Programme (October 2006 - May 2007)

Recent efforts to improve the quality of automated Question Answering (QA) have targetted such areas as

• better understanding of questions in order to better retrieve and assess answer candidates;
• richer indexing of corpora to reduce false positives among answer candidates;
• creating off-line resources to answer particular types of common questions (a return to database QA!);
• greater use of semantics to prune answer candidates;
• methods for re-ranking answer candidates -- e.g., by assessing the likelihood that an answer is correct.

But with few exceptions, answer candidates have been analysed and assessed independently of each other, as if one candidate was simply correct (or several, in the case of list questions) and the others, simply incorrect. What is actually the case, however, which is discernable from both TReC answer patterns and analysis of the answer candidate sets themselves, is more complex. Our research has shown that recognizing and assessing patterns of relationships between answer candidates allows significant gains for such QA-related tasks as answer candidate re-ranking and question and answer disambiguation. Such ``answer models'' also holds promise than do sets of answer strings, for supporting more cooperative or personalised answers to questions and the emerging area of interactive QA.

In this talk, I will describe this work, carried out by Tiphaine Dalmas as part of her PhD dissertation research.

In the seminar I intend to present the main criteria followed in the planning and partial realization of the ALTR - Archivio lessicale dei dialetti trentini, a lexical data-bank (whose elaboration has already provided about 35,000 records), based on the different linguistic varieties (Lombard, Venetian, Ladin) spoken in Trentino, in Northern Italy. The corpus consists in twenty dialectal dictionaries concerning thirteen linguistic areas of the province. A standard form of record is proposed, to show how the new electronic hyper-dictionary maintains the original data of the sources and, at the same time, how it gives uniformity to their presentation. Three fields of information that raise interesting questions are briefly discussed: phonetical transcriptions, italian translations of the dialectal lemmata, and semantic domain indicators.

This paper proposes a combined syntax and semantics for intonation in English. The semantics is surface-compositional under the generalized definition of surface-syntactic derivation afforded by Combinatory Categorial Grammar (CCG, see \citealt{Stee:99}, hereafter {\em SP}). This theory of grammar unites intonation structure and information structure with syntactic structure and Montague-style compositional semantics, even when intonation structure deviates from traditional surface structure. It revises and extends earlier accounts of this kind, grounding semantic notions like ``theme'' and ``rheme'' (a.k.a. ``topic'' and ``comment,'' ``presupposition'' and ``focus,'' etc.) in a logic of speaker/hearer supposition and update, and eliminating overgeneration in the CCG syntax. I will pay some attention to some well-known differences between English and the intonational system of Italian.

We propose an approach for extracting relations between entities from natural-language documents. The approach is based solely on shallow linguistic processing, such as tokenization, sentence splitting, Part-of-Speech tagging and lemmatization. It uses a combination of kernel functions to integrate two different information sources: (i) the whole sentence where the relation appears, and (ii) the local contexts around the interacting entities. We present experiments on extracting five different types of relations from a data set of newswire documents and show that each information source provides a useful contribution to the recognition task. The experiments were performed using both the ``correct'' Named Entities (i.e. those manually annotated in the corpus) and the ``noisy'' Named Entities (produced by a Machine Learning based Named Entity recognizer). The results obtained show that our approach is portable across domains and significantly improves previous results both in ideal and in realistic/noisy conditions.

The language variability problem is well known in Computational Linguistics. A general unifying framework - textual entailment has been proposed recently. The problem is addressed by defining entailment as relation between two language expressions text (T) and hypothesis(H) that holds if the meaning of H, as interpreted in the context of T, can be inferred from T. The Recognizing Textual Entailment (RTE) task takes as input T - H pair and consists in automatically determining whether an entailment relation holds between T and H or not.

We designed a RTE system based on the intuition that the probability of an entailment relation between T and H is related to the ability to show that the whole content of H can be mapped into the content of T. The more straightforward the mapping can be established, the more probable is the entailment relation. Since a mapping can be described as the sequence of editing operations needed to transform T into H, where each edit operation has a cost associated with it, we assign an entailment relation if the overall cost of the transformation is below a certain threshold, empirically estimated on training data. The core engine of our approach is an implementation of a tree edit distance algorithm applied to the syntactic representations (i.e. dependency trees) of both T and H.

After discussing some of the reasons why corpus linguistics is still awaiting the break-through as an instrument for language didactics, basic strategies of language learning activities in their connection to corpus queries will be discussed. Examples will come from Italian and German, the two languages studied not as "foreign", but as "second languages" in South Tyrolian schools. Corpus queries will be presented as innovative tools fostering self-instructed learning in settings of collaborative learning environments. Aspects of language learning in terms of strategies regarding the learning process as well as corpus searching strategies will be discussed in some detail.
Key-words:
The core-periphery-discussion in recent linguistics; language teaching and linguistics; construction grammar; formulaic language; corpus tools and language learning; affordance in regard to language, language teaching and corpus queries.

The advances in disciplines such as audio and visual scene analysis, speech recognition and understanding, haptics, etc., have made it possible to start addressing the problem of fusing together the information produced by different sensory channels into higher-level descriptions of what is going on in complex environments. An important application is to human behaviour, broadly conceived: researchers can now investigate to what extent simple and complex behavioural sequences can be automatically detected and analysed. Both individual and group/social behaviour are being actively investigated, with examples such as the automatic recognition of: social roles, of complex events performed by humans, of facial and bodily expressions and posture to infer internal states (emotions and moods, etc.), and so on. Application areas are also many, ranging from automatic support to group productivity (coaching, facilitation, multimedial minutes), to support to individuals in their natural environment (e.g., the elderly).

In this talk we will discuss: the motivation for such efforts and the scientific challenges, focusing on the level where the sensorial information are to be fused into higher level concepts. We will also discuss psychological and social implications of these systems, trying to understand the impact of this new scenario on dimensions such as trust and intrusion, usability and acceptability, etc.

We discuss the motivation and possible applications for a corpus of South Tyrolean written German, and focus on the practical aspects of corpus construction. We describe how we select texts, and how we work with publishers to obtain the data. We receive the data in many forms, often on obsolete media and in archaic or idiosyncratic file formats; our daunting task is to somehow convert these multifarious data to a standardized XML form. Working out a solution for each kind of data involves a good bit of detective work and ingenuity. The case studies in this talk will illustrate the detective tools that we use, and the sorts of practical problems which one encounters in the conversion phase of corpus development. We also discuss the challenges concerning text metadata; we must develop a metadata scheme which is as complete as possible while also not posing an unrealistic burden on those assembling the information. We conclude with a description of the process of linguistic annotation and construction of a corpus query system.

Sign languages are visual languages used in deaf communities, mainly. They are essentially tempo-spatial languages: signs are made of manual components, e.g., hand movements, and non-manual components, e.g., facial expressions. In this presentation, we overview the work done at FUB and EURAC on e-LIS, a project lead by EURAC for Italian sign language (LIS, Lingua Italiana dei Segni). More precisely, we focus on the web dictionary from LIS to verbal Italian. We start surveying several electronic dictionaries for sign languages (SLs). In general, their consultation requires that the users should be expert of a specific transcription system for SLs. On the contrary, the e-LIS dictionary does not require this of users; with the help of an ontology representing the way signs are composed, the dictionary interface should expertly guide users in "recomposing" the sign they are looking up for. We conclude this presentation by introducing an offspring of e-LIS, namely, LODE: a LOgic-based e-tool for DEaf children. LODE is conceived as a web interactive tool. The users of LODE are Italian deaf children with a scarce mastery of verbal Italian. With the help of automated constraint-based reasoners, LODE aims at tackling comprehension issues related to global causal relations in narratives written in verbal Italian.

Parthood constructions are quite frequent in natural language. In English, a variety of expressions (is part of, is a component of, is a portion of...) and many genitive constructions refer to parthood relations. Although linguistic and psycholinguistic studies have stressed the fact that these expressions refer to several different parthood relations, there is no general consensus on what these relations are from a semantic point of view. Accordingly, in computational linguistic more often than not a generic, only shallowly formally characterized, "part-of" relation is used. On the other hand, Mereology, the study of parthood, is the central pillar of formal ontology. But formal ontology points at the general character of this relation, not at its diversity. So the range of mereologies available do not really address the semantic variations observable in language, they correspond to diverse answers on fundamental ontological choices such as mereological extensionality (two entities sharing the same parts are identical). Nevertheless, the fields of formal semantics and formal ontology, that have benefited from each other in some occasions (for instance in the temporal domain), appear to meet again on the controversial issue of the transitivity of the most prominent parthood relation, functional parthood (the carburetor is part of the engine, my finger is part of my hand). In this talk I will present an analysis of functional parthood that both uses formal ontology tools (mereology, dependence) and takes language seriously. This proposal completes an earlier one in which the main parthood relations are given a formal account. It also gives an explanation of the apparently chaotic distribution of transitivity and non-transitivity patterns, such as:

• The cuff is part of the sleeve, the sleeve is part of the jacket, and the cuff is part of the jacket.
• The handle is part of the door, the door is part of the house, but the handle is not part of the house.

The availability of large language corpora makes the study of linguistic phenomena more and more possible also on a quantitative level. In the domain of language contact the analysis of texts enables the researcher to folllow the process of integration of foreign elements from variable use in discourse to actual integration into the phonological and morphosyntactic frame of a language. In this sense corpus analysis can integrate the more traditional research on dictionaries and also help to quantify more realistically the actual component of the foreign component in a language. In the presentation, examples of English loanwords into Italian will be discussed on the basis of data taken from dictionaries and from corpora. Finally, the case study of a German minority dialect surrounded by several Romance varieties will be considered, too.

The aim of this talk is to discuss the notion of "restrictive" vs. "non restrictive" modification, particularly in its application to attributive adjectives. In Romance languages, many adjectives can have a pre- and a post-N position, where the post-N position is characterized as "restrictive" and the pre-N as "non-restrictive" . In other cases (e.g. 2a,b) "restrictive/non restrictive" does not seem to be the right dimension of variation, and the distinction seems to be more connected with focus or information structure. In yet other cases we have quasi systematic meaning changes (3)

1) a. Una sorgente fresca.
b. Una fresca sorgente.

2) a. Il presidente attuale.
b. L' attuale presidente.

3) a. Un (vero) documento (vero)
b. Un (discreto) maggiordomo (discreto)
c. Le (numerose) famiglie (numerose)
d. Quel (grand) uomo (grande)

In the talk I will address questions such as: (i) whether the notion of "(non) restrictive" can always be reanalyzed in terms of focus/information structure (ii) how it interacts with (in)definiteness and demonstrativity (since in e.g. (4) the whole nominal restriction, "vehicle" seems to be non restrictive, and superfluous for the purpose of fixing reference), and in case (iii) which theories of definiteness can be best adapted to deal with these cases.

(4) That (vehicle) is a car.

In my talk I will introduce the field of machine translation and present our recent research activities and results on spoken language translation. I will\240 describe advances in the use of confusion networks as interface between automatic speech recognition and machine translation. In particular, I will presents a decoding algorithm for confusion networks which results as an extension of a state-of-the-art phrase-based text translation decoder.
The confusion network decoder significantly improves both in efficiency and performance previous work along this direction, and outperforms the background text translation system. Experimental results in terms of translation accuracy and decoding efficiency are reported on two tasks, namely the translation of Plenary Speeches at the European Parliament, from Spanish to English, and the translation of speech utterances in a tourism domain from Chinese to English. Finally, I will talk about our involvement in Moses, an open source project that is likely the most popular translation engine used within the scientific community.

The presentation discusses two sociolinguistic projects applying quantitative and qualitative approaches to the study of language use and multilingualism in South Tyrol.
The first part of the talk (Patterns of language choice. Asking for directions in South Tyrol) discuss some results of a sociolinguistic survey on language choice carried out in South Tyrol. Real interactions between a fieldworker and passers-by have been collected, recorded and analyzed. These verbal dialogues are focused on the request for street directions to pedestrians in Bozen and Brixen. The main goal of the research is to isolate the factors affecting the use of dialect in situations where language choice is uncertain and not completely predictable a priori. A logistic regression has been performed on a sample of 700 of such short verbal interactions, aiming at understanding under what circumstances dialect (the local choice) is selected against “colloquial German”, Standard German (Hochdeutsch), or Italian.
The second part (Linguistic representations and conceptual metaphors: examples from language biographies in South Tyrol), connecting language biography studies - which has emerged in the last decade as a new field of research and which has proved its fruitfulness when investigating the relationship speakers establish with languages during their lives - with conceptual metaphor theory, presents the analysis of eight South Tyrolean speakers’ narratives in terms of metaphors used to talk about standard languages, language varieties, language learning and use. It is thus discussed how conceptual metaphor theory can be relevant for the investigation of speaker’s representations of linguistic phenomena and experiences and provide a further tool of analysis within language biography studies.

This paper presents an application of PageRank, a random-walk model originally devised for ranking Web search results, to ranking WordNet synsets in terms of how strongly they possess a given semantic property.
The semantic properties we use for exemplifying the approach are positivity and negativity, two properties of central importance in sentiment analysis.
The rationale of applying PageRank to detecting the semantic properties of synsets lies in the fact that the space of WordNet synsets may be seen as a graph, in which synsets are connected through the binary relation "a term belonging to synset s_k occurs in the gloss of synset s_i". The data for this relation can be obtained from eXtended WordNet, a publicly available sense-disambiguated version of WordNet. We argue that this relation is structurally akin to the relation between hyperlinked Web pages, and thus lends itself to PageRank analysis. We report experimental results supporting our intuitions. (joint work with Andrea Esuli)

Recommender systems are intelligent applications that assist the users in a decision-making process by giving personalized product recommendations. Various techniques have been proposed in the literature but all of them reduce to the prediction of "what" will be the products or information most liked by the user. Constrained by this model, recommendation technologies have largely ignored the importance of the recommendation "process", i.e., the characteristics of the human/computer interaction and how the recommendation process itself, and not only the final recommendations, can influence the user acceptance of the system suggestions. In this talk, I will present a new type of recommender system capable of learning an interaction strategy by observing the consequences of its actions on the user and on the final outcome of the recommendation session. We view the recommendation process as a sequential decision problem, and we model it as a Markov Decision Process. I will present the results of some computer simulations and the application of this technology to the design of a recommender system for a major tourism web portal.

It has been long believed that Categorial Grammar (CG) offers a particularly attractive view on natural language syntax-semantics interface. This belief was strengthened by the discovery of the Curry-Howard Correspondence between Lambek Calculus (NL, the type logic version of CG) and intuitionistic lambda calculus. However, big disappointment came when people start realizing that NL fails to handle discontinuos phenomena and in-situ binding, namely those phenomena where the syntax-semantic interface is more challenged. Many have been the attempts to increase the expressivity of NL to overcome its limitation, all of them stayed within the regime of intuitionistic logic in a way or another, and still proved to be disappointing. In this talk, after giving an introduction to the Logic approach to natural language analysis, I will present new research lines, within the Lambek framework, inspired by Programming Languages. I will show that by moving towards classical logic, namely to a Symmetric Calculus (Lambek and Bi-Lambek Calculi together with Grishin interaction postulates, LG), the right flexibility to capture Natural Language in-situ binding is reached. As a consequence of the move from NL to LG, the semantic side is extended to the lambda mu calculus following Curien & Herbelin 2000/2005 and the Continuation Passing Style Transformers there proposed. We discuss the differences with respect to natural language interpretation obtained by using Call by Name vs. Call by Value evaluation strategies, and in particular the limitations of the latter with respect to the former when quantifier phrases are investigated.

Integration of new utterances into context is a central task in any model for rational (human-machine) dialogues in natural language. In this talk, a pragmatics-first approach to specifying the meaning of utterances in terms of plans is presented. These plans are computed during a dialogue on the basis of information about the current situation that is updated continually. A rational dialogue is driven by the reaction of dialogue participants on how they find their expectations on changes in the environment satisfied by their observations of the outcome of performed actions. We present a computational model for this view on dialogues and illustrate it with examples from a real-world application.

In this talk I will present our efforts in terms of building a bridge between an Intelligent Query Interface and Natural Language Generation technologies. The current version of our query interface enables users to access data sources by means of an ontology representing the knowledge of a domain in a well defined formal semantics. The main challenge we are facing now is that the underlying conjunctive query is to be presented to the user in natural language.
I will explain the steps needed to translate a conjunctive query into a text plan, focusing primarily on the problem of maximizing the local referential coherence of the natural language query that will be generated given the text plan. This is done by seeing this problem as a topological sort of a directed acyclic graph (the query), more precisely finding the linear ordering of the predicates that will maximize the local coherence and therefore the readability of the generated text, according to the coherence measures offered by Centering Theory.

Patents are complex multimedia objects which are searched for and closely scrutinized by hundreds of professionals on a daily basis, with the aim of detecting patent infringements, but also to assess the state of the art of specific technology areas. For these reasons, patents are receiving increasing attention from both the text and image processing research communities. Two European projects (WHISPER and PatExpert) have been funded with the aim of exploring advanced techniques for the automatic analysis of patents. Patent retrieval is one of the tasks proposed in the series of NTCIR evaluation workshops. The present seminar will show some results obtained from the on-going PatExpert project, which is trying to shift the paradigm currently followed for patent processing from textual (patents as sequences of morpho-syntactic units, or at best syntactic structures) to semantic (patents as multimedia knowledge objects), in the framework of the Semantic Web technologies. We will present two approaches to content distillery from patent documents. A shallow approach based on key-concepts extraction, and a deep approach based on relation extraction, ontology learning and ontology population. We will also show that knowledge extracted from text can be useful for image processing.

The talk will present the approach to Natural Language Interpretation adopted in the HOPS European Project. This is one of the first attempts to use full parsing and semantic analysis in a practical domain. The parser produces a dependency tree and is the same parser used for full text processing during the development of the Turin Treebank. It has been extended to cover the four languages involved in the project, i.e. Catalan, English, Italian and Spanish. The semantic analysis is based on an ontology that has been developed for describing the application domains of the project. The result of the interpretation is a semantic query that expresses, in terms of the concept in the ontology, the contents of the user's request.

In this talk I introduce some first explorations of the connection between lexical semantics and emotions.

In the first part of the talk I will speak about our experience in computational humor. Although deep modeling of humor in all of its facets is not something for the near future (humor is one of the most sophisticated forms of human intelligence), some steps can be followed to achieve results. Regarding humor production, I will introduce HAHAcronym, a system developed in the context of a European project. The goal is this system that makes fun of existing acronyms, or, starting from concepts provided by the user, produces a new funny acronym. Regarding verbal humor recognition, I will show that automatic classification techniques, through experiments performed on very large data sets, can be effectively used to distinguish between humorous and non-humorous texts, with significant improvements observed over apriori known baselines.

In the second part of the talk I will talk about emotion recognition in texts. The automatic detection of emotion in texts is becoming increasingly important from an applicative point of view. Consider for example the tasks of opinion mining and market analysis, affective computing, NLP and user-interface (e.g. e-learning environments, such as educational/edutainment games). All words can potentially convey affective meaning. Each of them, even those more apparently neutral, can evoke pleasant or painful experiences. While some words have emotional meaning with respect to the individual story, for many others the affective power is part of the collective imagination (e.g. words ``mum'', ``ghost'', ``war'' etc.). Therefore, it is interesting to individuate a way to measure the affective power of a generic term, through the study of the use of words in textual productions.

Languages share many aspects of argument structure organisation, generally preferring to place given, prominent or otherwise salient information towards the start of the clause, and new information towards the end. They also mirror the causal chain in tending to place active participants before and acted-upon participants after the verb. The differences emerge in how these two tendencies are resolved when they come into conflict - hence the use of constructions such as the passive (English, German, Spanish), middle voice (German, Spanish), inchoatives (English, German), detransitives (Spanish) and ba/bei structures (Chinese). These tendencies are also reflected in the various argument structures seen within and between languages for verbs that describe situations without any clearly dominant direction of causation (such as psych situations, or the balanced situations of "be same", "contain", "meet"). Here I present a simple linear model based on correlations with syntactic, semantic and pragmatic features suggested by the theories of Charles Fillmore, Edward Keenan, Susumo Kuno and David Dowty among others. The relative contributions of each factor are derived from annotation exercises presented to over 20 linguists. The model is shown to make predictions that are broadly consistent with results from a larger acceptability experiment performed with non-linguists.

Nowadays, databases are the standard technology for storing vast amount of data about a variety of domains. This constantly increasing amount of knowledge cannot be easily accessed by users with no competence of formal query languages (such as SQL). For this reason, the development of so called Natural Language Interfaces to Databases (NLIDBs) seems to be a crucial, yet very challenging, task. The approach I am going to introduce is about the creation of linguistic and ontological resources to perform database retrieval with requests in natural language. Adomain ontology is used as a knowledge representation interface between the logical structure of the DB and the natural language semantics. In a NLIDB built on the basis of this approach the ontology model is intended to represent an abstract description of the DB structure and provide, at the same time, the "conceptual" level on which natural language queries are interpreted. The lecture will focus on the design features of the ontology and the way it interacts with an automatically derived linguistic representation of natural language queries so as to produce logical forms to be eventually translated into SQL.

I'll describe a dialogue sub-genre which we have information-seeking , which is distinguished from other kinds of information-seeking dialogue (e.g. travel information) by its more exploratory and less (single) task-oriented nature.
I'll present an approach to modelling this kind of dialogue, based on the notion of weighted structures topic a single data structure that represents both the domain knowledge and the dialogue history, and sketch an implementa- tion of this approach in a typed (i.e., non-speech) dialogue system.
Reference: Stede, M. and David Schlangen [2004] Information-Seeking Chat: Dialogues Driven by Topic-Structure, in Proceedings of Catalog '04 (The 8th Workshop on the Semantics and Pragmatics of Dialogue, SemDial04), Barcelona, Spain, July. Available from: http://www.ling.uni-potsdam.de/~das/papers/stedeschlangen_final.pdf