Un exposé d'un ancien du cursus LI, qui applique des méthodes de TAL assez puissantes (construction semi-automatique de graphes lexicaux) dans une problématique de linguistique historique. C'est mercredi prochain 11 mars, au CRLAO, de 16h à 18h, à l'INaLCO, Salle des Plaques, 2, rue de Lille, 75007 Paris.
Pierre Magistry (Paris Diderot):
Stratification du réseau lexical du hokkien de Taïwan
Taiwanese Hokkien presents strong evidences of multiple historical
layers of borrowing through the pervasiveness of 「多音字」duoyinzi,
sinograms with multiple readings. For a given sinogram, traditional
analysis distinguish between wenduyin 文讀音 and baiduyin 白讀音 (so-called
"literary" and "colloquial" readings) but it is well accepted that more
than two strata are to be found and described.
We will both stress the limits of such analysis and explain how we can still benefit from it.
Fortunately,
a large amount of lexical data for Taiwanese Hokkien is available as
Open Data. We will propose a method to (semi-) automatically model and
explore the Taiwanese lexicon in search of such strata.
Our method is based on the modelisation of the lexicon as a complex network .
We
will first introduce all needed theoretical aspects of our modelisation
(so that no prior knowledge in graph theory is required to attend this
presentation). Then we will explain how we can rely on graph theory to
create a model of the lexical data as a graph that takes into account
the traditional analysis of 文讀音 and 白讀音. Once the model is created, we
will show how it can be analysed and explored in search of lexical
strata using community detection and advanced visualisation tools.
Actualités et discussions autour du cursus de Linguistique Informatique de l'université Paris Diderot
Affichage des articles dont le libellé est confs. Afficher tous les articles
Affichage des articles dont le libellé est confs. Afficher tous les articles
jeudi 5 mars 2015
Exposé au colloquium du DEC (ENS): Mark Steedman
Mark Steedman, personnalité reconnue en linguistique formelle et en TAL, va présenter des travaux dans le cadre du colloquium du DEC (ENS), mardi prochain 10 mars. L'exposé est intitulé "Statistical Parsing and Interpretation for Jazz Chord Sequences".
Abstract:
The paper describes work with Mark Granroth-Wilding on the problem of parsing and interpreting chord sequences using a grammar of tonal harmony. The grammar is drawn from the same "nearly context-free" class that is strongly adequate for natural languages, and supports a model theory of harmony. Though the grammar is much smaller than language grammars, the degree of ambiguity is much greater, in the sense that any chord can have an unboundedly large number of increasingly preposterous interpretations. Parsing requires the use of a supervised statistical model to limit search, of the same kind that is required for parsing natural language at scale. We evaluate its performance on held-out data against a baseline Markov model, which it outperforms, showing that the grammar is doing useful cognitive work.
http://www.cognition.ens.fr/ColloquiumAgenda.html
Mardi 10 mars, 11h30-13h00, salle Langevin, 29, rue d'Ulm.
Abstract:
The paper describes work with Mark Granroth-Wilding on the problem of parsing and interpreting chord sequences using a grammar of tonal harmony. The grammar is drawn from the same "nearly context-free" class that is strongly adequate for natural languages, and supports a model theory of harmony. Though the grammar is much smaller than language grammars, the degree of ambiguity is much greater, in the sense that any chord can have an unboundedly large number of increasingly preposterous interpretations. Parsing requires the use of a supervised statistical model to limit search, of the same kind that is required for parsing natural language at scale. We evaluate its performance on held-out data against a baseline Markov model, which it outperforms, showing that the grammar is doing useful cognitive work.
http://www.cognition.ens.fr/ColloquiumAgenda.html
mardi 10 septembre 2013
Exposés de Mark Johnson
Un des meilleurs spécialistes mondiaux de linguistique computationnelle, Mark Johnson, va donner deux exposés à l'université Paris Diderot. Tous les étudiants intéressés par le traitement automatique des langues son fortement invités à y assister:
*
*Language acquisition as statistical inference**
**
**Mark Johnson**
**Macquarie University**
**
**noon, 12th September, LingLunch*
Linglunch Paris Diderot
Thursday, 12th septembre 2013
12h-13h, salle 103
bâtiment Olympe de Gouges
(8) rue Albert Einstein, 75013
http://www.linguist.univ-paris-diderot.fr/linglunch.html
This talk argues that language acquisition -- in particular, syntactic
parameter setting -- is profitably viewed as a statistical inference
problem. I discuss some issues associated with statistical inference
that linguists might be concerned about, including the possibility of
"Zombie" parameter settings. The bulk of the talk focuses on estimating
parameters in a Stabler-style Minimalist Grammar framework. Building on
recent results of Hunter and Dyer (2013), we show how estimating weights
associated with lexical entries -- including the empty functional
categories that control parametric syntactic variation -- can be reduced
to estimating weights in what appears to be a new grammar formalism
called "feature-weighted context-free grammars", which is a MaxEnt
generalisation of the "tied context-free grammars" of Headden et al
(2009). Importantly, the partition function and its derivatives of a
feature-weighted context-free grammar can be calculated using a
generalisation inspired by the Inside-Outside algorithm of the
algorithms for calculating partition functions in Nederhof and Satta
(2009). We show how this can be used to learn lexical entries and verb
movement and XP movement parameters in three toy corpora.
*
*
*Grammars and Topic Models**
**
**Mark Johnson**
**Macquarie University**
**
**11am, 20th September, Alpage Group*
Séminaire ALPAGE
Friday, 20th september, 11h-12h30
salle 127
bâtiment Olympe de Gouges
(8) rue Albert Einstein, 75013
https://www.rocq.inria.fr/alpage-wiki/tiki-index.php?page=seminaire
Context-free grammars have been a cornerstone of theoretical computer
science and computational linguistics since their inception over half a
century ago. Topic models are a newer development in machine learning
that play an important role in document analysis and information
retrieval. It turns out there is a surprising connection between the
two that suggests novel ways of extending both grammars and topic
models. After explaining this connection, I go on to describe
extensions which identify topical multiword collocations and
automatically learn the internal structure of named-entity phrases.
These new models have applications in text data mining and information
retrieval.
****
**
*
Cette série d'exposés est financée par:
Research in Paris Programme - Mairie de Paris
Ecole Normale Supérieure
Ecole des Hautes Etudes en Sciences Sociales
Fondation Pierre Gilles de Gennes
*
**
****
*
*Language acquisition as statistical inference**
**
**Mark Johnson**
**Macquarie University**
**
**noon, 12th September, LingLunch*
Linglunch Paris Diderot
Thursday, 12th septembre 2013
12h-13h, salle 103
bâtiment Olympe de Gouges
(8) rue Albert Einstein, 75013
http://www.linguist.univ-paris-diderot.fr/linglunch.html
This talk argues that language acquisition -- in particular, syntactic
parameter setting -- is profitably viewed as a statistical inference
problem. I discuss some issues associated with statistical inference
that linguists might be concerned about, including the possibility of
"Zombie" parameter settings. The bulk of the talk focuses on estimating
parameters in a Stabler-style Minimalist Grammar framework. Building on
recent results of Hunter and Dyer (2013), we show how estimating weights
associated with lexical entries -- including the empty functional
categories that control parametric syntactic variation -- can be reduced
to estimating weights in what appears to be a new grammar formalism
called "feature-weighted context-free grammars", which is a MaxEnt
generalisation of the "tied context-free grammars" of Headden et al
(2009). Importantly, the partition function and its derivatives of a
feature-weighted context-free grammar can be calculated using a
generalisation inspired by the Inside-Outside algorithm of the
algorithms for calculating partition functions in Nederhof and Satta
(2009). We show how this can be used to learn lexical entries and verb
movement and XP movement parameters in three toy corpora.
*
*
*Grammars and Topic Models**
**
**Mark Johnson**
**Macquarie University**
**
**11am, 20th September, Alpage Group*
Séminaire ALPAGE
Friday, 20th september, 11h-12h30
salle 127
bâtiment Olympe de Gouges
(8) rue Albert Einstein, 75013
https://www.rocq.inria.fr/alpage-wiki/tiki-index.php?page=seminaire
Context-free grammars have been a cornerstone of theoretical computer
science and computational linguistics since their inception over half a
century ago. Topic models are a newer development in machine learning
that play an important role in document analysis and information
retrieval. It turns out there is a surprising connection between the
two that suggests novel ways of extending both grammars and topic
models. After explaining this connection, I go on to describe
extensions which identify topical multiword collocations and
automatically learn the internal structure of named-entity phrases.
These new models have applications in text data mining and information
retrieval.
****
**
*
Cette série d'exposés est financée par:
Research in Paris Programme - Mairie de Paris
Ecole Normale Supérieure
Ecole des Hautes Etudes en Sciences Sociales
Fondation Pierre Gilles de Gennes
*
**
****
lundi 24 septembre 2012
Talk @ #ParisJS: bibliothèque Fullproof
Un exposé dans le cadre de ParisJS (ParisJS est un user-group Javascript qui se réunit tous les mois, et présente les dernières technologies Web/JS du moment), présenté par un ancien du cursus LI, à propos d'une bibliothèque open-source de recherche full text, Fullproof.
Fullproof est une bibliothèque open-source fournissant une pile complète dédiée à la recherche plein texte (fulltext) en javascript, utilisant le stockage local lorsqu'il est disponible, intégrant une gestion de l'unicode, des algorithmes de lemmatisation et de simplification phonétique, et permettant ainsi des recherches textes de très bonne qualité pour les applications déconnectées. Je présenterai les enjeux, les avantages, et les limitations actuelles de cette solution.
Ca peut intéresser des étudiants/personnes du cursus, s'ils ont un petit bagage minimum en javascript (sinon ça risque d'être un peu abstrait, mais pourquoi pas).
La conférence est gratuite, mais il faut s'inscrire sur le site du user group pour avoir une place: http://parisjs.org/
Fullproof est une bibliothèque open-source fournissant une pile complète dédiée à la recherche plein texte (fulltext) en javascript, utilisant le stockage local lorsqu'il est disponible, intégrant une gestion de l'unicode, des algorithmes de lemmatisation et de simplification phonétique, et permettant ainsi des recherches textes de très bonne qualité pour les applications déconnectées. Je présenterai les enjeux, les avantages, et les limitations actuelles de cette solution.
présenté par Rodrigo Reyès
Ca peut intéresser des étudiants/personnes du cursus, s'ils ont un petit bagage minimum en javascript (sinon ça risque d'être un peu abstrait, mais pourquoi pas).
La conférence est gratuite, mais il faut s'inscrire sur le site du user group pour avoir une place: http://parisjs.org/
lundi 17 septembre 2012
Séminaires Edward Gibson
Linguistique expérimentale, théorie de l'information, ça n'est pas vraiment du TAL, mais cela devrait intéresser les linguistes-informaticiens, et c'est une grande chance que de pouvoir bénéficier de la présence de ce grand spécialiste, invité par l'EFL et LLF pendant ce semestre.
| Le Labex EFL est heureux d'annoncer le cours de Edward Gibson, MIT, Brain and Cognitive Sciences, qui commence lundi, 17 septembre, au LLF, Paris Diderot, Salle 4C92. Cours tous les lundis du 17 septembre au 10 décembre, 16-18 h 175 rue du Chevaleret, 4e étage, salle 4C92 Résumé The proposed seminar is inspired by Shannon's (1948) seminal work on information theory and will focus on communicative properties of language representations and processing. Three broad questions will be addressed: 1. How do communicative pressures shape the lexicon? 2. How do communicative pressures shape the syntax of languages? 3. How do communicative properties of language affect interpretation of the linguistic signal? In the first section of the course I will discuss Zipf's (1949) original ideas on the relationship between word length and frequency and recent findings that extend those ideas and show that word length is better explained by how predictable a word is in the context in which it appears than by word frequency. I will then talk about lexical ambiguity and present evidence that ambiguity is not a weakness of the language system, as has been argued by e.g., Chomsky. Instead, ambiguity is a desirable feature of any communication system because it allows for a more efficient lexicon. In particular, the same 'easy' (e.g., short or easily pronounceable) words can be reused in the language, referring to different things, because context can and does robustly disambiguate the intended meaning. I will also discuss ambiguity in the context of word learning. In the second section of the course I will focus on language structure. A recent paradigm, in which participants are asked to gesture the meanings of simple events (Goldin-Meadow et al., 2008), has been shown to reflect word-order preferences somewhat independent of native language and is thus promising for revealing fundamental properties of our communication system and perhaps shedding light on how different word orders arose in the course of language evolution. For example, when shown a picture or animation of a boy kicking a ball, speakers of languages with either a subject-verb-object (SVO) order or a subject-object-verb (SOV) order gesture the event as subject-object-verb (i.e., 'boy', 'ball', 'kick'). This finding has been interpreted as evidence for SOV order being somehow fundamental or ?base? in how we think about events. Critically, however, we have shown that when asked to gesture meanings of 'reversible' events (e.g., a boy pushing a girl), where either entity can be the agent or the patient of the action, the gesture sequences shift to the SVO order, across languages. This finding fits well with the idea that communication takes place over a noisy channel, where the signal may get corrupted in some way. The use of SVO gesture sequences for reversible events makes the meaning more robust to noise thus maximizing the chances of the comprehender recovering the intended meaning. I will further connect these experimental findings to typological patterns across languages. The noisy-channel hypothesis makes several attested predictions about cross-linguistic variation, such as prevalence of casemarking in SOV languages, and lack of case-marking in SVO languages, suggesting that a shift from SOV to SVO word order makes the language sufficiently robust to noise. However, if a language keeps its base SOV word order, then another 'device' is needed to make the signal more robust to noise, and case-marking is one way of marking the agent and patient of an event. In the third section of the course I will talk about how communicative properties of the language system affect the interpretation of the linguistic signal. Until recently, it has been assumed that the input to the sentence comprehension mechanism is an error-free sequence of words. However, given that noise (speaker errors, perception errors, or noise in the environment) can corrupt the linguistic signal, as discussed above, the human sentence comprehension mechanism is plausibly adapted to process sentences that contain errors. A noisy-channel model predicts that the interpretation of a corrupted linguistic signal will depend on how easily recoverable the plausibly intended signal is. If the comprehender can easily explain how the original signal got corrupted, then s/he will rely on semantic cues for interpretation. If, on the other hand, there is no clear explanation for how the signal got corrupted, then the comprehender is predicted to more closely follow the actual syntactic form of the utterance in deriving an interpretation. These predictions are in contrast to those of some earlier models, according to which the final interpretation is always determined by the syntax of an utterance (cf. MacWhinney et al., 1984), even though meaning may guide initial interpretation in the face of temporary syntactic ambiguity. I will discuss several recent sets of findings in sentence comprehension that provide support for the noisy-channel model as well as re-interpret some classical findings in the context of this model. |
lundi 2 avril 2012
Séminaire: Roser Saurí (linglunch)
Le prochain exposé du linglunch Paris Diderot est recommandé aux étudiants LI: il sera donné par Roser Saurí (chercheuse au `Voice and Language Group', Barcelona Media), actuellement invitée au laboratoire Alpage.
jeudi 5 avril, 12:00 – 13:00
UFR Linguistique. 175, rue du Chevaleret, 75013. Salle 4C92 (4e étage) (plan)
Words Talking Factually. Lexical Resources for Event FactualityWhenever we talk about situations, we express our degree of certainty about their factual status. We characterize them as unquestionable facts, or qualify them with some degree of uncertainty if we are not sure whether the situation holds, or will hold, in the world. Identifying the factuality status of event mentioned in discourse is fundamental for reasoning about eventualities. Inferences derived from events judged as not having happened, or as being only possible, are different from those derived from events evaluated as factual.Factuality is expressed through a complex interaction of different aspects in the overall linguistic expression. It involves polarity (events can be presented as positive or negative) as well as epistemic modality, which expresses the degree of certainty of a source regarding what is asserted. Other information at play is evidentiality (e.g., a seen event is presented with a factuality degree stronger than that of an event reported by someone else), and mood (e.g., indicative vs. subjunctive). Factuality is also a component in the semantics of certain types of predicates, most notoriously the so-called factive and implicative predicates, among many others. In this talk, I will present the effort of designing and compiling lexical resources for event factuality in English, and exemplify its practical usage in a tool for automatically identifying the factuality degree of eventualities in discourse. Together with Prof. Laurence Danlos, we are currently exploring venues for creating an equivalent resource for French. |
dimanche 18 mars 2012
Séminaires: Linguistique computationnelle du Sanskrit
Monsieur Peter Scharf,
titulaire d'une chaire Blaise Pascal
et hébergé (2012-13), à ce titre,
par le laboratoire d'histoire des théories linguistiques (CNRS, Paris
7, Paris 3), donnera une série de
conférences - en binôme avec un
autre intervenant - sur le thème de
la linguistique computationnelle du
sanskrit.
[Edit (09/05): le programme mis à jour est désormais disponible à la page
http://htl.linguist.univ-paris-diderot.fr/scharf.htm
à noter aussi un exposé dans le cadre du linglunch Paris Diderot, le jeudi 31 mai à 12:00.]
Researching the linguistic treatises of India to enrich formal and computational linguistics and utilizing techniques of contemporary linguistics to formalize Indian linguistic theories
Programme complet des interventions (dates et lieux precisés ultérieurement)
[Edit (09/05): le programme mis à jour est désormais disponible à la page
http://htl.linguist.univ-paris-diderot.fr/scharf.htm
à noter aussi un exposé dans le cadre du linglunch Paris Diderot, le jeudi 31 mai à 12:00.]
Peter Scharf, Université Paris Diderot,
"Introduction to Sanskrit Computational Linguistics"
Researching the linguistic treatises of India to enrich formal and computational linguistics and utilizing techniques of contemporary linguistics to formalize Indian linguistic theories
Programme complet des interventions (dates et lieux precisés ultérieurement)
- Linguistic Issues in the Character-Encoding of Sanskrit: clarifying the axes of information encoding: graphic-phonetic, sequential-featural, and contrastive- non-contrastive.
- Vedic recitation: The relationship between the description in phonetic treatises, manuscripts, and contemporary oral recitation.
- Indic lexical resources: The historical evolution of root lists (dhâtupâ.tha) and their relation to rules in Indian grammars.
- Ancient Indic versus modern European morphology: Linguistic perspectives on ontology.
- Non-linear Sanskrit syntax: Developing language-neutral syntactic representation.
- Indian semantic theory: Competing perspectives on speech cognition in grammar (vyâkara.na), logic (nyâya), ritual exegesis (karmamîmâ.msâ).
- Indian ontology and aesthetics: The categorization of what is, and the human experience of it.
- Developing interchange formats and interoperability protocols for computational linguistics of Sanskrit and other Indic languages.
vendredi 16 mars 2012
Séminaire: Petit guide du linguiste empirique
Dans le cadre du séminaire 2011-2012 du laboratoire "Histoire des Théories Linguistiques", intitulé « Expérience, empiricité, expérimentation en linguistique », les étudiants de L3 et de M2 (ceux de M1 ont malheureusement cours avec moi...) pourront sûrement tirer profit de la prochaine conférence, prévue le mardi 20 mars 2012, de 16h00 à 18h00:
Barbara Hemforth (& Jean-Marie Marandin)
Université Paris Diderot, laboratoire LLF
Un petit guide du linguiste empirique :
les questionnaires
Université Paris Diderot,
Bâtiment Lamarck,
rue Marie-Andrée Lagroua Weill-Hallé,75013 Paris
Salle RB15, RdC.
Inscription à :
Articles (Atom)