS mechanism employed in the downlink direction, the VoIP packets may suffer from serious queueing delay problem in the DSLAM buffer. This problem can be solved by introducing TCP packet dropping at the receiver. But the drop ratio should be adjusted depending on the network conditions such as network load. In this paper, an adaptive packet dropping algorithm based on the RED mechanism is proposed to guarantee the VoIP quality in ADSL downlink under different network conditions. Simulations are conducted to compare our approach with the fixed drop ratio approach considering different network traffic. The results show that our algorithm can successfully adapt to the network conditions. Moreover, compared with the fixed drop ratio one, our approach can achieve better VoIP quality with fewer TCP packet dropping.

Landsat TM/ETM+ data in May or June are used to depict the Land use changes from 1988,1994,2001,2003,2005,2006 respectively inner Beijing six ring express road and specially to watch the land use changes from the beginning of Beijing Olympic playground area housebrealdng in 2002. The great changes are identified in terms of land use change and usage replacements.

Recent years, the locust hazard is a great threat to the agriculture in the east coast of China. A Bayesian network is a probabilistic representation for uncertain relationships, the directed acyclic graph (DAG) of Bayesian network explicitly represents causes and affects assumptions among the affair happening variables, which can predict the event happening through probability inference. Based on the remote sensing monitoring and experimental conclusion, consideration of the causal relationship among the factors of locust environmental roosts, we select the label factor indicating locust hazard, construct Bayesian network and predict the East Asia locust hazard. The study area is located in Tianjin Dagang and HuangHua of HeBei province, surrounding the BOHAI Gulf area. We got a good agreement between prediction factors and practical factors in the change trend along with time and the correlation coefficient between them show their strong correlation. The results presented in figure1-5.

This paper addresses the adjacency effect estimation in ground spectrum and Landsat-5 pixels. The adjacency effect influences a pixel DN value by adding surface surrounding scatter signal and path scatter signals at sensor. Based on the theory of radiation transfer we designed a system that measuring the reflectance from the surface target material and the material in the box of 1.5m above the target to avoid upwelling reflectance. At every 9 /spl times/ 9 sites; the measurement was carried out during the period of 10:30 to 1:30 at the Guanting remote sensing test ground of the North Beijing. The results show that the adjacency influence becomes stronger from visible, near infrared to short wave infrared regions; the adjacency effect weakens with the increase of distance between testing site and than the corresponding adjacency influence was identified in Landsat-5 image.

Beijing will hold the 2008 Olympic Games. The main venue locates in northern Beijing, where there were villages, residential houses and commercial buildings. In order to help Beijing Olympic committee on the planning of housebreak and monitoring the construction implementation, the Institute of Remote Sensing Applications (IRSA, CAS), used the airplane to obtain true color photos from 2001 to 2003 in order to monitor the progress of the housebreaking. This work introduced the change detection methods and the land use changes were statistically analyzed, which show clearly the distribution of the housebreaking area, as well as the amount of houses or buildings. Until now 718 mu of houses or buildings remain to be broken down in near future.

Currently the function of operational locust monitor system mainly focused on after-hazards monitoring and assessment, and to found the way effectively to perform early warning and prediction has more practical meaning. Through 2001, 2002 two years continuously field sample and statistics for locusts eggs hatching, nymph growth, adults 7 phases observation, sample statistics and calculation, spectral measurements as well as synchronically remote sensing data processing we raise the view point of remote sensing three stage monitor the locust hazards.

ASTER data with 15 m resolution in visible and near infrared bands are proper data for urban change detection. Often proper data (without cloud) cannot be acquired conveniently. So cloud removal is a very important step before making comparisons. Traditional methods have some shortcomings. Thus, based on statistical characters of image information, an improved homomorphism filtering method is put forward. Instead of the filtering in frequency field, it isolates the low frequency component of the image representing cloud information with calculating neighborhood average in spatial field. An experiment indicates that the method can obtain good results and performs faster than traditional methods.

Traditionally in China the ring mark method is used for bird migratory road tracking average ten percentages of ring marks can be found after migratory. Any changes of wetland in migratory road or stopover can cause living problems for the birds. The migratory road correspondingly to the economy and population fast developing area of east part of China so that the result of the study can provide scientific information for bird protection and sustainable development of the region. In recent years, Institute of Remote Sensing Applications, CAS, Hokkaido Institute of Environment and the University of Tokyo organized a research project to study migratory roads and monitoring surface condition change using satellite tracking data (NOAA/NNT), NOAA/NDVI, surface temperature data, wetland change GIS data. Some results are introduced in this article.