Annotation Model Research Articles

Adaptive Attention Annotation Model: Optimizing the Prediction Path through Dependency Fusion

Previous methods build image annotation model by leveraging three basic dependencies: relations between image and label (image/label), between images (image/image) and between labels (label/label). Even though plenty of researches show that multiple dependencies can to improve annotation performance, different dependencies actually do not work jointly in their diagram, whose performance is largely depending on the result predicted by image/label section. To address this problem, we propose the adaptive attention annotation model (AAAM) to associate these dependencies with the prediction path, which is composed of a series of labels (tags) in the order they are detected. In particular, we optimize the prediction path by detecting the relevant labels from the easy-to-detect to the hard-to-detect, which are found using Binary Cross-Entropy (BCE) and Triplet Margin (TM) losses, respectively. Besides, in order to capture the inforamtion of each label, instead of explicitly extracting regional featutres, we propose the self-attention machanism to implicitly enhance the relevant region and restrain those irrelevant. To validate the effective of the model, we conduct experiments on three well-known public datasets, COCO 2014, IAPR TC-12 and NUSWIDE, and achieve better performance than the state-of-the-art methods.

Model annotation and discovery with the Physiome Model Repository

BackgroundMathematics and Phy sics-based simulation models have the potential to help interpret and encapsulate biological phenomena in a computable and reproducible form. Similarly, comprehensive descriptions of such models help to ensure that such models are accessible, discoverable, and reusable. To this end, researchers have developed tools and standards to encode mathematical models of biological systems enabling reproducibility and reuse, tools and guidelines to facilitate semantic description of mathematical models, and repositories in which to archive, share, and discover models. Scientists can leverage these resources to investigate specific questions and hypotheses in a more efficient manner.ResultsWe have comprehensively annotated a cohort of models with biological semantics. These annotated models are freely available in the Physiome Model Repository (PMR). To demonstrate the benefits of this approach, we have developed a web-based tool which enables users to discover models relevant to their work, with a particular focus on epithelial transport. Based on a semantic query, this tool will help users discover relevant models, suggesting similar or alternative models that the user may wish to explore or use.ConclusionThe semantic annotation and the web tool we have developed is a new contribution enabling scientists to discover relevant models in the PMR as candidates for reuse in their own scientific endeavours. This approach demonstrates how semantic web technologies and methodologies can contribute to biomedical and clinical research. The source code and links to the web tool are available at https://github.com/dewancse/model-discovery-tool

Power-law Behavior of the Alternative Splicing of Exons in Human Transcriptome

Aims: To establish the common rules of exon combinatorics during RNA splicing.
 Study Design: Inferring a plausible statistical model of exon combinatorics from the annotated models of human genes during RNA splicing.
 Place and Duration of Study: Department of Genetics (Belarusian State University), Proteome and Genome Research Unit (Luxembourg Institute of Health), Department of Genetics (Lomonosov Moscow State University) and Moscow Center of Experimental Embryology and Reproductive Biotechnologies, between January 2017 and July 2019.
 Methodology: We used human mRNA and EST sequences from GenBank (1093522 unique records in total) and linear models of the human genes from Ensembl (58051 genes), AceView (72384 genes), ECgene (57172 genes), NCBI RefSeq (54262 genes), UCSC Genome Browser (58037 genes) and VEGA (54950 genes) to calculate a combinatorial index of human exons. We inferred the most plausible statistical model describing the distribution of combinatorial index of human exons using Clauset’s mathematical formalism. Predictors of the combinatorial index values and functional outcomes of the predefined behavior of exons during splicing were also determined.
 Results: Power-law is the most plausible statistical model describing the combinatorics of exons during RNA splicing. The combinatorial index of human exons is defined by more than 90% by the 138 features that have different importance. The most important of these features are the abundance of exon in transcripts, the strength of splice sites, the rank of exon in transcripts and the type of exon. Analysis of the marginal effects shows that different values of the same feature have unequal influence on the combinatorial index of human exons. Power-law behavior of exons during RNA splicing pre-determines structural diversity of transcripts, low sensitivity of splicing process to random perturbations and its high vulnerability to manipulation with highly combinative exons.
 Conclusion: Exons widely involved in alternative splicing are a part of the common power-law phenomenon in human cells. The power-law behavior of exons during RNA splicing gives the unique characteristics to human genes.

Genes and gene models, an important distinction

Genome sequencing has fundamentally changed how plant biologists think about genes. All or nearly all genes can ultimately be associated with a gene model. However, many gene models appear to play little or no role in the traits of an organism. A range of structural, molecular, population and evolutionary features all show a separation between genes with known phenotypes and the overall set of annotated gene models. These different features could be combined to develop models to distinguish the genes that determine the traits of plants from the subset gene other annotated gene models which are unlikely to play a role in doing so. Efforts to identify the subset of annotated gene models likely involved in specifying the characteristics of plants would help aid a wide range of researchers.

20. RELATIONSHIP BETWEEN EUPLOIDY RATES AND D5 KIDSCORE™ OF BLASTOCYSTS DERIVED FROM EMBRYOSCOPE

Introduction Time-lapse microscopy (TLM) offers continuous monitoring of embryo development in an undisturbed culture condition, minimalising the stress imposed by changes in temperature, oxygen and pH in the culture medium; hence optimising culture environment. Moreover, the availability of morphokinetics assessment and annotation models such as KIDScore™ in TLM can help predict the implantation potential of an embryo. Several studies have also shown a reduction in pregnancy loss in IVF with TLM. However, the association between annotation models and euploidy rates has not been reported. This is a retrospective study to assess the relationship between euploidy rate and D5 KIDScore™ of blastocysts derived from IVF in Alpha Fertility Centre, Malaysia between January to December 2018. Material and Methods Eight hundred and seven (807) patients (age ranged 25 – 45) had their embryos cultured to blastocyst stage. All blastocysts were cultured in Embryoscope™ according to the manufacturer's specifications (Vitrolife, Sweden). Any utilisable blastocysts were annotated and scored using D5 KIDScore™. Blastocysts scored between 1.0 – 5.9 were categorised as Group A while those scored between 6.0 – 9.9 were categorised as Group B. Three hundred and seventy-one (371) blastocysts from Group A and 1783 blastocysts from Group B which was at least fair grade (using Gardner blastocyst grading system) were biopsied and screened using Next Generation Sequencing (Ion Torrent, USA). The mean age of patients from Group A and B were 34.4 and 34.1 respectively (p>0.05). Results The euploidy rate of blastocysts from Group A and B were 48.8 % and 59.3% respectively whereas the aneuploidy rate of blastocysts from Group A and B were 51.2% and 40.7% respectively. Both euploidy and aneuploidy rates between Group A and Group B were statistically significant (p=0.0002). Conclusions This retrospective study demonstrates that blastocysts with a D5 KIDScore™ of 6.0 – 9.9 have a higher euploidy rate compared to blastocysts with a D5 KIDScore™ 1.0 – 5.9. It is thus recommended for patients who had time-lapse embryo monitoring without PGT-A to transfer blastocysts with a D5 KIDScore™ 6.0 – 9.9 in lieu of the higher euploidy rate. Nevertheless, further studies with a larger sample size should be performed to confirm the above.

5. EUPLOID BLASTOCYSTS HAVE SIMILAR CLINICAL OUTCOME REGARDLESS OF THEIR D5 KIDSCORE™

Introduction Time-lapse monitoring (TLM) uses morphokinetics assessment and annotation models such as KIDScore™ to help predict the implantation potential of a blastocyst. Meanwhile, preimplantation Genetic Testing for Aneuploidy (PGT-A) determines the chromosomal status of an embryo, allowing the identification of embryos with the correct number of chromosomes. Theoretically, combining both of these advanced technologies in IVF may enable the selection of a competent blastocyst for transfer, thus maximising the chance of a successful pregnancy. To the best of our knowledge, the significance of KIDScore™ annotation model in euploid blastocysts has not been studied. This is a retrospective study to assess the clinical outcomes of elective frozen embryo transfer (eFET) of euploid blastocysts in various D5 KIDScore™ groups in Alpha Fertility Centre, Malaysia, between January to November 2018. Material and Methods Two hundred and twenty-one (221) patients aged 18 – 45 years (mean age 30.8 years) had their embryos cultured to blastocyst stage in the Embryoscope™ incubator (Vitrolife, Sweden) and screened using Next Generation Sequencing (Ion Torrent, USA). In order to accurately determine the D5 KIDScore™ of the euploid blastocysts transferred, only blastocysts intended for single frozen embryo transfer were included in this study. All the 221 blastocysts transferred were categorised according to D5 KIDScore™ ranged between 1.0 – 4.0 (Group A, n = 4), 4.1 – 5.0 (Group B, n = 4), 5.1 – 6.0 (Group C, n = 15), 6.1 – 7.0 (Group D, n = 49), 7.1 – 8.0 (Group E, n = 83), 8.1 – 9.0 (Group F, n = 53), and 9.1 – 9.9 (Group G, n = 13) respectively. Clinical pregnancy and number of gestational sacs were determined by ultrasound. Results The clinical pregnancy rates and implantation rates for Group A, B, C, D, E, F and G were 25.0%, 75.0%, 66.7%, 46.9%, 60.2%, 60.4% and 53.8% respectively. The miscarriage rates for Group A, B, C, D, E, F and G were 0.0%, 33.3%, 20.0%, 4.3%, 6.0%, 12.5% and 14.3% respectively. There was no significant difference in clinical pregnancy rate, implantation rate and miscarriage rate between all groups (p>0.05). Conclusions Our study shows no significant difference in clinical pregnancy rate, implantation rate and miscarriage rate across all D5 KIDScore™ groups for euploid blastocysts. Euploid blastocysts with lower D5 KIDScore™ gives comparable clinical outcomes of that to blastocyst with higher D5 KIDScore™ and should be considered in the selection for transfer. D5 KIDScore™ may serve as a guide for selection of blastocyst with high implantation potential. However, PGT-A still remains the more reliable method compared to solely relying on D5 KIDScore™ in TLM for blastocyst selection. Nevertheless, further studies with a larger sample size should be carried out to ascertain the significance of D5 KIDScore™ in blastocyst that was screened euploid using PGT-A.

End-to-End Automatic Image Annotation Based on Deep CNN and Multi-Label Data Augmentation

Automatic image annotation is a key step in image retrieval and image understanding. In this paper, we present an end-to-end automatic image annotation method based on a deep convolutional neural network (CNN) and multi-label data augmentation. Different from traditional annotation models that usually perform feature extraction and annotation as two independent tasks, we propose an end-to-end automatic image annotation model based on deep CNN (E2E-DCNN). E2E-DCNN transforms the image annotation problem into a multi-label learning problem. It uses a deep CNN structure to carry out the adaptive feature learning before constructing the end-to-end annotation structure using multiple cross-entropy loss functions for training. It is difficult to train a deep CNN model using small-scale datasets or scale up multi-label datasets using traditional data augmentation methods; hence, we propose a multi-label data augmentation method based on Wasserstein generative adversarial networks (ML-WGAN). The ML-WGAN generator can approximate the data distribution of a single multi-label image. The images generated by ML-WGAN can assist in the reduction of the over-fitting problem of training a deep CNN model and enhance the generalization ability of the trained CNN model. We optimize the network structure by using deformable convolution and spatial pyramid pooling. We experiment the proposed E2E-DCNN model with data augmentation by the proposed ML-WGAN on several public datasets. The experimental results demonstrate that the proposed model outperforms the state-of-the-art automatic image annotation models.

Learning From eHealth Implementations Through "Implementomics": A Multidimensional Annotation Model Applied to eHealth Projects of the RAFT Network.

The implementation of digital health technologies has increased globally, producing substantial amounts of information and knowledge. While there are still areas in digital health that are understudied, concurrently there is an exponential increase in published articles, guidelines, methods, projects, and experiences, many of which fail to reach critical mass (pilotitis). Semantically describing and documenting this implementation knowledge and the effectiveness of these tools will help to avoid the duplication of efforts, to reduce preventable implementation obstacles, and to assure that investments are targeted to the most important technological innovations. The RAFT annotation model, presented in this paper, enables to semantically describe all elements of various outputs and implementation projects that were developed, are used, or are part of the RAFT network. This model was initially developed to annotate various implementations and outputs of the RAFT network to facilitate knowledge documentation and sharing, and to be used as a proof of concept for the Implementome. The Implementome will be an interconnected knowledge system that enables the user to navigate on multiple dimensions through metadata annotated projects, people, and information, and can serve as base for consensus building, best practices and guidelines. The RAFT annotation model can be further developed to enable the annotation of outputs, implementations, people, initiatives, and projects of the digital health domain in general.

Towards a Flexible Deep Learning Method for Automatic Detection of Clinically Relevant Multi-Modal Events in the Polysomnogram.

Much attention has been given to automatic sleep staging algorithms in past years, but the detection of discrete events in sleep studies is also crucial for precise characterization of sleep patterns and possible diagnosis of sleep disorders. We propose here a deep learning model for automatic detection and annotation of arousals and leg movements. Both of these are commonly seen during normal sleep, while an excessive amount of either is linked to disrupted sleep patterns, excessive daytime sleepiness impacting quality of life, and various sleep disorders. Our model was trained on 1,485 subjects and tested on 1,000 separate recordings of sleep. We tested two different experimental setups and found optimal arousal detection was attained by including a recurrent neural network module in our default model with a dynamic default event window (F1 = 0.75), while optimal leg movement detection was attained using a static event window (F1 = 0.65). Our work show promise while still allowing for improvements. Specifically, future research will explore the proposed model as a general-purpose sleep analysis model.

Towards cross-platform interoperability for machine-assisted text annotation.

In this paper, we investigate cross-platform interoperability for natural language processing (NLP) and, in particular, annotation of textual resources, with an eye toward identifying the design elements of annotation models and processes that are particularly problematic for, or amenable to, enabling seamless communication across different platforms. The study is conducted in the context of a specific annotation methodology, namely machine-assisted interactive annotation (also known as human-in-the-loop annotation). This methodology requires the ability to freely combine resources from different document repositories, access a wide array of NLP tools that automatically annotate corpora for various linguistic phenomena, and use a sophisticated annotation editor that enables interactive manual annotation coupled with on-the-fly machine learning. We consider three independently developed platforms, each of which utilizes a different model for representing annotations over text, and each of which performs a different role in the process.

Model Law – Prohibiting Reparative Practices

This paper proposes a detailed and annotated model law on prohibiting conversion practices. Conversion practices, also known as conversion or reparative therapy, attempt to alter people’s gender identities or sexual orientation and/or make them cisgender and/or heterosexual. The proposed model law seeks to improve upon existing legislation, which is insufficiently specific. The paper includes details annotations explaining the rationale for including each provision of the model law and wording them in this manner.

An annotation and modeling schema for prescription regimens

BackgroundWe introduce TranScriptML, a semantic representation schema for prescription regimens allowing various properties of prescriptions (e.g. dose, frequency, route) to be specified separately and applied (manually or automatically) as annotations to patient instructions. In this paper, we describe the annotation schema, the curation of a corpus of prescription instructions through a manual annotation effort, and initial experiments in modeling and automated generation of TranScriptML representations.ResultsTranScriptML was developed in the process of curating a corpus of 2914 ambulatory prescriptions written within the Partners Healthcare network, and its schema is informed by the content of that corpus. We developed the representation schema as a novel set of semantic tags for prescription concept categories (e.g. frequency); each tag label is defined with an accompanying attribute framework in which the meaning of tagged concepts can be specified in a normalized fashion. We annotated a subset (1746) of this dataset using cross-validation and reconciliation between multiple annotators, and used Conditional Random Field machine learning and various other methods to train automated annotation models based on the manual annotations. The TranScriptML schema implementation, manual annotation, and machine learning were all performed using the MITRE Annotation Toolkit (MAT). We report that our annotation schema can be applied with varying levels of pairwise agreement, ranging from low agreement levels (0.125 F for the relatively rare REFILL tag) to high agreement levels approaching 0.9 F for some of the more frequent tags. We report similarly variable scores for modeling tag labels and spans, averaging 0.748 F-measure with balanced precision and recall. The best of our various attribute modeling methods captured most attributes with accuracy above 0.9.ConclusionsWe have described an annotation schema for prescription regimens, and shown that it is possible to annotate prescription regimens at high accuracy for many tag types. We have further shown that many of these tags and attributes can be modeled at high accuracy with various techniques. By structuring the textual representation through annotation enriched with normalized values, the text can be compared against the pharmacist-entered structured data, offering an opportunity to detect and correct discrepancies.

A probabilistic topic model for event-based image classification and multi-label annotation

We propose an enhanced latent topic model based on latent Dirichlet allocation and convolutional neural nets for event classification and annotation in images. Our model builds on the semantic structure relating events, objects and scenes in images. Based on initial labels extracted from convolution neural networks (CNNs), and possibly user-defined tags, we estimate the event category and final annotation of an image through a refinement process based on the expectation–maximization (EM) algorithm. The EM steps allow to progressively ascertain the class category and refine the final annotation of the image. Our model can be thought of as a two-level annotation system, where the first level derives the image event from CNN labels and image tags and the second level derives the final annotation consisting of event-related objects/scenes. Experimental results show that the proposed model yields better classification and annotation performance in the two standard datasets: UIUC-Sports and WIDER.

Ontological Semantic Annotation of an English Corpus Through Condition Random Fields

One way to increase the understanding of texts by machines is through adding semantic information to lexical items by including metadata tags, a process also called semantic annotation. There are several semantic aspects that can be added to the words, among them the information about the nature of the concept denoted through the association with a category of an ontology. The application of ontologies in the annotation task can span multiple domains. However, this particular research focused its approach on top-level ontologies due to its generalizing characteristic. Considering that annotation is an arduous task that demands time and specialized personnel to perform it, much is done on ways to implement the semantic annotation automatically. The use of machine learning techniques are the most effective approaches in the annotation process. Another factor of great importance for the success of the training process of the supervised learning algorithms is the use of a sufficiently large corpus and able to condense the linguistic variance of the natural language. In this sense, this article aims to present an automatic approach to enrich documents from the American English corpus through a CRF model for semantic annotation of ontologies from Schema.org top-level. The research uses two approaches of the model obtaining promising results for the development of semantic annotation based on top-level ontologies. Although it is a new line of research, the use of top-level ontologies for automatic semantic enrichment of texts can contribute significantly to the improvement of text interpretation by machines.

Association of Variants in PINX1 and TREM2 With Late-Onset Alzheimer Disease

Genetic causes of late-onset Alzheimer disease (LOAD) are not completely explained by known genetic loci. Whole-exome and whole-genome sequencing can improve the understanding of the causes of LOAD and provide initial steps required to identify potential therapeutic targets. To identify the genetic loci for LOAD across different ethnic groups. This multicenter cohort study was designed to analyze whole-exome sequencing data from a multiethnic cohort using a transethnic gene-kernel association test meta-analysis, adjusted for sex, age, and principal components, to identify genetic variants associated with LOAD. A meta-analysis was conducted on the results of 2 independent studies of whole-exome and whole-genome sequence data from individuals of European ancestry. This group of European American, African American, and Caribbean Hispanic individuals participating in an urban population-based study were the discovery cohort; the additional cohorts included affected individuals and control participants from 2 publicly available data sets. Replication was achieved using independent data sets from Caribbean Hispanic families with multiple family members affected by LOAD and the International Genetics of Alzheimer Project. Late-onset Alzheimer disease. The discovery cohort included 3595 affected individuals, while the additional cohorts included 5931 individuals with LOAD and 5504 control participants. Of 3916 individuals in the discovery cohort, we included 3595 individuals (1397 with LOAD and 2198 cognitively healthy controls; 2451 [68.2%] women; mean [SD] age, 80.3 [6.83] years). Another 321 individuals (8.2%) were excluded because of non-LOAD diagnosis, age younger than 60 years, missing covariates, duplicate data, or genetic outlier status. Gene-based tests that compared affected individuals (n = 7328) and control participants (n = 7702) and included only rare and uncommon variants annotated as having moderate-high functional effect supported PINX1 (8p23.1) as a locus with gene-wide significance (P = 2.81 × 10-6) after meta-analysis across the 3 studies. The PINX1 finding was replicated using data from the family-based study and the International Genetics of Alzheimer Project. Full meta-analysis of discovery and replication cohorts reached a P value of 6.16 × 10-7 for PINX1 (in 7620 affected individuals vs 7768 control participants). We also identified TREM2 in an annotation model that prioritized highly deleterious variants with a combined annotation dependent depletion greater than 20 (P= 1.0 × 10-7). This gene-based, transethnic approach identified PINX1, a gene involved in telomere integrity, and TREM2, a gene with a product of an immune receptor found in microglia, as associated with LOAD. Both genes have well-established roles in aging and neurodegeneration.

IntFOLD: an integrated web resource for high performance protein structure and function prediction.

The IntFOLD server provides a unified resource for the automated prediction of: protein tertiary structures with built-in estimates of model accuracy (EMA), protein structural domain boundaries, natively unstructured or disordered regions in proteins, and protein–ligand interactions. The component methods have been independently evaluated via the successive blind CASP experiments and the continual CAMEO benchmarking project. The IntFOLD server has established its ranking as one of the best performing publicly available servers, based on independent official evaluation metrics. Here, we describe significant updates to the server back end, where we have focused on performance improvements in tertiary structure predictions, in terms of global 3D model quality and accuracy self-estimates (ASE), which we achieve using our newly improved ModFOLD7_rank algorithm. We also report on various upgrades to the front end including: a streamlined submission process, enhanced visualization of models, new confidence scores for ranking, and links for accessing all annotated model data. Furthermore, we now include an option for users to submit selected models for further refinement via convenient push buttons. The IntFOLD server is freely available at: http://www.reading.ac.uk/bioinf/IntFOLD/.

Heritage image annotation via collective knowledge

The automatic image annotation can provide semantic illustrations to understand image contents, and builds a foundation to develop algorithms that can search images within a large database. However, most current methods focus on solving the annotation problem by modeling the image visual content and tag semantic information, which overlooks the additional information, such as scene descriptions and locations. Moreover, the majority of current annotation datasets are visually consistent and only annotated by common visual objects and attributes, which makes the classic methods vulnerable to handle the more diverse image annotation. To address above issues, we propose to annotate images via collective knowledge, that is, we uncover relationships between the image and its neighbors by measuring similarities among metadata and conduct the metric learning to obtain the representations of image contents, we also generate semantic representations for images given collective semantic information from their neighbors. Two representations from different paradigms are embedded together to train an annotation model. We ground our model on the heritage image collection we collected from the library online open data. Annotations on the heritage image collection are not limited to common visual objects, and are highly relevant to historical events, and the diversity of the heritage image content is much larger than the current datasets, which makes it more suitable for this task. Comprehensive experimental results on the benchmark dataset indicate that the proposed model achieves the best performance compared to baselines and state-of-the-art methods.

OpinionML—Opinion Markup Language for Sentiment Representation

It is the age of the social web, where people express themselves by giving their opinions about various issues, from their personal life to the world’s political issues. This process generates a lot of opinion data on the web that can be processed for valuable information, and therefore, semantic annotation of opinions becomes an important task. Unfortunately, existing opinion annotation schemes have failed to satisfy annotation challenges and cannot even adhere to the basic definition of opinion. Opinion holders, topical features and temporal expressions are major components of an opinion that remain ignored in existing annotation schemes. In this work, we propose OpinionML, a new Markup Language, that aims to compensate for the issues that existing typical opinion markup languages fail to resolve. We present a detailed discussion about existing annotation schemes and their associated problems. We argue that OpinionML is more robust, flexible and easier for annotating opinion data. Its modular approach while implementing a logical model provides us with a flexible and easier model of annotation. OpinionML can be considered a step towards “information symmetry”. It is an effort for consistent sentiment annotations across the research community. We perform experiments to prove robustness of the proposed OpinionML and the results demonstrate its capability of retrieving significant components of opinion segments. We also propose OpinionML ontology in an effort to make OpinionML more inter-operable. The ontology proposed is more complete than existing opinion ontologies like Marl and Onyx. A comprehensive comparison of the proposed ontology with existing sentiment ontologies Marl and Onyx proves its worth.

Rhea, an expert curated resource of biochemical reactions for enzyme annotation and genome-scale metabolic modeling

Rhea, an expert curated resource of biochemical reactions for enzyme annotation and genome-scale metabolic modeling

Web image annotation based on Tri-relational Graph and semantic context analysis

Web image annotation has became a hot research topic owing to massive image data and abundant semantic context. In this paper, we propose a Tri-relational Graph (TG) model for web image annotation, which comprises the image data graph, the region data graph and the label graph as subgraphs, and connects them by an additional tripartite graph induced from image segmentation results and label assignments. Through analyzing the global visual similarity between images, the visual similarity between regions, the semantic correlations between labels and the relationships between the three subgraphs by TG model, we perform multilevel Random Walk with Restart algorithm on TG to produce vertex-to-vertex relevance, including image-to-region, region-to-label and image-to-label relevances. Then semi-supervised learning is used to predict labels for unannotated image regions by inserting unlabeled images and their regions into TG. In addition, we also analyze the text context information of web image and achieve the semantic and proper nouns for the further label expansion through WordNet. Experiments on public web images datasets demonstrate that our proposed TG model and multilevel RWR algorithm can achieve good performance on image region annotation and outperform the similar image annotation methods. Moreover label expansion by web semantic context analysis can achieve more accurate and abundant annotation results.