Search results for: embeddings

Authors: Colin Smith, Linsey S Passarella

Abstract:

Due to the proliferation and decentralized nature of academic publication, no widely accepted scheme exists for organizing papers by their scientific field of study (FoS) to the author’s best knowledge. While many academic journals require author provided keywords for papers, these keywords range wildly in scope and are not consistent across papers, journals, or field domains, necessitating alternative approaches to paper classification. Past attempts to perform field-of-study (FoS) classification on scientific texts have largely used a-hierarchical FoS schemas or ignored the schema’s inherently hierarchical structure, e.g. by compressing the structure into a single layer for multi-label classification. In this paper, we introduce an application of a Layered Neural Network (LNN) to the problem of performing supervised hierarchical classification of scientific fields of study (FoS) on research papers. In this approach, paper embeddings from a pretrained language model are fed into a top-down LNN. Beginning with a single neural network (NN) for the highest layer of the class hierarchy, each node uses a separate local NN to classify the subsequent subfield child node(s) for an input embedding of concatenated paper titles and abstracts. We compare our LNN-FOS method to other recent machine learning methods using the Microsoft Academic Graph (MAG) FoS hierarchy and find that the LNN-FOS offers increased classification accuracy at each FoS hierarchical level.

Keywords: hierarchical classification, layer neural network, scientific field of study, scientific taxonomy

Procedia PDF Downloads 103

Authors: Joseph Gear, Yue Xu, Ernest Foo, Praveen Gauravaran, Zahra Jadidi, Leonie Simpson

Abstract:

Deep learning methods have been seeing an increasing application to the long-standing security research goal of automatic vulnerability detection for source code. Attention, however, must still be paid to the task of producing vector representations for source code (code embeddings) as input for these deep learning models. Graphical representations of code, most predominantly Abstract Syntax Trees and Code Property Graphs, have received some use in this task of late; however, for very large graphs representing very large code snip- pets, learning becomes prohibitively computationally expensive. This expense may be reduced by intelligently pruning this input to only vulnerability-relevant information; however, little research in this area has been performed. Additionally, most existing work comprehends code based solely on the structure of the graph at the expense of the information contained by the node in the graph. This paper proposes Semantic-enhanced Code Embedding for Vulnerability Discovery (SCEVD), a deep learning model which uses semantic-based feature selection for its vulnerability classification model. It uses information from the nodes as well as the structure of the code graph in order to select features which are most indicative of the presence or absence of vulnerabilities. This model is implemented and experimentally tested using the SARD Juliet vulnerability test suite to determine its efficacy. It is able to improve on existing code graph feature selection methods, as demonstrated by its improved ability to discover vulnerabilities.

Keywords: code representation, deep learning, source code semantics, vulnerability discovery

Procedia PDF Downloads 131

Authors: Zunaira Zaman, David Bohannon

Abstract:

This study explores the application of machine learning (ML) for detecting security vulnerabilities in source code. The research aims to assist organizations with large application portfolios and limited security testing capabilities in prioritizing security activities. ML-based approaches offer benefits such as increased confidence scores, false positives and negatives tuning, and automated feedback. The initial approach using natural language processing techniques to extract features achieved 86% accuracy during the training phase but suffered from overfitting and performed poorly on unseen datasets during testing. To address these issues, the study proposes using the abstract syntax tree (AST) for Java and C++ codebases to capture code semantics and structure and generate path-context representations for each function. The Code2Vec model architecture is used to learn distributed representations of source code snippets for training a machine-learning classifier for vulnerability prediction. The study evaluates the performance of the proposed methodology using two datasets and compares the results with existing approaches. The Devign dataset yielded 60% accuracy in predicting vulnerable code snippets and helped resist overfitting, while the Juliet Test Suite predicted specific vulnerabilities such as OS-Command Injection, Cryptographic, and Cross-Site Scripting vulnerabilities. The Code2Vec model achieved 75% accuracy and a 98% recall rate in predicting OS-Command Injection vulnerabilities. The study concludes that even partial AST representations of source code can be useful for vulnerability prediction. The approach has the potential for automated intelligent analysis of source code, including vulnerability prediction on unseen source code. State-of-the-art models using natural language processing techniques and CNN models with ensemble modelling techniques did not generalize well on unseen data and faced overfitting issues. However, predicting vulnerabilities in source code using machine learning poses challenges such as high dimensionality and complexity of source code, imbalanced datasets, and identifying specific types of vulnerabilities. Future work will address these challenges and expand the scope of the research.

Keywords: code embeddings, neural networks, natural language processing, OS command injection, software security, code properties

Procedia PDF Downloads 79

Authors: Despoina Mouratidis, Katia Lida Kermanidis

Abstract:

Machine translation (MT) is a very important task in Natural Language Processing (NLP). MT evaluation is crucial in MT development, as it constitutes the means to assess the success of an MT system, and also helps improve its performance. Several methods have been proposed for the evaluation of (MT) systems. Some of the most popular ones in automatic MT evaluation are score-based, such as the BLEU score, and others are based on lexical similarity or syntactic similarity between the MT outputs and the reference involving higher-level information like part of speech tagging (POS). This paper presents a language-independent machine learning framework for classifying pairwise translations. This framework uses vector representations of two machine-produced translations, one from a statistical machine translation model (SMT) and one from a neural machine translation model (NMT). The vector representations consist of automatically extracted word embeddings and string-like language-independent features. These vector representations used as an input to a multi-layer neural network (NN) that models the similarity between each MT output and the reference, as well as between the two MT outputs. To evaluate the proposed approach, a professional translation and a "ground-truth" annotation are used. The parallel corpora used are English-Greek (EN-GR) and English-Italian (EN-IT), in the educational domain and of informal genres (video lecture subtitles, course forum text, etc.) that are difficult to be reliably translated. They have tested three basic deep learning (DL) architectures to this schema: (i) fully-connected dense, (ii) Convolutional Neural Network (CNN), and (iii) Long Short-Term Memory (LSTM). Experiments show that all tested architectures achieved better results when compared against those of some of the well-known basic approaches, such as Random Forest (RF) and Support Vector Machine (SVM). Better accuracy results are obtained when LSTM layers are used in our schema. In terms of a balance between the results, better accuracy results are obtained when dense layers are used. The reason for this is that the model correctly classifies more sentences of the minority class (SMT). For a more integrated analysis of the accuracy results, a qualitative linguistic analysis is carried out. In this context, problems have been identified about some figures of speech, as the metaphors, or about certain linguistic phenomena, such as per etymology: paronyms. It is quite interesting to find out why all the classifiers led to worse accuracy results in Italian as compared to Greek, taking into account that the linguistic features employed are language independent.

Keywords: machine learning, machine translation evaluation, neural network architecture, pairwise classification

Procedia PDF Downloads 105

Authors: Kien Vu, Jien Min Soh, Mohamed Jahangir Abubacker, Piyawut Pattamanon, Soojin Lee, Suvro Banerjee

Abstract:

Recent advances in Generative Artificial Intelligence (GenAI), in particular Large Language Models (LLMs) have shown promise to disrupt multiple industries at scale. However, LLMs also present unique challenges, notably, these so-called "hallucination" which is the generation of outputs that are not grounded in the input data that hinders its adoption into production. Common practice to mitigate hallucination problem is utilizing Retrieval Agmented Generation (RAG) system to ground LLMs'response to ground truth. RAG converts the grounding documents into embeddings, retrieve the relevant parts with vector similarity between user's query and documents, then generates a response that is not only based on its pre-trained knowledge but also on the specific information from the retrieved documents. However, the RAG system is not suitable for tabular data and subsequent data analysis tasks due to multiple reasons such as information loss, data format, and retrieval mechanism. In this study, we have explored a novel methodology that combines planning-and-execution and code generation agents to enhance LLMs' data analysis capabilities. The approach enables LLMs to autonomously dissect a complex analytical task into simpler sub-tasks and requirements, then convert them into executable segments of code. In the final step, it generates the complete response from output of the executed code. When deployed beta version on DataSense, the property insight tool of PropertyGuru, the approach yielded promising results, as it was able to provide market insights and data visualization needs with high accuracy and extensive coverage by abstracting the complexities for real-estate agents and developers from non-programming background. In essence, the methodology not only refines the analytical process but also serves as a strategic tool for real estate professionals, aiding in market understanding and enhancement without the need for programming skills. The implication extends beyond immediate analytics, paving the way for a new era in the real estate industry characterized by efficiency and advanced data utilization.

Keywords: large language model, reasoning, planning and execution, code generation, natural language processing, prompt engineering, data analysis, real estate, data sense, PropertyGuru

Procedia PDF Downloads 46

Authors: Szymon Rutkowski, Tomasz Korbak

Abstract:

We present a reinforcement-learning-based method of training neural machine translation models without parallel corpora. The standard encoder-decoder approach to machine translation suffers from two problems we aim to address. First, it needs parallel corpora, which are scarce, especially for low-resource languages. Second, it lacks psychological plausibility of learning procedure: learning a foreign language is about learning to communicate useful information, not merely learning to transduce from one language’s 'encoding' to another. We instead pose the problem of learning to translate as learning a policy in a communication game between two agents: the translator and the classifier. The classifier is trained beforehand on a natural language inference task (determining the entailment relation between a premise and a hypothesis) in the target language. The translator produces a sequence of actions that correspond to generating translations of both the hypothesis and premise, which are then passed to the classifier. The translator is rewarded for classifier’s performance on determining entailment between sentences translated by the translator to disciple’s native language. Translator’s performance thus reflects its ability to communicate useful information to the classifier. In effect, we train a machine translation model without the need for parallel corpora altogether. While similar reinforcement learning formulations for zero-shot translation were proposed before, there is a number of improvements we introduce. While prior research aimed at grounding the translation task in the physical world by evaluating agents on an image captioning task, we found that using a linguistic task is more sample-efficient. Natural language inference (also known as recognizing textual entailment) captures semantic properties of sentence pairs that are poorly correlated with semantic similarity, thus enforcing basic understanding of the role played by compositionality. It has been shown that models trained recognizing textual entailment produce high-quality general-purpose sentence embeddings transferrable to other tasks. We use stanford natural language inference (SNLI) dataset as well as its analogous datasets for French (XNLI) and Polish (CDSCorpus). Textual entailment corpora can be obtained relatively easily for any language, which makes our approach more extensible to low-resource languages than traditional approaches based on parallel corpora. We evaluated a number of reinforcement learning algorithms (including policy gradients and actor-critic) to solve the problem of translator’s policy optimization and found that our attempts yield some promising improvements over previous approaches to reinforcement-learning based zero-shot machine translation.

Keywords: agent-based language learning, low-resource translation, natural language inference, neural machine translation, reinforcement learning

Procedia PDF Downloads 103

Authors: Varun Agarwal

Abstract:

Introduction: With the advent of personalized medicine, histopathological review of whole slide images (WSIs) for cancer diagnosis presents an exceedingly time-consuming, complex task. Specifically, detecting metastatic regions in WSIs of sentinel lymph node biopsies necessitates a full-scanned, holistic evaluation of the image. Thus, digital pathology, low-level image manipulation algorithms, and machine learning provide significant advancements in improving the efficiency and accuracy of WSI analysis. Using Camelyon16 data, this paper proposes a deep learning pipeline to automate and ameliorate breast cancer metastasis localization and WSI classification. Methodology: The model broadly follows five stages -region of interest detection, WSI partitioning into image tiles, convolutional neural network (CNN) image-segment classifications, probabilistic mapping of tumor localizations, and further processing for whole WSI classification. Transfer learning is applied to the task, with the implementation of Inception-ResNetV2 - an effective CNN classifier that uses residual connections to enhance feature representation, adding convolved outputs in the inception unit to the proceeding input data. Moreover, in order to augment the performance of the transfer learning CNN, a stack of convolutional denoising autoencoders (CDAE) is applied to produce embeddings that enrich image representation. Through a saliency-detection algorithm, visual training segments are generated, which are then processed through a denoising autoencoder -primarily consisting of convolutional, leaky rectified linear unit, and batch normalization layers- and subsequently a contrast-normalization function. A spatial pyramid pooling algorithm extracts the key features from the processed image, creating a viable feature map for the CNN that minimizes spatial resolution and noise. Results and Conclusion: The simplified and effective architecture of the fine-tuned transfer learning Inception-ResNetV2 network enhanced with the CDAE stack yields state of the art performance in WSI classification and tumor localization, achieving AUC scores of 0.947 and 0.753, respectively. The convolutional feature retention and compilation with the residual connections to inception units synergized with the input denoising algorithm enable the pipeline to serve as an effective, efficient tool in the histopathological review of WSIs.

Keywords: breast cancer, convolutional neural networks, metastasis mapping, whole slide images

Procedia PDF Downloads 107

Authors: Zhao Gao, Eran Edirisinghe

Abstract:

The production of an accurate sketch of a suspect based on a verbal description obtained from a witness is an essential task for most criminal investigations. The criminal investigation system employs specifically trained professional artists to manually draw a facial image of the suspect according to the descriptions of an eyewitness for subsequent identification. Within the advancement of Deep Learning, Recurrent Neural Networks (RNN) have shown great promise in Natural Language Processing (NLP) tasks. Additionally, Generative Adversarial Networks (GAN) have also proven to be very effective in image generation. In this study, a trained GAN conditioned on textual features such as keywords automatically encoded from a verbal description of a human face using an RNN is used to generate photo-realistic facial images for criminal investigations. The intention of the proposed system is to map corresponding features into text generated from verbal descriptions. With this, it becomes possible to generate many reasonably accurate alternatives to which the witness can use to hopefully identify a suspect from. This reduces subjectivity in decision making both by the eyewitness and the artist while giving an opportunity for the witness to evaluate and reconsider decisions. Furthermore, the proposed approach benefits law enforcement agencies by reducing the time taken to physically draw each potential sketch, thus increasing response times and mitigating potentially malicious human intervention. With publically available 'CelebFaces Attributes Dataset' (CelebA) and additionally providing verbal description as training data, the proposed architecture is able to effectively produce facial structures from given text. Word Embeddings are learnt by applying the RNN architecture in order to perform semantic parsing, the output of which is fed into the GAN for synthesizing photo-realistic images. Rather than the grid search method, a metaheuristic search based on genetic algorithms is applied to evolve the network with the intent of achieving optimal hyperparameters in a fraction the time of a typical brute force approach. With the exception of the ‘CelebA’ training database, further novel test cases are supplied to the network for evaluation. Witness reports detailing criminals from Interpol or other law enforcement agencies are sampled on the network. Using the descriptions provided, samples are generated and compared with the ground truth images of a criminal in order to calculate the similarities. Two factors are used for performance evaluation: The Structural Similarity Index (SSIM) and the Peak Signal-to-Noise Ratio (PSNR). A high percentile output from this performance matrix should attribute to demonstrating the accuracy, in hope of proving that the proposed approach can be an effective tool for law enforcement agencies. The proposed approach to criminal facial image generation has potential to increase the ratio of criminal cases that can be ultimately resolved using eyewitness information gathering.

Keywords: RNN, GAN, NLP, facial composition, criminal investigation

Procedia PDF Downloads 138

Authors: Gertraud Koch, Teresa Stumpf, Alejandra Tijerina García

Abstract:

Discourse analysis research approaches belong to the central research strategies applied throughout the humanities; they focus on the countless forms and ways digital texts and images shape present-day notions of the world. Despite the constantly growing number of relevant digital, multimodal discourse resources, digital humanities (DH) methods are thus far not systematically developed and accessible for discourse analysis approaches. Specifically, the significance of multimodality and meaning plurality modelling are yet to be sufficiently addressed. In order to address this research gap, the D-WISE project aims to develop a prototypical working environment as digital support for the sociology of knowledge approach to discourse analysis and new IT-analysis approaches for the use of context-oriented embedding representations. Playing an essential role throughout our research endeavor is the constant optimization of hermeneutical methodology in the use of (semi)automated processes and their corresponding epistemological reflection. Among the discourse analyses, the sociology of knowledge approach to discourse analysis is characterised by the reconstructive and accompanying research into the formation of knowledge systems in social negotiation processes. The approach analyses how dominant understandings of a phenomenon develop, i.e., the way they are expressed and consolidated by various actors in specific arenas of discourse until a specific understanding of the phenomenon and its socially accepted structure are established. This article presents insights and initial findings from D-WISE, a joint research project running since 2021 between the Institute of Anthropological Studies in Culture and History and the Language Technology Group of the Department of Informatics at the University of Hamburg. As an interdisciplinary team, we develop central innovations with regard to the availability of relevant DH applications by building up a uniform working environment, which supports the procedure of the sociology of knowledge approach to discourse analysis within open corpora and heterogeneous, multimodal data sources for researchers in the humanities. We are hereby expanding the existing range of DH methods by developing contextualized embeddings for improved modelling of the plurality of meaning and the integrated processing of multimodal data. The alignment of this methodological and technical innovation is based on the epistemological working methods according to grounded theory as a hermeneutic methodology. In order to systematically relate, compare, and reflect the approaches of structural-IT and hermeneutic-interpretative analysis, the discourse analysis is carried out both manually and digitally. Using the example of current discourses on digitization in the healthcare sector and the associated issues regarding data protection, we have manually built an initial data corpus of which the relevant actors and discourse positions are analysed in conventional qualitative discourse analysis. At the same time, we are building an extensive digital corpus on the same topic based on the use and further development of entity-centered research tools such as topic crawlers and automated newsreaders. In addition to the text material, this consists of multimodal sources such as images, video sequences, and apps. In a blended reading process, the data material is filtered, annotated, and finally coded with the help of NLP tools such as dependency parsing, named entity recognition, co-reference resolution, entity linking, sentiment analysis, and other project-specific tools that are being adapted and developed. The coding process is carried out (semi-)automated by programs that propose coding paradigms based on the calculated entities and their relationships. Simultaneously, these can be specifically trained by manual coding in a closed reading process and specified according to the content issues. Overall, this approach enables purely qualitative, fully automated, and semi-automated analyses to be compared and reflected upon.

Keywords: entanglement of structural IT and hermeneutic-interpretative analysis, multimodality, plurality of meaning, sociology of knowledge approach to discourse analysis

Procedia PDF Downloads 204