Latent Semantic Analysis

This paper explores three methodological issues related to content analysis of online asynchronous discussions: unitizing, reliability, and manifest versus latent content. Unitizing involves balancing feasibility, reliability, identifiability, and discriminant capability of semantic versus syntactic units. Reliability is discussed in relation to differences between tasks, discussants, and number of coding decisions, as well as between coders. Manifest versus latent content contrasts observed versus intended behaviors in content analysis. For each of the three issues, the paper presents a brief theoretical overview. Each issue is subsequently contextualized and illustrated using empirical results. Finally, for each issue, the paper provides a discussion of lessons learned and implications. The paper concludes with suggestions for future studies.

The current dominant paradigm in artificial intelligence, which anchors logical derivation and world knowledge in monolithic, statically compressed model weights, is mathematically non-composable and computationally destructive. This structural approach leads to systemic fragility, difficulty in debugging internal errors, and redundant compute expenditures across the ecosystem. In response, this analysis explores a fundamental shift: treating logical derivation as a dynamic control process executed in a continuous latent space. The Latent-Space Reasoning (LSR) framework, which utilizes training-free continuous soft prompts to precondition a model’s initial state, demonstrates that latent-state dynamics are the primary substrate for complex logical processing. Empirical benchmarks show that minor continuous perturbations can systematically unlock immense latent domain knowledge—such as achieving a +19.6pp lift in arithmetic accuracy on a quantized four-billion parameter model—by steering the model’s trajectory past early attention-sink barriers. This paradigm reframes intelligence from a proprietary commodity to a modular infrastructure, enabling highly parameter-efficient systems like the Tiny Recursive Model (TRM) to outperform massive corporate architectures. Ultimately, the shift toward lightweight continuous preconditioning and open-source models ensures complete digital sovereignty and self-reliance for emerging economies, allowing advanced, data-compliant AI to run efficiently on consumer-grade hardware.

El procesamiento de lenguaje natural (PLN) es un área de estudio que surge de la intersección entre inteligencia artificial y lingüística. La traducción palabra por palabra es una de las primeras soluciones propuestas. No obstante, los aspectos morfológicos más comunes de las palabras representan una de las causas del mal desempeño de las primeras soluciones a las tareas que involucran lenguaje natural (Nadkarni, Ohno-Machado & Chapman, 2011). Si bien el PLN es un área activa en investigación, no existe consenso respecto de la definición del término. Se toma una de las propuestas aquí: PLN es un conjunto de técnicas computacionales que tienen el propósito de analizar y representar textos ocurridos naturalmente, en uno o más niveles de análisis lingüístico, con la finalidad de realizar procesamiento de lenguaje a nivel humano en un amplio rango de aplicaciones (Liddy, 2001). Existen varios enfoques para realizar las tareas de PLN, tales como el simbólico basado en reglas (Lally & Fodor, 2011) y el simbólico basado en redes semánticas (Sowa, 2006).

Contemporary multi-agent systems (MAS) operating via text-mediated message passing suffer from critical limitations, including high recurrent decoding costs, context window saturation, semantic quantization errors, and end-to-end gradient vanishing due to the softmax bottleneck. This paper presents a comprehensive architectural blueprint for the immediate next generation of agentic systems, advocating for a paradigm shift to latent-space recursive collaboration governed by an Information-Geometric Framework. By treating internal hidden states as points on a structured semantic manifold defined by the Fisher Information Metric (FIM), the framework formalizes reasoning as geodesic trajectories, enabling precise measurement of semantic divergence and stable co-optimization. Key components include the Latent Recursive Multi-Agent System (RecursiveMAS), which uses Inner and Outer RecursiveLinks to achieve stable back-propagation and a 2.4x speedup, and Semantic Tube Prediction (STP), which confines reasoning to optimal geodesics to achieve baseline accuracy with 16 times less training data. The architecture further integrates hyperbolic shaping for branching search and uses cellular sheaves and cohomology to topologically govern coordination, providing an absolute limit on communication cost and resolving ambiguity, even transitioning to Quantum Semantic Communication when irreducible ambiguity is encountered. The resulting system replaces heuristic agent swarms with a geometrically constrained computational graph, demonstrating an average 8.3% accuracy improvement and up to a 75.6% reduction in token usage over traditional text-based MAS.

Histopathological images are an important resource for clinical diagnosis and biomedical research. From an image understanding point of view, the automatic annotation of these images is a challenging problem. This paper presents a new method for automatic histopathological image annotation based on three complementary strategies, first, a part-based image representation, called the bag of features, which takes advantage of the natural redundancy of histopathological images for capturing the fundamental patterns of biological structures, second, a latent topic model, based on non-negative matrix factorization, which captures the high-level visual patterns hidden in the image, and, third, a probabilistic annotation model that links visual appearance of morphological and architectural features associated to 10 histopathological image annotations. The method was evaluated using 1,604 annotated images of skin tissues, which included normal and pathological architectural and morphological features, obtaining a recall of 74% and a precision of 50%, which improved a baseline annotation method based on support vector machines in a 64% and 24%, respectively.

Large language models (LLMs) exhibit strong local fluency and short-horizon conversational competence, but often degrade over extended interactions. Common failure modes include forgetting previously introduced entities, abandoning delayed user requests, confusing concurrent conversational threads, and generating responses that are locally plausible but globally inconsistent with previous turns. These failures are usually treated as consequences of limited context length or inadequate retrieval. We argue that this diagnosis is incomplete. Longhorizon dialog is not only a problem of preserving more text; it is a problem of preserving the geometry and dynamics of conversational state.
This paper presents Memory Transformers, a framework that models conversational memory as a trajectory through structured latent subspaces. Rather than representing dialog history purely as a token sequence, retrieved passages, or a single textual summary, the proposed method constructs checkpoint states from pooled hidden representations drawn from selected transformer layers together with cross-layer interaction features. These checkpoint states are interpreted as positions in a structured representation space whose geometry encodes topical persistence, semantic drift, subtopic branching, and recurrent visitation of discourse-relevant subspaces. A structured semantic memory state Ψt tracks entities, threads, intentions, topic associations, salience, and style-related signals, while context-conditioned low rank transition operators govern how the latent checkpoint state evolves over time.
The paper makes four contributions. First, it reformulates long-horizon dialog as trajectory modeling over structured latent subspaces, thereby shifting the central object of memory from stored text to geometry-preserving state evolution. Second, it introduces a multi-layer checkpoint representation, a recurrent subspace visitation mechanism, and a context-conditioned low rank transition operator that jointly model continuity, branching discourse, and persistent persona or style tendencies. Third, it specifies a training objective that combines language modeling with state reconstruction, behavioral consistency, and continuity oriented auxiliary losses for entities, intents, and threads. Fourth, it proposes a comprehensive empirical program spanning synthetic long-dialog stress tests and real benchmarks including MultiWOZ 2.2 and Schema-Guided Dialog with ablations, human evaluation, and significance analysis. The central claim is that conversational continuity is better modeled as geometry-preserving latent state evolution than as transcript expansion alone.

The present era of large-scale foundation models is defined by a philosophical and practical tension: the successful, scalable application of frequentist Maximum Likelihood Estimation (MLE) and Maximum A Posteriori (MAP) optimization, which bypasses the theoretically superior but computationally intractable Bayesian ideal of full posterior marginalisation. While MLE-driven models achieve remarkable empirical results, they inherently lack the capacity to quantify epistemic uncertainty, leading to vulnerabilities like overconfidence and brittle generalization when encountering out-of-distribution data. This report examines the foundational advantages of Bayesian inference—automatic regularisation, model averaging, and calibrated probabilities—and the "curse of dimensionality" that necessitated the pivot to point-estimation scaling laws. The analysis further reveals a profound mathematical equivalence between Bayesian neural networks and quantum physics, specifically the Feynman path integral, suggesting that training a neural network is analogous to finding the ground state of a quantum many-body system. Finally, the paper evaluates cutting-edge solutions, including hybrid algorithmic frameworks like PoLAR-VBLL that surgically reintroduce uncertainty into Large Language Models, and argues that the ultimate resolution lies in a hardware paradigm shift toward stochastic neuromorphic memristors and quantum computing, which are physically engineered to perform probabilistic inference.

Recent research in text processing has emphasized the importance of the cohesion of a text in comprehension (e.g., . Cohesion is the degree to which ideas in the text are explicitly related to each other and facilitate a unified situation model for the reader. Such research has led to the development of a computational tool, Coh-Metrix, ) that delivers over 300 indices of textual cohesion and difficulty. We hypothesized that a Coh-Metrix analysis of texts would indicate that cohesion indices -more so than traditional, shallow difficulty indices such as Flesch-Kincaid Grade Level (FKGL, Klare, 1974-75) -would identify characteristics of texts. Specifically, we hypothesized that within the expository domain, science texts would demonstrate greater cohesion than history texts, as the former dealt with less familiar subjects and would be likely to employ greater redundancy. We further hypothesized that as the parts of a text (beginning, middle, and end) serve different rhetorical purposes, that the sophisticated indices of Coh-Metrix would identify these differences. To test our hypothesis, we sampled three representative 1000-word sections from the beginning, middle and end of each chapter of seven commonly used high-school text books (three from science and four from history). Each section was analyzed using Coh-Metrix indices of Cohesion (argument overlap, latent semantic analysis (LSA), and number of connectives) as well as FKGL to assess difficulty.

In this paper we describe an information extraction and text mining system which identifies key information components from text documents. The information components are centered on domain entities and their relationships. The components mined from a repository are chained using an n-gram-based algorithm. The information chains provide a comprehensive view of the collection and can be also used for inferential reasoning.

A number of techniques such as information extraction, document classification, document clustering and information visualization have been developed to ease extraction and understanding of information embedded within text documents. However, knowledge that is embedded in natural language texts is difficult to extract using simple pattern matching techniques and most of these methods do not help users directly understand key concepts and their semantic relationships in document corpora, which are critical for capturing their conceptual structures. The problem arises due to the fact that most of the information is embedded within unstructured or semi-structured texts that computers can not interpret very easily. In this paper, we have presented a novel Biomedical Knowledge Extraction and Visualization framework, BioKEVis to identify key information components from biomedical text documents. The information components are centered on key concepts. BioKEVis applies linguistic analysis and Latent Semantic Analysis (LSA) to identify key concepts. The information component extraction principle is based on natural language processing techniques and semanticbased analysis. The system is also integrated with a biomedical named entity recognizer, ABNER, to tag genes, proteins and other entity names in the text. We have also presented a method for collating information extracted from multiple sources to generate semantic network. The network provides distinct user perspectives and allows navigation over documents with similar information components and is also used to provide a comprehensive view of the collection. The system stores the extracted information components in a structured repository which is integrated with a query-processing module to handle biomedical queries over text documents. We have also proposed a document ranking mechanism to present retrieved documents in order of their relevance to the user query.

This study is a follow-up to Pynte, New and Kennedy (2008), Journal of Eye Movement Research . 2(1):4, 1-11. A new series of multiple regression analyses were conducted on the French part of the Dundee corpus, using a new set of syntactic and semantic predictors. In line with our prior study, quite different patterns of results were obtained for function and content words. We conclude that syntactic processing operations during reading mainly concern function words and are carried out ahead of semantic processing.

Semantic and syntactic influences during reading normal text were examined in a series of multiple regression analyses conducted on a large-scale corpus of eyemovement data. Two measures of contextual constraints, based on the syntactic descriptions provided by Abeillé, Clément et Toussenel (2003) and one measure on semantic constraint, based on Latent Semantic Analysis, were included in the regression equation, together with a set of properties (length, frequency, etc.), known to affect inspection times. Both syntactic and semantic constraints were found to exert a significant influence, with less time spent inspecting highly constrained target words, relative to weakly constrained ones. Semantic and syntactic properties apparently exerted their influence independently from each other, as suggested by the lack of interaction.

On-line contextual influences during reading were examined in a series of multiple-regression analyses conducted on a large-scale corpus of eye-movement data, using Latent Semantic Analysis (LSA) to assess the degree of contextual constraints exerted on a given target word by the immediately prior word and by the prior sentence fragment. A decrease in inspection time was observed as contextual constraints increased. Word-level constraints exerted their influence both forward (on both single-fixation and gaze durations) and backward (on gaze duration only). An independent sentence-level effect was only visible in the forward direction, and only for gaze duration. Gaze duration was also sensitive to the depth of embedding of the target word in the syntactic structure. We conclude that both low-level and high-level contextual constraints can translate in the eye-movement record.

Matériel et préparation du pied ……………………………………64 b. Examen du pied à l'appui …………………………………………65 c. Examen du pied au soutien ……………………………………….72 3-Examen dynamique ……………………………………………………….74 DEUXIEME PARTIE : LA GRILLE D'ÉVALUATION DE LA SANTÉ DU PIED DU CHEVAL I-La relation vétérinaire -maréchal-ferrant -propriétaire ……………………………...76 A-Mise en place d'un partenariat vétérinaire / maréchal-ferrant …………………….77 B-Les difficultés de communication fréquemment rencontrées ……………………..79 C-Les éléments clés d'une bonne communication …………………………………..81 D-Mise en place de nouveaux moyens de communication …………………………..83 II-Création d'une grille d'évaluation de la santé du pied …………………………………84 A-Choix des critères d'évaluation …………………………………………………...85 B-Choix du système de notation ……………………………………………………..88 C-Le carnet de santé du pied …………………………………………………………89 III-Validation théorique de la grille .

Recently, singular value decomposition (SVD) and its variants, which are singular value rescaling (SVR), approximation dimension equalization (ADE) and iterative residual rescaling (IRR), were proposed to conduct the job of latent semantic indexing (LSI). Although they are all based on linear algebraic method for tem-document matrix computation, which is SVD, the basic motivations behind them concerning LSI are different from each other. In this paper, a series of experiments are conducted to examine their effectiveness of LSI for the practical application of text mining, including information retrieval, text categorization and similarity measure. The experimental results demonstrate that SVD and SVR have better performances than other proposed LSI methods in the above mentioned applications. Meanwhile, ADE and IRR, because of the too much difference between their approximation matrix and original term-document matrix in Frobenius norm, can not derive good performances for text mining applications using LSI.

The legacy of Nisbett and Wilson's classic article, Telling More Than We Can Know: Verbal Reports on Mental Processes (1977), is mixed. It is perhaps the most cited article in the recent history of consciousness studies, yet no empirical research program currently exists that continues the work presented in the article. To remedy this, we have introduced an experimental paradigm we call choice blindness [Johansson, P., Hall, L., Sikstrom, S., & Olsson, A. (2005). Failure to detect mismatches between intention and outcome in a simple decision task. Science, 310(5745), 116-119.]. In the choice blindness paradigm participants fail to notice mismatches between their intended choice and the outcome they are presented with, while nevertheless offering introspectively derived reasons for why they chose the way they did. In this article, we use word-frequency and latent semantic analysis (LSA) to investigate a corpus of introspective reports collected within the choice blindness paradig...

This paper provides a comprehensive review of the article “Architecting Trust in Artificial Epistemic Agents” by Marchal et al. (2026), which posits that large language models (LLMs) are evolving into epistemic agents capable of autonomously pursuing knowledge-related goals and reshaping shared information ecosystems. Key postulates include the definition of epistemic AI agents, their potential roles (e.g., scientists, educators), and associated risks such as cognitive deskilling and epistemic homogenization. The review highlights the article’s normative framework for trustworthiness, encompassing epistemic competence, falsifiability, and virtuous behaviors, alongside alignment with human epistemic goals and resilient socio-technical infrastructure. Similar research in AI ethics, epistemology, and human-computer interaction is surveyed, revealing ongoing trends toward agentic AI governance and debates on trust calibration. Algebraic interpretations model trust as a multidimensional vector space, while geometric views represent knowledge ecosystems as graphs with agents as nodes. Hypotheses from the original work are tested against empirical evidence from related studies, yielding results that affirm the need for proactive ethical interventions. The conclusion synthesizes directions for future AI development, emphasizing inclusive, verifiable systems to mitigate epistemic risks.

This paper provides a framework for optimally representing student written essays in a vector space, based upon Latent Semantic Analysis and instructor evaluated grades. Comparing student essays to an authoritative source, a ranking scheme is optimized that allows for a unique vector space representation on the unit circle. Once such a representation has been found, traditional methods of circular data analysis and inference can be applied, as we demonstrate.

Many modem data analysis methods involve computing a matrix singular value decomposition (SVD) or eigenvalue decomposition (EVD). Principal components analysis is the time-honored example, but more recent applications include latent semantic indexing, hypertext induced topic selection (HITS), clustering, classification, etc. Though the SVD and EVD are well-established and can be computed via state-of-the-art algorithms, it is not commonly mentioned that there is an intrinsic sign indeterminacy that can significantly impact the conclusions and interpretations drawn from their results. Here we provide a solution to the sign ambiguity problem and show how it leads to more sensible solutions.

Exploring the textual emotional value of the Reader is to help readers understand the Reader's embedded emotions in an all-around way. In this paper, two text analysis methods, latent semantic analysis and probabilistic latent semantic analysis are described, starting from the analysis model of text research, and the automatic text classification technique is illustrated. The principles of support vector machines are explained, the optimal decision function selection solution is performed using kernel functions, and a PLSA-SVM text analysis model is also constructed using the PLSA method jointly with SVM. The structural features of Reader magazine were analyzed and explained, and the PLSA-SVM text analysis model was used to analyze the data on the narrative features of Reader magazine. In terms of narrative themes, the main narrative themes of Reader magazine were distributed as affection and love, which increased by 16.84% and 17.42% from 2012 to 2021, respectively. In terms of narrative perspectives, the proportion of first-person and third-person narrative perspectives is comparable, increasing by 9.03% and 12.3% from 2012 to 2021, respectively. Thus, in the context of big data, the PLSA-SVM text analysis model can be used to analyze the text narrative features of Reader magazine effectively, and Reader magazine can use the analysis to optimize further the text content and issue texts that are attuned to readers' emotions.

PedaBot is a new discussion scaffolding application designed to aid student knowledge acquisition, promote reflection about course topics and encourage student participation in discussions. It dynamically processes student discussions and presents related discussions from a knowledge base of past discussions. This paper describes the system and presents a comparative analysis of the information retrieval techniques used to respond to free-form student discussions, a combination of topic profiling, term frequency-inverse document frequency, and latent semantic analysis. Responses are presented as annotated links that students can follow and rate. We report a pilot study of PedaBot based on student viewings, student ratings, and a small survey. Initial results indicate that there is a high level of student interest in the feature and that its responses are moderately relevant to student discussions.

The animation found in frame/video content is usually characterized by specific color distribution and texture. The existence of animation within a frame can be described by carefully selected features. We analyzed both fully and partially animated frames. Here, we present our initial results related to automatic animation interpretation using different types of video sequences.

The recent developments in the field of Information and Communication Technology (ICT) has resulted in a major paradigm shift in pedagogy and teaching learning. No longer restricted to the four walls of a classroom, ubiquitous learning is the current trend and universities are gearing up for larger intakes. In the wake of such an upsurge in volume, automatic evaluation of student answers is the need of the hour. The limitations of close ended questions notwithstanding, the popularity of such types are on the rise due to implementation ease. This paper presents a technique to evaluate free text answers of students which will augment and popularize the use of open ended questions for evaluation in online learning.

The rapid evolution of large language models (LLMs) has accelerated the development of autonomous agents capable of multi-turn reasoning, tool-use, and environmental interaction. Reinforcement learning (RL) remains a central paradigm for training such agents, yet its scalability is constrained by the scarcity of diverse, reliable, and executable environments. The Agent World Model (AWM) represents a major advance in synthetic environment generation, offering 1,000 code-driven, database-backed environments optimised for tool-use RL (Wang et al., 2026a). This paper synthesises two complementary strands of research: a comprehensive review of AWM's architecture, empirical performance, and comparative position within the broader ecosystem of synthetic environment frameworks; and an extended proposal for multi-agent generalisation of AWM, incorporating collaborative, competitive, and mixed-motive interactions. Drawing from recent developments in multi-agent reinforcement learning (MARL), we outline methodological extensions to scenario generation, database design, interface synthesis, and reward verification. We further discuss challenges including non-stationarity, credit assignment, and ethical safeguards. The result is a unified roadmap for scaling synthetic environments toward generalist, socially capable, and robust agentic intelligence. This expanded analysis integrates insights from contemporaneous works on synthetic data generation and MARL, providing a holistic view for advancing agentic AI in dynamic, real-world-like settings.

Essays are paramount for of assessing the academic excellence along with linking the different ideas with the ability to recall but are notably time consuming when they are assessed manually. Manual grading takes significant amount of evaluator's time and hence it is an expensive process. Automated grading if proven effective will not only reduce the time for assessment but comparing it with human scores will also make the score realistic. The project aims to develop an automated essay assessment system by use of machine learning techniques by classifying a corpus of textual entities into small number of discrete categories, corresponding to possible grades. Linear regression technique will be utilized for training the model along with making the use of various other classifications and clustering techniques. We intend to train classifiers on the training set, make it go through the downloaded dataset, and then measure performance our dataset by comparing the obtained values wit...

This review examines the seminal work by Kaddour et al. (2026) on agentic uncertainty, which reveals pervasive overconfidence in AI agents during coding tasks. We summarize the paper's key contributions, including the elicitation of success probabilities at pre-, mid-, and post-execution stages, and the mitigating role of adversarial prompting. Comparisons with related literature highlight alignments with LLM uncertainty estimation (e.g., Kadavath et al., 2022) and divergences from self-correction studies (e.g., Huang et al., 2024). A critical analysis underscores methodological strengths and limitations, such as domain specificity. We extend the framework to non-coding domains like healthcare decision-making and multi-agent systems, proposing hybrid human-AI uncertainty propagation models. This coherent synthesis advances scalable oversight for autonomous AI, emphasizing calibration's role in safety. Mamun, SM (2026)………………………………………1

This review critically examines the FEDI (Fion-Enhanced Distributed Intelligence) architecture, a novel physics-inspired framework for epistemic AI proposed by Chomiuk (2026). We summarize its key components, including Fion Units, the Loop-5 validation cycle, and mutual-information-based coherence metrics, highlighting its claims of reducing hallucinations by 75% compared to transformer baselines. Comparisons with related research in active inference, hallucination mitigation, multi-agent systems, and reproducible AI reveal FEDI's strengths in architectural enforcement of coherence but also limitations in scalability and multimodal integration. We provide a detailed academic critique, evaluating methodological rigor, empirical validity, and theoretical foundations. Extending beyond the original work, we propose integrations with quantum computing, real-time adaptive learning, and ethical governance frameworks to address uncovered areas such as dynamic environments and societal impacts. This coherent extension positions FEDI as a foundational step toward verifiable, robust AI systems. Mamun, SM (2026)…………………………………1

Aborder la pensée de Kimura en japonais pour ensuite la traduire dans une autre langue, en l'occurrence le français, fait souvent perdre de vue la tentative qui s'exprime dans la langue originale à savoir l'aïda, l'entre même, du à dire et du vouloir dire,-à dire de ce qui est et vouloir dire qui anime l'auteur. Cette distance entre l'originale et la traduction, tient à ce que la langue japonaise échappe, au moins partiellement, au principe des langues que Johannes Lohmann définit en tant que « système objectif de signes se rapportant à l'objectité par des entités appelées mots, ceux-ci ne consistant que de ce rapport qui les doue nativement d'une ou plusieurs significations .» 1 Cela permet en général d'échanger ces significations d'une langue à une autre, si différentes soientelles, car elles ont en commun le même principe constitutif : signifier l'être des choses dans la forme de l'objectité. Cette possibilité d'échange est valide entre les langues de signification dont c'est la fonction et l'être même, mais il faut garder à l'esprit qu'objectité n'est pas réalité. Comme l'ont reconnu au IIIe siècle les Stoïciens, une langue de signification permet d'énoncer du discible : les exprimables (λεκτά) qui sont pré-signifiés dans l'ensemble fermé qu'elle est. À parler une langue de signification, nous n'avons pas accès aux choses mêmes avec lesquelles nous avons réellement à faire, aux (πράγµατα), à ce qui surgit dans une rencontre, c'est-à-dire à l'étant en tant que tel, en co-présence avec laquelle nous nous surprenons à être. À l'opposé, la langue japonaise tend à guider vers une telle rencontre qu'annonce le terme même d'aïda traduit, et c'est le problème nous le verrons, par entre en français. Cette tendance est facilitée par la langue japonaise du fait qu'elle ne comprend pas d'article qui spécifie le nom, comme en français, en l'opposant au verbe à l'intérieur du système du mot. L'article, qui est dans notre langue le tenseur de l'extension nominale, n'existe pas en japonais. Il en résulte que les noms dits communs, en français, ont en japonais un statut proche de celui des noms propres en français. Pour bien saisir cela, je vais à nouveau citer Johannes Lohmann qui disait dans sa conférence intitulée Le concept de nom « L'événement principal de l'histoire du concept de nom, c'est la scission en deux de ce concept, opérée par Chrysippe au commencement du IIIe siècle. Depuis lors, on est accoutumé à distinguer en logique, puis dans la vie quotidienne, entre le nom propre et le nom commun des choses, c'est-à-dire le mot par lequel la chose est appelée 1

Implementing technology in a modern-day classroom is an ongoing challenge. In this paper, we created a system for an automatic assessment of student answers using Latent Semantic Analysis (lsa) -a method with an underlying assumption that words with similar meanings will appear in the same contexts. The system will be used within digital lexical flashcards for L2 vocabulary acquisition in a clil classroom. Results presented in this paper indicate that while lsa does well in creating semantic spaces for longer texts, it fell somewhat short of detecting topics in answers and word definitions. The answers were classified using knn, for both binary and multinomial classification. The results of knn classification are as follows: precision P = 0.73, recall R = 1.00, F1 = 0.85 for binary classification, and P = 0.50, R = 0.47, F1 = 0.46 score for the multinomial classifier. The results are to be taken with a grain of salt, due to a small test and training dataset.

Electronic assessment (e-assessment) becomes more popular in educational tools especially in e-learning environment.This is because it has some advantages such as reducing the staffs needed for assessment tasks, automated marking is not prone to human error and it gives instant feedback to the students.However, the e-assessment existed only more on assessing essays and lack of implementation in assessing computer program in computer science environment [Haley, et.al, 2003]. This paper will briefly describe on the techniques that have been used to implement e-assessment for essays such as Project Essay Grade (PEG), E-rater, Intelligent Essay Assessor (IEA) and Latent Semantic Analysis (LSA).Mainly, the discussion will concentrate on LSA to build the eassessment for computer programming corpus.One great advantage of LSA over the others is its ability to make absolute relative comparisons, where set of documents can be compared to each other, or to an answer schema.Another reason is th...

Traditionally, computer programming assignments are graded manually by educators. As this task is tedious, timeconsuming and prone to bias, the need for automated grading tool is necessary to reduce the educators' burden and avoid inconsistency and favoritism. Recent researches have claimed that Latent Semantic Analysis (LSA) has the ability to represent human cognitive knowledge to assess essays, retrieving information, classification of documents and indexing. In this paper, we adapt LSA technique to grade computer programming assignments and observe how far it can be applied as an alternative approach to traditional grading methods by human. The grades of the assignments are generated from the cosine similarity that shows how close students' assignments to the model answers in the latent semantic vector space. The results show that LSA is not able to detect orders of computer programming and symbols; however, LSA is able to grade assignments faster and consistently, which avoid bias and reduces the time spent by human.

Dimensionality reduction techniques address a relevant problem of Vector Space Models that is the size of involved dictionaries. Certain geometrical transformations applied over the original feature space, like the Latent Semantic Analysis (LSA), aim at preserving and discovering semantic relations between documents within small dimensional spaces. In this paper, a linear transformation method, named Locality Preserving Projections (LPP), is evaluated with respect to a document clustering task and results are compared with LSA. LPP is here applied directly on the original space, through an efficient C-based implementation, and different parameterizations are investigated. Experimental results suggest that LPP is an effective technique able to account for the availability of a priori knowledge within an unsupervised learning framework.

Recent work about textual entailment or paraphrasing emphasizes the role of automatic learning of inference rules. Major weakness of these repositories is the low accuracy reachable in applying the rules in operational settings (e.g. textual entailment challenges or question answering). In this paper a robust method for automatic learning of inference rules is presented. As opposed to existing proposals, it relies on a geometrical model of similarity, based on a form of latent semantic analysis applied to the source text collection. The result is a non merely distributional notion of lexical similarity that implies also selectional preference for the individual rule arguments. Experiments on a large data set show that selectional restrictions, applied conjunctively to all arguments in a pattern, are able to better select correct vs. incorrect cases. As the designed learning process is completely unsupervised and widely applicable, the method provides a very useful tool for different application and domains. It expresses rules as necessary conditions for equivalence or entailment to hold in unseen texts.

Techniques for data-mining, latent semantic analysis, contextual search of databases, etc. have long ago been developed by computer scientists working on information retrieval (IR). Experimental scientists, from all disciplines, having to analyse large collections of raw experimental data (astronomical, physical, biological, etc.) have developed powerful methods for their statistical analysis and for clustering, categorising, and classifying objects. Finally, physicists have developed a theory of quantum measurement, unifying the logical, algebraic, and probabilistic aspects of queries into a single formalism. The purpose of this paper is twofold: first to show that when formulated at an abstract level, problems from IR, from statistical data analysis, and from physical measurement theories are very similar and hence can profitably be cross-fertilised, and, secondly, to propose a novel method of fuzzy hierarchical clustering, termed semantic distillation -strongly inspired from the theory of quantum measurement -, we developed to analyse raw data coming from various types of experiments on DNA arrays. We illustrate the method by analysing DNA arrays experiments and clustering the genes of the array according to their specificity.