Systematic reviews (SR) summarise the knowledge available in the literature on a specific topic. Keeping SRs up to date with new publications as soon as they become available is fundamental to avoid their early obsolescence. SR updating is particularly challenging in "living evidences", which comprise large numbers of SRs across entire domains of research. Living evidences are particularly useful in life sciences, where organizations like Cochrane maintain libraries with thousands of domain specific SRs. As the typical workflows for SR updating usually focus on one SR at a time, they are not adequate for living evidences. Specifically, they are extremely specific for their target SR, in that (i) they leverages ad-hoc search queries to find all the potentially relevant new citations, (ii) they train SR-specific machine learning models to automatically classify the relevant abstracts, and (iii) they still require huge efforts to manually assess a large number of abstracts being false positives. This research proposes "ContReviews", an automated system to manage living evidences in the health care domain. Specifically, ContReviews is based on an academic knowledge graph and a content-based recommendation model. The academic knowledge graph allows the quick identification of new publications for entire domains of research, without the need to identify SR-specific bibliographic databases one by one. The content-based recommendation model leverages publications already included in the living evidence, to learn a unique model for assessing the new publications' relevance to each SR in the living evidence. This lets to avoid designing and testing complex search queries over bibliographic databases, and developing, training, and evaluating SR-specific relevance assessment models. Finally, to represent publications and SRs more faithfully, ContReviews leverages multiple of their features, which includes titles, abstracts, citations and authors. ContReviews leverages both bag of words and embeddings to represent textual features (i.e., titles and abstracts). For embeddings, two fine-tuning methods have are proposed, to align the language model to the specific living evidence domain. ContReviews combines bag of words and the fine-tuned embeddings with binary vector representations of entities (i.e., authors and citations). ContReviews has been evaluated over a large dataset of Cochrane Reviews, and compared to two baseline models. Evaluation results show that ContReviews is both efficient and effective (i.e., precision above 97\% with 100\% recall), beating the traditional methods which compromise aggressively on efficiency to be effective. In addition, ContReviews shows that a single content-based recommendation model can be used to manage an entire living evidence, as opposed to the traditional models which are SR-specific.
Le "systematic review" (SR) riassumono la conoscienza disponibile in letteratura su un argomento specifico. Mantenere le SR aggiornate con i risultati delle nuove ricerche è fondamentale per evitare la loro obsolescenza. L'aggiornamento delle SR è particolarmente sfidante nelle "living evidence": queste comprendono moltissime SR e riguardano interi ambiti di ricerca. Le living evidence sono particolarmente utili in medicina, dove organizzazioni come Cochrane mantengono migliaia di SR di dominio. I tipici approcci all'aggiornamento delle SR, che solitamente si focalizzano su una sola SR, non sono adatti alle living evidence. In particolare, essi sono specifici per la loro SR di riferimento: (i) sfruttano query di ricerca ad-hoc per trovare tutte le nuove citazioni potenzialmente utili; (ii) addestrano modelli di machine learning specifici per una SR, per classificare i nuovi abstract in base alla loro rilevanza; (iii) richiedono sforzi ingenti per valutare manualmente un gran numero di abstract che risultano come falsi positivi. Questa ricerca propone "ContReviews", un sistema automatizzato per gestire le living evidence nel settore della medicina. In particolare, ContReviews si basa su un "academic knowledge graph" e su un "content-based recommendation model". Il academic knowledge graph consente di identificare le nuove evidenze scientifiche per interi ambiti di ricerca, senza dover individuare a priori i database specifici di dominio. Il content-based recommendation model sfrutta le pubblicazioni già incluse nella living evidence per imparare un unico modello di valutazione della rilevanza delle nuove pubblicazioni per ciascuna SR nella living evidence. Ciò consente di evitare la progettazione e il test di complesse query di ricerca e lo sviluppo, l'addestramento e la valutazione di modelli specifici per la valutazione della rilevanza delle SR. Infine, per rappresentare le pubblicazioni e le SR in modo più fedele, ContReviews sfrutta diverse feature delle pubblicazioni, tra cui titolo, abstract, citazioni e autori. ContReviews utilizza sia i bag of word che gli embedding per rappresentare le feature testuali (cioè titoli e abstract). Per gli embedding, sono proposti due metodi di fine-tuning per allineare il language model al dominio specifico della living evidence. ContReviews combina i bag of word e gli embedding con rappresentazioni vettoriali binarie delle entità (cioè autori e citazioni). ContReviews è stato valutato su un ampio dataset di Cochrane Reviews e confrontato con due modelli di base. I risultati della valutazione mostrano che ContReviews è sia efficiente che efficace (precision superiore al 97\% con un richiamo del 100\%), superando i metodi tradizionali che sacrificano l'efficienza per l'efficacia. Inoltre, ContReviews dimostra che un unico content-based recommendation model può essere utilizzato per gestire un'intera living evidence, a differenza dei modelli tradizionali che sono specifici per le SR.
(2024). A CONTENT-BASED RECOMMENDATION MODEL FOR LIVING EVIDENCE IN THE HEALTH CARE DOMAIN. (Tesi di dottorato, Università degli Studi di Milano-Bicocca, 2024).
A CONTENT-BASED RECOMMENDATION MODEL FOR LIVING EVIDENCE IN THE HEALTH CARE DOMAIN
TENTI, PAOLO
2024
Abstract
Systematic reviews (SR) summarise the knowledge available in the literature on a specific topic. Keeping SRs up to date with new publications as soon as they become available is fundamental to avoid their early obsolescence. SR updating is particularly challenging in "living evidences", which comprise large numbers of SRs across entire domains of research. Living evidences are particularly useful in life sciences, where organizations like Cochrane maintain libraries with thousands of domain specific SRs. As the typical workflows for SR updating usually focus on one SR at a time, they are not adequate for living evidences. Specifically, they are extremely specific for their target SR, in that (i) they leverages ad-hoc search queries to find all the potentially relevant new citations, (ii) they train SR-specific machine learning models to automatically classify the relevant abstracts, and (iii) they still require huge efforts to manually assess a large number of abstracts being false positives. This research proposes "ContReviews", an automated system to manage living evidences in the health care domain. Specifically, ContReviews is based on an academic knowledge graph and a content-based recommendation model. The academic knowledge graph allows the quick identification of new publications for entire domains of research, without the need to identify SR-specific bibliographic databases one by one. The content-based recommendation model leverages publications already included in the living evidence, to learn a unique model for assessing the new publications' relevance to each SR in the living evidence. This lets to avoid designing and testing complex search queries over bibliographic databases, and developing, training, and evaluating SR-specific relevance assessment models. Finally, to represent publications and SRs more faithfully, ContReviews leverages multiple of their features, which includes titles, abstracts, citations and authors. ContReviews leverages both bag of words and embeddings to represent textual features (i.e., titles and abstracts). For embeddings, two fine-tuning methods have are proposed, to align the language model to the specific living evidence domain. ContReviews combines bag of words and the fine-tuned embeddings with binary vector representations of entities (i.e., authors and citations). ContReviews has been evaluated over a large dataset of Cochrane Reviews, and compared to two baseline models. Evaluation results show that ContReviews is both efficient and effective (i.e., precision above 97\% with 100\% recall), beating the traditional methods which compromise aggressively on efficiency to be effective. In addition, ContReviews shows that a single content-based recommendation model can be used to manage an entire living evidence, as opposed to the traditional models which are SR-specific.File | Dimensione | Formato | |
---|---|---|---|
phd_unimib_744757.pdf
accesso aperto
Descrizione: Tesi
Tipologia di allegato:
Doctoral thesis
Dimensione
1.25 MB
Formato
Adobe PDF
|
1.25 MB | Adobe PDF | Visualizza/Apri |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.