Linguistic Modeling and its Interfaces
Oberseminar, Detmar Meurers, Winter Semester 2011/2012
The OS features presentations and discussions of current issues in linguistic modeling and its interfaces. This includes linguistic modeling in computational linguistics, language acquisition research, Intelligent Computer-Assisted Language Learning, as well as theoretical linguistic research with a focus on the interfaces of syntax and information structure. It is open to advanced students and anyone interested in this interdisciplinary enterprise.
Slides of talk, her recent PhD thesis on this topic
Abstract: Assessing language learners’ language is an invaluable source of information especially with respect to learners’ linguistic strengths and weaknesses. Such learners’ particulars generally constitute the basis of any language learner model. Typically integrated into intelligent computer-assisted language learning applications, learner models enable a representation of the learner knowledge, which is generally used to adapt teaching strategies and support learning activities. However, most representations of learners’ knowledge depend on observations drawn from learners’ performance, assuming that an incorrect grammatical or lexical word refers by default to a lack of competence, as opposed to a possible gap in performance, errors and mistakes, respectively. In effect, predictability with regard to learning accomplishments may be compromised due to non-germane variations in a learner’s knowledge estimation. While the relationship between learners’ interlanguage variability and factors, such as task type, time pressure, task complexity, or planning, to name but a few, is well documented in second language acquisition research, and more specifically in the psycholinguistic and sociolinguistic domains, the impact of the distinction between errors and mistakes on the variability of a learner’s interlanguage has not been significantly addressed.
This research is located at the centre of three discrete, yet interconnected domains:
learner interlanguage variability, learner knowledge modeling, and learner language
assessment. This presentation suggests a dynamic assessment approach grounded in
sociocultural theory to (a) model the knowledge of intermediate learners of French from their
unrestricted written language by differentiating between competence-dependent errors and
performance-related mistakes, and (b) investigate the extent to which interlanguage
competence varies across time, text types, and students. Results demonstrate that
interlanguage competence is subject to not only systematic, but also unsystematic
variations.
extra meeting: Wednesday, October 26, 9:00st-10:00
Martí Quixal (UPF, Barcelona)
Making ICALL viable in real-life schools: Supporting teacher autonomy in
building NLP enhanced activities
Abstract: The talk will present the work and the results obtained during the first year of the project ”Integrating CALL in Early Education Environments (ICE3)”. The ICE3 project aims at the integration of CALL and particularly CALL activities integrating Natural Language Processing (ICALL) in language teaching and learning processes in early education by facilitating teachers the creation and management of ICALL activities. To do so ICE3 promotes teacher autonomy: Teachers move from using prefabriacted teaching materials to designing and developing their own, better adapted to the target learner needs. Moreover, ICALL technology fosters students’ autonomy: Students can work at their own pace and be actively involved in learning, as the software generates immediate, scaffolding feedback.
We will go over the materials created by one of the teachers to exemplify the achievements teachers made and the difficulties they found. We will also analyse the extent to which teachers, upon reflection, seem to have understood the ”problem” of developing NLP resources for automated evaluation. We will include an analysis of the system feedback by comparing it to the feedback that teachers would have provided.
This research is funded by the Lifelong Learning Programme 2007-2013. Project number:
510653-LLP-1-2010-1-ES-COMENIUS-CMP.
October 28:
Bernd Bohnet (Uni. Stuttgart)
Efficient Syntactic and Semantic Dependency Parsing
Abstract: This talk introduces the task of syntactic and semantic parsing and the design of efficient
parsing systems. Semantic analyzers consist of processing pipelines to tokenize, lemmatize, tag, and
parse sentences, where all the steps are crucial to their overall performance. In practice, however,
while code of dependency parsers and semantic role labelers is available, few systems can be run as
standalone applications and even fewer with a fast processing times. We address decisions which
have to made when systems are composed and important factors that contribute to a
accurate and fast system. We compare approaches and their performance on different
languages for both dependency parsing and semantic role labeling. Finally, we address the
problem of domain adaptation and give an outlook on further developments in the
field.
November 4:
SFB 833 CoMiC Project
Dependency annotation of learner language
November 11:
Patrick Rebuschat (Bangor University)
Implicit and Explicit Learning of Language
Abstract: The process of implicit learning, essentially the ability to acquire unconscious knowledge, is one of the central topics in cognitive psychology. The term implicit learning was first employed by Arthur Reber (1967) to describe a process during which subjects acquire knowledge about a complex, rule-governed stimulus domain without intending to and without becoming aware of the knowledge they have acquired. In contrast, the term explicit learning is usually applied to learning scenarios in which subjects are instructed to actively look for patterns, i.e. learning is intentional, a process which tends to result conscious knowledge. Many essential skills, including language comprehension and production, social interaction, music perception, and intuitive decision making, are largely dependent on implicit knowledge.
In this talk, I will review a series of experiments that investigated the implicit and explicit learning
of languages. The linguistic focus will be on the acquisition of L2 word order (German verb
placement). This research addressed questions such as the following: Is there implicit learning in
the case of second language (L2) acquisition? If so, how is this knowledge represented in the
mind (rules, patterns, chunks...)? How do task instructions affect implicit and explicit
learning? Is there an implicit-explicit interface? And what is the role of individual
differences (e.g. working memory capacity), in the implicit and explicit learning of
languages?
November 25:
Update on collaborations/projects:
December 2:
Manfred Sailer (Universität Göttingen)
The Anti-C-Command Condition in Head-Driven Phrase Structure
Grammar
Abstract: The existence of multiple gap constructions (MGC) as in (1) follows automatically from the basic mechanism of HPSG’s nonderivational treatment of unbounded dependencies.
(1) (I wonder) which booki Pat gave (a copy of ti) (to a fan of ti).
There is, however, a constraint on MGCs that posed an unsuperable problem for earlier nonderivational accounts of MGCs: Engdahl’s Anti-C-Command Condition (ACC), which says that a gap may not c-command a co-indexed gap, see (2).
(2) * (I wonder) which booki Pat gave ti to a fan of ti.
Being a configurationally formulated constraint the ACCs looks equally problematic for HPSG. Levine & Hukari 2006 show that it follows from HPSG’s binding theory for the cases discussed in the literature so far. The traces in (2) are subject to Principle C and, consequently, must not be o-commanded. The sentence is ruled out because the second trace is o-commanded by the first. While this account is extremely elegant, I will show that it does not capture the ACC facts completely, nor does it account for constructions which involve multiple fillers as in (3).
(3) * You need to work out wherei whoj to assign tj to ti. (Levine & Hukari 2006)
I will propose a Generalized ACC that holds for fillers with little descriptive content (pronouns or wh-phrases consisting of a wh-proform only), whereas the ACC holds in its traditional form for fillers with rich descriptive content (full NPs or ”which”-phrases). I will present a conservatively modified version of HPSG’s theory of unbounded dependencies that incorporates the Generalized ACC as a principle on local trees.
References: Levine, R.D. and Hukari, T.E. (2006): The Unity of Unbounded Dependency
Constructions. Stanford: CSLI Publications.
December 5 and 6:
Second Tübingen-Berlin Meeting on Analyzing Learner Language
January 13:
Niels Beuck and Arne Köhn (NATS, Uni Hamburg)
Incremental Dependency Parsing with WCDG
Abstract: Human language processing is incremental in the sense that we start processing sentences and have an interpretation of them before they are finished. As for many NLP applications whole sentences are available at all times, most parsers don’t provide incremental processing. However, it is needed for systems that react on language input that is still being produced.
In this talk we will present the Weighted Constraint Dependency Grammar (WCDG) Framework and how incremental syntactic dependency parsing can be done with it. We will present different strategies for incremental processing in a pipeline consisting of a part-of-speech tagger and a parser. The main challenge in incremental processing is that decisions need to be made before all relevant information is available. Delaying the assignment of a head to a word until both of them are available can lead to very fragmented intermediate analyses, e.g., for the verb final constellation seen in (1).
(1) dass der Händler dem Kunden den Wagen [verb]
We will show how structural prediction can be used to build more connected structures in
WCDG.
January 20:
Yi Xu (University College London)
Prosodic focus from an experimental perspective
Abstract: In this talk I will present experimental evidence that prosodic focus is a stylized,
morpheme-like prosodic structure with an asymmetrical configuration. This structure interacts
extensively with other functions like lexical contrast, bound ary marking and modality signaling,
resulting in various allomorph-like forms. The focus construction is highly categorical in the
languages where it is found, but its characteristic asymmetrical pattern is absent in many other
languages, and this uneven cross-linguistic distribution may have a historical origin.
This view of focus has many potential implications. First, focus projection is likely an
epiphenomenon of the asymmetrical form of focus. Second, there needs to be a conceptual
separation of focus construction from its functional triggers. The latter may be universal
while the former is likely language-dependent. Third, contrary to the notion of nuclear
accent, focus is unlikely to be omnipresent in every sentence. Finally, I will discuss
ways to enhance experimental rigor in the study of focus as well as other prosodic
phenomena.
_________________________________________________________________________________
Last updated: January 27, 2012