Guido Zuccon

Integrating and Evaluating Neural Word Embeddings in Information Retrieval

Recent advances in neural language models have contributed new methods for learning distributed vector representations of words (also called word embeddings). Two such methods are the continuous bag-of-words model and the skipgram model. These methods have been shown to produce embeddings that capture higher order relationships between words that are highly effective in natural language processing tasks involving the use of word similarity and word analogy. Despite these promising results, there has been little analysis of the use of these word embeddings for retrieval. Motivated by these observations, in this paper, we set out to determine how these word embeddings can be used within a retrieval model and what the benefit might be. To this aim, we use neural word embeddings within the well known translation language model for information retrieval. This language model captures implicit semantic relations between the words in queries and those in relevant documents, thus producing more accurate estimations of document relevance. The word embeddings used to estimate neural language models produce translations that differ from previous translation language model approaches; differences that deliver improvements in retrieval effectiveness. The models are robust to choices made in building word embeddings and, even more so, our results show that embeddings do not even need to be produced from the same corpus being used for retrieval.

Medical Semantic Similarity with a Neural Language Model

Advances in neural network language models have demonstrated that these models can effectively learn representations of words meaning. In this paper, we explore a variation of neural language models that can learn on concepts taken from structured ontologies and extracted from free-text, rather than directly from terms in free-text. This model is employed for the task of measuring semantic similarity between medical concepts, a task that is central to a number of techniques in medical informatics and information retrieval. The model is built with two medical corpora (journal abstracts and patient records) and empirically validated on two ground-truth datasets of human-judged concept pairs assessed by medical professionals. Empirically, our approach correlates closely with expert human assessors (≈0.9) and outperforms a number of state-of-the-art benchmarks for medical semantic similarity. The demonstrated superiority of this model for providing an effective semantic similarity measure is promising in that this may translate into effectiveness gains for techniques in medical information retrieval and medical informatics (e.g., query expansion and literature-based discovery).

Crowdsourcing interactions: using crowdsourcing for evaluating interactive information retrieval systems

In the field of information retrieval (IR), researchers and practitioners are often faced with a demand for valid approaches to evaluate the performance of retrieval systems. The Cranfield experiment paradigm has been dominant for the in-vitro evaluation of IR systems. Alternative to this paradigm, laboratory-based user studies have been widely used to evaluate interactive information retrieval (IIR) systems, and at the same time investigate users’ information searching behaviours. Major drawbacks of laboratory-based user studies for evaluating IIR systems include the high monetary and temporal costs involved in setting up and running those experiments, the lack of heterogeneity amongst the user population and the limited scale of the experiments, which usually involve a relatively restricted set of users. In this paper, we propose an alternative experimental methodology to laboratory-based user studies. Our novel experimental methodology uses a crowdsourcing platform as a means of engaging study participants. Through crowdsourcing, our experimental methodology can capture user interactions and searching behaviours at a lower cost, with more data, and within a shorter period than traditional laboratory-based user studies, and therefore can be used to assess the performances of IIR systems. In this article, we show the characteristic differences of our approach with respect to traditional IIR experimental and evaluation procedures. We also perform a use case study comparing crowdsourcing-based evaluation with laboratory-based evaluation of IIR systems, which can serve as a tutorial for setting up crowdsourcing-based IIR evaluations.

Understandability biased evaluation for information retrieval

Although relevance is known to be a multidimensional concept, information retrieval measures mainly consider one dimension of relevance: topicality. In this paper we propose a method to integrate multiple dimensions of relevance in the evaluation of information retrieval systems. This is done within the gain-discount evaluation framework, which underlies measures like rank-biased precision (RBP), cumulative gain, and expected reciprocal rank. Albeit the proposal is general and applicable to any dimension of relevance, we study specific instantiations of the approach in the context of evaluating retrieval systems with respect to both the topicality and the understandability of retrieved documents. This leads to the formulation of understandability biased evaluation measures based on RBP. We study these measures using both simulated experiments and real human assessments. The findings show that considering both understandability and topicality in the evaluation of retrieval systems leads to claims about system effectiveness that differ from those obtained when considering topicality alone.

Automatic ICD-10 classification of cancers from free-text death certificates.

OBJECTIVE Death certificates provide an invaluable source for cancer mortality statistics; however, this value can only be realised if accurate, quantitative data can be extracted from certificates--an aim hampered by both the volume and variable nature of certificates written in natural language. This paper proposes an automatic classification system for identifying cancer related causes of death from death certificates. METHODS Detailed features, including terms, n-grams and SNOMED CT concepts were extracted from a collection of 447,336 death certificates. These features were used to train Support Vector Machine classifiers (one classifier for each cancer type). The classifiers were deployed in a cascaded architecture: the first level identified the presence of cancer (i.e., binary cancer/nocancer) and the second level identified the type of cancer (according to the ICD-10 classification system). A held-out test set was used to evaluate the effectiveness of the classifiers according to precision, recall and F-measure. In addition, detailed feature analysis was performed to reveal the characteristics of a successful cancer classification model. RESULTS The system was highly effective at identifying cancer as the underlying cause of death (F-measure 0.94). The system was also effective at determining the type of cancer for common cancers (F-measure 0.7). Rare cancers, for which there was little training data, were difficult to classify accurately (F-measure 0.12). Factors influencing performance were the amount of training data and certain ambiguous cancers (e.g., those in the stomach region). The feature analysis revealed a combination of features were important for cancer type classification, with SNOMED CT concept and oncology specific morphology features proving the most valuable. CONCLUSION The system proposed in this study provides automatic identification and characterisation of cancers from large collections of free-text death certificates. This allows organisations such as Cancer Registries to monitor and report on cancer mortality in a timely and accurate manner. In addition, the methods and findings are generally applicable beyond cancer classification and to other sources of medical text besides death certificates.

Information retrieval as semantic inference: a Graph Inference model applied to medical search

This paper presents a Graph Inference retrieval model that integrates structured knowledge resources, statistical information retrieval methods and inference in a unified framework. Key components of the model are a graph-based representation of the corpus and retrieval driven by an inference mechanism achieved as a traversal over the graph. The model is proposed to tackle the semantic gap problem—the mismatch between the raw data and the way a human being interprets it. We break down the semantic gap problem into five core issues, each requiring a specific type of inference in order to be overcome. Our model and evaluation is applied to the medical domain because search within this domain is particularly challenging and, as we show, often requires inference. In addition, this domain features both structured knowledge resources as well as unstructured text. Our evaluation shows that inference can be effective, retrieving many new relevant documents that are not retrieved by state-of-the-art information retrieval models. We show that many retrieved documents were not pooled by keyword-based search methods, prompting us to perform additional relevance assessment on these new documents. A third of the newly retrieved documents judged were found to be relevant. Our analysis provides a thorough understanding of when and how to apply inference for retrieval, including a categorisation of queries according to the effect of inference. The inference mechanism promoted recall by retrieving new relevant documents not found by previous keyword-based approaches. In addition, it promoted precision by an effective reranking of documents. When inference is used, performance gains can generally be expected on hard queries. However, inference should not be applied universally: for easy, unambiguous queries and queries with few relevant documents, inference did adversely affect effectiveness. These conclusions reflect the fact that for retrieval as inference to be effective, a careful balancing act is involved. Finally, although the Graph Inference model is developed and applied to medical search, it is a general retrieval model applicable to other areas such as web search, where an emerging research trend is to utilise structured knowledge resources for more effective semantic search.

An Analysis of Theories of Search and Search Behavior

Theories of search and search behavior can be used to glean insights and generate hypotheses about how people interact with retrieval systems. This paper examines three such theories, the long standing Information Foraging Theory, along with the more recently proposed Search Economic Theory and the Interactive Probability Ranking Principle. Our goal is to develop a model for ad-hoc topic retrieval using each approach, all within a common framework, in order to (1) determine what predictions each approach makes about search behavior, and (2) show the relationships, equivalences and differences between the approaches. While each approach takes a different perspective on modeling searcher interactions, we show that under certain assumptions, they lead to similar hypotheses regarding search behavior. Moreover, we show that the models are complementary to each other, but operate at different levels (i.e., sessions, patches and situations). We further show how the differences between the approaches lead to new insights into the theories and new models. This contribution will not only lead to further theoretical developments, but also enables practitioners to employ one of the three equivalent models depending on the data available.

Active learning: a step towards automating medical concept extraction.

OBJECTIVE This paper presents an automatic, active learning-based system for the extraction of medical concepts from clinical free-text reports. Specifically, (1) the contribution of active learning in reducing the annotation effort and (2) the robustness of incremental active learning framework across different selection criteria and data sets are determined. MATERIALS AND METHODS The comparative performance of an active learning framework and a fully supervised approach were investigated to study how active learning reduces the annotation effort while achieving the same effectiveness as a supervised approach. Conditional random fields as the supervised method, and least confidence and information density as 2 selection criteria for active learning framework were used. The effect of incremental learning vs standard learning on the robustness of the models within the active learning framework with different selection criteria was also investigated. The following 2 clinical data sets were used for evaluation: the Informatics for Integrating Biology and the Bedside/Veteran Affairs (i2b2/VA) 2010 natural language processing challenge and the Shared Annotated Resources/Conference and Labs of the Evaluation Forum (ShARe/CLEF) 2013 eHealth Evaluation Lab. RESULTS The annotation effort saved by active learning to achieve the same effectiveness as supervised learning is up to 77%, 57%, and 46% of the total number of sequences, tokens, and concepts, respectively. Compared with the random sampling baseline, the saving is at least doubled. CONCLUSION Incremental active learning is a promising approach for building effective and robust medical concept extraction models while significantly reducing the burden of manual annotation.

Relevation!: An open source system for information retrieval relevance assessment

Relevation! is a system for performing relevance judgements for information retrieval evaluation. Relevation! is web-based, fully configurable and expandable; it allows researchers to effectively collect assessments and additional qualitative data. The system is easily deployed allowing assessors to smoothly perform their relevance judging tasks, even remotely. Relevation! is available as an open source project at: http://ielab.github.io/relevation.

Graph-based concept weighting for medical information retrieval

This paper presents a graph-based method to weight medical concepts in documents for the purposes of information retrieval. Medical concepts are extracted from free-text documents using a state-of-the-art technique that maps n-grams to concepts from the SNOMED CT medical ontology. In our graph-based concept representation, concepts are vertices in a graph built from a document, edges represent associations between concepts. This representation naturally captures dependencies between concepts, an important requirement for interpreting medical text, and a feature lacking in bag-of-words representations. We apply existing graph-based term weighting methods to weight medical concepts. Using concepts rather than terms addresses vocabulary mismatch as well as encapsulates terms belonging to a single medical entity into a single concept. In addition, we further extend previous graph-based approaches by injecting domain knowledge that estimates the importance of a concept within the global medical domain. Retrieval experiments on the TREC Medical Records collection show our method outperforms both term and concept baselines. More generally, this work provides a means of integrating background knowledge contained in medical ontologies into data-driven information retrieval approaches.