PICO entity extraction for preclinical animal literature.

BACKGROUND: Natural language processing could assist multiple tasks in systematic reviews to reduce workflow, including the extraction of PICO elements such as study populations, interventions, comparators and outcomes. The PICO framework provides a basis for the retrieval and selection for inclusion of evidence relevant to a specific systematic review question, and automatic approaches to PICO extraction have been developed particularly for reviews of clinical trial findings. Considering the difference between preclinical animal studies and clinical trials, developing separate approaches is necessary. Facilitating preclinical systematic reviews will inform the translation from preclinical to clinical research. METHODS: We randomly selected 400 abstracts from the PubMed Central Open Access database which described in vivo animal research and manually annotated these with PICO phrases for Species, Strain, methods of Induction of disease model, Intervention, Comparator and Outcome. We developed a two-stage workflow for preclinical PICO extraction. Firstly we fine-tuned BERT with different pre-trained modules for PICO sentence classification. Then, after removing the text irrelevant to PICO features, we explored LSTM-, CRF- and BERT-based models for PICO entity recognition. We also explored a self-training approach because of the small training corpus. RESULTS: For PICO sentence classification, BERT models using all pre-trained modules achieved an F1 score of over 80%, and models pre-trained on PubMed abstracts achieved the highest F1 of 85%. For PICO entity recognition, fine-tuning BERT pre-trained on PubMed abstracts achieved an overall F1 of 71% and satisfactory F1 for Species (98%), Strain (70%), Intervention (70%) and Outcome (67%). The score of Induction and Comparator is less satisfactory, but F1 of Comparator can be improved to 50% by applying self-training. CONCLUSIONS: Our study indicates that of the approaches tested, BERT pre-trained on PubMed abstracts is the best for both PICO sentence classification and PICO entity recognition in the preclinical abstracts. Self-training yields better performance for identifying comparators and strains.

Risk of bias assessment in preclinical literature using natural language processing.

We sought to apply natural language processing to the task of automatic risk of bias assessment in preclinical literature, which could speed the process of systematic review, provide information to guide research improvement activity, and support translation from preclinical to clinical research. We use 7840 full-text publications describing animal experiments with yes/no annotations for five risk of bias items. We implement a series of models including baselines (support vector machine, logistic regression, random forest), neural models (convolutional neural network, recurrent neural network with attention, hierarchical neural network) and models using BERT with two strategies (document chunk pooling and sentence extraction). We tune hyperparameters to obtain the highest F1 scores for each risk of bias item on the validation set and compare evaluation results on the test set to our previous regular expression approach. The F1 scores of best models on test set are 82.0% for random allocation, 81.6% for blinded assessment of outcome, 82.6% for conflict of interests, 91.4% for compliance with animal welfare regulations and 46.6% for reporting animals excluded from analysis. Our models significantly outperform regular expressions for four risk of bias items. For random allocation, blinded assessment of outcome, conflict of interests and animal exclusions, neural models achieve good performance; for animal welfare regulations, BERT model with a sentence extraction strategy works better. Convolutional neural networks are the overall best models. The tool is publicly available which may contribute to the future monitoring of risk of bias reporting for research improvement activities.

Disambiguating Visual Verbs.

In this article, we introduce a new task, visual sense disambiguation for verbs: given an image and a verb, assign the correct sense of the verb, i.e., the one that describes the action depicted in the image. Just as textual word sense disambiguation is useful for a wide range of NLP tasks, visual sense disambiguation can be useful for multimodal tasks such as image retrieval, image description, and text illustration. We introduce a new dataset, which we call VerSe (short for Verb Sense) that augments existing multimodal datasets (COCO and TUHOI) with verb and sense labels. We explore supervised and unsupervised models for the sense disambiguation task using textual, visual, and multimodal embeddings. We also consider a scenario in which we must detect the verb depicted in an image prior to predicting its sense (i.e., there is no verbal information associated with the image). We find that textual embeddings perform well when gold-standard annotations (object labels and image descriptions) are available, while multimodal embeddings perform well on unannotated images. VerSe is publicly available at https://github.com/spandanagella/verse.

Visually Grounded Meaning Representations.

In this paper we address the problem of grounding distributional representations of lexical meaning. We introduce a new model which uses stacked autoencoders to learn higher-level representations from textual and visual input. The visual modality is encoded via vectors of attributes obtained automatically from images. We create a new large-scale taxonomy of 600 visual attributes representing more than 500 concepts and 700 K images. We use this dataset to train attribute classifiers and integrate their predictions with text-based distributional models of word meaning. We evaluate our model on its ability to simulate word similarity judgments and concept categorization. On both tasks, our model yields a better fit to behavioral data compared to baselines and related models which either rely on a single modality or do not make use of attribute-based input.

Incremental Bayesian Category Learning From Natural Language.

Models of category learning have been extensively studied in cognitive science and primarily tested on perceptual abstractions or artificial stimuli. In this paper, we focus on categories acquired from natural language stimuli, that is, words (e.g., chair is a member of the furniture category). We present a Bayesian model that, unlike previous work, learns both categories and their features in a single process. We model category induction as two interrelated subproblems: (a) the acquisition of features that discriminate among categories, and (b) the grouping of concepts into categories based on those features. Our model learns categories incrementally using particle filters, a sequential Monte Carlo method commonly used for approximate probabilistic inference that sequentially integrates newly observed data and can be viewed as a plausible mechanism for human learning. Experimental results show that our incremental learner obtains meaningful categories which yield a closer fit to behavioral data compared to related models while at the same time acquiring features which characterize the learned categories. (An earlier version of this work was published in Frermann and Lapata .).

Automatic caption generation for news images.

This paper is concerned with the task of automatically generating captions for images, which is important for many image-related applications. Examples include video and image retrieval as well as the development of tools that aid visually impaired individuals to access pictorial information. Our approach leverages the vast resource of pictures available on the web and the fact that many of them are captioned and colocated with thematically related documents. Our model learns to create captions from a database of news articles, the pictures embedded in them, and their captions, and consists of two stages. Content selection identifies what the image and accompanying article are about, whereas surface realization determines how to verbalize the chosen content. We approximate content selection with a probabilistic image annotation model that suggests keywords for an image. The model postulates that images and their textual descriptions are generated by a shared set of latent variables (topics) and is trained on a weakly labeled dataset (which treats the captions and associated news articles as image labels). Inspired by recent work in summarization, we propose extractive and abstractive surface realization models. Experimental results show that it is viable to generate captions that are pertinent to the specific content of an image and its associated article, while permitting creativity in the description. Indeed, the output of our abstractive model compares favorably to handwritten captions and is often superior to extractive methods.

An experimental study of graph connectivity for unsupervised word sense disambiguation.

Word sense disambiguation (WSD), the task of identifying the intended meanings (senses) of words in context, has been a long-standing research objective for natural language processing. In this paper, we are concerned with graph-based algorithms for large-scale WSD. Under this framework, finding the right sense for a given word amounts to identifying the most "important" node among the set of graph nodes representing its senses. We introduce a graph-based WSD algorithm which has few parameters and does not require sense-annotated data for training. Using this algorithm, we investigate several measures of graph connectivity with the aim of identifying those best suited for WSD. We also examine how the chosen lexicon and its connectivity influences WSD performance. We report results on standard data sets and show that our graph-based approach performs comparably to the state of the art.

Composition in distributional models of semantics.

Vector-based models of word meaning have become increasingly popular in cognitive science. The appeal of these models lies in their ability to represent meaning simply by using distributional information under the assumption that words occurring within similar contexts are semantically similar. Despite their widespread use, vector-based models are typically directed at representing words in isolation, and methods for constructing representations for phrases or sentences have received little attention in the literature. This is in marked contrast to experimental evidence (e.g., in sentential priming) suggesting that semantic similarity is more complex than simply a relation between isolated words. This article proposes a framework for representing the meaning of word combinations in vector space. Central to our approach is vector composition, which we operationalize in terms of additive and multiplicative functions. Under this framework, we introduce a wide range of composition models that we evaluate empirically on a phrase similarity task.

Evidence for serial coercion: a time course analysis using the visual-world paradigm.

Metonymic verbs like start or enjoy often occur with artifact-denoting complements (e.g., The artist started the picture) although semantically they require event-denoting complements (e.g., The artist started painting the picture). In case of artifact-denoting objects, the complement is assumed to be type shifted (or coerced) into an event to conform to the verb's semantic restrictions. Psycholinguistic research has provided evidence for this kind of enriched composition: readers experience processing difficulty when faced with metonymic constructions compared to non-metonymic controls. However, slower reading times for metonymic constructions could also be due to competition between multiple interpretations that are being entertained in parallel whenever a metonymic verb is encountered. Using the visual-world paradigm, we devised an experiment which enabled us to determine the time course of metonymic interpretation in relation to non-metonymic controls. The experiment provided evidence in favor of a non-competitive, serial coercion process.