Krister Lindén

HFST Tools for Morphology – An Efficient Open-Source Package for Construction of Morphological Analyzers

Morphological analysis of a wide range of languages can be implemented efficiently using finite-state transducer technologies. Over the last 30 years, a number of attempts have been made to create tools for computational morphologies. The two main competing approaches have been parallel vs. cascaded rule application. The parallel rule application was originally introduced by Koskenniemi [1] and implemented in tools like TwolC and LexC. Currently many applications of morphologies could use dictionaries encoding the a priori likelihoods of words and expressions as well as the likelihood of relations to other representations or languages. We have made the choice to create open-source tools and language descriptions in order to let as many as possible participate in the effort. The current article presents some of the main tools that we have created such as HFST-LexC, HFST-TwolC and HFST-Compose-Intersect. We evaluate their efficiency in comparison to some similar tools and libraries. In particular, we evaluate them using several full-fledged morphological descriptions. Our tools compare well with similar open source tools, even if we still have some challenges ahead before we can catch up with the commercial tools. We demonstrate that for various reasons a parallel rule approach still seems to be more efficient than a cascaded rule approach when developing finite-state morphologies.

HFST—Framework for Compiling and Applying Morphologies

HFST–Helsinki Finite-State Technology ( hfst.sf.net ) is a framework for compiling and applying linguistic descriptions with finite-state methods. HFST currently connects some of the most important finite-state tools for creating morphologies and spellers into one open-source platform and supports extending and improving the descriptions with weights to accommodate the modeling of statistical information. HFST offers a path from language descriptions to efficient language applications in key environments and operating systems. HFST also provides an opportunity to exchange transducers between different software providers in order to get the best out of each finite-state library.

A probabilistic model for guessing base forms of new words by analogy

Language software applications encounter new words, e.g., acronyms, technical terminology, loan words, names or compounds of such words. Looking at English, one might assume that they appear in base form, i.e., the lexical look-up form. However, in more highly inflecting languages like Finnish or Swahili only 40-50 % of new words appear in base form. In order to index documents or discover translations for these languages, it would be useful to reduce new words to their base forms as well. We often have access to analyzes for more frequent words which shape our intuition for how new words will inflect. We formalize this into a probabilistic model for lemmatization of new words using analogy, i.e., guessing base forms, and test the model on English, Finnish, Swedish and Swahili demonstrating that we get a recall of 89- 99 % with an average precision of 76-94 % depending on language and the amount of training material.

Multilingual modeling of cross-lingual spelling variants

Technical term translations are important for cross-lingual information retrieval. In many languages, new technical terms have a common origin rendered with different spelling of the underlying sounds, also known as cross-lingual spelling variants (CLSV).To find the best CLSV in a text database index, we contribute a formulation of the problem in a probabilistic framework, and implement this with an instance of the general edit distance using weighted finite-state transducers. Some training data is required when estimating the costs for the general edit distance. We demonstrate that after some basic training our new multilingual model is robust and requires little or no adaptation for covering additional languages, as the model takes advantage of language independent transliteration patterns.We train the model with medical terms in seven languages and test it with terms from varied domains in six languages. Two test languages are not in the training data. Against a large text database index, we achieve 64–78 % precision at the point of 100% recall. This is a relative improvement of 22% on the simple edit distance.

Is it possible to create a very large wordnet in 100 days? An evaluation

Wordnets are large-scale lexical databases of related words and concepts, useful for language-aware software applications. They have recently been built for many languages by using various approaches. The Finnish wordnet, FinnWordNet (FiWN), was created by translating the more than 200,000 word senses in the English Princeton WordNet (PWN) 3.0 in 100 days. To ensure quality, they were translated by professional translators. The direct translation approach was based on the assumption that most synsets in PWN represent language-independent real-world concepts. Thus also the semantic relations between synsets were assumed mostly language-independent, so the structure of PWN could be reused as well. This approach allowed the creation of an extensive Finnish wordnet directly aligned with PWN and also provided us with a translation relation and thus a bilingual wordnet usable as a dictionary. In this paper, we address several concerns raised with regard to our approach, many of them for the first time. We evaluate the craftsmanship of the translators by checking the spelling and translation quality, the viability of the approach by assessing the synonym quality both on the lexeme and concept level, as well as the usefulness of the resulting lexical resource both for humans and in a language-technological task. We discovered no new problems compared with those already known in PWN. As a whole, the paper contributes to the scientific discourse on what it takes to create a very large wordnet. As a side-effect of the evaluation, we extended FiWN to contain 208,645 word senses in 120,449 synsets, effectively making version 2.0 of FiWN currently the largest wordnet in the world by these statistics.

Entry Generation for New Words by Analogy for Morphological Lexicons

Language software applications encounter new words, e.g., acronyms, technical terminology, loan words, names or compounds of such words. To add new words to a lexicon, we need to indicate their base form and inflectional paradigm. In this article, we evaluate a combination of corpus-based and lexicon-based methods for assigning the base form and inflectional paradigm to new words in Finnish, Swedish and English finite-state transducer lexicons. The methods have been implemented with the open-source Helsinki Finite-State Technology (Lindén & al., 2009). As an entry generator often produces numerous suggestions, it is important that the best suggestions be among the first few, otherwise it may become more efficient to create the entries by hand. By combining the probabilities calculated from corpus data and from lexical data, we get a more precise combined model. The combined method has 77-81 % precision and 89-97 % recall, i.e. the first correctly generated entry is on the average found as the first or second candidate for the test languages. A further study demonstrated that a native speaker could revise suggestions from the entry generator at a speed of 300-400 entries per hour.

FinnPos: an open-source morphological tagging and lemmatization toolkit for Finnish

This paper describes FinnPos, an open-source morphological tagging and lemmatization toolkit for Finnish. The morphological tagging model is based on the averaged structured perceptron classifier. Given training data, new taggers are estimated in a computationally efficient manner using a combination of beam search and model cascade. The lemmatization is performed employing a combination of a rule-based morphological analyzer, OMorFi, and a data-driven lemmatization model. The toolkit is readily applicable for tagging and lemmatization of running text with models learned from the recently published Finnish Turku Dependency Treebank and FinnTreeBank. Empirical evaluation on these corpora shows that FinnPos performs favorably compared to reference systems in terms of tagging and lemmatization accuracy. In addition, we demonstrate that our system is highly competitive with regard to computational efficiency of learning new models and assigning analyses to novel sentences.

HFST — A System for Creating NLP Tools

The paper presents and evaluates various NLP tools that have been created using the open source library HFST – Helsinki Finite-State Technology and outlines the minimal extensions that this has required to a pure finite-state system. In particular, the paper describes an implementation and application of Pmatch presented by Karttunen at SFCM 2011.

Using HFST for Creating Computational Linguistic Applications

HFST-HelsinkiFinite-StateTechnology (http://hfst.sf.net/) is a framework for compiling and applying linguistic descriptions with finitestatemethods. HFST currently collects some of the most important finite-state tools for creatingmorphologies and spellcheckers into one open-source platform and supports extending and improving the descriptions with weights to accommodate the modeling of statistical information. HFST offers a path from language descriptions to efficient language applications. In this article, we focus on aspects of HFST that are new to the end user, i.e. new tools, new features in existing tools, or new language applications, in addition to some revised algorithms that increase performance.

Part-of-Speech Tagging using Conditional Random Fields: Exploiting Sub-Label Dependencies for Improved Accuracy

We discuss part-of-speech (POS) tagging in presence of large, fine-grained label sets using conditional random fields (CRFs). We propose improving tagging accuracy by utilizing dependencies within sub-components of the fine-grained labels. These sub-label dependencies are incorporated into the CRF model via a (relatively) straightforward feature extraction scheme. Experiments on five languages show that the approach can yield significant improvement in tagging accuracy in case the labels have sufficiently rich inner structure.