Biomedical literature incorporates millions of figures, which are a rich and important knowledge resource for biomedical researchers. Scientists need access to the figures and the knowledge they represent in order to validate research findings and to generate new hypotheses. By themselves, these figures are nearly always incomprehensible to both humans and machines and their associated texts are therefore essential for full comprehension. The associated text of a figure, however, is scattered throughout its full-text article and contains redundant information content. In this paper, we report the continued development and evaluation of several figure summarization systems, the FigSum+ systems, that automatically identify associated texts, remove redundant information, and generate a text summary for every figure in an article. Using a set of 94 annotated figures selected from 19 different journals, we conducted an intrinsic evaluation of FigSum+. We evaluate the performance by precision, recall, F1, and ROUGE scores. The best FigSum+ system is based on an unsupervised method, achieving F1 score of 0.66 and ROUGE-1 score of 0.97. The annotated data is available at figshare.com (http://figshare.com/articles/Figure_Associated_Text_Summarization_and_Evaluation/858903).
Citation: Polepalli Ramesh B, Sethi RJ, Yu H (2015) Figure-Associated Text Summarization and Evaluation. PLoS ONE 10(2): e0115671. https://doi.org/10.1371/journal.pone.0115671
Academic Editor: Indra Neil Sarkar, University of Vermont, UNITED STATES
Received: August 27, 2013; Accepted: November 26, 2014; Published: February 2, 2015
Copyright: © 2015 Polepalli Ramesh et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited
Funding: Research reported in this publication was supported in part by 1R01GM095476 to Hong Yu, by a start-up fund from University of Massachusetts Medical School to Hong Yu, and by the National Center for Advancing Translational Sciences of the National Institutes of Health under award number UL1TR000161. The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
Figures in biomedical publications are an essential part of biomedical knowledge. Futrelle  found that nearly 50% of article content in the biological domain is figure related. Figures assist researchers by providing evidence to support their finding, report their discovery, and generate new research hypotheses. On the other hand, hundreds of millions of figures are available in biomedical literature, which makes it difficult for biomedical researchers to search for figures. Therefore, we are developing an intelligent figure search engine (http://figuresearch.askhermes.org). Currently our figure search engine is available as a SciVerse API and has indexed over 4 million full-text biomedical journal articles published by Elsevier.
Given the enormous number of figures in biomedical literature, a key aspect in building an effective figure search engine is the ability to automatically interpret figure content. A number of studies have examined various approaches for the analysis and retrieval of relevant figures from literature [2–13]. The ImageCLEF (http://www.imageclef.org/) competition for automatic annotation and retrieval of images from literature has been held annually for the last 10 years. But, there is very limited research on extracting information related to figures from the full paper text in the biomedical domain . Demner-Fushman  emphasized the importance of analyzing the text associated with the figure for its comprehension.
Our initial evaluation  showed that for a figure to be comprehended, it must be interpreted in conjunction with the text that refers to it in the article. We evaluated figure comprehension when a figure was presented (1) with its caption only, (2) with its caption along with the article title and abstract, and (3) with the article full text. The study found that presentation of the figure to biomedical researchers with just the title and abstract failed to convey 30% of the information related to the figure, compared to comprehension of the figure with the full text article. For example, Fig. 1 shows a figure along with its caption. The caption information alone is not sufficient for complete comprehension of the figure. Hence, the associated text from the full-text of the article is required to completely understand figures . However, the associated text can be scattered throughout the full-text article and, moreover, can be redundant .
We therefore developed a preliminary figure summarization system called FigSum  that automatically generates a summary for every figure by extracting summary sentences from a full-text article. FigSum selects sentences to be included in the summary based on word-level similarities between the sentences and figure captions. A pilot evaluation showed biologists prefer the generated summaries , as they provide users with a new way for comprehending figure content without spending time navigating through the full-text article.
In our previous work we did not explore and evaluate other text summarization approaches. In this study, we compare and evaluate several summarization approaches, which we implemented as FigSum+ systems:
- The baseline FigSum approach, which is an information retrieval (IR) based approach wherein we find the sentences associated with a figure by finding sentences that are most similar to the figure caption.
- The surface-cue approach, in which we generate a figure summary by identifying sentences and paragraphs that explicitly refer the figure.
- A hybrid approach, in which we first identify paragraphs that explicitly refer the figure using the surface-cue approach and then we rank sentences by the centroid-based summarization algorithm.
Summarization is one of the most extensively studied fields in natural language processing (NLP). The summarization approaches can be broadly classified as extractive and abstractive [20, 21]. Extractive approaches extract and concatenate sentences from a text corpus to construct a summary, whereas abstractive summarization relies on natural language generation approaches that build new sentences representing the content of a text corpus to be summarized. In this work, we focus on the task of extractive summarization based on the text associated with a biomedical figure. The following sections review related work in open-domain text summarization, text summarization in the biomedical domain, and figure summarization.
Extractive summarization identifies sentences/paragraphs that subsume the key points of a text or a collection of texts. An early work by Luhn  proposed a simple idea based on the intuition that words occurring frequently in a document tend to describe the main topic and therefore sentences containing those frequent words shall be selected. Later studies improved this strategy by adding weight to words, using different techniques [23–27]. For example, Brunn et al.  used syntactic parsing to identify important words for summarization. Approaches that identify summary sentences based on location or other structural characteristics were also developed. For example, Nakov et al.  used citance (text that surrounds a citation reference) to summarize a document.
Edmundson  applied a linear function that combines different factors, including resemblance to the title, indicative context cues (e.g., in summary), keywords, and sentence location. Myaeng and Jang  extended this work by adding centrality of the sentence to the document to select summary sentences.
Later studies explored various information retrieval (IR) techniques, such as the TF × IDF weighting scheme, which alleviates the negative impact of overweighting of some common words [24, 31–33], and latent semantic analysis, which derives an implicit representation of text semantics based on observed word co-occurrences for summarization [34, 35]. For instance, Hovy and Lin  developed SUMMARIST, which integrates IR approaches, topic signatures (words that are highly descriptive of a document), dictionaries, and semantic knowledge derived from WordNet  to generate a summary. Inspired by link analyses and page rank algorithms for Web document retrieval, Mihalcea et al.  and Erkan et al.  applied a graph-based ranking method to select important sentences based on the graph derived from words and sentences. Radev et al.  developed a MEAD summarizer that generates summaries based on a cluster centroid calculated by TF × IDF word similarity.
Studies also explored supervised machine learning approaches for summarization , [40–43]. Kupiec et al.  developed a Naïve Bayes classifier using the following five features to select summary sentences from 188 documents: (1) length of the sentence, (2) occurrence of common phrases (such as “In conclusion”) or phrases appearing after sections such as “results” and “discussion”, (3) location of the paragraph in the document, (4) occurrence of high frequency words as in , and (5) sentences containing proper nouns and acronyms. Wang et al.  and Hirao et al.  ranked sentences using a support vector machine classifier to generate summaries. Leskovec et al.  built semantic graphs to extract subject–object–predicate triplets from sentences and then trained a support vector machine classifier to extract salient sentence triplets for summarization.
Evaluation is important in all NLP tasks. Mani  discussed various summarization evaluation criteria, including coherence, informativeness, relative utility, and relevance of the summary. Evaluation methods include word similarity measures such as cosine similarity , the overlap of a sequence of words that include n-grams (sequences of n number of word tokens) and longest common subsequence [46, 47], and the Bleu  machine translation evaluation measure for summarization . The Document Understanding Conference (DUC) adopted the ROUGE package for content-based evaluation . Among various summarization evaluation metrics[51, 52], ROUGE score is widely used and is calculated based on n-gram overlap between the gold standard and the summary generated. The scores range between zero and one, with a higher score indicating a summary closer to the gold standard. In our study we apply ROUGE to evaluate the quality of the summary generated by our system by comparing it to the gold standard.
Open-domain summarization approaches are based on similarity and term occurrence approaches and would not be the optimal choice for biomedical text due to domain-specific characteristics. Biomedical summarization systems are frequently built upon biomedical knowledge resources, including the Medical Subject Headings (MeSH), the Unified Medical Language System (UMLS), and the Gene Ontology (GO) project, to overcome the challenge of domain-specific jargons.
Chiang, et al.  developed GeneLibrarian, which generates a viewgraph of genes related to the input query based on GO similarity. The system also generates a summary of a gene by selecting sentences based on term occurrences. Ling, et al.  developed approaches to automatically generate a structured gene summary by first retrieving gene-related documents and then extracting sentences containing factual information about the target gene. Jin, et al.  developed a query-based gene summarization system that integrates the page rank algorithm, sentence similarity, and the function of the gene represented by GO.
Many studies focused on summarizing the content in biomedical text using semantic resources. Bhattacharya, et al.  developed a method that computes similarities between the MeSH terms assigned to an article in addition to its word tokens and then returns the top N-ranked sentences as summary sentences. Plaza  generated summaries based on where the sentence resides. For example, the first few sentences are typical summary sentences. Reeve, et al.  developed the BioChain system using the concept chaining technique which links semantically related concepts in text using the UMLS . Sentences with strong concept chains (where strength is based on the number of concepts) are used to form the summary. Fiszman, et al.  applied hand-crafted transformation rules to the output of SemRep (http://semrep.nlm.nih.gov/) to summarize content. SemRep is a system that extracts biomedical concepts and relations relevant to a given query from the MEDLINE records. Workman, et al.  later modified this work to generate domain-specific summaries to support database curation. Workman and Hurdle  applied SemRep to citations obtained from PubMed. They analyzed the outputs using statistical methods to automatically identify salient data in bibliographic text for summarization. Shang et al.  extended the work of Fiszman, et al.  to develop a multi-document summarizer for a given biomedical concept. Concepts and relations in sentences are extracted using SemRep. The sentences that contain high-frequency relations are then extracted as a summary. Other studies [64, 65] explored knowledge from the UMLS to construct a graph and then selecting summary sentences based on node clustering.
Futrelle  proposed the idea of figure summarization. He described the challenges related to summarizing figures and emphasized the importance of captions and referring text. Bhatia and Mitra  applied a supervised approach to summarize document objects such as figures, tables and algorithms on a set of 290 document elements. Wu and Carberry  identified relevant paragraphs for images in news domain articles.
We developed a preliminary summarization system, FigSum , for the biomedical domain. FigSum first classifies sentences into the introduction, methods, results, and discussion categories using a supervised machine learning classifier . Each sentence is then scored based on its TF × IDF weighted cosine similarity with the figure caption and the article’s central theme. The top-scoring sentence in each category is included in the summary. The FigSum system is integrated into our larger figure search system (http://figuresearch.askhermes.org). An online survey revealed that 65.2% participants found that FigSum summaries improved figure comprehension . The current study explores additional figure summarization methods and performs an intrinsic evaluation to compare the performance of all systems.
We explored several different summarization systems in which we explored different features. In the following, we first describe features and then systems.
Features used for Summarization
We explored a number of features to build figure summarization systems.
- 1). IR based features
- a). Caption similarity feature—The cosine similarity value between each of the candidate sentences in the full text and the figure caption.
- b). Title similarity feature—The cosine similarity between each of the candidate sentences in the full text and the article title.
- c). Reference sentence similarity feature—The cosine similarity between each of the candidate sentences and sentences referring to the figure.
- d). TFIDF feature—The text association between each of the candidate sentences in the full text and the figure caption is computed by calculating the TF × IDF vector for every candidate sentence and figure caption. A score is calculated as the cosine similarity of the TF × IDF vectors of candidate sentences and the figure caption.
- 2). Reference Features
- a). Figure reference sentence feature—This feature represents if the sentence is figure referring (i.e., a sentence that incorporates figure reference cues such as Fig. X).
- b). Figure reference paragraph feature—This feature represents if the sentence belongs to the paragraph referring to the figure.
- 3). Hybrid feature—We first identify paragraphs in the full text article that contain figure reference sentences. We apply MEAD , a centroid-based text summarizer as described earlier on these sentences that are a part of the figure referring paragraphs. The n top scoring sentences are selected as summary sentences.
- 4). Position
- a). Distance from start feature—The position of the sentence from the start of the article.
- b). Distance from end feature—The position of the sentence from the end of the article.
- c). Distance from reference sentence feature—This is a binary feature that indicates if the candidate sentence is within 10 sentences of the reference sentence.
- 5). Sentence length feature—The length of the sentence.
- 6). Cue words and phrase feature—Authors of articles use certain cue words and phrases to describe document elements such as figures, as discussed in . We use the list of 140 cue words and phrases listed in  and add the presence or absence of these cue words in the sentence as a binary feature.
Figure Summarization Systems
In this section we describe a total of 23 figure summarization systems, which include our unsupervised FigSum+ methods, and other unsupervised and supervised systems we built for comparison with our system.
FigSum+ Systems. Fig. 3 shows the general pipeline of the unsupervised FigSum+ systems. Given a full text article, the Text Extractor module extracts individual sentences from the article. If the article is in XML file format, an XML parser module will process the text to extract sentences from the XML file. If the article is in PDF format, the PDF to text converter (PDFTextStream—http://snowtide.com) tool extracts the text from the PDF document and then we split the text to individual sentences using an in-house sentence splitter, which splits sentences by determining sentence boundaries such as period. The figure summarization module utilizes five unsupervised techniques, as described below, to summarize figures in the article and generate a summary for each figure.
Each implementation of the FigSum+ system differs by including only one of the five modules described in Section 3.2.1 and shown in the Figure Summarization component above: Similarity, TFIDF, SurfaceCue, Paragraph, or Hybrid.
We describe five different implementations of our unsupervised FigSum+ systems, which differ on the features used in the figure summarization module. Each implementation of the FigSum+ system differs by including one of the following five figure summarization modules: 1a, 1b, 2a, 2b, or 3.
- (1). IR-based approaches: We explore two IR-based approaches for summarization.
- (a). Similarity—We select the top scoring sentences for the caption similarity feature as figure summary.
- (b). TFIDF—We select the top scoring sentences for the TFIDF feature as figure summary.
- (2). Surface-cue approaches: We identify summary content using surface cues.
- (a). SurfaceCue—We use the figure reference sentence feature. It extracts all figure referring sentences in the full text as figure summary.
- (b). Paragraph—We use the figure reference paragraph feature. It extracts all paragraphs containing figure referring sentences as figure summary.
- (3). Hybrid—We select the top scoring sentences for the hybrid feature as figure summary.
Unsupervised baseline systems. For comparison, we built three additional unsupervised systems as the baseline systems: RandomSent, RandomPara and MEAD. The RandomSent system randomly selects n sentences from the article as the summary for the figure. The RandomPara system randomly selects n paragraphs and then includes the first sentence of every randomly selected paragraph as the summary for the figure. For the last baseline system, MEAD, we applied the centroid summarizer MEAD to the entire full text article and select n top scoring sentences as the summary for each figure.
Supervised baseline systems. In FigSum+, we use five features as described in section 3.2.1; namely: caption similarity feature, TFIDF feature, figure reference sentence feature, figure reference paragraph feature and hybrid feature. We explored each of these features individually and trained baseline supervised machine learning models to generate figure summaries for evaluation. Each individual feature was used with both a naïve bayes (NB) and Support Vector Machine (SVM) classifier, thus resulting in 10 baseline supervised systems: NBSimilarity, NBTFIDF, NBSurfaceCue, NBParagraph, NBHybrid, SVMSimilarity, SVMTFIDF, SVMSurfaceCue, SVMParagraph and SVMHybrid.
Unsupervised state-of-the-art system. We also implemented the state of the art unsupervised system, FigSum, which summarizes the figure as described earlier, for performance comparison with FigSum+.
Supervised state-of-the-art system. We implemented the state-of-the-art system described in  by building two systems, NBSOTA and SVMSOTA, using the NB and SVM models respectively, with the features described in . The features used are: figure reference sentence, figure reference paragraph, caption similarity, reference sentence similarity, distance from reference sentence and cue words.
We then extended the state-of-the-art system and build two more systems, NBSOTA+ and SVMSOTA+ using NB and SVM respectively, that incorporate all the features described in Section 3.1.
Since the datasets are of different sizes, we calculate the micro-average of recall (R), precision (P), and F1 (F) scores to evaluate the summaries generated by each of the figure summarization systems described in Section 3.2. Recall is defined as the ratio of the number of sentences correctly identified by the system to the total number of sentences in the gold standard, precision is defined as the ratio of the number of sentences correctly identified by the system to the total number of sentences identified by the system, and the F1 score is the harmonic mean of recall and precision: (1) (2) (3)
We also compute the ROUGE score using the parameters established by DUC 2007 . Eq (4) gives the formula to calculate ROUGE-N, where n stands for the length of the n-gram, gramn, and Countmatch(gramn) is the maximum number of n-grams co-occurring in a candidate summary and a set of reference summaries. For every sentence in the summary generated by the FigSum+ implementation, we calculate the ROUGE score against every sentence in the gold standard using the formula in Eq (4) and retain the best scores. Then we calculate the average of the best ROUGE score sentences for every figure: ROUGE-1 (R1) compares summaries based on the co-occurrence of unigrams (single words), ROUGE-2 (R2) compares summaries based on the co-occurrence of bigrams (two consecutive words), and ROUGE-SU4 (RSU4) compares summaries based on the co-occurrence of skip bigrams with a maximum gap length of four .(4)
We evaluated all the systems we built (FigSum+, baseline, and state-of-the-art) on a set of 19 full-text biomedical articles. Nine articles were randomly selected from our BioDRB corpus, a collection of 24 GENIA full-text articles fully annotated by us for discourse connectives and relations . Four biologists with expertise in the biology domain each selected either two or three additional articles from various biomedical journals, for a total of 10 additional articles. The combined dataset of 19 articles comprises 94 figures and is made publicly available on figshare.com. The five FigSum+ implementations are evaluated against the following two gold standards developed on these full-text articles; we selected two gold standards built using different approaches to show the robustness and efficacy of the five different techniques for figure summarization:
- a). FigSumGS1 dataset—A gold standard of 94 figures from 19 articles from various biomedical journals was created as follows: four biologists (B1—B4) read two papers each, for a sub-total of 8 articles, and then selected sentences within each article that summarized figure content. In addition, two (B1 and B2) of the four biologists, read and selected sentences from 11 additional articles, thus yielding a total of 19 articles in the gold standard. The two biologists (B1 and B2) identified 303 and 383 sentences, respectively. They had an inter-annotator agreement (IAA) of 0.68 Cohen’s κ value on the subset of 11 articles, which indicates a fair agreement between the annotators. The gold standard consists of a total of 678 sentences from 19 articles with a micro average of 7.21 sentences per figure and a macro average of 7.73 sentences per figure.
- b). FigSumGS2 dataset—A second gold standard consisting of a subset of 17 articles from the 19 articles collected in (a) was created using the guideline that was developed to evaluate the FigSum system . Seven annotators with advanced degrees (MS and above) selected three to four sentences that best described the background of the figure, the methods used to generate the figure, the outcome of the figure, and the conclusion inferred from the figure on the subset of 17 articles consisting of 84 figures; this subset was chosen from the 19 articles due to constraints of manual annotation. Hence, for each figure, a summary consisting of 12 to 16 sentences was obtained. All seven annotators together identified 869 unique sentences from the 17 articles with a micro average of 10.34 unique sentences per figure and a macro average of 10.44 unique sentences per figure.
We conducted an intrinsic evaluation to compare the performance of all five FigSum+ implementations against baseline and state of the art unsupervised and supervised systems. Table 2 and Table 3 show the average performance of the various systems we built for summarization on the FigSumGS1 and FigSumGS2 datasets respectively. We chose the value of top n to be equal to the average number of sentences per figure in the gold standard. Hence, n is equal to 8 and 11 sentences per figure for FigSumGS1 and FigSumGS2 datasets respectively.
Baseline Systems Result
For unsupervised baseline case, the RandomSent system had an F1 score performance of 0.06 and 0.08 and R1 scores of 0.28 and 0.32 on FigSumGS1 and FigSumGS2 datasets. The RandomPara system had an F1 score performance of 0.01 on both gold standards and R1 scores of 0.22 and 0.32 on FigSumGS1 and FigSumGS2 datasets respectively. The MEAD system achieved an F1 score performance of 0.05 and 0.07 and R1 scores of 0.30 and 0.36 on FigSumGS1 and FigSumGS2 datasets respectively. Whereas the state of the art unsupervised method FigSum system had an F1 score performance of 0.22 and 0.18 and R1 score of 0.51 and 0.55 on FigSumGS1 and FigSumGS2 datasets respectively.
For supervised baseline case, all the implementations of the baseline SVM systems, except for the system using the hybrid feature, failed to generate summaries. Both the NB and SVM based systems using the hybrid feature, NBHybrid and SVMHybrid, performed similarly and had the best baseline F1 score performance of 0.49 and 0.26 and R1 performance of 0.95 and 0.75 on the FigSumGS1 and FigSumGS2 datasets respectively.
State-of-the-art Systems Result
For unsupervised state-of-the-art case, the unsupervised method FigSum system had an F1 score performance of 0.22 and 0.18 and R1 score of 0.51 and 0.55 on FigSumGS1 and FigSumGS2 datasets respectively.
For supervised state-of-the-art case, the NB-based supervised systems performed well compared to the SVM-based model similar to performance in article . On FigSumGS1 dataset, the NB-based state-of-the-art systems NBSOTA and NBSOTA+ had an F1 score performance of 0.53 but SVMSOTA+ achieved the second best R1 score of 0.95. Similarly, on FigSumGS2 dataset, NBSOTA and NBSOTA+ had the best F1 score performance of 0.38 and SVMSOTA+ achieved the best R1 score of 0.76.
Our FigSum+ Systems Result
The SurfaceCue implementation of FigSum+ achieves the highest precision on both gold standards (0.96 and 0.63 on FigSumGS1 and FigSumGS2 datasets respectively) and the Paragraph implementation results in the highest recall (0.82 and 0.42 on FigSumGS1 and FigSumGS2 datasets respectively) and the highest F1 score (0.66 and 0.41 on FigSumGS1 FigSumGS2 datasets respectively). The Hybrid implementation performs second best, yielding F1 scores of 0.62 and 0.39, respectively, on FigSumGS1 and FigSumGS2 datasets.
The ROUGE score evaluation of SurfaceCue resulted in the highest R1, R2, and RSU4 scores, all above 0.97, on FigSumGS1 dataset. Similarly, SurfaceCue resulted in the highest R1 score of 0.76 on FigSumGS2 dataset.
In this study, we developed and investigated five implementations of FigSum+ to automatically summarize every figure in a full-text biomedical article. Our summarization approaches remove redundant information by extracting sentences associated with the figure, reducing the redundancy and generating a succinct summary for every figure. We evaluated the performance of these approaches against two sets of gold standards. The first gold standard was comprised of 94 figures from 19 PMC articles (FigSumGS1 dataset) and the second, a subset of 84 figures from 17 articles in the FigSumGS1 dataset (FigSumGS2 dataset). The FigSumGS1 dataset showed a good IAA of 0.68 Cohen’s κ for a subset of 11 articles.
We first compared the performance of the five FigSum+ systems against unsupervised baseline (RandomSent, RandomPara and MEAD) and unsupervised state-of-the-art (FigSum) systems. The improvement in both F1 score and ROUGE performance of SurfaceCue, Paragraph, Hybrid compared to all unsupervised systems was statistically significant (t-test, p < 0.05) on the FigSumGS1 dataset. Whereas, for the FigSumGS2 dataset comparison of unsupervised baseline systems, the improvement in the ROUGE score performance of SurfaceCue, Paragraph and Hybrid was statistically significant (t-test, p < 0.05) but the F1 score performance of only Paragraph and Hybrid was statistically significant (t-test, p < 0.05).
Supervised baseline systems using the same individual features as in the FigSum+ systems were built using the NB and SVM machine learning models. All baseline SVM systems except for the system using the hybrid feature failed to generate figure summaries on both datasets. Among the supervised baseline systems based on NB, the system using the reference paragraph feature achieved an F1 score performance of 0.59 and 0.40 on FigSumGS1 and FigSumGS2 datasets respectively. The NB system using the hybrid feature had the highest R1 performance of 0.95 and 0.76 on FigSumGS1 and FigSumGS2 datasets respectively. The difference in F1 and ROUGE score performance of NB based systems was statistically significant over the Paragraph and Hybrid (t-test, p < 0.05).
We also compared the performance of the FigSum+ systems against state-of-the-art supervised systems as described in section 3.2.5. The F1 score performance of the Paragraph and Hybrid systems were statistically significantly better than all state-of-the-art supervised systems (t-test, p < 0.05). In addition, the F1 score performance of SurfaceCue was statistically significantly better than the SVMSOTA system (t-test, p < 0.05) on FigSumGS1 dataset. In terms of the ROUGE score performance, SVMSOTA+ achieved the best scores using supervised approaches and the difference in performance against the best performing SurfaceCue was not statistically significant. The systems performed similarly on FigSumGS2 dataset but the improvement of state-of-the-art supervised systems over Paragraph and Hybrid systems were not statistically significant.
The unsupervised FigSum+ systems performed better than the state-of-the-art supervised systems  (NBSOTA and SVMSOTA). Although this is an interesting result, previous studies have also demonstrated that unsupervised methods often have comparable, if not better, performance than supervised techniques [72–74]. In our case, this could be attributed to a number of reasons.
First, our systems were limited to the biomedical domain. Hence, these features could be better tuned to outperform in our domain. Second, although we used the same set of features as described in , the implementation of the similarity feature between our systems and  was different. We used the cosine similarity instead of the Okapi BM25 similarity, which we will explore in our future work. Third, the evaluation data used in  were different from the data used in our experiments.
We explored both supervised and unsupervised methods for figure summarization and concluded that the unsupervised techniques performed better. This is not surprising. Our annotated data size is small and therefore prone to the data sparseness challenge. One way to improve the performance is to increase the robustness of word representation. Word embedding  clusters similar words and therefore reduces the dimensionality of word features and may improve the performance of supervised learning.
As shown in Table 2 and 3, the feature of word similarity between the sentences to be included in a summary and the caption of a figure does not always improve the performance. This is not surprising. We found that frequently figure captions contain detailed methodological descriptions of experiments while summary sentences tend to interpret the results. This may explain why our first figure summarization FigSum does not perform as well as our FigSum+ systems and why we need to explore additional features for optimal performance.
The FigSum+ approaches SurfaceCue, Paragraph, and Hybrid had average F1 scores of 0.79 and 0.26, 0.84 and 0.27, and 0.82 and 0.21, respectively, for the FigSumGS1 dataset and of 0.62 and 0.10, 0.62 and 0.24, and 0.64 and 0.21, respectively, for the FigSumGS2 dataset. Human-generated summaries often show such variations as well [76, 77]. The performance differences of the various FigSum+ techniques can be attributed to variations in the quality of the gold standard generated by the annotators.
Further analysis of the FigSum+ performance on FigSumGS1 dataset using Spearman Rank Correlation showed that there was no correlation between the F1 score and the length of the article or the number of figures. However, the F1 score of SurfaceCue showed moderate negative correlation (rho = -0.51, p < 0.05) with the average number of sentences per figure. For FigSumGS2 dataset, the length of the article had a moderate negative correlation with the performance of Paragraph (rho = -0.52, p < 0.05) and Hybrid (rho = -0.50, p < 0.05) implementations and the average number of sentences per figure and had a negative correlation with the performance of Paragraph (rho = -0.71, p < 0.05) and the Hybrid (rho = -0.74, p < 0.05) implementations. This finding suggests that longer summaries tend to have lower quality.
The SurfaceCue system had a near perfect ROUGE score for FigSumGS1 dataset, since the annotators picked figure-referring sentences as part of the gold standard. Although the SurfaceCue approach had a very high ROUGE score, it also had a very low recall (0.41 for FigSumGS1 and 0.16 for FigSumGS2 datasets) compared to the Paragraph and Hybrid approaches. There was no correlation between the ROUGE score performance and the length of the article, the number of figures, or the average number of sentences per figure for the FigSumGS1 dataset. Similarly, there was no correlation between the number of figures or the average number of sentences per figure except length of the article, which had a negative correlation with SurfaceCue (rho = -0.72, p < 0.05) for FigSumGS2 dataset.
The FigSum+ approaches performed well against two different gold standards constructed using different criteria, demonstrating the robustness of the approaches and their efficacy in rendering comprehensive figure summaries. It was also interesting that one article in FigSumGS1 dataset had an F1 score of 0.44 for the Hybrid approach but achieved an R1 score of 0.85, indicating that the quality of the summaries extracted by the FigSum+ implementations were as good as human-generated summaries.
One of the inherent problems of extractive summaries is that they lack coherence and certain sentences do not make sense when taken out of context (e.g., as in the SurfaceCue implementation). For example, Fig. 4 shows a figure along with its caption and the sentence extracted by the SurfaceCue method. The sentence “The summary risk difference was 0.27% (−0.10% to 0.63%, P = 0.15, I2 = 0%; fig. 2) with no indication of publication bias in the funnel plot”, provides very little context for the figure. To overcome this problem, we extracted whole paragraphs where figure-referring sentences appeared, as in the Paragraph approach. Fig. 5 shows the summary extracted by the Paragraph method for the figure shown in Fig. 4. The summary provides more information and context to help understand the figure better. We believe this method provides users with the sentence context and improves the overall comprehension of the figure while reducing user information overload.
Limitations and Future Work
There are, however, certain limitations to the study. The current results are based on only 94 figures from 19 biomedical articles. Although this number of figures is small, it is on a par with other studies that also require extensive manual annotation . The results indicate that the FigSum+ approaches—especially Paragraph and Hybrid—can generate summaries that are closely related to the information deemed important by experts to explain the content of figures. Our annotation may have limitations as well. Although biomedical experts annotated the gold standard summaries, we found all of whom selected sentences from paragraphs referenced a figure and bias may be introduced. A future work for creating a summary gold standard is to extract all relevant sentences and then to ask experts to pick out the most informative and representative ones. As stated earlier, in the future, we will explore class-based language modeling approaches (word embedding) to overcome the data sparseness challenge. We also evaluate the system’s utility by comparing it with other systems in an extrinsic evaluation.
This study explored a number of supervised and unsupervised approaches to summarize figures in biomedical articles by aggregating sentences associated with a figure and removing redundant sentences. Our evaluation results show that a simple unsupervised FigSum+ system that is based on surface cues achieved the best F1 score of 0.66 and ROUGE-1 score of 0.97. FigSum+ can be readily implemented with minimum computation cost and thereby maximizing its speed. These results demonstrate that the FigSum+ approaches present a promising approach for figure summarization by reducing information overload while improving users’ information-seeking behavior and maintaining information content.
Conceived and designed the experiments: BPR RS HY. Performed the experiments: BPR. Analyzed the data: BPR. Wrote the paper: BPR RS HY.
- 1. Futrelle RP (2004) Handling figures in document summarization. Proceedings of the ACL-04 Workshop: Text Summarization Branches Out: 61–65.
- 2. Rowe NC (1993) Efficient caption-based retrieval of multimedia information. Monterey, California, Naval Postgraduate School.
- 3. Guglielmo EJ, Rowe NC (1996) Natural-language retrieval of images based on descriptive captions. ACM Transactions on Information Systems (TOIS), 14(3): 237.
- 4. Smith JR, Chang SF (1997) VisualSEEk: a fully automated content-based image query system. Proceedings of the fourth ACM international conference on Multimedia: 87–98.
- 5. Rowe NC (1999) Precise and efficient retrieval of captioned images: The MARIE project. Library Trends, 48(2): 475–495.
- 6. Rui Y, Huang TS, Ortega M, Mehrotra S (1998) Relevance feedback: a power tool for interactive content-based image retrieval. Circuits and Systems for Video Technology, IEEE Transactions on, 8(5): 644–655.
- 7. Jeon J, Lavrenko V, Manmatha R (2003) Automatic image annotation and retrieval using cross-media relevance models. Proceedings of the 26th annual international ACM SIGIR conference on Research and development in informaion retrieval: 119–126.
- 8. Swets DL, Weng JJ (1996) Using discriminant eigen features for image retrieval. Pattern Analysis and Machine Intelligence, IEEE Transactions, 18(8): 831–836.
- 9. Kekre HB, Thepade SD (2010) Image Retrieval using Color-Texture Features Extracted from Walshlet Pyramid. ICGST International Journal on Graphics, Vision and Image Processing (GVIP), 10: 9–18.
- 10. Müller H, Michoux N, Bandon D, Geissbuhler A (2004) A review of content-based image retrieval systems in medical applications—clinical benefits and future directions. International journal of medical informatics, 73(1): 1–23. pmid:15036075
- 11. Tufte ER (1991) Envisioning information. Optometry & Vision Science, 68(4): 322–324.
- 12. Hammond T, Eoff B, Paulson B, Wolin A, Dahmen, K et al. (2008) Free-sketch recognition: putting the chi in sketching, in CHI’08 extended abstracts on Human factors in computing systems: 3027–3032.
- 13. Paulson B, Hammond T (2008) PaleoSketch: accurate primitive sketch recognition and beautification. Proceedings of the 13th international conference on Intelligent user interfaces: 1–10.
- 14. Agarwal S, Yu H (2009) FigSum: automatically generating structured text summaries for figures in biomedical literature. AMIA Annual Symposium Proceedings, 2009: 6–10. pmid:20351812
- 15. Demner-Fushman D, Antani S, Simpson M, Thoma GR (2009) Annotation and retrieval of clinically relevant images. Int J Med Inform, 78(12): 59–67.
- 16. Yu H, Agarwal S, Johnston M, Cohen A (2009) Are figure legends sufficient? Evaluating the contribution of associated text to biomedical figure comprehension. Journal of biomedical discovery and collaboration, 4(1): 1. pmid:19126221
- 17. Yu H (2006) Towards answering biological questions with experimental evidence: automatically identifying text that summarize image content in full-text articles. AMIA Annual Symposium Proceedings: 834.
- 18. Lipinski R, Bijlsma M, Gipp J, Podhaizer D, Bushman W (2008) Establishment and characterization of immortalized Gli-null mouse embryonic fibroblast cell lines. BMC Cell Biology, 13(9).
- 19. Agarwal S, Yu H (2011) Figure summarizer browser extensions for PubMed Central. Bioinformatics, 27(12): 1723–1724. pmid:21493658
- 20. Mani I (2001) Automatic summarization. John Benjamin’s Publishing Company, vol 3.
- 21. Nenkova A, Maskey S, Liu Y (2011) Automatic Summarization. Tutorial Abstracts of ACL 2011: 3.
- 22. Luhn HP (1958) The automatic creation of literature abstracts. IBM Journal of research and development 2(2): 159–165.
- 23. Brunn M, Chali Y, Pinchak C (2001) Text Summarization Using Lexical Chains. Document Understanding Conference 2001: 135–140.
- 24. Hovy E, Lin CY (1999) Automated text summarization in SUMMARIST. Advances in Automatic Text Summarization: 81–94.
- 25. Kupiec J, Pedersen J, Chen F (1995) A trainable document summarizer. Proceedings of the 18th annual international ACM SIGIR conference on Research and development in information retrieval: 68–73.
- 26. Conroy JM, Schlesinger JD, O’Leary DP (2006) Topic-focused multi-document summarization using an approximate oracle score. Proceedings of the COLING/ACL on Main conference poster sessions, Stroudsburg, PA, USA: 152–159.
- 27. Gupta S, Nenkova A, Jurafsky D (2007) Measuring importance and query relevance in topic-focused multi-document summarization. Proceedings of the 45th Annual Meeting of the ACL on Interactive Poster and Demonstration Sessions, Stroudsburg, PA, USA: 193–196.
- 28. Nakov PI, Schwartz AS, Hearst MA (2004) Citances: Citation Sentences for Semantic Analysis of Bioscience Text. Proceedings of the SIGIR’04 workshop on Search and Discovery in Bioinformatics.
- 29. Edmundson HP (1969) New methods in automatic extracting. Journal of the ACM, 16(2): 264–285.
- 30. Myaeng SH, Jang DH (1999) Development and evaluation of a statistically-based document summarization system. Advances in automatic text summarization: 61–70.
- 31. Aone C, Okurowski M, Gorlinsky J, Larsen B (1999) A trainable summarizer with knowledge acquired from robust nlp techniques. Advances in Automatic Text Summarization, MIT Press: 71–80.
- 32. Filatova E, Hatzivassiloglou V (2004) A formal model for information selection in multi-sentence text extraction. Proceedings of the 20th international conference on Computational Linguistics, Stroudsburg, PA, USA.
- 33. Kikuchi T, Furui S, Hori C (2003) Automatic speech summarization based on sentence extraction and compaction. IEEE International Conference on Acoustics, Speech, and Signal Processing, 1: 384–387.
- 34. Gong Y (2001) Generic text summarization using relevance measure and latent semantic analysis. Proceedings of the 24th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval.
- 35. Yeh JY, Ke HR, Yang WP, Meng I (2005) Text summarization using a trainable summarizer and latent semantic analysis. Information processing & management, 41(1): 75–95.
- 36. Miller GA (1995) WordNet: a lexical database for English. Communications of the ACM, 38(11): 39–41.
- 37. Mihalcea R (2004) Graph-based ranking algorithms for sentence extraction, applied to text summarization. Proceedings of the ACL 2004 on Interactive poster and demonstration sessions, Stroudsburg, PA, USA.
- 38. Erkan G, Radev DR (2004) LexRank: graph-based lexical centrality as salience in text summarization. J. Artif. Int. Res., 22(1): 457–479.
- 39. Radev DR, Jing H, Sty M, Tam D (2004) Centroid-based summarization of multiple documents. Information Processing & Management, 40(6): 919–938.
- 40. Wang C, Jing F, Zhang L, Zhang HJ (2007) Learning query-biased web page summarization. Proceedings of the sixteenth ACM conference on Conference on information and knowledge management, New York, NY, USA: 555–562.
- 41. Hirao T, Isozaki H, Maeda E, Matsumoto Y (2002) Extracting important sentences with support vector machines. Proceedings of the 19th international conference on Computational linguistics, Stroudsburg, PA, USA: 1–7.
- 42. Conroy JM, O’leary DP (2001) Text summarization via hidden markov models. Proceedings of the 24th annual international ACM SIGIR conference on Research and development in information retrieval: 406–407.
- 43. Leskovec J, Grobelnik M, Milic-Frayling N (2004) Learning sub-structures of document semantic graphs for document summarization. LinkKDD Workshop: 133–138.
- 44. Mani I (2001) Summarization Evaluation: An Overview. Proceedings of the NTCIR Workshop, vol. 2.
- 45. Donaway RL, Drummey KW, Mather LA (2000) A comparison of rankings produced by summarization evaluation measures. Proceedings of the 2000 NAACL-ANLPWorkshop on Automatic summarization, 4: 69–78.
- 46. Saggion H, Teufel S, Radev D, Lam W (2002) Meta-evaluation of summaries in a cross-lingual environment using content-based metrics. Proceedings of the 19th international conference on Computational linguistics, 1: 1–7.
- 47. Radev DR, Teufel S, Saggion H, Lam W, Blitzer J et al. (2003) Evaluation challenges in large-scale document summarization. Proceedings of the 41st Annual Meeting on Association for Computational Linguistics, Stroudsburg, PA, USA: 375–382.
- 48. Papineni K, Roukos S, Ward T, Zhu WJ (2002) BLEU: a method for automatic evaluation of machine translation. Proceedings of the 40th annual meeting on association for computational linguistics: 311–318.
- 49. Pastra K, Saggion H (2003) Colouring summaries BLEU. Proceedings of the EACL 2003 Workshop on Evaluation Initiatives in Natural Language Processing: are evaluation methods, metrics and resources reusable?: 35–42.
- 50. Lin C (2004) ROUGE: A package for automatic evaluation of summaries. Proceedings of the ACL Workshop: Text Summarization Braches Out 2004: 74–81.
- 51. Nenkova A, Passonneau R, McKeown K (2007) The pyramid method: Incorporating human content selection variation in summarization evaluation. ACM Transactions on Speech and Language Processing (TSLP), 4(2): 4.
- 52. Radev DR, Tam D (2003) Summarization evaluation using relative utility. Proceedings of the twelfth international conference on Information and knowledge management: 508–511.
- 53. Chiang JH, Shin JW, Liu HH, Chin CL (2006) GeneLibrarian: an effective gene-information summarization and visualization system. BMC bioinformatics, 7(1): 392. pmid:16939640
- 54. Ling X, Jiang J, He X, Mei Q, Zhai C et al. (2007) Generating gene summaries from biomedical literature: A study of semi-structured summarization. Information Processing & Management, 43(6): 1777–1791.
- 55. Jin F, Huang M, Lu Z, Zhu X (2009) Towards automatic generation of gene summary. Proceedings of the Workshop on BioNLP 2009: 97–105.
- 56. Bhattacharya S, Ha V, Srinivasan P (2011) MeSH: a window into full text for document summarization. Bioinformatics 27: i120–i128. pmid:21685060
- 57. Plaza L, Carrillo-de-Albornoz J (2013) Evaluating the use of different positional strategies for sentence selection in biomedical literature summarization. BMC Bioinformatics, 14(1): 71. pmid:23445074
- 58. Reeve L, Han H, Brooks AD (2006) BioChain: lexical chaining methods for biomedical text summarization. Proceedings of the 2006 ACM symposium on Applied computing: 180–184.
- 59. Humphrey B, Lindberg DAB, Schoolman HM, Barnett GO (1998) The Unified Medical Language System: An Informatics Research Collaboration. Journal of the American Medical Association, 5: 1–11.
- 60. Fiszman M, Rindflesch TC, Kilicoglu H (2004) Abstraction summarization for managing the biomedical research literature. Proceedings of the HLT-NAACL Workshop on Computational Lexical Semantics, Stroudsburg, PA, USA: 76–83.
- 61. Workman TE, Fiszman M, Hurdle JF, Rindflesch TC (2010) Biomedical text summarization to support genetic database curation: using Semantic MEDLINE to create a secondary database of genetic information. J Med Libr Assoc, 98(4): 273–281. pmid:20936065
- 62. Workman TE, Hurdle JF (2011) Dynamic summarization of bibliographic-based data. BMC Med Inform Decision Making, 11: 6.
- 63. Shang Y, Li Y, Lin H, Yang Z (2011) Enhancing Biomedical Text Summarization Using Semantic Relation Extraction. PLoS ONE, 6(8): e23862. pmid:21887336
- 64. Plaza L, Díaz A, Gervás P (2011) A semantic graph-based approach to biomedical summarisation. Artif Intell Med, 53(1): 1–14. pmid:21752612
- 65. Yoo I, Hu X, Song IY (2007) A coherent graph-based semantic clustering and summarization approach for biomedical literature and a new summarization evaluation method. BMC Bioinformatics, 8(9): S4. pmid:18047705
- 66. Futrelle RP (1999) Summarization of diagrams in documents. Advances in Automated Text Summarization: 403–421.
- 67. Bhatia S, Mitra P (2012) Summarizing figures, tables, and algorithms in scientific publications to augment search results. ACM Transactions on Information Systems (TOIS), 30(1): 3.
- 68. Wu P, Carberry S (2011) Toward extractive summarization of multimodal documents. Proceedings of the Workshop on Text Summarization at the Canadian Conference on Artificial Intelligence: 53–61.
- 69. Agarwal S, Yu H (2009) Automatically classifying sentences in full-text biomedical articles into Introduction, Methods, Results and Discussion. Bioinformatics, 25(23): 3174. pmid:19783830
- 70. DUC Task (2007) Documents for Summarization, and Measures. Available: http://duc.nist.gov/duc2007/tasks.html. Accessed: 19 December 2014.
- 71. Prasad R, McRoy S, Frid N, Joshi A, Yu H (2011) The biomedical discourse relation bank. BMC Bioinformatics, 12(1): 188. pmid:21605399
- 72. Yarowsky D (1995) Unsupervised word sense disambiguation rivaling supervised methods. Proceedings of 33rd annual meeting of ACL: 189–196.
- 73. Alegre E, Gonzalez-Castro V, Suarez S, Castejon M (2009) Comparison of supervised and unsupervised methods to classify boar acrosomes using texture descriptors. Proceedings of 51st International Symposium ELMAR: 65–70.
- 74. Japkowicz N (2001) Supervised Versus Unsupervised Binary-Learning by Feed forward Neural Networks. Machine Learning, 42(1–2): 97–122.
- 75. Collobert R, Weston J, Bottou L, Karlen M, Kavukcuoglu K et al. (2011) Natural Language Processing (Almost) from Scratch. Journal of Machine Learning Research, 12: 2493–2537.
- 76. Salton G, Singhal A, Mitra M, Buckley C (1997) Automatic text structuring and summarization. Information Processing & Management, 33(2): 193–207.
- 77. Nomoto T, Matsumoto Y (1997) Data Reliability and Its Effects on Automatic Abstracting. In Proceedings of the Fifth Workshop on Very Large Corpora, Beijing/Hong Kong.
- 78. Prochaska JJ, Hilton JF (2012) Risk of cardiovascular serious adverse events associated with varenicline use for tobacco cessation: systematic review and meta-analysis. BMJ, 344: e2856. pmid:22563098