Darja Fišer

Combining Multiple Resources to Build Reliable Wordnets

This paper compares automatically generated sets of synonyms in French and Slovene wordnets with respect to the resources used in the construction process. Polysemous words were disambiguated via a five-language word-alignment of the SEERA.NET parallel corpus, a subcorpus of the JRC Acquis. The extracted multilingual lexicon was disambiguated with the existing wordnets for these languages. On the other hand, a bilingual approach sufficed to acquire equivalents for monosemous words. Bilingual lexicons were extracted from different resources, including Wikipedia, Wiktionary and EUROVOC thesaurus. A representative sample of the generated synsets was evaluated against the goldstandards.

Leveraging Parallel Corpora and Existing Wordnets for Automatic Construction of the Slovene Wordnet

The paper reports on a series of experiments conducted in order to test the feasibility of automatically generating synsets for Slovene wordnet. The resources used were the multilingual parallel corpus of George Orwells Nineteen Eighty-Four and wordnets for several languages. First, the corpus was word-aligned to obtain multilingual lexicons and then these lexicons were compared to the wordnets in various languages in order to disambiguate the entries and attach appropriate synset ids to Slovene entries in the lexicon. Slovene lexicon entries sharing the same attached synset id were then organized into a synset. The results obtained by the different settings in the experiment are evaluated against a manually created gold standard and also checked by hand.

Standardizing Tweets with Character-Level Machine Translation

This paper presents the results of the standardization procedure of Slovene tweets that are full of colloquial, dialectal and foreign-language elements. With the aim of minimizing the human input required we produced a manually normalized lexicon of the most salient out-of-vocabulary OOV tokens and used it to train a character-level statistical machine translation system CSMT. Best results were obtained by combining the manually constructed lexicon and CSMT as fallback with an overall improvement of 9.9% increase on all tokens and 31.3% on OOV tokens. Manual preparation of data in a lexicon manner has proven to be more efficient than normalizing running text for the task at hand. Finally we performed an extrinsic evaluation where we automatically lemmatized the test corpus taking as input either original or automatically standardized wordforms, and achieved 75.1% per-token accuracy with the former and 83.6% with the latter, thus demonstrating that standardization has significant benefits for upstream processing.

Bootstrapping bilingual lexicons from comparable corpora for closely related languages

In this paper we present an approach to bootstrap a Croatian-Slovene bilingual lexicon from comparable news corpora from scratch, without relying on any external bilingual knowledge resource. Instead of using a dictionary to translate context vectors, we build a seed lexicon from identical words in both languages and extend it with context-based cognates and translation candidates of the most frequent words. By enlarging the seed dictionary for only 7% we were able to improve the baseline precision from 0.597 to 0.731 on the mean reciprocal rank for the ten top-ranking translation candidates with a 50.4% recall on the gold standard of 500 entries.

A Global Analysis of Emoji Usage.

Emojis are a quickly spreading and rather unknown communication phenomenon which occasionally receives attention in the mainstream press, but lacks the scientific exploration it deserves. This paper is a first attempt at investigating the global distribution of emojis. We perform our analysis of the spatial distribution of emojis on a dataset of ∼17 million (and growing) geo-encoded tweets containing emojis by running a cluster analysis over countries represented as emoji distributions and performing correlation analysis of emoji distributions and World Development Indicators. We show that emoji usage tends to draw quite a realistic picture of the living conditions in various parts of our world.

Constructing a poor man's wordnet in a resource-rich world

In this paper we present a language-independent, fully modular and automatic approach to bootstrap a wordnet for a new language by recycling different types of already existing language resources, such as machine-readable dictionaries, parallel corpora, and Wikipedia. The approach, which we apply here to Slovene, takes into account monosemous and polysemous words, general and specialised vocabulary as well as simple and multi-word lexemes. The extracted words are then assigned one or several synset ids, based on a classifier that relies on several features including distributional similarity. Finally, we identify and remove highly dubious (literal, synset) pairs, based on simple distributional information extracted from a large corpus in an unsupervised way. Automatic, manual and task-based evaluations show that the resulting resource, the latest version of the Slovene wordnet, is already a valuable source of lexico-semantic information.

Legal Framework, Dataset and Annotation Schema for Socially Unacceptable Online Discourse Practices in Slovene

In this paper we present the legal framework, dataset and annotation schema of socially unacceptable discourse practices on social networking platforms in Slovenia. On this basis we aim to train an automatic identification and classification system with which we wish contribute towards an improved methodology, understanding and treatment of such practices in the contemporary, increasingly multicultural information society.

Adapting a State-of-the-Art Tagger for South Slavic Languages to Non-Standard Text.

In this paper we present the adaptations of a state-of-the-art tagger for South Slavic languages to non-standard texts on the example of the Slovene language. We investigate the impact of introducing in-domain training data as well as additional supervision through external resources or tools like word clusters and word normalization. We remove more than half of the error of the standard tagger when applied to nonstandard texts by training it on a combination of standard and non-standard training data, while enriching the data representation with external resources removes additional 11 percent of the error. The final configuration achieves tagging accuracy of 87.41% on the full morphosyntactic description, which is, nevertheless, still quite far from the accuracy of 94.27% achieved on standard text.

Using WordNet-Based Word Sense Disambiguation to Improve MT Performance

We report on a series of experiments aimed at improving the machine translation of ambiguous lexical items by using WordNet-based unsupervised Word Sense Disambiguation (WSD) and comparing its results to three MT systems. Our experiments are performed for the English-Slovene language pair using UKB, a freely available graph-based word sense disambiguation system. Since the fine granularity of WordNet is often reported as problematic, we compare the performance of UKB using all WordNet senses with using sense clusters. Results are evaluated in three ways: a manual evaluation of WSD performance from MT perspective, an analysis of agreement between the WSD-proposed equivalent and those suggested by the three systems, and finally by computing BLEU, NIST and METEOR scores for all translation versions. Our results show that WSD performs with a MT-relevant precision of 71 % and that 21 % of sense-related MT errors could be prevented by using unsupervised WSD. We also show that sense clusters improve MT-relevant precision.

Language-independent Gender Prediction on Twitter.

In this paper we present a set of experiments and analyses on predicting the gender of Twitter users based on language-independent features extracted either from the text or the metadata of users’ tweets. We perform our experiments on the TwiSty dataset containing manual gender annotations for users speaking six different languages. Our classification results show that, while the prediction model based on language-independent features performs worse than the bag-of-words model when training and testing on the same language, it regularly outperforms the bag-of-words model when applied to different languages, showing very stable results across various languages. Finally we perform a comparative analysis of feature effect sizes across the six languages and show that differences in our features correspond to cultural distances.