Archived Seminars by Year

Abstract

Visual processing of complex natural environments requires animals to combine, in a highly dynamic and adaptive manner, sensory signals that originate from the environment (bottom-up) with behavioral goals and priorities dictated by the task at hand (top-down). Together, bottom-up and top-down influences combine to serve the many tasks which require that we direct attention to the most ''relevant'' entities in our visual environment. While much progress has been made in investigating experimentally how humans and other primates may operate such goal-based attentional selection, very little is understood of the general mathematical principles and neuro-computational architectures that subserve the observed behavior. I will describe recent computational work which attacks the problem of developing models of visual attentional selection that are more flexible and can be strongly modulated by the task at hand. I will back the proposed architectures up by comparing their predictions to behavioral recordings from humans and monkeys. I will show examples of applications of these models to real-world vision challenges, using complex stimuli from television programs or modern immersive video games.

Abstract

A translation is generally taken to be a text that expresses the same meaning as another text in a different language. But the products of the best translators reflects a different, if more illusive, goal. I will seek a somewhat more adequate characterization of translation as it is actually practiced and discuss its consequences for machine translation.

Abstract

Language understanding is hard because it requires a lot of knowledge. However, the only cost-effective way to acquire a lot of knowledge is by extracting it from text. The best (only?) hope for solving this "chicken and egg" problem is bootstrapping: start with a small knowledge base, use it to process some text, add the extracted knowledge to the KB, process more text, etc. Doing this requires a modeling language that can incorporate noisy knowledge and seamlessly combine it with statistical NLP algorithms. Markov logic accomplishes this by attaching weights to first-order formulas and viewing them as templates for features of Markov random fields. In this talk, I will describe some of the main inference and learning algorithms for Markov logic, and the progress we have made so far in applying them to NLP. For example, we have developed a system for unsupervised coreference resolution that outperforms state-of-the-art supervised ones on MUC and ACE benchmarks.

Abstract

State-of-the-art NLP models are anything but compact. Parsers have huge grammars, machine translation systems have huge transfer tables, and so on across a range of tasks. With such complexity comes two challenges. First, how can we learn highly complex models? Second, how can we efficiently infer optimal structures within them? Hierarchical coarse-to-fine (CTF) methods address both questions. CTF approaches exploit sequences of models which introduce complexity gradually. At the top of the sequence is a trivial model in which learning and inference are both cheap. Each subsequent model refines the previous one, until a final, full-complexity model is reached. Because each refinement introduces only limited complexity, both learning and inference can be done in an incremental fashion. In this talk, I describe several coarse-to-fine NLP systems. In the domain of syntactic parsing, complexity comes from the grammar. I present a latent-variable approach which begins with an X-bar grammar and learns by iteratively splitting grammar symbols. For example, noun phrases might be split into subjects and objects, singular and plural, and so on. This splitting process admits an efficient incremental inference scheme which reduces parsing times by orders of magnitude. I also present a multiscale variant which splits grammar rules rather than grammar symbols. In the multiscale approach, complexity need not be uniform across the entire grammar, providing orders of magnitude of space savings. These approaches produce the best parsing accuracies in a variety of languages, in a fully language-general fashion. In the domain of syntactic machine translation, complexity arises from both the translation model and the language model. In short, there are too many transfer rules and too many target language word types. To manage the translation model, we compute minimizations which drop rules that have high computational cost but low importance. To manage the language model, we translate into target language clusterings of increasing vocabulary size. These approaches give dramatic speed-ups, while actually increasing final translation quality.

Abstract

Over the last years finite state transducer technology has found its way in many speech applications, from text processing for synthesizers to the core search algorithms used in speech recognition systems. In this talk we present a novel application of finite state transducers and acoustic modeling techniques to the problem of music fingerprinting*. We will show how the power of FST's can be applied to this problem with great results. In the talk I will also give an overview of current activities in google speech team. * This is joint work with Prof. Mehriar Mohri and Eugene Weinstein at NYU.

Abstract

Supervised learning technology has led to systems for part of speech tagging, parsing, named entity recognition with accuracies in the high 90%s. Unfortunately, the performance of these systems degrades drastically when they are applied on text outside their training domain (typically, newswire). Machine translation systems work fantastically for translating Parliamentary proceedings, but fall down when applied to alternate domains. I'll discuss research that aims to understand what goes wrong when models are applied outside their domain, and some (partial) solutions to this problem. I'll focus on named entity recognition and machine translation tasks, where we'll see a range of different sources of error (some of which are quite counter-intuitive!).

Abstract

Several optimization techniques are vastly used today in the speech and language community for estimating model parameters. The Extended Baum-Welch (EBW) is one such technique that is extensively used for estimating the parameters of Gaussian mixture models based on a discriminative criteria (like Maximum Mutual Information). In this talk, we present EBW as a consistent, theoretical framework for parameter estimation and show how other common parameter estimation techniques (for example, based on Constrained Line Search) belong to this family of model update rules. We introduce a general family of parameter updates that generalizes a Baum-Welch recursive process to an arbitrary objective function of Gaussian Mixture Models or Poisson Processes. In the second part of this talk we introduce an extension of the EBW for estimating sparse signals from a sequence of noisy observations. As part of this, the underlining EBW algorithms are compared with recently introduced Kalman filtering-based compressed sensing methods. This is joint work with Avishy Carmi, David Nahamoo, Bhuvana Ramabhadran and Tara Sainath.

Abstract

Via four applications: distribution modeling, probability estimation, data compression, and classification, we argue that when learning from data, discrete values should be ignored except for just their appearance-order pattern. Along the way, we encounter Laplace, Good, Turing, Hardy, Ramanujan, Fisher, Shakespeare, and Shannon. The talk is self contained and based on work with P. Santhanam, K. Viswanathan, J. Zhang, and others.

Abstract

This talk describes three recent models of how the brain visually understands the world. The models use hierarchical and parallel processes within and across the What and Where cortical streams to accumulate information that cannot in principle be fully computed at a single processing stage. The models hereby raise basic questions about the functional brain units that are selected by the evolutionary process, and challenge all models that use non-local information to explain vision. The ARTSCAN model (Fazl, Grossberg, & Mingolla, 2008, Cognitive Psychology) clarifies the following issues: What is an object? How does the brain learn to bind multiple views of an object into a view-invariant object category, during both unsupervised and supervised learning, while scanning its various parts with active eye movements? In particular, how does the brain avoid the problem of erroneously classifying views of different objects as belonging to a single object, and how does the brain direct the eyes to explore an object's surface even before it has a concept of the object? How does the brain coordinate object and spatial attention during object learning and recognition? ARTSCAN proposes an answer to these questions by modeling interactions between cortical areas V1, V2, V3A, V4, ITp, ITa, PPC, LIP, and PFC. The ARTSCENE model (Grossberg & Huang, 2008, Journal of Vision) also uses attentional shrouds. It clarifies the following issues: How do humans rapidly recognize a scene? How can neural models capture this biological competence to achieve state-of-the-art scene classification? ARTSCENE classifies natural scene photographs better than competing models by using multiple spatial scales to efficiently accumulate evidence for gist and texture. The model can incrementally learn and rapidly predict scene identity by gist information alone (defining gist computatationally along the way), and then accumulate learned evidence from scenic textures to refine this hypothesis. The MODE model (Grossberg & Pilly, 2008, Vision Research) clarifies the following basic issue: How does the brain make decisions? Speed and accuracy of perceptual decisions covary with certainty in the input, and correlate with the rate of evidence accumulation in parietal and frontal cortical "decision neurons." MODE models interactions within and between Retina/LGN and cortical areas V1, MT, MST, and LIP, gated by basal ganglia, to simulate dynamic properties of decision-making in response to ambiguous visual motion stimuli used by Newsome, Shadlen, and colleagues in their neurophysiological experiments. The model shows how the brain can carry out probabilistic decisions without using Bayesian mechanisms.

Abstract

Speech production involves coordinated processing in many regions of the brain. To better understand these processes, our laboratory has designed, tested, and refined a neural network model whose components correspond to brain regions involved in speech. Babbling and imitation phases are used to train neural mappings between phonological, articulatory, auditory, and somatosensory representations. After learning, the model can produce syllables and words it has learned by commanding movements of an articulatory synthesizer. Because the model’s components correspond to neurons and are given precise anatomical locations, activity in the model’s cells can be compared to neuroimaging data. Computer simulations of the model account for a wide range of experimental findings, including data on acquisition of speaking skills, articulatory kinematics, and brain activity during speech. "Impaired" versions of the model are being used to investigate several communication disorders, and the model has been used to guide development of a neural prosthesis aimed at restoring speech output to profoundly paralyzed individuals.

Abstract

In this talk I'll outline current work at the University of Edinburgh to model machine translation (MT) as a probabilistic machine learning problem. Although MT systems have made large gains in translation quality in recent years, most are currently induced using a hand engineered pipeline of disparate models linked by heuristics. Although such techniques are effective for translating between related languages, they fail to capture the latent structure necessary to translate between languages which diverge significantly in word order, such as Chinese and English. I'll present a non-parametric Bayesian model for inducing synchronous context free grammars capable of learning the latent structure of translation equivalence from a corpus of parallel string pairs. I'll discuss the efficacy of both variational Bayes and Gibbs sampling inference procedures for this model and present experiments demonstrating competitive results on full scale translation evaluations.

Abstract

Many acoustic events, particularly those associated with speech events, can be thought of as events in a rich descriptive subspace where the dimensions of the subspace can be thought of as a sort of decomposition of the original event space. In phonetic terms, we can think of how phonological features can be integrated to determine phonetic identity; for auditory scene analysis we can look how features like harmonic energy and cross-channel correlation come together to determine whether a particular frequency corresponds to target speech versus background noise. Some success has been achieved by thinking of these problems as probabilistic detection of acoustic (sub-)events. However, event detectors are typically local in nature, and need to be smoothed out by looking at neighboring events in time. In this talk, I describe current work in the Speech and Language Technologies Lab at OSU where we are looking at Conditional Random Fields models for both automatic speech recognition and computational auditory scene analysis problems. The talk will explore some of the successes and limitations of this log-linear method which integrates local evidence over time sequences. Joint work with Jeremy Morris, Ilana Heintz, Rohit Prabhavalkar, Zhaozhang Jin.

Abstract

Proper representation of the acoustic speech signal is crucial for almost every speech processing application. We often use short-time Fourier transform to convert the time-domain speech waveform to a new signal that is a function of both time and frequency by applying a moving time window of about 20 ms in duration. There are many issues, such as the size and shape of the window, that remain unresolved. The use of a relatively short window is widespread. In early development of the sound spectrograph, both narrow and wideband analysis were used, but the narrow-band analysis faded away. In digital speech coding applications (multipulse and code-excited linear prediction), high-quality speech is produced at low bit rates only when prediction using both short and long intervals is used. Recently Hermansky and others have argued that speech window for automatic speech recognition should be long, perhaps extending to as much as 1 s. What are the issues that arise in using a short or a long window? What are the relative advantages or disadvantages? In this talk, we will discuss these topics and present results that suggest that a short-time Fourier transform using long windows has advantages. In most speech representations, the Fourier components are not used directly but converted to their magnitude spectrum; the so-called phase is considered to be irrelevant. There are open questions regarding the use of phase information and we will discuss this important issue in the talk.

Abstract

Fillmore's study "The Grammar of Hitting and Breaking" demonstrated the importance of semantically coherent verb classes as descriptive devices for understanding the organization of the verb lexicon and for capturing patterns of shared verb behavior. Much subsequent work has confirmed and extended the findings of this study. For example, in my 1993 book "English Verb Classes and Alternations", verbs are essentially classified in two ways: according to their semantic content (e.g., verbs of manner of motion, verbs of sound emission) and according to their participation in particular argument alternations (e.g., dative alternation, causative alternation). The first approach yields a fairly fine-grained semantic classification, while the second yields a coarser-grained classification, which appears to have more grammatical relevance than the first. This and other work suggests that verb classes should not be taken as primitive, as they have sometimes been. This position is reinforced by the considerable number of verbs which show complex patterns of behavior that have been handled by positing multiple semantic class membership. Previous studies of verb classes, then, raise important questions: What are the most useful dimensions for classifying verbs? What is the appropriate grain size for the description of verb classes? What determines whether a given verb shows multiple class membership? In this talk I ask what is behind verb classes that makes them so appealing as a research tool, yet explains their limitations. I show that many phenomena falling under the label "verb class" can be understood in the context of three levels of linguistic description and the relations between them: (i) the meaning lexicalized by the verb itself (its "root"), (ii) the set of event schemas, and (iii) the morphosyntactic devices that languages make available for the realization of arguments (e.g., grammatical relations, case markers, serial verb constructions). Each provides a way of grouping verbs into classes that can be helpful for certain facets of both language-specific and crosslinguistic studies.

Abstract

It’s easy to have a $50 idea. Innovation at a scale large enough to be material to a big company like Microsoft is a completely different story. I’ll discuss some of the interesting challenges and opportunities innovating at the scale of a billion dollars, and how different types of innovation play a role in driving financial value. I’ll also share other things I’ve learned in my decade at Microsoft, including: tips to young scientists on how to have a great life-long career in a corporate setting, and how to make basic research much more valuable/impactful/profitable.

Abstract

Conceptually sequence kernels map variable length sequences into a fixed dimensional feature-space. In this feature space, for example, an inner-product can be computed. The ability to handle variable length sequences means that these kernels are suitable for speech signals which are by nature time varying. In the speech processing area, sequence kernels have been succesfully applied in speaker verification, where they are used in combination with support vector machines (SVMs) for classification. This talk will concentrate on a particular class of sequence kernels, generative kernels and how they can be used for speaker and speech recognition. Generative kernels, and score-spaces, make use of generative models such as hidden Markov models (HMMs) and Gaussian mixture models (GMMs). By taking first and higher-order derivatives of the log likelihood with respect to the model paarameters fixed dimenesional feature vectors can be extracted. An example of this form of kernel is the Fisher Kernel successfully applied to a range of biological sequences. The relationship of this form of kernel to schemes such as the GMM mean-Supervector kernel, commonly used in speaker verification, will be discussed. In addition, how these kernels and associated feature-spaces can be used for speech recognition and how they can handle speaker and environment changes will be looked at.

Abstract

Events occur daily that challenge the security, health and sustainable growth of our nation, and often find our government agencies unprepared for the catastrophic outcomes. These events involve the interaction of complex processes such as climate change, energy reliability, terrorism, nuclear proliferation, natural and man-made disasters, social/political and economic vulnerability. If we are to help our nation to meet the challenges that emerge from these events, we must develop novel methods for predictive analysis that support a concerted decision-making effort by analyst and policymakers to anticipate and counter strategic surprise. There is now increased awareness among subject-matter experts, analysts, and decision makers that a combined understanding of interacting physical and human factors is essential in addressing strategic surprise proactively. The Technosocial Predictive Analytics (TPA) framework provides an operational advancement of this insight through the development of new methods for anticipatory analysis and response that · implement a multi-perspective approach to predictive modeling through the integration of human and physical models · facilitate the achievement of knowledge/evidence inputs to support the modeling task and promote inferential transparency · enable analysts and policymakers to stress-test the quality of their intelligence products and planned responses without waiting for history to prove them right or wrong. Human Language Technologies (HLT) play an important role in the realization of this framework with specific reference to evidence extraction, but must be augmented to support TPA’s knowledge requirements properly. In presenting TPA, I will discuss an approach which provides such an extension of HLT through the integration of insights from specific domains of expertise and content analysis processes.

Abstract

Computational advertising is an emerging new scientific sub-discipline, at the intersection of large scale search and text analysis, information retrieval, statistical modeling, machine learning, classification, optimization, and microeconomics. The central challenge of computational advertising is to find the "best match" between a given user in a given context and a suitable advertisement. The context could be a user entering a query in a search engine ("sponsored search") , a user reading a web page ("content match" and "display ads"), a user watching a movie on a portable device, and so on. The information about the user can vary from scarily detailed to practically nil. The number of potential advertisements might be in the billions. Thus, depending on the definition of "best match" this challenge leads to a variety of massive optimization and search problems, with complicated constraints. This talk will give an introduction to this area focusing on the IR and NLP connections.

Abstract

Anyone who has used an automatic speech recognition (ASR) system, either on a customer support line or on their own personal computer, knows firsthand there is vast room for improvement. While state-of-the-art commercial systems perform very well in near-ideal environments, system robustness remains far below human levels. The prevailing hidden Markov model (HMM) based paradigm will undoubtedly see gains in future decades as increased computing capacity admits more complex acoustic models that encompass a range of acoustic environments. In the meantime, there is a wealth of scientific understanding of production and perceptual mechanisms that has yet to be fully exploited by engineers and technologists. In this talk, I will present the main results of a research program that takes scientific inspiration from linguistics, speech perception, and neuroscience as starting points for alternative directions in automatic speech recognition. First, I consider the implications speech production have on the geometric structure of speech sounds and the role this perspective can play in speech technology. Second, I consider the hypothesis that the linguistic content underlying human speech may be more efficiently and robustly coded in the pattern of timings of various acoustic events (landmarks) present in the speech signal. I will present a point process-based statistical framework for phonetic recognition and keyword spotting that matches the performance of equivalent frame-based systems. This approach suggests a new unsupervised adaptation strategy for improving recognizer robustness that outperforms maximum likelihood linear regression adaptation of a continuous density keyword-filler HMM system.

Abstract

Anyone who has used an automatic speech recognition (ASR) system, either on a customer support line or on their own personal computer, knows firsthand there is vast room for improvement. While state-of-the-art commercial systems perform very well in near-ideal environments, system robustness remains far below human levels. The prevailing hidden Markov model (HMM) based paradigm will undoubtedly see gains in future decades as increased computing capacity admits more complex acoustic models that encompass a range of acoustic environments. In the meantime, there is a wealth of scientific understanding of production and perceptual mechanisms that has yet to be fully exploited by engineers and technologists. In this talk, I will present the main results of a research program that takes scientific inspiration from linguistics, speech perception, and neuroscience as starting points for alternative directions in automatic speech recognition. First, I consider the implications speech production have on the geometric structure of speech sounds and the role this perspective can play in speech technology. Second, I consider the hypothesis that the linguistic content underlying human speech may be more efficiently and robustly coded in the pattern of timings of various acoustic events (landmarks) present in the speech signal. I will present a point process-based statistical framework for phonetic recognition and keyword spotting that matches the performance of equivalent frame-based systems. This approach suggests a new unsupervised adaptation strategy for improving recognizer robustness that outperforms maximum likelihood linear regression adaptation of a continuous density keyword-filler HMM system.

Abstract

NLP systems typically consist of a series of components where the output of one module (e.g., a word segmenter) serves as input to another (e.g., a translator). Integration between the components is often achieved using only the 1-best analysis from an upstream component as the input to a downstream component. Unfortunately, this naive integration strategy results in compounding error propagation (cf. Finkel et al. 2006, Dyer et al. 2008). In this talk, I briefly review the effects of this problem in machine translation, where examples of upstream uncertainty include not only the noisy outputs of statistical preprocessors (such as word segmenters and STT systems), but also "development-time" decisions (such as determining what the appropriate granularity of the lexical units is or how much text normalization to do). I show that by encoding input alternatives in a word lattice, translation quality can be improved over a 1-best baseline, with only a slight runtime performance cost. I then explore in more detail the implications of modeling development-time uncertainty jointly with translation, focusing on the problem of source language word segmentation. I tackle this problem in two ways. First, I present a Markov random field model of word segmentation and describe how to use it to generate lattices appropriate for translation by training it to maximize the (conditional) probability of a collection of segmentation alternatives, rather than maximizing the probability of a single correct analysis. Second, I describe generalized alignment models that align lattices in one language to strings in another, enabling the joint modeling of segmentation (or other noisy processes) and translation. Since lattice inputs break the Markov assumptions that enable the efficient inference made in many common word alignment models, I also present novel Monte Carlo techniques for performing word and lattice alignment.

Abstract

EM (the Expectation Maximization Algorithm) is a well known technique for unsupervised learning (where one does not have any hand labeled solutions available, but instead one must learn from the raw text). Unfortunately EM is known to fail to find good solutions in many (most?) applications on which it is tried. In this talk we present some recent work on using EM to learn how to resolve pronoun-anaphora, e.g., determining that "the dog" is the antecedent of "he" and "his" in "When Sally fed the dog he wagged his tail". For this application EM works strikingly well, determining tens of thousands of parameters and resulting in a program that produces state of the art performance.

Abstract

For centuries, the deep connection between human languages has fascinated scholars, and driven many important discoveries in linguistics and anthropology. In this talk, I will show that this connection can empower unsupervised methods for language analysis. The key insight is that joint learning from several languages reduces uncertainty about the linguistic structure of each individual language. I will present multilingual generative unsupervised models for morphological segmentation, part-of-speech tagging, and parsing. In all of these instances we model the multilingual data as arising through a combination of language-independent and language-specific probabilistic processes. This feature allows the model to identify and learn from recurring cross-lingual patterns to improve prediction accuracy in each language. I will also discuss ongoing work on unsupervised decoding of ancient Ugaritic tablets using data from related Semitic languages. This is joint work with Benjamin Snyder, Tahira Naseem and Jacob Eisenstein.

Abstract

I will discuss a couple of non-standard features that I believe could be useful for working with comparable corpora. Dotplots have been used in biology to find interesting DNA sequences. Biology is interested in ordered matches, which show up as (possibly broken) diagonals in dotplots. Information Retrieval is more interested in unordered matches (e.g., cosine similarity), which shows up as squares in dotplots. Parallel corpora have both squares and diagonals multiplexed together. The diagonals tell us what is a translation of what, and the squares tell us what is in the same language. There is also an opportunity to take advantage of repetition in comparable corpora. Repetition is very common. Standard bag-of-word models in Information Retrieval do not attempt to model discourse structure such as given/new. The first mention in a news article (e.g., "Manuel Noriega, former President of Panama") is different from subsequent mentions (e.g., "Noriega"). Adaptive language models were introduced in Speech Recognition to capture the fact that probabilities change or adapt. After we see the first mention, we should expect a subsequent mention. If the first mention has probability p, then under standard (bag-of words) independence assumptions, two mentions ought to have probability p^2, but we find the probability is actually closer to p/2. Adaptation matters more for meaningful units of text. In Japanese, words (meaningful sequences of characters) are more likely to be repeated than fragments (meaningless sequences of characters from words that happen to be adjacent). In newswire, we find more adaptation for content words (proper nouns, technical terminology, out of vocabulary (OOV) words and good keywords for information retrieval), and less adaptation for function words, clichés and ordinary first names. There is more to meaning than frequency. Content words are not only low frequency, but likely to be repeated.

Abstract

AI is beginning to make some dents in the "knowledge acquisition bottleneck", the problem of acquiring large amounts of general world knowledge to support language understanding and commonsense reasoning. Two text-based approaches to the problem are (1) to abstract such knowledge from patterns of predication and modification in miscellaneous texts, and (2) to derive such knowledge by direct interpretation of general statements in ordinary language, such as are found in lexicons and resources like Open Mind. I will discuss the status of our efforts in these directions (currently centered around the KNEXT system), and the problems that are encountered. Among these problems are what exactly is meant by generalities such as "Cats land on their feet", and how this meaning should be formalized. One particular difficulty is that such statements typically involve ``donkey anaphora". I will suggest a "dynamic Skolemization" approach that leads naturally to script- or frame-like representations, of the sort that have been developed in AI independently of linguistic considerations.

Abstract

When the number of sources exceeds the number of microphones, acoustic source separation is an underconstrained problem that must rely on additional constraints for solution. In a single-channel environment the expected behavior of the source -- i.e. an acoustic model -- is the only feasible basis for separation. I will describe our recent work in monaural speech separation based on fitting parametric "eigenvoice" speaker-adapted models to both voices in a mixture. In a binaural, reverberant environment, the interaural characteristics of an acoustic source exhibit structure that can be used to separate, even without prior knowledge of location or room characteristics. I will present MESSL, our EM-based system for source separation and localization. MESSL's probabilistic foundation facilitates the incorporation of more specific source models; I will also describe MESSL-EV, which incorporates the eigenvoice speech models for improved binaural separation in reverberant environments. Joint work with Ron Weiss and Mike Mandel.

Abstract

The parallels between language change and biological changes were noted by none other than Darwin himself. However, the development of a mathematical foundation for evolution has not taken place in the study of language change, even though tools from quantitative genetics have seen applications in the linguistic arena.This work attempts to develop a series of models of language change, drawing insights from population genetics on the one hand, and modern theories of linguistic structures, language acquisition and language processing on the other. The dynamics of language learning over generations turn out to bear strong resemblance to the process of Natural Selection. In some cases, this allows one to quantitatively measure the "fitness" of grammatical hypotheses and thus predict the directionality of language change. I will discuss the general use of population models in language, and present two specific case studies: the word order change from Old French to Modern French, and the cot-caught merger recently documented at the Massachusetts and Rhode Island border. The outcome of both changes is shown to be entirely predictable from the statistical composition of linguistic data in the environment.

Abstract

From the perspective of a linguist, today's vast archives of digital text and speech, along with new analysis techniques from language engineering, look like a wonderful new scientific instrument, a modern equivalent of the 17th-century invention of the telescope and microscope. We can now observe linguistic patterns in space, time, and cultural context, on a scale three to five orders of magnitude greater than in the past, and simultaneously in much greater detail than before. Scientific use of these new instruments remains mainly potential, especially in phonetics and related disciplines, but the next decade is likely to be a new "golden age" of research. This talk will discuss some of the barriers to be overcome, present some successful examples, and speculate about future directions.

Abstract

Vector-based models of word meaning have become increasingly popular in natural language processing and 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. In this talk we propose 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 which we evaluate empirically on a phrase similarity task. We also propose a novel statistical language model that is based on vector composition and can capture long-range semantic dependencies. Joint work with Jeff Mitchell

Abstract

For the last quarter century (measured in person years), the KnowItAll project has investigated information extraction at Web scale. If successful, this effort will begin to address the long-standing "Knowledge Acquisition Bottleneck" in Artificial Intelligence, and will enable a new generation of search engines that extract and synthesize information from text to answer complex user queries. To date, we have generalized information extraction methods to process arbitrary Web text, to handle unanticipated concepts, and to leverage the redundancy inherent in the Web corpus, but many challenges remain. One of the most formidable challenges is moving from extracting isolated nuggets of information to capturing a coherent body of knowledge that can support automatic inference. My talk will describe the lessons we have learned and identify directions for future work.

Abstract

Within the machine learning and data mining communities, there has been a growing interest in learning structured models from input data that is itself structured or semi-structured. Graph identification refers to methods that transform observational data described as a noisy input graph into an inferred, "clean" information graph. Examples include inferring social networks from online, noisy, communication data, identifying gene regulatory networks from protein-protein interactions, and extracting semantic graphs from noisy and ambiguous co-occurrence information. Some of the key processes in graph identification are: entity resolution, collective classification, and link prediction. I will overview algorithms for these tasks, discuss the need for integrating the methods to solve the overall problem jointly. Time permitting, I will also give quick overviews of some of the other research projects in my group.

Abstract

HiFST is a lattice-based decoder for hierarchical phrase-based translation and alignment. The decoder is implemented with standard Weighted Finite-State Transducer (WFST) operations as an alternative to the well-known cube pruning procedure. We find that the use of WFSTs rather than k-best lists requires less pruning in translation search, resulting in fewer search errors, better parameter optimization, and improved translation performance. The direct generation of translation lattices in the target language can improve subsequent rescoring procedures, yielding further gains when applying long-span language models and Minimum Bayes Risk decoding.

Abstract

In this talk we will give an overview of the different kinds of communication disorders. We will concentrate on communication disorders related to language and speech (i.e., not look at disorders like blindness or deafness). Speech and language disorders can range from simple sound substitution to the inability to understand or use language. Thus, a disorder may affect one or several linguistic levels: A patient with an articulation disorder cannot correctly produce speech sounds (phonemes) because of voice disorders or imprecise placement, timing, pressure, speed, or flow of movement of the lips, tongue, or throat. His speech may be acoustically unintelligible, yet the syntactic, semantic, and pragmatic level are not affected. With other pathologies, e.g. Wernicke’s aphasia, the acoustics of the speech signal might be intelligible, yet the patient is – due to mixup of words (semantic paraphasia) or sounds (phonematic paraphasia) – unintelligible. We will look at what linguistic knowledge has to be modeled in order to analyze different pathologies with speech technology, how difficult the task is, and how speech technology is able to support the speech therapist for the tasks diagnosis, therapy control, comparison of therapies, and screening. Joint work with Andreas Maier, Tino Haderlein, Stefan Steidl, and Maria Schuster.