| { |
| "paper_id": "U12-1004", |
| "header": { |
| "generated_with": "S2ORC 1.0.0", |
| "date_generated": "2023-01-19T03:07:32.651556Z" |
| }, |
| "title": "Semantic Judgement of Medical Concepts: Combining Syntagmatic and Paradigmatic Information with the Tensor Encoding Model", |
| "authors": [ |
| { |
| "first": "Mike", |
| "middle": [], |
| "last": "Symonds", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "Queensland University of Technology", |
| "location": {} |
| }, |
| "email": "" |
| }, |
| { |
| "first": "Guido", |
| "middle": [], |
| "last": "Zuccon", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "CSIRO", |
| "location": { |
| "settlement": "Brisbane", |
| "country": "Australia" |
| } |
| }, |
| "email": "guido.zuccon@csiro.au" |
| }, |
| { |
| "first": "Bevan", |
| "middle": [], |
| "last": "Koopman", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "Queensland University of Technology", |
| "location": {} |
| }, |
| "email": "bevan.koopman@csiro.au" |
| }, |
| { |
| "first": "Peter", |
| "middle": [], |
| "last": "Bruza", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "Queensland University of Technology", |
| "location": {} |
| }, |
| "email": "p.bruza@qut.edu.au" |
| }, |
| { |
| "first": "Anthony", |
| "middle": [ |
| "2012" |
| ], |
| "last": "Nguyen", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "CSIRO", |
| "location": { |
| "settlement": "Brisbane", |
| "country": "Australia" |
| } |
| }, |
| "email": "anthony.nguyen@csiro.au" |
| }, |
| { |
| "first": "Michael", |
| "middle": [], |
| "last": "Symonds", |
| "suffix": "", |
| "affiliation": {}, |
| "email": "michael.symonds@qut.edu.au" |
| }, |
| { |
| "first": "", |
| "middle": [], |
| "last": "Semantic", |
| "suffix": "", |
| "affiliation": {}, |
| "email": "" |
| } |
| ], |
| "year": "", |
| "venue": null, |
| "identifiers": {}, |
| "abstract": "This paper outlines a novel approach for modelling semantic relationships within medical documents. Medical terminologies contain a rich source of semantic information critical to a number of techniques in medical informatics, including medical information retrieval. Recent research suggests that corpus-driven approaches are effective at automatically capturing semantic similarities between medical concepts, thus making them an attractive option for accessing semantic information. Most previous corpus-driven methods only considered syntagmatic associations. In this paper, we adapt a recent approach that explicitly models both syntagmatic and paradigmatic associations. We show that the implicit similarity between certain medical concepts can only be modelled using paradigmatic associations. In addition, the inclusion of both types of associations overcomes the sensitivity to the training corpus experienced by previous approaches, making our method both more effective and more robust. This finding may have implications for researchers in the area of medical information retrieval.", |
| "pdf_parse": { |
| "paper_id": "U12-1004", |
| "_pdf_hash": "", |
| "abstract": [ |
| { |
| "text": "This paper outlines a novel approach for modelling semantic relationships within medical documents. Medical terminologies contain a rich source of semantic information critical to a number of techniques in medical informatics, including medical information retrieval. Recent research suggests that corpus-driven approaches are effective at automatically capturing semantic similarities between medical concepts, thus making them an attractive option for accessing semantic information. Most previous corpus-driven methods only considered syntagmatic associations. In this paper, we adapt a recent approach that explicitly models both syntagmatic and paradigmatic associations. We show that the implicit similarity between certain medical concepts can only be modelled using paradigmatic associations. In addition, the inclusion of both types of associations overcomes the sensitivity to the training corpus experienced by previous approaches, making our method both more effective and more robust. This finding may have implications for researchers in the area of medical information retrieval.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Abstract", |
| "sec_num": null |
| } |
| ], |
| "body_text": [ |
| { |
| "text": "Semantic similarity measures are central to several techniques used in medical informatics, including: medical search (Voorhees and Tong, 2011; Cohen and Widdows, 2009) , literaturebased discovery (e.g., drug discovery (Agarwal and Searls, 2009) ), clustering (e.g., gene clustering (Glenisson et al., 2003) ), and ontology construction or maintenance (Cederberg and Widdows, 2003) .", |
| "cite_spans": [ |
| { |
| "start": 118, |
| "end": 143, |
| "text": "(Voorhees and Tong, 2011;", |
| "ref_id": "BIBREF13" |
| }, |
| { |
| "start": 144, |
| "end": 168, |
| "text": "Cohen and Widdows, 2009)", |
| "ref_id": "BIBREF4" |
| }, |
| { |
| "start": 219, |
| "end": 245, |
| "text": "(Agarwal and Searls, 2009)", |
| "ref_id": "BIBREF0" |
| }, |
| { |
| "start": 283, |
| "end": 307, |
| "text": "(Glenisson et al., 2003)", |
| "ref_id": "BIBREF5" |
| }, |
| { |
| "start": 352, |
| "end": 381, |
| "text": "(Cederberg and Widdows, 2003)", |
| "ref_id": "BIBREF3" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1" |
| }, |
| { |
| "text": "Automatically determining the similarity between medical concepts presents a number of specific challenges, including vocabulary mismatch. For example, the phrases heart attack and myocardial infarction are synonymous, referring to the same medical concept. Beyond vocabulary mismatch are situations where semantic similarity is based on implied relationships, for example the mention of an organism (e.g. Varicella zoster virus) suggests the presence of a disease (e.g. chickenpox).", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1" |
| }, |
| { |
| "text": "Existing approaches for measuring medical semantic similarities fall into two major categories: (i) those that utilise path-based measures between concepts in medical thesauri/ontologies, and (ii) corpus-based approaches that derive similarity judgements from the occurrence and cooccurrence of concepts within text, e.g., using Latent Semantic Analysis (LSA).", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1" |
| }, |
| { |
| "text": "Research comparing path-based methods with corpus-based methods highlighted the ability for corpus-based methods to provide superior performance on medical concept similarity judgements (Pedersen et al., 2007) . However, research evaluating eight different corpus-based approaches found that the performance was sensitive to the choice of training corpus (Koopman et al., 2012) . This finding means it is difficult to apply a corpus-based approach which is both robust and effective.", |
| "cite_spans": [ |
| { |
| "start": 186, |
| "end": 209, |
| "text": "(Pedersen et al., 2007)", |
| "ref_id": "BIBREF9" |
| }, |
| { |
| "start": 355, |
| "end": 377, |
| "text": "(Koopman et al., 2012)", |
| "ref_id": "BIBREF6" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1" |
| }, |
| { |
| "text": "It is important to note that the corpus based approaches referred to in this paper do not rely on syntactic information found in extra-linguistic resources, and use solely the co-occurrence statistics of words found in natural language to model word associations. Therefore, research modelling semantic associations using part of speech (POS) taggers, parsers or hand-coded resources are not within the scope of this work.", |
| "cite_spans": [ |
| { |
| "start": 337, |
| "end": 342, |
| "text": "(POS)", |
| "ref_id": null |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1" |
| }, |
| { |
| "text": "Within this paper we adapt a novel corpusbased approach, known as the tensor encoding (TE) model (Symonds et al., 2011a) , for use in judging the similarity of medical concepts. The TE approach explicitly models the two types of word associations argued to give words their meaning within structural linguistic theory. These are: (i) syntagmatic and (ii) paradigmatic associations.", |
| "cite_spans": [ |
| { |
| "start": 97, |
| "end": 120, |
| "text": "(Symonds et al., 2011a)", |
| "ref_id": "BIBREF10" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1" |
| }, |
| { |
| "text": "A syntagmatic association exists between two words if they co-occur more frequently than expected from chance. Two medical concepts that likely have a strong syntagmatic association include bone and x-ray.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1" |
| }, |
| { |
| "text": "A paradigmatic association exists between two words if they can substitute for one another in a sentence without affecting the grammaticality or acceptability of the sentence. Medical concepts that display synonymy, like heart attack and myocardial infarction display a strong paradigmatic association.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1" |
| }, |
| { |
| "text": "The TE model combines measures of syntagmatic and paradigmatic association within a single, formal framework. In this paper we demonstrate that not only does the TE model provide robust, superior performance across a wide variety of data sets when compared to past corpusbased approaches, but offers a flexible framework whose performance is not sensitive to the choice of training corpus. Our findings provide a robust and effective model for predicting semantic similarity between medical concepts, and also draws out useful statistical behavior relating to the modelling of syntagmatic and paradigmatic associations that exist within medical documents.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1" |
| }, |
| { |
| "text": "The remainder of this paper is set out as follows. Section 2 provides background on corpusbased approaches previously evaluated on medical concept similarity judgements. In Section 3 we describe the TE model and outline our novel variant for use in judging the similarity of medical concepts. Section 4 details the experiments to be used in evaluating the performance of the TE approach, with the results and their discussion following in Section 5. Concluding remarks and suggestions for future work are presented in Section 6.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1" |
| }, |
| { |
| "text": "Corpus-based models that learn the relationships between words based on their distribution in natural language have a strong history in the field of natural language processing. Some of the most well-known include LSA (Latent Semantic Analysis (Landauer and Dumais, 1997)) and HAL (Hyperspace to Analogue of Language (Lund and Burgess, 1996) ).", |
| "cite_spans": [ |
| { |
| "start": 317, |
| "end": 341, |
| "text": "(Lund and Burgess, 1996)", |
| "ref_id": "BIBREF8" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Background", |
| "sec_num": "2" |
| }, |
| { |
| "text": "A rigorous evaluation of eight different corpusbased approaches on the task of judging medical concept similarity found that the best performance was achieved using a positive pointwise mutual information (PPMI) measure. This measure had an average correlation of \u2248 0.7 with judgements made by expert human assessors (Koopman et al., 2012) .", |
| "cite_spans": [ |
| { |
| "start": 317, |
| "end": 339, |
| "text": "(Koopman et al., 2012)", |
| "ref_id": "BIBREF6" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Background", |
| "sec_num": "2" |
| }, |
| { |
| "text": "PPMI is a variation of PMI where negative values are substituted by zero-values. The strength of PMI between word q and w within a stream of text can be expressed as:", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Background", |
| "sec_num": "2" |
| }, |
| { |
| "text": "EQUATION", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [ |
| { |
| "start": 0, |
| "end": 8, |
| "text": "EQUATION", |
| "ref_id": "EQREF", |
| "raw_str": "S ppmi (q, w) = \ufffd log \ufffd p(q,w) p(q)p(w) \ufffd if log \ufffd p(q,w) p(q)p(w) \ufffd > 0 0 otherwise,", |
| "eq_num": "(1)" |
| } |
| ], |
| "section": "Background", |
| "sec_num": "2" |
| }, |
| { |
| "text": "where p(q, w) is the joint probability of q and w, and p(q), p(w) are the expected probabilities of q and w respectively. In practice, these probabilities are computed as:", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Background", |
| "sec_num": "2" |
| }, |
| { |
| "text": "p(q, w) = |D q \u2229 D w | |D| , p(q) = |D q | |D| , p(w) = |D w | |D| ,", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Background", |
| "sec_num": "2" |
| }, |
| { |
| "text": "where D q is the set of documents containing term q and D is the set of documents in the collection. Although the PPMI measure achieved the best average performance across a number of test sets, it displayed a high degree of sensitivity when the training corpus was changed, thus reducing its overall utility.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Background", |
| "sec_num": "2" |
| }, |
| { |
| "text": "Next, we introduce the tensor encoding model as a robust and effective alternative to previous proposed measures.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Background", |
| "sec_num": "2" |
| }, |
| { |
| "text": "A recent corpus-based approach, known as the tensor encoding (TE) model, was originally presented as a model of word meaning (Symonds et al., 2011a) , and later used to provide a flexible approach to semantic categorisation (Symonds et al., 2012 ). The TE model provides a formal framework for combining two measures that explicitly model syntagmatic and paradigmatic associations between words.", |
| "cite_spans": [ |
| { |
| "start": 125, |
| "end": 148, |
| "text": "(Symonds et al., 2011a)", |
| "ref_id": "BIBREF10" |
| }, |
| { |
| "start": 224, |
| "end": 245, |
| "text": "(Symonds et al., 2012", |
| "ref_id": "BIBREF12" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "The Tensor Encoding Model", |
| "sec_num": "3" |
| }, |
| { |
| "text": "As the TE model has a strong theoretical basis in linguistics, it has potential applications in other areas that deal with natural language, including e.g. information retrieval. To demonstrate, the TE model was used to perform similarity judgements within the query expansion process of an ad-hoc information retrieval task (Symonds et al., 2011b) . This method, known as tensor query expansion achieved robust and significant performance improvements over a strong benchmark model. This result was attributed to the inclusion of information about paradigmatic associations, which are not effectively modelled in existing information retrieval systems. Similarly, these paradigmatic associations are not explicitly modelled in previous corpus-based measures of medical concept similarity. We hypothesise that the inclusion of paradigmatic associations in semantic similarity measures would better capture similarities between medical concepts.", |
| "cite_spans": [ |
| { |
| "start": 325, |
| "end": 348, |
| "text": "(Symonds et al., 2011b)", |
| "ref_id": "BIBREF11" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "The Tensor Encoding Model", |
| "sec_num": "3" |
| }, |
| { |
| "text": "To support this insight, consider how the PPMI measure in Equation 1is oblivious to paradigmatic associations that may exist between two words. If word q and word w do not co-occur in any documents (i.e., |D q \u2229 D w | = 0) then S ppmi (q, w) = 0. This result suggests q and w are unrelated. However, consider a toy example using heart attack(q) and myocardial infarction(w). One clinician may use the first concept exclusively in a document, while another may use the second term exclusively. If the PPMI score between heart attack and myocardial infarction was calculated using these two example documents, the score would be zero and the two concepts considered unrelated.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "The Tensor Encoding Model", |
| "sec_num": "3" |
| }, |
| { |
| "text": "From a structural linguistic viewpoint, one might say there are no syntagmatic associations between the two concepts, as they do not co-occur in the same context (i.e., medical report). Therefore, PPMI only captures syntagmatic associations, and hence fails to model any paradigmatic information that may exist. However, consider the same example using the TE model's paradigmatic measure, and hence a pure paradigmatic perspective. Within the TE model, the strength of paradigmatic associations between two medical concepts, q and w can be defined as:", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "The Tensor Encoding Model", |
| "sec_num": "3" |
| }, |
| { |
| "text": "S par (q, w) = N \ufffd i=1 f iq .f iw max( f iq , f iw , f wq ) 2 , (2)", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "The Tensor Encoding Model", |
| "sec_num": "3" |
| }, |
| { |
| "text": "where f iq is the unordered co-occurrence frequency of concepts i and q, f iw is the unordered co-occurrence frequency of concepts i and w, and N is the number of concepts in the vocabulary.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "The Tensor Encoding Model", |
| "sec_num": "3" |
| }, |
| { |
| "text": "Intuitively, Equation 2enhances the score for concept w if q and w co-occur with the same concepts, independent of whether q and w occur within the same document. In fact this measure has a factor, 1 f wq , that reduces the paradigmatic score if concepts q and w occur within the same document often. In our simple example, this would mean that if heart attack and myocardial infarction co-occurred with any of the same terms (e.g., CPR, chest pain, etc.) then they would have a paradigmatic score greater than 0.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "The Tensor Encoding Model", |
| "sec_num": "3" |
| }, |
| { |
| "text": "The fact that pure paradigmatic information is not currently utilised within most corpus-based approaches leads us to hypothesise that more robust performance on medical concept similarity judgements can be achieved by adding paradigmatic information to the similarity estimate. In the remainder of this paper the measure in Equation (2) will be referred to as PARA.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "The Tensor Encoding Model", |
| "sec_num": "3" |
| }, |
| { |
| "text": "The TE model uses a Markov random field to formalise the estimate of observing one concept w given a second concept q:", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "The Tensor Encoding Model", |
| "sec_num": "3" |
| }, |
| { |
| "text": "EQUATION", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [ |
| { |
| "start": 0, |
| "end": 8, |
| "text": "EQUATION", |
| "ref_id": "EQREF", |
| "raw_str": "P (w|q) = 1 Z [\u03b3S par (q, w) + (1 \u2212 \u03b3)S ppmi (q, w)] ,", |
| "eq_num": "(3)" |
| } |
| ], |
| "section": "The Tensor Encoding Model", |
| "sec_num": "3" |
| }, |
| { |
| "text": "where \u03b3 \u2208 [0, 1] mixes the paradigmatic S par () and syntagmatic S ppmi () measures, and Z normalises the resulting distribution. We refer the interested reader to Symonds et al. (2012) for details. The estimate in Equation (3) can be reduced to the following rank equivalent measure of semantic similarity between q and w:", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "The Tensor Encoding Model", |
| "sec_num": "3" |
| }, |
| { |
| "text": "S TE (q, w) \u221d \u03b3S par (q, w) + (1 \u2212 \u03b3)S ppmi (q, w). (4)", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "The Tensor Encoding Model", |
| "sec_num": "3" |
| }, |
| { |
| "text": "In the remainder of this paper the model defined in Equation (4) will be referred to as TE.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "The Tensor Encoding Model", |
| "sec_num": "3" |
| }, |
| { |
| "text": "It is worth noting that the TE model formally supports the combining of any measure of syntagmatic and paradigmatic information. Therefore, if a more effective measure of syntagmatic or paradigmatic information is developed, it can be applied within the TE framework. ", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "The Tensor Encoding Model", |
| "sec_num": "3" |
| }, |
| { |
| "text": "In this section we outline the experimental setup used to evaluate the performance of the TE approach on two separate medical concept similarity judgement data sets. The first data set involves judging the similarity of 29 1 UMLS medical concept pairs. These were first developed by Pedersen et al. (Pedersen et al., 2007) and human assessments of semantic similarity were produced by 9 clinical terminologists (coders) and 3 physicians, with inter-coded relatedness equal to 0.85. Assessors scored each pair between 1 and 4, with 1 being unrelated and 4 being highly synonymous. This data set is indicated as Ped in the remainder of this paper.", |
| "cite_spans": [ |
| { |
| "start": 299, |
| "end": 322, |
| "text": "(Pedersen et al., 2007)", |
| "ref_id": "BIBREF9" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Experimental Setup", |
| "sec_num": "4" |
| }, |
| { |
| "text": "The second data set is comprised of 45 UMLS concept pairs, developed by Caviedes and Cimino (2004) , for which semantic similarity assessments were performed by three physicians. Similarities were scored between 1 and 10, with higher scores indicating a stronger similarity between concepts. This data set is indicated as Cav in the remainder of this paper.", |
| "cite_spans": [ |
| { |
| "start": 72, |
| "end": 98, |
| "text": "Caviedes and Cimino (2004)", |
| "ref_id": "BIBREF2" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Experimental Setup", |
| "sec_num": "4" |
| }, |
| { |
| "text": "Two separate corpora were used as data to prime all models; corpus statistics are shown in Table 1 . The TREC MedTrack collection consists of documents created from concatenating clinical patient records for a single visit, while the OHSUMED collection is based on MEDLINE journal abstracts.", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 91, |
| "end": 98, |
| "text": "Table 1", |
| "ref_id": "TABREF1" |
| } |
| ], |
| "eq_spans": [], |
| "section": "Experimental Setup", |
| "sec_num": "4" |
| }, |
| { |
| "text": "Following the procedure outlined by Koopman et al. (2012) the original textual documents for both corpora were translated into UMLS medi-cal concept identifiers using MetaMap, a biomedical concept identification system (Aronson and Lang, 2010). After processing, the individual documents contained only UMLS concept ids. For example, the phrase Congestive heart failure in the original document will be replaced with C0018802 in the new document. Both data sets (Ped and Cav) contained UMLS concept pairs (which may actually represent term phrases rather than single terms); converting the corpora to concepts thus allows direct comparison of the single concept pairs contained in the two data sets.", |
| "cite_spans": [ |
| { |
| "start": 36, |
| "end": 57, |
| "text": "Koopman et al. (2012)", |
| "ref_id": "BIBREF6" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Experimental Setup", |
| "sec_num": "4" |
| }, |
| { |
| "text": "When modelling paradigmatic associations it is common to consider only those terms close to the target term, i.e. within a window of text centred around the target term. However, here we used the whole document as the context window. In this way we aim to capture in MedTrack the associations that exists within the context of a single patient record and in OHSUMED the associations that exists within the context of a single medical abstract.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Experimental Setup", |
| "sec_num": "4" |
| }, |
| { |
| "text": "As the TE model in Equation (4) is parameterized over \u03b3, i.e. the mix between paradigmatic and syntagmatic information, this parameter was tuned to maximise the correlation with human similarity judgements (gold standard/ground truth labels). To fairly tune \u03b3 and also provide insight into the robustness of the TE model, a split train/test methodology was used. This was done by training on one data set and corpus to find the best value of \u03b3 and then testing on another data set/corpus, ensuring a cross corpus and cross data set combination was done for each. Table 2 summarises the results obtained following this methodology and also reports the per- formance of PPMI for comparison. Note that PPMI was shown to be the strongest performing measure when compared to thirteen other corpus based measures (Koopman et al., 2012) .", |
| "cite_spans": [ |
| { |
| "start": 807, |
| "end": 829, |
| "text": "(Koopman et al., 2012)", |
| "ref_id": "BIBREF6" |
| } |
| ], |
| "ref_spans": [ |
| { |
| "start": 563, |
| "end": 570, |
| "text": "Table 2", |
| "ref_id": "TABREF2" |
| } |
| ], |
| "eq_spans": [], |
| "section": "Experimental Setup", |
| "sec_num": "4" |
| }, |
| { |
| "text": "The effect of explicitly modelling both syntagmatic and paradigmatic information when estimating the similarity of medical concepts is shown in Figure 1 . This graph shows that the TE model achieves a much higher correlation with human judged similarity scores (with an average correlation of 0.74 over all datasets and corpora) than both the paradigmatic (PARA: 0.57) and syntagmatic (PPMI: 0.66) approaches alone. To gain a broader understanding of how each of these measures compares to our TE variant, an updated graph showing the average performance of each across all data sets is provided in Figure 2 . We refer the reader to Koopman et. al., (Koopman et al., 2012) for more details on the settings used for each of the other measures.", |
| "cite_spans": [ |
| { |
| "start": 633, |
| "end": 672, |
| "text": "Koopman et. al., (Koopman et al., 2012)", |
| "ref_id": "BIBREF6" |
| } |
| ], |
| "ref_spans": [ |
| { |
| "start": 144, |
| "end": 152, |
| "text": "Figure 1", |
| "ref_id": "FIGREF0" |
| }, |
| { |
| "start": 599, |
| "end": 607, |
| "text": "Figure 2", |
| "ref_id": "FIGREF3" |
| } |
| ], |
| "eq_spans": [], |
| "section": "Experimental Results", |
| "sec_num": "5" |
| }, |
| { |
| "text": "The sensitivity to the mixing parameter \u03b3 of the TE approach is shown in Figure 3 . This illustrates that for all datasets the best performance is achieved by some mix of both syntagmatic (PPMI) and paradigmatic (PARA) information. The robustness of the PPMI and PAR measures across datasets and corpora can be inferred by comparing the distance between the end points of the TE lines drawn in Figure 3 . The left hand side of the graph (where \u03b3 = 0) illustrates the performance of TE when only syntagmatic associations are considered, i.e. when TE uses only the PPMI measure (as Equation (4) reduces to only S ppmi () when \u03b3 = 0). The right hand side of the graph (\u03b3 = 1) shows the performance of TE when considering only paradigmatic information, i.e. when only the PARA measure is used. With most lines converging to the same point on the right hand side, this demonstrates the increased robustness the PARA measure (and therefore paradigmatic associations) brings to the overall model.", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 73, |
| "end": 81, |
| "text": "Figure 3", |
| "ref_id": "FIGREF2" |
| }, |
| { |
| "start": 394, |
| "end": 402, |
| "text": "Figure 3", |
| "ref_id": "FIGREF2" |
| } |
| ], |
| "eq_spans": [], |
| "section": "Sensitivity to the Mixing Parameter \u03b3", |
| "sec_num": "5.1" |
| }, |
| { |
| "text": "To illustrate why the combination of both paradigmatic and syntagmatic measures can achieve such robust results across all datasets and corpora we compare the correlation of PPMI, PAR and TE against human assessments on a per concept-pair basis. Figure 4 illustrates the normalised similarity scores (on log scale) of human assessors, PPMI, PARA and TE on the Caviedes and Cimino (Cav) dataset when using the OHSUMED corpora. The concept-pairs are placed in descending order of similarity as assessed by human judges, i.e. from the most similar human judged pairs to the least from left to right. The performance of a measure can be visualised by comparing its trend line with that of the descending human judged trend line. If the measure's trend line is parallel to that of the Table 3 : Example concept pairs for which the PARA measure diverges from the human judgements on the OHSUMED corpus. Document frequencies showing the prevalence of the concepts in the corpus are reported. We conclude that the PARA measure is unable to estimate accurate semantic similarity when insufficient occurrence statistics are available for either concept.", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 246, |
| "end": 254, |
| "text": "Figure 4", |
| "ref_id": "FIGREF4" |
| }, |
| { |
| "start": 780, |
| "end": 787, |
| "text": "Table 3", |
| "ref_id": null |
| } |
| ], |
| "eq_spans": [], |
| "section": "Analysis of Paradigmatic and Syntagmatic Behaviour", |
| "sec_num": "5.2" |
| }, |
| { |
| "text": "human judges, then this indicates a strong correlation.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Analysis of Paradigmatic and Syntagmatic Behaviour", |
| "sec_num": "5.2" |
| }, |
| { |
| "text": "To better understand why the paradigmatic based measure differs from human assessors in Figure 4 , the document frequencies of concept pairs 11, 16, 21, 34, 36, 37 and 38 from the Cav data set are reported in Table 3 .", |
| "cite_spans": [ |
| { |
| "start": 141, |
| "end": 144, |
| "text": "11,", |
| "ref_id": null |
| }, |
| { |
| "start": 145, |
| "end": 148, |
| "text": "16,", |
| "ref_id": null |
| }, |
| { |
| "start": 149, |
| "end": 152, |
| "text": "21,", |
| "ref_id": null |
| }, |
| { |
| "start": 153, |
| "end": 156, |
| "text": "34,", |
| "ref_id": null |
| }, |
| { |
| "start": 157, |
| "end": 160, |
| "text": "36,", |
| "ref_id": null |
| }, |
| { |
| "start": 161, |
| "end": 163, |
| "text": "37", |
| "ref_id": null |
| } |
| ], |
| "ref_spans": [ |
| { |
| "start": 88, |
| "end": 96, |
| "text": "Figure 4", |
| "ref_id": "FIGREF4" |
| }, |
| { |
| "start": 209, |
| "end": 216, |
| "text": "Table 3", |
| "ref_id": null |
| } |
| ], |
| "eq_spans": [], |
| "section": "Analysis of Paradigmatic and Syntagmatic Behaviour", |
| "sec_num": "5.2" |
| }, |
| { |
| "text": "This table shows that for these concept pairs at least one concept occurs in a very small number of documents. This provides little evidence for the accurate estimation of paradigmatic associations between the concept pairs. We therefore conclude that the PARA measure requires that concepts occur in a sufficient number of documents for an effective semantic similarity estimation.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Analysis of Paradigmatic and Syntagmatic Behaviour", |
| "sec_num": "5.2" |
| }, |
| { |
| "text": "Similar observations are valid across datasets and corpora. For example, consider the correlation of PARA with human judgements for the Pedersen et al. (Ped) data set and the MedTrack corpus, as shown in Figure 5 . The document frequencies for a number of concept pairs that show divergence from the Ped data set are shown in Table 4 .", |
| "cite_spans": [ |
| { |
| "start": 136, |
| "end": 157, |
| "text": "Pedersen et al. (Ped)", |
| "ref_id": null |
| } |
| ], |
| "ref_spans": [ |
| { |
| "start": 204, |
| "end": 212, |
| "text": "Figure 5", |
| "ref_id": "FIGREF5" |
| }, |
| { |
| "start": 326, |
| "end": 333, |
| "text": "Table 4", |
| "ref_id": "TABREF5" |
| } |
| ], |
| "eq_spans": [], |
| "section": "Analysis of Paradigmatic and Syntagmatic Behaviour", |
| "sec_num": "5.2" |
| }, |
| { |
| "text": "For these concept pairs where PARA is inconsistent with human judges, the PPMI measure effectively estimates semantic similarity. Thus the TE model, which mixes the two form of associations, is still effective even when the PARA measure is unreliable. This further supports the inclusion of both paradigmatic and syntagmatic associations for assessing semantic similarity between medical concepts. Figure 4 also illustrates a large number of discontinuities in the PPMI graph. A discontinuity, i.e. the absence of the data-point within the plot, is due to a PPMI score of zero for the concept pair 2 .", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 398, |
| "end": 406, |
| "text": "Figure 4", |
| "ref_id": "FIGREF4" |
| } |
| ], |
| "eq_spans": [], |
| "section": "Analysis of Paradigmatic and Syntagmatic Behaviour", |
| "sec_num": "5.2" |
| }, |
| { |
| "text": "In practice, these discontinuities represent instances where the concept pair never co-occurs within any document. The same situation applies across other datasets and corpora, for example the While PPMI discontinuities for concept pairs judged as unrelated by human assessors are correct estimates (as PPMI= 0 implies unrelatedness), discontinuities for concept pairs judged similar (e.g. pairs 11, 16, etc. in Figure 4 ) indicate a failure of the PPMI measure. These situations may provide the reason why the performance of PPMI, and indeed of many existing corpus-based approaches (Koopman et al., 2012) , are sensitive to the choice of priming corpus. This may indicate that the ability to successfully model syntagmatic associations is sensitive to the corpus used for priming. However, because of the results obtained by the TE model we can conclude that appropriately mixing both syntagmatic and paradigmatic associations overcomes corpus sensitivity issues.", |
| "cite_spans": [ |
| { |
| "start": 384, |
| "end": 411, |
| "text": "(e.g. pairs 11, 16, etc. in", |
| "ref_id": null |
| }, |
| { |
| "start": 584, |
| "end": 606, |
| "text": "(Koopman et al., 2012)", |
| "ref_id": "BIBREF6" |
| } |
| ], |
| "ref_spans": [ |
| { |
| "start": 412, |
| "end": 420, |
| "text": "Figure 4", |
| "ref_id": "FIGREF4" |
| } |
| ], |
| "eq_spans": [], |
| "section": "Analysis of Paradigmatic and Syntagmatic Behaviour", |
| "sec_num": "5.2" |
| }, |
| { |
| "text": "In summary, the performance (both in terms of robustness and effectiveness) of the TE model is achieved by including both syntagmatic and paradigmatic associations between medical concepts; this is due to the diversification of the type of information used to underpin the semantic similarity estimation process.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Analysis of Paradigmatic and Syntagmatic Behaviour", |
| "sec_num": "5.2" |
| }, |
| { |
| "text": "This paper has presented a novel variant of a robust and effective corpus-based approach that estimates similarity between medical concepts that strongly correlates with human judges. This approach is based on the tensor encoding (TE) model. By explicitly modelling syntagmatic and paradigmatic associations the TE model is able to outperform state of the art corpus-based approaches. Furthermore, the TE model is robust across corpora and datasets, in particular overcoming corpus sensitivity issues experienced by previous approaches. A significant contribution of this paper is to highlight the important role of paradigmatic associations. Our results suggest that paradigmatic information provides an alternative source of evidence from which semantic similarity judgements can be drawn. It is this diversity of both syntagmatic and paradigmatic information that allows the TE model to be robust and effective.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Conclusion", |
| "sec_num": "6" |
| }, |
| { |
| "text": "A possible area of future work is the development of an adaptive TE approach. An adaptive approach would determine the best mix of syntagmatic and paradigmatic information on a caseby-case basis, using corpus statistic features. Our analysis in fact has shown that paradigmatic associations require a minimum number of occurrences of concepts within documents. While, syntagmatic associations require a minimum number of co-occurrences of concept pairs within documents. These corpus statistics could represent features for a machine learning approach that predicts the optimal mix of syntagmatic and paradigmatic information for the TE model.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Conclusion", |
| "sec_num": "6" |
| }, |
| { |
| "text": "Finally, because of its effectiveness and robustness, the TE model has other potential applications beyond semantic similarity measures. One relevant application may include using the TE model within query expansion tasks in ad-hoc medical information retrieval (as this process already relies heavily on similarity judgements).", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Conclusion", |
| "sec_num": "6" |
| }, |
| { |
| "text": "The pair Lymphoid hyperplasia was removed from the original set of 30 as neither concept existed in the test collections shown in table 1.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "As the graph is in log scale, log(0) = \u2212\u221e cannot be plotted.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| } |
| ], |
| "back_matter": [], |
| "bib_entries": { |
| "BIBREF0": { |
| "ref_id": "b0", |
| "title": "Can literature analysis identify innovation drivers in drug discovery?", |
| "authors": [ |
| { |
| "first": "Pankaj", |
| "middle": [], |
| "last": "Agarwal", |
| "suffix": "" |
| }, |
| { |
| "first": "B", |
| "middle": [], |
| "last": "David", |
| "suffix": "" |
| }, |
| { |
| "first": "", |
| "middle": [], |
| "last": "Searls", |
| "suffix": "" |
| } |
| ], |
| "year": 2009, |
| "venue": "Nature reviews. Drug discovery", |
| "volume": "8", |
| "issue": "11", |
| "pages": "865--78", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Pankaj Agarwal and David B Searls. 2009. Can literature analysis identify innovation drivers in drug discovery? Nature reviews. Drug discovery, 8(11):865-78, November.", |
| "links": null |
| }, |
| "BIBREF1": { |
| "ref_id": "b1", |
| "title": "An overview of MetaMap: historical perspective and recent advances", |
| "authors": [ |
| { |
| "first": "Fran\u00e7ois-Michel", |
| "middle": [], |
| "last": "Alan R Aronson", |
| "suffix": "" |
| }, |
| { |
| "first": "", |
| "middle": [], |
| "last": "Lang", |
| "suffix": "" |
| } |
| ], |
| "year": 2010, |
| "venue": "JAMIA", |
| "volume": "17", |
| "issue": "3", |
| "pages": "229--236", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Alan R Aronson and Fran\u00e7ois-Michel Lang. 2010. An overview of MetaMap: historical perspective and recent advances. JAMIA, 17(3):229-236.", |
| "links": null |
| }, |
| "BIBREF2": { |
| "ref_id": "b2", |
| "title": "Towards the development of a conceptual distance metric for the UMLS", |
| "authors": [ |
| { |
| "first": "E", |
| "middle": [], |
| "last": "Jorge", |
| "suffix": "" |
| }, |
| { |
| "first": "James J", |
| "middle": [], |
| "last": "Caviedes", |
| "suffix": "" |
| }, |
| { |
| "first": "", |
| "middle": [], |
| "last": "Cimino", |
| "suffix": "" |
| } |
| ], |
| "year": 2004, |
| "venue": "Journal of biomedical informatics", |
| "volume": "37", |
| "issue": "2", |
| "pages": "77--85", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Jorge E Caviedes and James J Cimino. 2004. Towards the development of a conceptual distance metric for the UMLS. Journal of biomedical informatics, 37(2):77-85, April.", |
| "links": null |
| }, |
| "BIBREF3": { |
| "ref_id": "b3", |
| "title": "Using LSA and noun coordination information to improve the precision and recall of automatic hyponymy extraction", |
| "authors": [ |
| { |
| "first": "Scott", |
| "middle": [], |
| "last": "Cederberg", |
| "suffix": "" |
| }, |
| { |
| "first": "Dominic", |
| "middle": [], |
| "last": "Widdows", |
| "suffix": "" |
| } |
| ], |
| "year": 2003, |
| "venue": "Proc of CoNLL'03", |
| "volume": "", |
| "issue": "", |
| "pages": "111--118", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Scott Cederberg and Dominic Widdows. 2003. Using LSA and noun coordination information to improve the precision and recall of automatic hyponymy ex- traction. In Proc of CoNLL'03, pages 111-118, NJ, USA.", |
| "links": null |
| }, |
| "BIBREF4": { |
| "ref_id": "b4", |
| "title": "Empirical distributional semantics: Methods and biomedical applications", |
| "authors": [ |
| { |
| "first": "Trevor", |
| "middle": [], |
| "last": "Cohen", |
| "suffix": "" |
| }, |
| { |
| "first": "Dominic", |
| "middle": [], |
| "last": "Widdows", |
| "suffix": "" |
| } |
| ], |
| "year": 2009, |
| "venue": "Journal of Biomedical Informatics", |
| "volume": "42", |
| "issue": "2", |
| "pages": "390--405", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Trevor Cohen and Dominic Widdows. 2009. Empiri- cal distributional semantics: Methods and biomed- ical applications. Journal of Biomedical Informat- ics, 42(2):390-405.", |
| "links": null |
| }, |
| "BIBREF5": { |
| "ref_id": "b5", |
| "title": "Evaluation Of The Vector Space Representation In Text-Based Gene Clustering", |
| "authors": [ |
| { |
| "first": "P", |
| "middle": [], |
| "last": "Glenisson", |
| "suffix": "" |
| }, |
| { |
| "first": "", |
| "middle": [], |
| "last": "Antal", |
| "suffix": "" |
| }, |
| { |
| "first": "Y", |
| "middle": [], |
| "last": "Mathys", |
| "suffix": "" |
| }, |
| { |
| "first": "B De", |
| "middle": [], |
| "last": "Moreau", |
| "suffix": "" |
| }, |
| { |
| "first": "", |
| "middle": [], |
| "last": "Moor", |
| "suffix": "" |
| } |
| ], |
| "year": 2003, |
| "venue": "Proc Pacific Symposium of Biocomputing", |
| "volume": "", |
| "issue": "", |
| "pages": "391--402", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "P Glenisson, P Antal, J Mathys, Y Moreau, and B De Moor. 2003. Evaluation Of The Vector Space Rep- resentation In Text-Based Gene Clustering. In Proc Pacific Symposium of Biocomputing, pages 391- 402.", |
| "links": null |
| }, |
| "BIBREF6": { |
| "ref_id": "b6", |
| "title": "An Evaluation of Corpus-driven Measures of Medical Concept Similarity for Information Retrieval", |
| "authors": [ |
| { |
| "first": "Bevan", |
| "middle": [], |
| "last": "Koopman", |
| "suffix": "" |
| }, |
| { |
| "first": "Guido", |
| "middle": [], |
| "last": "Zuccon", |
| "suffix": "" |
| }, |
| { |
| "first": "Peter", |
| "middle": [], |
| "last": "Bruza", |
| "suffix": "" |
| }, |
| { |
| "first": "Laurianne", |
| "middle": [], |
| "last": "Sitbon", |
| "suffix": "" |
| }, |
| { |
| "first": "Michael", |
| "middle": [], |
| "last": "Lawley", |
| "suffix": "" |
| } |
| ], |
| "year": 2012, |
| "venue": "The 21st ACM International Conference on Information Knowledge Management", |
| "volume": "", |
| "issue": "", |
| "pages": "2439--2442", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Bevan Koopman, Guido Zuccon, Peter Bruza, Lauri- anne Sitbon, and Michael Lawley. 2012. An Eval- uation of Corpus-driven Measures of Medical Con- cept Similarity for Information Retrieval. In The 21st ACM International Conference on Information Knowledge Management 2012, pages 2439-2442.", |
| "links": null |
| }, |
| "BIBREF7": { |
| "ref_id": "b7", |
| "title": "Solution to Plato's Problem: The Latent Semantic Analysis Theory of Acquisition, Induction and Representation of Knowledge", |
| "authors": [ |
| { |
| "first": "K", |
| "middle": [], |
| "last": "Thomas", |
| "suffix": "" |
| }, |
| { |
| "first": "Susan", |
| "middle": [ |
| "T" |
| ], |
| "last": "Landauer", |
| "suffix": "" |
| }, |
| { |
| "first": "", |
| "middle": [], |
| "last": "Dumais", |
| "suffix": "" |
| } |
| ], |
| "year": 1997, |
| "venue": "Psychological Review", |
| "volume": "104", |
| "issue": "2", |
| "pages": "211--240", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Thomas K. Landauer and Susan T. Dumais. 1997. Solution to Plato's Problem: The Latent Semantic Analysis Theory of Acquisition, Induction and Rep- resentation of Knowledge. Psychological Review, 104(2):211-240.", |
| "links": null |
| }, |
| "BIBREF8": { |
| "ref_id": "b8", |
| "title": "Producing high-dimensional semantic spaces from lexical cooccurrence. Behavior research methods, instruments and computers", |
| "authors": [ |
| { |
| "first": "Kevin", |
| "middle": [], |
| "last": "Lund", |
| "suffix": "" |
| }, |
| { |
| "first": "Curt", |
| "middle": [], |
| "last": "Burgess", |
| "suffix": "" |
| } |
| ], |
| "year": 1996, |
| "venue": "", |
| "volume": "28", |
| "issue": "", |
| "pages": "203--208", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Kevin Lund and Curt Burgess. 1996. Producing high-dimensional semantic spaces from lexical co- occurrence. Behavior research methods, instru- ments and computers, 28:203-208.", |
| "links": null |
| }, |
| "BIBREF9": { |
| "ref_id": "b9", |
| "title": "Measures of semantic similarity and relatedness in the biomedical domain", |
| "authors": [ |
| { |
| "first": "Ted", |
| "middle": [], |
| "last": "Pedersen", |
| "suffix": "" |
| }, |
| { |
| "first": "Sarguei", |
| "middle": [], |
| "last": "Pakhomov", |
| "suffix": "" |
| } |
| ], |
| "year": 2007, |
| "venue": "Journal of Biomedical Informatics", |
| "volume": "40", |
| "issue": "3", |
| "pages": "288--299", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Ted Pedersen, Sarguei Pakhomov, Siddharth Patward- han, and Christopher Chute. 2007. Measures of semantic similarity and relatedness in the biomed- ical domain. Journal of Biomedical Informatics, 40(3):288-299.", |
| "links": null |
| }, |
| "BIBREF10": { |
| "ref_id": "b10", |
| "title": "Modelling Word Meaning using Efficient Tensor Representations", |
| "authors": [ |
| { |
| "first": "Michael", |
| "middle": [], |
| "last": "Symonds", |
| "suffix": "" |
| }, |
| { |
| "first": "Peter", |
| "middle": [], |
| "last": "Bruza", |
| "suffix": "" |
| }, |
| { |
| "first": "Laurianne", |
| "middle": [], |
| "last": "Sitbon", |
| "suffix": "" |
| }, |
| { |
| "first": "Ian", |
| "middle": [], |
| "last": "Turner", |
| "suffix": "" |
| } |
| ], |
| "year": 2011, |
| "venue": "Pacific Asia Conference on Language, Information and Computation 2011", |
| "volume": "", |
| "issue": "", |
| "pages": "313--322", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Michael Symonds, Peter Bruza, Laurianne Sitbon, and Ian Turner. 2011a. Modelling Word Meaning using Efficient Tensor Representations. In Pacific Asia Conference on Language, Information and Compu- tation 2011, pages 313-322.", |
| "links": null |
| }, |
| "BIBREF11": { |
| "ref_id": "b11", |
| "title": "Tensor Query Expansion: a cognitive based relevance model", |
| "authors": [ |
| { |
| "first": "Michael", |
| "middle": [], |
| "last": "Symonds", |
| "suffix": "" |
| }, |
| { |
| "first": "Peter", |
| "middle": [], |
| "last": "Bruza", |
| "suffix": "" |
| }, |
| { |
| "first": "Laurianne", |
| "middle": [], |
| "last": "Sitbon", |
| "suffix": "" |
| }, |
| { |
| "first": "Ian", |
| "middle": [], |
| "last": "Turner", |
| "suffix": "" |
| } |
| ], |
| "year": 2011, |
| "venue": "Australasian Document Computing Symposium", |
| "volume": "", |
| "issue": "", |
| "pages": "87--94", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Michael Symonds, Peter Bruza, Laurianne Sitbon, and Ian Turner. 2011b. Tensor Query Expansion: a cog- nitive based relevance model. In Australasian Doc- ument Computing Symposium 2011, pages 87-94.", |
| "links": null |
| }, |
| "BIBREF12": { |
| "ref_id": "b12", |
| "title": "A Tensor Encoding Model for Semantic Processing", |
| "authors": [ |
| { |
| "first": "Michael", |
| "middle": [], |
| "last": "Symonds", |
| "suffix": "" |
| }, |
| { |
| "first": "Peter", |
| "middle": [], |
| "last": "Bruza", |
| "suffix": "" |
| }, |
| { |
| "first": "Laurianne", |
| "middle": [], |
| "last": "Sitbon", |
| "suffix": "" |
| }, |
| { |
| "first": "Ian", |
| "middle": [], |
| "last": "Turner", |
| "suffix": "" |
| } |
| ], |
| "year": 2012, |
| "venue": "The 21st ACM International Conference on Information Knowledge Management 2012", |
| "volume": "", |
| "issue": "", |
| "pages": "2267--2270", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Michael Symonds, Peter Bruza, Laurianne Sitbon, and Ian Turner. 2012. A Tensor Encoding Model for Semantic Processing. In The 21st ACM Interna- tional Conference on Information Knowledge Man- agement 2012, pages 2267-2270.", |
| "links": null |
| }, |
| "BIBREF13": { |
| "ref_id": "b13", |
| "title": "Overview of the TREC Medical Records Track", |
| "authors": [ |
| { |
| "first": "Ellen", |
| "middle": [], |
| "last": "Voorhees", |
| "suffix": "" |
| }, |
| { |
| "first": "Richard", |
| "middle": [], |
| "last": "Tong", |
| "suffix": "" |
| } |
| ], |
| "year": 2011, |
| "venue": "Proc of TREC'11", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Ellen Voorhees and Richard Tong. 2011. Overview of the TREC Medical Records Track. In Proc of TREC'11, MD, USA.", |
| "links": null |
| } |
| }, |
| "ref_entries": { |
| "FIGREF0": { |
| "text": "Correlation of medical concept judgements produced by PARA, PPMI and TE with those produced by human assessors.", |
| "uris": null, |
| "type_str": "figure", |
| "num": null |
| }, |
| "FIGREF2": { |
| "text": "TE sensitivity to the mixing parameter \u03b3. Results show that the TE model is robust across datasets and corpora.", |
| "uris": null, |
| "type_str": "figure", |
| "num": null |
| }, |
| "FIGREF3": { |
| "text": "Comparison of all corpus-based measures (fromKoopman et al., 2012), including TE and PARA; correlations averaged across datasets and corpora.", |
| "uris": null, |
| "type_str": "figure", |
| "num": null |
| }, |
| "FIGREF4": { |
| "text": "Normalised similarity scores (on log scale) of human assessors, PPMI, PARA and TE on Caviedes and Cimino dataset (Cav) when using the OHSUMED corpus for priming.Ped data set on MedTrack corpus shown in", |
| "uris": null, |
| "type_str": "figure", |
| "num": null |
| }, |
| "FIGREF5": { |
| "text": "Normalised similarity scores (on log scale) of human assessors, PPMI, PARA and TE on Pedersen et al dataset (Ped) when using the Medtrack corpus for priming.", |
| "uris": null, |
| "type_str": "figure", |
| "num": null |
| }, |
| "TABREF1": { |
| "content": "<table><tr><td colspan=\"3\">Test: Corpus (data set) Training: Corpus (data set) \u03b3</td><td>TE</td><td>PPMI</td></tr><tr><td>MedTrack (Ped)</td><td>OHSUMED (Ped)</td><td colspan=\"2\">0.5 r = 0.6706 r = 0.4674</td></tr><tr><td>MedTrack (Cav)</td><td>MedTrack (Ped)</td><td colspan=\"2\">0.5 r = 0.6857 r = 0.6154</td></tr><tr><td>OHSUMED (Ped)</td><td>OHSUMED (Cav)</td><td colspan=\"2\">0.2 r = 0.7698 r = 0.7427</td></tr><tr><td>OHSUMED (Cav)</td><td>MedTrack (Cav)</td><td colspan=\"2\">0.4 r = 0.8297 r = 0.8242</td></tr></table>", |
| "text": "Document collections (corpora) used.", |
| "html": null, |
| "type_str": "table", |
| "num": null |
| }, |
| "TABREF2": { |
| "content": "<table/>", |
| "text": "Performance of TE using the \u03b3 produced by the specified train/test splits; performance of PPMI included for comparison.", |
| "html": null, |
| "type_str": "table", |
| "num": null |
| }, |
| "TABREF5": { |
| "content": "<table><tr><td/><td>0.1000</td><td/><td/><td/></tr><tr><td/><td>0.0500</td><td/><td/><td/></tr><tr><td>Similarity Score</td><td>0.0010 0.0050 0.0100</td><td/><td/><td/></tr><tr><td/><td>0.0001 0.0005</td><td colspan=\"2\">human assessed PPMI PARA TE</td><td/></tr><tr><td/><td>0</td><td>10</td><td>20 concept pair</td><td>30</td><td>40</td></tr></table>", |
| "text": "Example concept pairs for which the PARA measure diverges from the human judgements on the MedTrack corpus. Document frequencies showing the prevalence of the concepts in the corpus are reported. We conclude that the PARA measure is unable to estimate accurate semantic similarity when insufficient occurrence statistics are available for either concept.", |
| "html": null, |
| "type_str": "table", |
| "num": null |
| } |
| } |
| } |
| } |
