Search

We present SEJF, a lexical resource of Polish nominal, adjectival and adverbial multiword expressions. It consists of an intensional module with about 4,700 multiword lemmas assigned to 160 inflection graphs, and an extensional module with 88,000 automatically generated inflected forms annotated with grammatical tags. We show the results of its coverage evaluation against an annotated corpus. The resource is freely available under the Creative Commons BY-SA license.

International audience; Parsing efficiency within the context of tree adjoining grammars (TAGs) depends not only on the size of the input sentence but also, linearly, on the size of the input TAG, which can attain several thousands of elementary trees. We propose a factorized, finite-state TAG representation which copes with this combinatorial explosion. The associated parsing algorithm substantially increases the parsing performance on a real-size French TAG grammar.

The computational treatment of multilingual language resources (as in the Prolex projet, cf (Grass et al. 2002)) should respect lexical conventions admitted by each language’s native speakers. These conventions may vary from one language to another, as in the case of alphabetical sorting algorithm. This algorithm must take a number of universal as well as language-dependent particularities into account, such as the distinction of upper- and lowercase letters, the sorting bi-directionality (from the left to the right or conversely), the role of diacritics (resulting either in variants of a letter, as é, è and ê in French, or in independent letters, as å in Danish or ą in Polish), the role of punctuation characters, the multi-character letters (as ch or ll in Spanish, or dzs in Hungarian) and the ligatures (as œ in French, or ß in German). We describe a Unicode-based sorting algorithm inspired by (LaBonté 1998) for proper names. In the particular case of the sorting of the proper names, three additional points are to be taken into account : the presence of numerical values (Arab numerals or Roman numerals), the variation of spelling of the ligatures and the permutation in the sorting of the multi-word units. Apart from the word list to be sorted, its input is a language-dependent code table which defines the language’s alphabet, the number of algorithm’s passes, the direction of each pass, and the order of letters or groups of letters in each pass. The implementation of the algorithm is done by a finite-state transducer which allows a fast assignment of sort keys to words. The algorithm proved correct for European languages such as English, French, and Polish, as well as for Thai. It outperforms other sorting algorithms, such as those implemented in Intex (Silberztein 1993) and Unitex (Paumier 2003) systems.

International audience; This paper reports on linguistic features and decisions that we findvital in the process of annotation and resolution of coreference for highly inflec-tional languages. The presented results have been collected during preparation ofa corpus of general direct nominal coreference of Polish. Starting from the notionof a mention, its borders and potential vs. actual referentiality, we discuss theproblem of complete and near-identity, zero subjects and dominant expressions.We also present interesting linguistic cases influencing the coreference resolutionsuch as the difference between semantic and syntactic heads or the phenomenonof coreference chains made of indefinite pronouns.

The ANR EmotiRob project aims at detecting emotions in an original application context: realizing an emotional companion robot for weakened children. This paper presents a system which aims at characterizing emotions by only considering the linguistic content of utterances. It is based on the assumption of compositionality: simple lexical words have an intrinsic emotional value, while verbal and adjectival predicates act as a function on the emotional values of their arguments. The paper describes the semantic component of the system, the algorithm of compositional computation of the emotion value and the lexical emotional norm used by this algorithm. A quantitative and qualitative analysis of the differences between system outputs and expert annotations is given, which shows satisfactory results, with the right detection of emotional valency in 90% of the test utterances.

Minimal perfect hashing provides a mapping between a set of n unique words and n consecutive numbers. When implemented with minimal finite-state automata, the mapping is determined only by the (usually alphabetical) order of words in the set. Addition of new words would change the order of words already in the language of the automaton, changing the whole mapping, and making it useless in many domains. Therefore, we call it static. Dynamic minimal perfect hashing assigns consecutive numbers to consecutive words as they are added to the language of the automaton. Dynamic perfect hashing is important in many domains, including text retrieval and databases. We investigate three methods for its implementation.

Pseudo-minimal automata ([1],[2]) are minimal acyclic automata that have a proper element (a transition or a state) for each word belonging to the language of the automaton. That proper element is not shared with any other word, and it can be used for implementing a function on words belonging to the language. For instance, dynamic perfect hashing (e.g. a mapping from n unique words to n consecutive numbers, such that addition of new elements does not change the order of the previous elements) can be implemented using a pseudo-minimal automaton ([3]).

A description of a system for identifying Verbal Multi-Word Expressions (VMWEs) in running text is presented. The system mainly exploits universal syntactic dependency features through a Conditional Random Fields (CRF) sequence model. The system competed in the Closed Track at the PARSEME VMWE Shared Task 2017, ranking 2nd place in most languages on full VMWE-based evaluation and 1st in three languages on token-based evaluation. In addition, this paper presents an option to re-rank the 10 best CRF-predicted sequences via semantic vectors, boosting its scores above other systems in the competition. We also show that all systems in the competition would struggle to beat a simple lookup base-line system and argue for a more purpose-specific evaluation scheme.