The faithful and interpretable opinion summarization aims to generate a summary that captures the diverse opinions expressed in a document set while providing explanations for the divergences between these opinions. In this paper, we propose an evidence-guided framework to enhance opinion coverage and provide divergence explanations. It first generates the majority opinion as an initial summary and partitions the source documents into multiple evidence sets based on their relevance to the majority opinion. Then, a summary extension strategy is employed to expand the initial summary by incorporating different opinions from these sets. The framework also employs a submodular optimization algorithm to select evidence from different evidence sets in order to reflect the divergences between opinions. Experiments on two benchmark datasets demonstrate that our method outperforms multiple baselines in terms of both the lexical and semantic consistency with reference summaries, while having low computational overhead.Ablation studies confirm that both the document partition and summary extension mechanisms contribute to the model performance.The LLM-based and human evaluation results also show that our method can identify more comprehensive evidence that better captures opinion divergences.

Instructions, as the primary means of using large language models (LLMs), significantly impact the results. Automatically inducing instructions from few-shot instances is meaningful, yet the induced instructions suffer from two inherent divergences: (1) systematic discrepancies across tasks, and (2) instance-level heterogeneity within a task. To address these challenges, inspired by human analogical reasoning, we propose AdaIn, an iteratively adaptive instruction induction framework that leverages instance-level structural similarities. AdaIn groups instances by data features to induce tailored instructions, and adaptively applies them to new samples. Instruction performance is further utilized to guide updates„ enabling iterative identification and refinement of ineffective instructions. We conducted experiments on different tasks to verify our method and the results demonstrate that it outperforms the SOTA results. Ablation experiments indicate that the adaptive strategy in induction and selection instruction contributes much to performance.

Text semantic hashing maps a text to a compact binary code, which is an important part of information retrieval and language processing. There are two main challenges for this task, one is to make the hash codes express the hierarchical category information for improving the retrieval accuracy, and the other is how to deal with the hard samples. In this paper, we adopt the Bernoulli VAE to encode the text semantics and design the parent and child level contrastive losses to learn the hierarchical information of the text. To find the hard samples, for each category, we introduce a latent sphere space to split the majority samples and hard samples, where the center and radius are dynamically calculated based on the semantic distance between samples. For the hard samples, we introduce the de-confusion loss to pull them close to the center. We conduct experiments on three datasets and the results show that the proposed model outperforms the SOTA baselines. The ablation experiments show that the category constraints and the de-confusion loss contribute to the model performance. The results of t-SNE also show that the hash codes learned by our model reflect high category differences.

In this paper, we propose an NER promotion method formed as an external discriminator. It learns the patterns about the contextual entity usages from the extensive web data and thus it can check whether the recognized entity by an NER model is correct. Different with the current popular methods on introducing the entity knowledge by gazetteers or labeled data, it can be used as the additional part to work with any NER method for promoting its performance. We adopt three widely adopted datasets for the empirical studies and the results show that our method significantly improves the NER performance. Besides, by using only a small proportion of labeled data, our method achieves a comparable performance against other models using the whole labeled data.

Diverse opinion summarization aims to generate a summary that captures multiple opinions in texts. Although large language models (LLMs) have become the main choice for this task, the performance is highly depend on prompts. In this paper, we propose a self-evaluation based prompt calibration framework to stimulate LLM for generating high quality summary. It adopts the reinforcement learning mechanism to calibrate prompts for maximizing the reward of summary. The framework contains three parts. In the prompt construction part, we design the prompt that contains topic, task instruction and key opinion reference. The topic indicates the main focus of documents, the instruction describes the task with natural language and the key opinion reference is the explicit constraint on the expected opinions. In the reward part, for each summary, its coverage score and diversity score are used to represent the semantic coverage to the source documents and the inter opinion differences, respectively. The prompt calibration part selects the sentences in generated summaries to calibrate the prompts for the next iteration. With this framework, we use a LLM with 7B parameters to generate summaries, which outperforms large GPT-4 and multiple strong baselines. The ablation studies indicate the effectiveness of the iterative calibration process. We analyze the opinion difference in terms of the tendencies of sentences in summaries and use the Natural Language Inference (NLI)-based method to evaluate the faithfulness of summaries. Experiment results show that our method generates summaries with high opinion difference and faithfulness.

The soundness of syntax is an important issue for the paraphrase generation task. Most methods control the syntax of paraphrases by embedding the syntax and semantics in the generation process, which cannot guarantee the syntactical correctness of the results. Different from them, in this paper we investigate the structural patterns of word usages termed as the word composable knowledge and integrate it into the paraphrase generation to control the syntax in an explicit way. This syntax knowledge is pretrained on a large corpus with the dependency relationships and formed as the probabilistic functions on the word-level syntactical soundness. For the sentence-level correctness, we design a hierarchical syntax structure loss to quantitatively verify the syntactical soundness of the paraphrase against the given dependency template. Thus, the generation process can select the appropriate words with consideration on both semantics and syntax. The proposed method is evaluated on a few paraphrase datasets. The experimental results show that the quality of paraphrases by our proposed method outperforms the compared methods, especially in terms of syntax correctness.

As huge numbers of academic papers are published every year, it is critical to be able to recommend high quality papers. The typical evaluation method for papers is to use citation information, which however is not applicable to new papers. To address such a shortcoming, in this paper, we consider a novel perspective on the association between the content difference of a paper, with respect to other papers, and its innovation. Since innovation has often domain-specific characteristics and forms, we introduce the concept of subspace to describe the commonly recognized aspects of paper contents, namely background, methods and results. A set of expert rules are formalized to annotate the differences between papers, based on which a twin-network is proposed for learning the embeddings of papers in different subspaces. A series of empirical studies show that there are clear correlations between a paper influence and its difference with others in those subspaces. The results also show the characteristics of innovation in different scientific disciplines. To take into account information about academic networks for paper recommendation, we propose a graph convolutional neural method to combine the paper content with other related elements, where user interests and academic influences are modeled asymmetric. Experimental results on real datasets show that our method is more effective than other baseline methods for new paper recommendation. We also discuss the characteristics of scientific disciplines and authors to show the effectiveness of modeling the asymmetric user interests and influences. Finally, we verify the reusability of our method on a patent dataset. The results show that it is also applicable to academic data with low-resource features.

In this paper, we consider the embedding problem of Chinese new word with respect to its conceptual definition or description, which is especially important for understanding specialty documents. We present a two-stage model to learn the Chinese new word embedding, where the first encodes the information of character components and context, and the second aggregates the semantics of multiple texts. We perform extensive experiments to verify the proposed method and the results outperform the state of art methods on both direct semantics verification and advanced NLP tasks. Comparing with previous methods that require a corpus or an elaborately designed dataset for learning a new word embedding, our method requires only a few pieces of text and supports the evolution of meanings. We also experimentally verify the effects of different parts of model, the number and types of conceptual texts. Finally, we present some biology texts to illustrate whether the specialty semantics are encoded in the word embedding.

Understanding how knowledge is technically transferred across academic disciplines is very relevant for understanding and facilitating innovation. There are two challenges for this purpose, namely the semantic ambiguity and the asymmetric influence across disciplines. In this paper we investigate knowledge propagation and characterize semantic correlations for cross discipline paper recommendation. We adopt a generative model to represent a paper content as the probabilistic association with an existing hierarchically classified discipline to reduce the ambiguity of word semantics. The semantic correlation across disciplines is represented by an influence function, a correlation metric and a ranking mechanism. Then a user interest is represented as a probabilistic distribution over the target domain semantics and the correlated papers are recommended. Experimental results on real datasets show the effectiveness of our methods. We also discuss the intrinsic factors of results in an interpretable way. Compared with traditional word embedding based methods, our approach supports the evolution of domain semantics that accordingly lead to the update of semantic correlation. Another advantage of our approach is its flexibility and uniformity in supporting user interest specifications by either a list of papers or a query of key words, which is suited for practical scenarios.

BACKGROUND: An enzyme activity is influenced by the external environment. It is important to have an enzyme remain high activity in a specific condition. A usual way is to first determine the optimal condition of an enzyme by either the gradient test or by tertiary structure, and then to use protein engineering to mutate a wild type enzyme for a higher activity in an expected condition. RESULTS: In this paper, we investigate the optimal condition of an enzyme by directly analyzing the sequence. We propose an embedding method to represent the amino acids and the structural information as vectors in the latent space. These vectors contain information about the correlations between amino acids and sites in the aligned amino acid sequences, as well as the correlation with the optimal condition. We crawled and processed the amino acid sequences in the glycoside hydrolase GH11 family, and got 125 amino acid sequences with optimal pH condition. We used probabilistic approximation method to implement the embedding learning method on these samples. Based on these embedding vectors, we design a computational score to determine which one has a better optimal condition for two given amino acid sequences and achieves the accuracy 80% on the test proteins in the same family. We also give the mutation suggestion such that it has a higher activity in an expected environment, which is consistent with the previously professional wet experiments and analysis. CONCLUSION: A new computational method is proposed for the sequence based on the enzyme optimal condition analysis. Compared with the traditional process that involves a lot of wet experiments and requires multiple mutations, this method can give recommendations on the direction and location of amino acid substitution with reference significance for an expected condition in an efficient and effective way.

Abstract We propose an edge‐based method for 6DOF pose tracking of rigid objects using a monocular RGB camera. One of the critical problem for edge‐based methods is to search the object contour points in the image corresponding to the known 3D model points. However, previous methods often produce false object contour points in case of cluttered backgrounds and partial occlusions. In this paper, we propose a novel edge‐based 3D objects tracking method to tackle this problem. To search the object contour points, foreground and background clutter points are first filtered out using edge color cue, then object contour points are searched by maximizing their edge confidence which combines edge color and distance cues. Furthermore, the edge confidence is integrated into the edge‐based energy function to reduce the influence of false contour points caused by cluttered backgrounds and partial occlusions. We also extend our method to multi‐object tracking which can handle mutual occlusions. We compare our method with the recent state‐of‐art methods on challenging public datasets. Experiments demonstrate that our method improves robustness and accuracy against cluttered backgrounds and partial occlusions.

In location based services, predicting users' temporal-spatial behavior is critical for accurate recommendation. In this paper, we adopt a joint embedding (JointE) model to learn the representations of user, location, and users' action in the same latent space. The functionality of a location is the critical factor influencing different elements of the behavior and is learned by an embedding vector encoding crowd behaviors. A user personalized preference is learned from the user historical behaviors and has two features. One is the combination of action and location, which is learned by maximizing the semantic consistency of the observed behaviors. The other is the periodic preference. Inspired by the notion of periodical temporal rules, we introduce the concept of temporal pattern to describe how often users visit places so as to reduce the high temporal variance of behaviors. A projection matrix is introduced to combine the temporal patterns with location functionality. A user behavior is predicted by the joint probability on behavior elements. We conduct experiments against two representative datasets. The results show that our approach outperforms other approaches.

With the popularity of electronic commerce, there are increasing requirements on product comparison services, which collect the similar products information on different platforms for a user reference. Since there are a large quantity of products on each platform, it is necessary to classify the products based on their short descriptions and to learn the relationships between the different categories on multiple platforms. In this paper, we propose the Rectified Topic Classification model to classify products into hierarchical categories based on their short text descriptions. We adopt the topic model to capture the latent features of products from the noisy short descriptions generated by merchants. To reduce the uncertainty of the inferring topic features of a new product, we invoke the topic model several times to get a set of probabilistic feature results and adopt the convolutional neural network for classification. To learn the correlations between two platform categories, the mapping matrix is learned by using a set of seed products. We crawled several real datasets from popular e-commerce platforms and perform experiments to verify our methods. The results show that our method outperforms the related methods.

In this paper we seek to utilize the behavior information and the attribute labels of users from an auxiliary platform to help make attribute prediction. We propose a cross-platform demographic attribute inference model (CDAIM for short), in which we first learn the user representations with the behavior data, and then regulate the vectors of the same users from both platform to be close via a transfer function, and finally train a classifier with the feature vectors and attribute labels of all the users. We conduct extensive experiments on real datasets and the results show that our CDAIM outperforms the baselines.

There are an increasing number of papers published every year. It is desired for researchers to find the new high-quality papers, which is also a challenging task due to the lack of citation information. In this paper, we propose a novel method to predicate a new paper influence by collaboratively learning the latent vectors of paper features and correlations. We propose the concept topic related authority to integrate the dynamic topic model with paper citations so as to learn how content and authors influence a paper quality. We adopt the Factorization Machine method to collaboratively learn the latent vectors of correlations between different paper features. Comparing with traditional methods, it does not require the citation information to evaluate a paper quality, which is appropriate for new published papers. We conduct extensive evaluation against a real dataset crawled from ACM Digital Library. The results show that our method outperforms the other methods.

Most users on social media have intrinsic characteristics, such as interests and political views, that can be exploited to identify and track them. It raises privacy and identity issues in online communities. In this paper we investigate the problem of user identity linkage on two behavior datasets collected from different experiments. Specifically, we focus on user linkage based on users' interaction behaviors with respect to content topics. We propose an embedding method to model a topic as a vector in a latent space so as to interpret its deep semantics. Then a user is modeled as a vector based on his or her interactions with topics. The embedding representations of topics are learned by optimizing the joint-objective: the compatibility between topics with similar semantics, the discriminative abilities of topics to distinguish identities, and the consistency of the same user's characteristics fromtwo datasets. The effectiveness of our method is verified on real-life datasets and the results show that it outperforms related methods.

It is meaningful to recommend appropriate works to a researcher. One important consideration is the relatedness to one's interests. Although it can be expressed by one's query on an academic dataset, there often exists some semantic ambiguity in relatedness computation that are caused by personalized vocabularies of authors and queriers. Another considered aspect is the quality of a publication, which is often justified by the number and quality of its citations. But it is difficult to estimate the potential influence of a new publication when it has few citation. In this paper, we try to solve the two problems in academic recommendation. To reduce the semantic ambiguity, domain knowledge is created by learning the inherit relativity of word usage from an academic dataset. To compute the potential influence of a new publication,we taking into account the contents and venue of a paper, as well as the reputation of its authors. A recommendation algorithm is designed to find the top k related and influential papers for a query from new publications. We verify the proposed method on real dataset.

In this paper, we focus on the link predication problem in social networks. Our approach is based on the observation that there is a large amount of social behavior taking place every day which contains substantial information about user intrinsic characteristics that influence the dynamics of social networks. In order to obtain a deeper understanding of user behavior, we introduce the concept of latent factor to capture the motivation behind social activities. Since user relationships are often asymmetric, we also take into account bilateral user wishes with respect to friend as preferences, which is beyond traditional approaches or overall measurements. Two combination modes are proposed, independent fusion and interdependent fusion, to integrate these hybrid metrics with traditional measurements for link inference. In order to quantify the sensitivity of each element in metrics we use information theory. Experimental results on several real datasets show that our approach has better performance than previous methods.

Insiders are often legal users who are authorized to access system and data. If they misuse their privileges, it would bring great threat to system security. In practice, we could not have any knowledge about fraud pattern in advance, and most malicious behaviors are often in accordance with security rules; thus, it is difficult to predefine regulations for preventing all kinds of frauds. In this paper, we propose a data-driven evaluation model to detect malicious insiders, which audits user behaviors from both parallel and incremental aspects. Users are grouped together according to their positions and responsibilities, based on which the normal pattern is learned. For each user, a routine behavior pattern is also learned for historical assessment. Then, users are evaluated against both group patterns and routine patterns by probabilistic methods. The deviation degree is adopted as an evidence to justify an anomaly. We also recognize the abnormal activities that often make a user behavior much deviate, which can help an administrator revisit security policies or update activity weights in assessment. At last, experiments are performed on several real dataset.

In social media, users are allowed to express their opinions by commenting on an item or rating an item with scores. The collection of user reviews would generate a positive or negative influence to the media audience. Some malicious users may create multiple variant accounts on the same social media so as to influence or manipulate public opinions for business or criminal purposes. To maintain good social environment, it is necessary to find those fake users. In this paper, we investigate the user variants identification problem using both user behavior and item related information. We study the characteristics of user behaviors on social media and introduce two concepts visibility and distingushibility to preliminarily quantify whether a fake user can be identified. To better understand user intention and characteristics, we profile a user with apparent and implicit features, which are extracted from three aspects: User Generated Contents (UGC), user behavior context and item information. Based on these features, we propose the user Variants Identification Problem (VIP) and an identification algorithm, which finds the top-k similar variants in a social media. We evaluate our methods against two real datasets MovieLens and Amazon and make comparison on the effectiveness against different features in identifying user variants.

Personal Health Record (PHR) systems store a large amount of users' health information, which is very sensitive for users. So privacy protection is an important issue for PHR systems. Although information technology audit can find out improper accesses to users' sensitive data that violate their privacy policies, it is difficult for inexperienced users to use audit commands. This paper presents an audit recommendation framework to help users appropriately audit their sensitive records so as to have a good understanding of their privacy situation. For an audit requester, we analyze doctors' historical queries to find the similar users of the requester. Then we analyze the audit commands of those similar users and make some recommendations to the requester. Experiments are performed to verify our method.

To adapt to the demand of agile reengineering of business process in Business Process Management System, a kind of software model which based on message computing is proposed. Aiming at conducting varied modeling of the business process variation type through the message transferred between functional activity and connector, the model is made up by business functional activities and logic calculation modules (connector). The process structure is controlled by the logic calculation between messages and connectors. Process change and stability can be limited through defining the replaceable and restrictive relations between functional activities or process structure. A demonstration of a partial process applied to insurance management information system illustrates the above-mentioned model's support for business process convenient adjustive.

With the development of mobile networks and positioning technologies, location based service is becoming more and more popular, such as location-aware emergency response, advertisement, and car navigation system etc. However, while people enjoy more convenient life provided by location based services, their location privacy may leak. To tackle this problem, many researchers propose different ways to make users' privacy preserving and most of them focus on how to confuse users' true identities, or hide users' exact position while a user is in a cloaking square. The representative method is K-anonymity method which requires not less than k users in the same square when a location based service request occurs. In case there are not enough users in this cloaking square, the condition of the k-anonymity is not satisfied and location privacy would be leak. In this paper, we would tackle the problem of location privacy in a different way by predicating a safe path to users. Our objective is to provide more guarantee of privacy preserving for users under the K-anonymity criteria. We propose a path predicting algorithm from a user's current position to his/her destination. Every point on such path satisfies a location privacy policy under K-anonymity criteria. We also consider an upper threshold so as to balance privacy requirement and performance. Furthermore, the path would be dynamic adjusted according to users' movement and environment changes. Experiments are performed to verify our method and the experiment results show that this method is correct and efficient.

In practice, assigning access permissions to users must satisfy a variety of constraints motivated by business and security requirements. Here, we focus on Role-Based Access Control (RBAC) systems, in which access permissions are assigned to roles and roles are then assigned to users. User-role assignment is subject to role-based constraints, such as mutual exclusion constraints, prerequisite constraints, and role-cardinality constraints. Also, whether a user is qualified for a role depends on whether his/her qualification satisfies the role's requirements. In other words, a role can only be assigned to a certain set of qualified users. In this paper, we study fundamental problems related to access control constraints and user-role assignment, such as determining whether there are conflicts in a set of constraints, verifying whether a user-role assignment satisfies all constraints, and how to generate a valid user-role assignment for a system configuration. Computational complexity results and/or algorithms are given for the problems we consider.

Workforce management is an important issue for human involved collaboration. When multiple urgent events simultaneously happen at different places with unexpected requirements, administrators should schedule workforce under the consideration of both system and user context, such as event requirements, user qualifications, etc. In this paper, we tackle this challenging problem of scheduling workforce for multiple urgent events. We study the Feasible Workforce Assignment Problem (WAP) that determines whether a system state satisfies the requirements of a given multiple-event scenario. Then, we present an efficient polynomial solution to solve it, with a case study to illustrate the effectiveness of our method.

In a multi-domain collaboration environment, an enterprise should authorize different access rights for sensitive information to partners according to its security policies and relationships with them, which may be changed dynamically with the development of transaction and business rules. So, it is emerging as one of the major concerns to effectively manage the authorizations while supporting flexible multi-level collaboration. In this work, we propose an active authorization model for multi-domain cooperation, which introduces the notions of business rules and context parameters to update security policies automatically and satisfy the dynamic context requirements. The algorithms of handling authorization queries and roles mapping are also presented. The system architecture is discussed in detail to implement this model and support interoperation among heterogeneous platforms.

Security and flexibility are two important issues in workflow management systems. The RBAC flexible workflow model is proposed recently which captures above needs, however there exist an open issue of its role hierarchies. With the numbers of users and roles increasing, how to reduce down the complexity of management is emerging concernful. A novel method of role management is presented in this paper that introduce the notation of partial inheritance into the role hierarchies. Compared with other dominating methods, it can efficiently achieve the objectives of role management inflexible workflow

Web-based workflow can be used to facilitate enterprise business process while the security and flexibility are two of the most important aspects in electronic business system. RBAC is regarded as a neutral policy and has been the most popular secure model in recent years. The flexible RBAC workflow model (FRWM) has been proposed recently. It has encapsulated RBAC and workflow together considering both security and flexibility features. In this paper the enforcement of FRWM is introduced. We also present the design and implementation of a practical system for Property Right Exchange (PRES) based on FRWM, in which the flexibility of workflow can be reached through definition and execution while considering the security with RBAC.

With the fast increase of electronic commerce, more and more enterprises and organizations are facilitating their business processes by workflow. To protect information secure and meet the requirement of frequent business changes over time, the security and flexibility become two of the most important aspects that attract attention both from academy and industry. Many related research work on flexible workflow and secure RBAC model were presented respectively. Unfortunately, the analysis and implementation of enforcing RBAC into Web-based flexible workflow have not been mentioned The intention of this paper is to extend RBAC framework further to flexible workflow to support the security, flexibility and expansibility of organization business. A model and its corresponding mechanism are introduced for establishment, dynamical customization and run-time management of the RBAC workflow. A practical system for Property Right Exchange (PRES) based on this model is implemented.