Elaboration and testing of new methodologies for automatic abstracting

The primary objective of this research was to understand what kinds of knowledge and skills people use in `extracting' relevant information from text and to assess the extent to which expert systems techniques could be applied to automate the process of abstracting. The approach adopted in this thesis is based on research in cognitive science, information science, psycholinguistics and textlinguistics. The study addressed the significance of domain knowledge and heuristic rules by developing an information extraction system, called INFORMEX. This system, which was implemented partly in SPITBOL, and partly in PROLOG, used a set of heuristic rules to analyse five scientific papers of expository type, to interpret the content in relation to the key abstract elements and to extract a set of sentences recognised as relevant for abstracting purposes. The analysis of these extracts revealed that an adequate abstract could be generated. Furthermore, INFORMEX showed that a rule based system was a suitable computational model to represent experts' knowledge and strategies. This computational technique provided the basis for a new approach to the modelling of cognition. It showed how experts tackle the task of abstracting by integrating formal knowledge as well as experiential learning. This thesis demonstrated that empirical and theoretical knowledge can be effectively combined in expert systems technology to provide a valuable starting approach to automatic abstracting.

The functional organisation of the fronto-temporal language system:evidence from syntactic and semantic ambiguity

Spoken language comprehension is known to involve a large left-dominant network of fronto-temporal brain regions, but there is still little consensus about how the syntactic and semantic aspects of language are processed within this network. In an fMRI study, volunteers heard spoken sentences that contained either syntactic or semantic ambiguities as well as carefully matched low-ambiguity sentences. Results showed ambiguity-related responses in the posterior left inferior frontal gyrus (pLIFG) and posterior left middle temporal regions. The pLIFG activations were present for both syntactic and semantic ambiguities suggesting that this region is not specialised for processing either semantic or syntactic information, but instead performs cognitive operations that are required to resolve different types of ambiguity irrespective of their linguistic nature, for example by selecting between possible interpretations or reinterpreting misparsed sentences. Syntactic ambiguities also produced activation in the posterior middle temporal gyrus. These data confirm the functional relationship between these two brain regions and their importance in constructing grammatical representations of spoken language.

Biomedical events extraction using the hidden vector state model

Objective: Biomedical events extraction concerns about events describing changes on the state of bio-molecules from literature. Comparing to the protein-protein interactions (PPIs) extraction task which often only involves the extraction of binary relations between two proteins, biomedical events extraction is much harder since it needs to deal with complex events consisting of embedded or hierarchical relations among proteins, events, and their textual triggers. In this paper, we propose an information extraction system based on the hidden vector state (HVS) model, called HVS-BioEvent, for biomedical events extraction, and investigate its capability in extracting complex events. Methods and material: HVS has been previously employed for extracting PPIs. In HVS-BioEvent, we propose an automated way to generate abstract annotations for HVS training and further propose novel machine learning approaches for event trigger words identification, and for biomedical events extraction from the HVS parse results. Results: Our proposed system achieves an F-score of 49.57% on the corpus used in the BioNLP'09 shared task, which is only 2.38% lower than the best performing system by UTurku in the BioNLP'09 shared task. Nevertheless, HVS-BioEvent outperforms UTurku's system on complex events extraction with 36.57% vs. 30.52% being achieved for extracting regulation events, and 40.61% vs. 38.99% for negative regulation events. Conclusions: The results suggest that the HVS model with the hierarchical hidden state structure is indeed more suitable for complex event extraction since it could naturally model embedded structural context in sentences.

A hybrid generative/discriminative framework to train a semantic parser from an un-annotated corpus

We propose a hybrid generative/discriminative framework for semantic parsing which combines the hidden vector state (HVS) model and the hidden Markov support vector machines (HM-SVMs). The HVS model is an extension of the basic discrete Markov model in which context is encoded as a stack-oriented state vector. The HM-SVMs combine the advantages of the hidden Markov models and the support vector machines. By employing a modified K-means clustering method, a small set of most representative sentences can be automatically selected from an un-annotated corpus. These sentences together with their abstract annotations are used to train an HVS model which could be subsequently applied on the whole corpus to generate semantic parsing results. The most confident semantic parsing results are selected to generate a fully-annotated corpus which is used to train the HM-SVMs. The proposed framework has been tested on the DARPA Communicator Data. Experimental results show that an improvement over the baseline HVS parser has been observed using the hybrid framework. When compared with the HM-SVMs trained from the fully-annotated corpus, the hybrid framework gave a comparable performance with only a small set of lightly annotated sentences. © 2008. Licensed under the Creative Commons.

A novel framework of training hidden Markov support vector machines from lightly-annotated data

Natural language understanding (NLU) aims to map sentences to their semantic mean representations. Statistical approaches to NLU normally require fully-annotated training data where each sentence is paired with its word-level semantic annotations. In this paper, we propose a novel learning framework which trains the Hidden Markov Support Vector Machines (HM-SVMs) without the use of expensive fully-annotated data. In particular, our learning approach takes as input a training set of sentences labeled with abstract semantic annotations encoding underlying embedded structural relations and automatically induces derivation rules that map sentences to their semantic meaning representations. The proposed approach has been tested on the DARPA Communicator Data and achieved 93.18% in F-measure, which outperforms the previously proposed approaches of training the hidden vector state model or conditional random fields from unaligned data, with a relative error reduction rate of 43.3% and 10.6% being achieved.

Formant-frequency variation and its effects on across-formant grouping in speech perception

How speech is separated perceptually from other speech remains poorly understood. In a series of experiments, perceptual organisation was probed by presenting three-formant (F1+F2+F3) analogues of target sentences dichotically, together with a competitor for F2 (F2C), or for F2+F3, which listeners must reject to optimise recognition. To control for energetic masking, the competitor was always presented in the opposite ear to the corresponding target formant(s). Sine-wave speech was used initially, and different versions of F2C were derived from F2 using separate manipulations of its amplitude and frequency contours. F2Cs with time-varying frequency contours were highly effective competitors, whatever their amplitude characteristics, whereas constant-frequency F2Cs were ineffective. Subsequent studies used synthetic-formant speech to explore the effects of manipulating the rate and depth of formant-frequency change in the competitor. Competitor efficacy was not tuned to the rate of formant-frequency variation in the target sentences; rather, the reduction in intelligibility increased with competitor rate relative to the rate for the target sentences. Therefore, differences in speech rate may not be a useful cue for separating the speech of concurrent talkers. Effects of competitors whose depth of formant-frequency variation was scaled by a range of factors were explored using competitors derived either by inverting the frequency contour of F2 about its geometric mean (plausibly speech-like pattern) or by using a regular and arbitrary frequency contour (triangle wave, not plausibly speech-like) matched to the average rate and depth of variation for the inverted F2C. Competitor efficacy depended on the overall depth of frequency variation, not depth relative to that for the other formants. Furthermore, the triangle-wave competitors were as effective as their more speech-like counterparts. Overall, the results suggest that formant-frequency variation is critical for the across-frequency grouping of formants but that this grouping does not depend on speech-specific constraints.

Reading without central vision:effects of text spacing on reading in patients with macular disease

Abstract: Loss of central vision caused by age-related macular degeneration (AMD) is a problem affecting increasingly large numbers of people within the ageing population. AMD is the leading cause of blindness in the developed world, with estimates of over 600,000 people affected in the UK . Central vision loss can be devastating for the sufferer, with vision loss impacting on the ability to carry out daily activities. In particular, inability to read is linked to higher rates of depression in AMD sufferers compared to age-matched controls. Methods to improve reading ability in the presence of central vision loss will help maintain independence and quality of life for those affected. Various attempts to improve reading with central vision loss have been made. Most textual manipulations, including font size, have led to only modest gains in reading speed. Previous experimental work and theoretical arguments on spatial integrative properties of the peripheral retina suggest that ‘visual crowding’ may be a major factor contributing to inefficient reading. Crowding refers to the phenomena in which juxtaposed targets viewed eccentrically may be difficult to identify. Manipulating text spacing of reading material may be a simple method that reduces crowding and benefits reading ability in macular disease patients. In this thesis the effect of textual manipulation on reading speed was investigated, firstly for normally sighted observers using eccentric viewing, and secondly for observers with central vision loss. Test stimuli mimicked normal reading conditions by using whole sentences that required normal saccadic eye movements and observer comprehension. Preliminary measures on normally-sighted observers (n = 2) used forced-choice procedures in conjunction with the method of constant stimuli. Psychometric functions relating the proportion of correct responses to exposure time were determined for text size, font type (Lucida Sans and Times New Roman) and text spacing, with threshold exposure time (75% correct responses) used as a measure of reading performance. The results of these initial measures were used to derive an appropriate search space, in terms of text spacing, for assessing reading performance in AMD patients. The main clinical measures were completed on a group of macular disease sufferers (n=24). Firstly, high and low contrast reading acuity and critical print size were measured using modified MNREAD test charts, and secondly, the effect of word and line spacing was investigated using a new test, designed specifically for this study, called the Equal Readability Passages (ERP) test. The results from normally-sighted observers were in close agreement with those from the group of macular disease sufferers. Results show that: (i) optimum reading performance was achieved when using both double line and double word spacing; (ii) the effect of line spacing was greater than the effect of word spacing (iii) a text size of approximately 0.85o is sufficiently large for reading at 5o eccentricity. In conclusion, the results suggest that crowding is detrimental to reading with peripheral vision, and its effects can be minimized with a modest increase in text spacing.

Formant-Frequency Variation and Informational Masking of Speech by Extraneous Formants : Evidence Against Dynamic and Speech-Specific Acoustical Constraints

How speech is separated perceptually from other speech remains poorly understood. Recent research indicates that the ability of an extraneous formant to impair intelligibility depends on the variation of its frequency contour. This study explored the effects of manipulating the depth and pattern of that variation. Three formants (F1+F2+F3) constituting synthetic analogues of natural sentences were distributed across the 2 ears, together with a competitor for F2 (F2C) that listeners must reject to optimize recognition (left = F1+F2C; right = F2+F3). The frequency contours of F1 − F3 were each scaled to 50% of their natural depth, with little effect on intelligibility. Competitors were created either by inverting the frequency contour of F2 about its geometric mean (a plausibly speech-like pattern) or using a regular and arbitrary frequency contour (triangle wave, not plausibly speech-like) matched to the average rate and depth of variation for the inverted F2C. Adding a competitor typically reduced intelligibility; this reduction depended on the depth of F2C variation, being greatest for 100%-depth, intermediate for 50%-depth, and least for 0%-depth (constant) F2Cs. This suggests that competitor impact depends on overall depth of frequency variation, not depth relative to that for the target formants. The absence of tuning (i.e., no minimum in intelligibility for the 50% case) suggests that the ability to reject an extraneous formant does not depend on similarity in the depth of formant-frequency variation. Furthermore, triangle-wave competitors were as effective as their more speech-like counterparts, suggesting that the selection of formants from the ensemble also does not depend on speech-specific constraints.

Formant-frequency variation and informational masking of speech by extraneous formants:evidence against dynamic and speech-specific acoustical constraints

How speech is separated perceptually from other speech remains poorly understood. Recent research indicates that the ability of an extraneous formant to impair intelligibility depends on the variation of its frequency contour. This study explored the effects of manipulating the depth and pattern of that variation. Three formants (F1+F2+F3) constituting synthetic analogues of natural sentences were distributed across the 2 ears, together with a competitor for F2 (F2C) that listeners must reject to optimize recognition (left = F1+F2C; right = F2+F3). The frequency contours of F1 - F3 were each scaled to 50% of their natural depth, with little effect on intelligibility. Competitors were created either by inverting the frequency contour of F2 about its geometric mean (a plausibly speech-like pattern) or using a regular and arbitrary frequency contour (triangle wave, not plausibly speech-like) matched to the average rate and depth of variation for the inverted F2C. Adding a competitor typically reduced intelligibility; this reduction depended on the depth of F2C variation, being greatest for 100%-depth, intermediate for 50%-depth, and least for 0%-depth (constant) F2Cs. This suggests that competitor impact depends on overall depth of frequency variation, not depth relative to that for the target formants. The absence of tuning (i.e., no minimum in intelligibility for the 50% case) suggests that the ability to reject an extraneous formant does not depend on similarity in the depth of formant-frequency variation. Furthermore, triangle-wave competitors were as effective as their more speech-like counterparts, suggesting that the selection of formants from the ensemble also does not depend on speech-specific constraints. © 2014 The Author(s).

Formant-frequency variation and its effects on across-formant grouping in speech perception

How speech is separated perceptually from other speech remains poorly understood. In a series of experiments, perceptual organisation was probed by presenting three-formant (F1+F2+F3) analogues of target sentences dichotically, together with a competitor for F2 (F2C), or for F2+F3, which listeners must reject to optimise recognition. To control for energetic masking, the competitor was always presented in the opposite ear to the corresponding target formant(s). Sine-wave speech was used initially, and different versions of F2C were derived from F2 using separate manipulations of its amplitude and frequency contours. F2Cs with time-varying frequency contours were highly effective competitors, whatever their amplitude characteristics, whereas constant-frequency F2Cs were ineffective. Subsequent studies used synthetic-formant speech to explore the effects of manipulating the rate and depth of formant-frequency change in the competitor. Competitor efficacy was not tuned to the rate of formant-frequency variation in the target sentences; rather, the reduction in intelligibility increased with competitor rate relative to the rate for the target sentences. Therefore, differences in speech rate may not be a useful cue for separating the speech of concurrent talkers. Effects of competitors whose depth of formant-frequency variation was scaled by a range of factors were explored using competitors derived either by inverting the frequency contour of F2 about its geometric mean (plausibly speech-like pattern) or by using a regular and arbitrary frequency contour (triangle wave, not plausibly speech-like) matched to the average rate and depth of variation for the inverted F2C. Competitor efficacy depended on the overall depth of frequency variation, not depth relative to that for the other formants. Furthermore, the triangle-wave competitors were as effective as their more speech-like counterparts. Overall, the results suggest that formant-frequency variation is critical for the across-frequency grouping of formants but that this grouping does not depend on speech-specific constraints. © Springer Science+Business Media New York 2013.

Technological enhancements to optometric clinical tests

A sizeable amount of the testing in eye care, requires either the identification of targets such as letters to assess functional vision, or the subjective evaluation of imagery by an examiner. Computers can render a variety of different targets on their monitors and can be used to store and analyse ophthalmic images. However, existing computing hardware tends to be large, screen resolutions are often too low, and objective assessments of ophthalmic images unreliable. Recent advances in mobile computing hardware and computer-vision systems can be used to enhance clinical testing in optometry. High resolution touch screens embedded in mobile devices, can render targets at a wide variety of distances and can be used to record and respond to patient responses, automating testing methods. This has opened up new opportunities in computerised near vision testing. Equally, new image processing techniques can be used to increase the validity and reliability of objective computer vision systems. Three novel apps for assessing reading speed, contrast sensitivity and amplitude of accommodation were created by the author to demonstrate the potential of mobile computing to enhance clinical measurement. The reading speed app could present sentences effectively, control illumination and automate the testing procedure for reading speed assessment. Meanwhile the contrast sensitivity app made use of a bit stealing technique and swept frequency target, to rapidly assess a patient’s full contrast sensitivity function at both near and far distances. Finally, customised electronic hardware was created and interfaced to an app on a smartphone device to allow free space amplitude of accommodation measurement. A new geometrical model of the tear film and a ray tracing simulation of a Placido disc topographer were produced to provide insights on the effect of tear film breakdown on ophthalmic images. Furthermore, a new computer vision system, that used a novel eye-lash segmentation technique, was created to demonstrate the potential of computer vision systems for the clinical assessment of tear stability. Studies undertaken by the author to assess the validity and repeatability of the novel apps, found that their repeatability was comparable to, or better, than existing clinical methods for reading speed and contrast sensitivity assessment. Furthermore, the apps offered reduced examination times in comparison to their paper based equivalents. The reading speed and amplitude of accommodation apps correlated highly with existing methods of assessment supporting their validity. Their still remains questions over the validity of using a swept frequency sine-wave target to assess patient’s contrast sensitivity functions as no clinical test provides the range of spatial frequencies and contrasts, nor equivalent assessment at distance and near. A validation study of the new computer vision system found that the authors tear metric correlated better with existing subjective measures of tear film stability than those of a competing computer-vision system. However, repeatability was poor in comparison to the subjective measures due to eye lash interference. The new mobile apps, computer vision system, and studies outlined in this thesis provide further insight into the potential of applying mobile and image processing technology to enhance clinical testing by eye care professionals.

Semi-supervised learning of statistical models for natural language understanding

Natural language understanding is to specify a computational model that maps sentences to their semantic mean representation. In this paper, we propose a novel framework to train the statistical models without using expensive fully annotated data. In particular, the input of our framework is a set of sentences labeled with abstract semantic annotations. These annotations encode the underlying embedded semantic structural relations without explicit word/semantic tag alignment. The proposed framework can automatically induce derivation rules that map sentences to their semantic meaning representations. The learning framework is applied on two statistical models, the conditional random fields (CRFs) and the hidden Markov support vector machines (HM-SVMs). Our experimental results on the DARPA communicator data show that both CRFs and HM-SVMs outperform the baseline approach, previously proposed hidden vector state (HVS) model which is also trained on abstract semantic annotations. In addition, the proposed framework shows superior performance than two other baseline approaches, a hybrid framework combining HVS and HM-SVMs and discriminative training of HVS, with a relative error reduction rate of about 25% and 15% being achieved in F-measure.

A statistical analysis of regional variation in adverb position in a corpus of written Standard American English

This paper investigates whether the position of adverb phrases in sentences is regionally patterned in written Standard American English, based on an analysis of a 25 million word corpus of letters to the editor representing the language of 200 cities from across the United States. Seven measures of adverb position were tested for regional patterns using the global spatial autocorrelation statistic Moran’s I and the local spatial autocorrelation statistic Getis-Ord Gi*. Three of these seven measures were indentified as exhibiting significant levels of spatial autocorrelation, contrasting the language of the Northeast with language of the Southeast and the South Central states. These results demonstrate that continuous regional grammatical variation exists in American English and that regional linguistic variation exists in written Standard English.

Automatic Identification of False Friends in Parallel Corpora: Statistical and Semantic Approach

False friends are pairs of words in two languages that are perceived as similar but have different meanings. We present an improved algorithm for acquiring false friends from sentence-level aligned parallel corpus based on statistical observations of words occurrences and co-occurrences in the parallel sentences. The results are compared with an entirely semantic measure for cross-lingual similarity between words based on using the Web as a corpus through analyzing the words’ local contexts extracted from the text snippets returned by searching in Google. The statistical and semantic measures are further combined into an improved algorithm for identification of false friends that achieves almost twice better results than previously known algorithms. The evaluation is performed for identifying cognates between Bulgarian and Russian but the proposed methods could be adopted for other language pairs for which parallel corpora and bilingual glossaries are available.

Learning higher-level features with convolutional restricted Boltzmann machines for sentiment analysis

In recent years, learning word vector representations has attracted much interest in Natural Language Processing. Word representations or embeddings learned using unsupervised methods help addressing the problem of traditional bag-of-word approaches which fail to capture contextual semantics. In this paper we go beyond the vector representations at the word level and propose a novel framework that learns higher-level feature representations of n-grams, phrases and sentences using a deep neural network built from stacked Convolutional Restricted Boltzmann Machines (CRBMs). These representations have been shown to map syntactically and semantically related n-grams to closeby locations in the hidden feature space. We have experimented to additionally incorporate these higher-level features into supervised classifier training for two sentiment analysis tasks: subjectivity classification and sentiment classification. Our results have demonstrated the success of our proposed framework with 4% improvement in accuracy observed for subjectivity classification and improved the results achieved for sentiment classification over models trained without our higher level features.