Michel Gagnon

Learned student models with item to item knowledge structures

Probabilistic and learned approaches to student modeling are attractive because of the uncertainty surrounding the student skills assessment and because of the need to automatize the process. Item to item structures readily lend themselves to probabilistic and fully learned models because they are solely composed of observable nodes, like answers to test questions. Their structure is also well grounded in the cognitive theory of knowledge spaces. We study the effectiveness of two Bayesian frameworks to learn item to item structures and to use the induced structures to predict item outcome from a subset of evidence. One approach, Partial Order Knowledge Structures (POKS), relies on a naive Bayes framework whereas the other is based on the Bayesian network (BN) learning and inference framework. Both approaches are assessed over their predictive ability and their computational efficiency in different experimental simulations. The results from simulations over three data sets show that they both can effectively perform accurate predictions, but POKS generally displays higher predictive power than the BN. Moreover, the simplicity of POKS translates to a time efficiency between one to three orders of magnitude greater than the BN runs. We further explore the use of the item to item approach for handling concepts mastery assessment. The approach investigated consist in augmenting an initial set of observations, based on inferences with the item to item structure, and feed the augmented set to a BN containing a number of concepts. The results show that augmented set can effectively improve predictive power of a BN for item outcome, but that improvement does not transfer to the concept assessment in this particular experiment. We discuss different explanations for the results and outline future research avenues.

Automatic semantic web annotation of named entities

This paper describes a method to perform automated semantic annotation of named entities contained in large corpora. The semantic annotation is made in the context of the Semantic Web. The method is based on an algorithm that compares the set of words that appear before and after the name entity with the content of Wikipedia articles, and identifies the more relevant one by means of a similarity measure. It then uses the link that exists between the selected Wikipedia entry and the corresponding RDF description in the Linked Data project to establish a connection between the named entity and some URI in the Semantic Web. We present our system, discuss its architecture, and describe an algorithm dedicated to ontological disambiguation of named entities contained in large-scale corpora. We evaluate the algorithm, and present our results.

An Assessment of Online Semantic Annotators for the Keyword Extraction Task

The task of keyword extraction aims at capturing expressions (or entities) that best represent the main topics of a document. Given the rapid adoption of these online semantic annotators and their contribution to the growth of the Semantic Web, one important task is to assess their quality. This article presents an evaluation of the quality and stability of semantic annotators on domain-specific and open domain corpora. We evaluate five semantic annotators and compare them to two state-of-the-art keyword extractors, namely KP-miner and Maui. Our evaluation demonstrates that semantic annotators are not able to outperform keyword extractors and that annotators perform best on domains having a high keyword density.

Severe radiation pneumonitis after lung stereotactic ablative radiation therapy in patients with interstitial lung disease

PURPOSE To investigate the incidence and predictive factors of severe radiation pneumonitis (RP) after stereotactic ablative radiation therapy (SABR) in early-stage lung cancer patients with preexisting radiological interstitial lung disease (ILD). METHODS AND MATERIALS A retrospective analysis of patients with stage I lung cancer treated with SABR from 2009 to 2014 was conducted. Interstitial lung disease diagnosis and grading was based on pretreatment high-resolution computed tomography imaging. A central review of pretreatment computed tomography by a single experienced thoracic radiologist was conducted. Univariate and multivariate analyses were conducted to determine potential predictors of severe RP in patients with ILD. RESULTS Among 504 patients treated with SABR in this period, 6% were identified as having preexisting ILD. There was a 4% rate of ≥ grade 3 RP in the entire cohort. Interstitial lung disease was associated with increased risk of ≥ grade 3 RP (32% in ILD+ vs 2% in ILD-, P < .001). Five patients (21%) with ILD developed grade 5 RP. Lower forced expiratory volume in 1 second and forced vital capacity, higher V5Gy and mean lung dose, presence of severe radiological ILD, and combined emphysema were significant predictors of ≥ grade 3 RP on univariate analysis; only forced expiratory volume in 1 second remained on multivariate analysis. CONCLUSION Interstitial lung disease is associated with an increased risk of severe RP after SABR. Chest imaging should be reviewed for ILD before SABR, and the risk of fatal RP should be carefully weighed against the benefits of SABR in this subgroup.

Bayesian student models based on item to item knowledge structures

Bayesian networks are commonly used in cognitive student modeling and assessment. They typically represent the item-concepts relationships, where items are observable responses to questions or exercises and concepts represent latent traits and skills. Bayesian networks can also represent concepts-concepts and concepts-misconceptions relationships. We explore their use for modeling item-item relationships, in accordance with the theory of knowledge spaces. We compare two Bayesian frameworks for that purpose, a standard Bayesian network approach and a more constrained framework that relies on a local independence assumption. Their performance is compared over their respective ability to predict item outcome and through simulations over two data sets. The simulation results show that both approaches can effectively perform accurate predictions, but the constrained approach shows higher predictive power than a Bayesian Network. We discuss the applications of item to item structure for cognitive modeling within different contexts.

Can we use linked data semantic annotators for the extraction of domain-relevant expressions?

Semantic annotation is the process of identifying expressions in texts and linking them to some semantic structure. In particular, Linked data-based Semantic Annotators are now becoming the new Holy Grail for meaning extraction from unstructured documents. This paper presents an evaluation of the main linked data-based annotators available with a focus on domain topics and named entities. In particular, we compare the ability of each tool to annotate relevant domain expressions in text. The paper also proposes a combination of annotators through voting methods and machine learning. Our results show that some linked-data annotators, especially Alchemy, can be considered as a useful resource for topic extraction. They also show that a substantial increase in recall can be achieved by combining the annotators with a weighted voting scheme. Finally, an interesting result is that by removing Alchemy from the combination, or by combining only the more precise annotators, we get a significant increase in precision, at the cost of a lower recall.

Collective Disambiguation and Semantic Annotation for Entity Linking and Typing

In this paper we present the WESTLAB system, the winner of the 2016 OKE challenge Task 1. Our approach combines the output of a semantic annotator with the output of a named entity recognizer, and applies some heuristics for merging and filtering the detected mentions. The approach also applies a collective disambiguation method that relies on all the previously linked entities to choose between multiple candidate entities for a given mention. Using this approach, we greatly improve the performance of all the semantic annotators that are used as baselines in our experiments and also outperform the best system of the OKE Challenge 2015.

Text compression by syntactic pruning

We present a method for text compression, which relies on pruning of a syntactic tree. The syntactic pruning applies to a complete analysis of sentences, performed by a French dependency grammar. Sub-trees in the syntactic analysis are pruned when they are labelled with targeted relations. Evaluation is performed on a corpus of sentences which have been manually compressed. The reduction ratio of extracted sentences averages around 70%, while retaining grammaticality or readability in a proportion of over 74%. Given these results on a limited set of syntactic relations, this shows promise for any application which requires compression of texts, including text summarization.

Prétexte: A Generator for the Expression of Temporal Information

In this paper we present a method for generating French texts conveying temporal information. This method integrates the Discourse Representation Theory (DRT) and the Systemic Grammar Theory. First, we show how the DRT is used to represent temporal information. We then show how this formalism is used to represent the temporal localization expressed by temporal adverbial phrases. Finally, we give a description of how this representation of adverbial phrases can be translated into a syntactic form, using Systemic Grammar Theory. Pretexte, our implementation of this method, is able to generate a great variety of temporal adverbial phrases.

Assessing the Quality of Domain Concepts Descriptions in DBpedia

With the increasing volume of datasets on the Linked Open Data (LOD) cloud, it becomes necessary to assess Linked Data quality. This is especially important for DBpedia, which has become a prominent resource on the LOD. In this paper, our aim is to evaluate the quality of the description of domain concepts in DBpedia. Using a data-driven approach on a sample of domain concepts from Wikipedia, we show that a) the resources in our sample are described mainly by facts in DBpedia and seldom refer to the DBpedia ontology, b) DBpedia models very poorly these sample domain concepts at the instance level and schema level, c) very few predicates can be used for inference purposes, and d) very few domain predicates (object properties) are used in the description of domain concepts. This highlights the importance of restructuring the DBpedia knowledge base and including domain knowledge at the schema and instance levels.