Search

The National Library of Finland (NLF) has digitized historical newspapers, journals and ephemera published in Finland since the late 1990s. The present collection consists of about 16.51 million pages mainly in Finnish and Swedish. Out of these about 7.64 million pages are freely available on the web site https://digi.kansalliskirjasto.fi/etusivu . The copyright restricted part of the collection can be used at six legal deposit libraries in different parts of Finland. The time period of the open collection is from 1771 to 1929. The last nine years, 1921–1929, were opened in January 2018. This paper presents briefly the ground truth Optical Character Recognition data of about 500 000 words that has been compiled at the NLF for development of an improved OCR process for the Finnish collection. We discuss compilation of the data generally and show results of the new OCR process in comparison to current OCR, using the ground truth data as an evaluation benchmark. We also show with real newspaper data of 30 years and 109 million words that the re-OCRing process is improving the quality of the OCRed data.

The current paper presents work that has been carried out in the National Library of Finland (NLF) to improve optical character recognition (OCR) quality of the historical Finnish newspaper collection 1771–1910. Evaluation results reported in the paper are based mainly on a 500 000 word sample of the Finnish language part of the whole collection. The sample has three different parallel parts: a manually corrected ground truth version, original OCR with ABBYY FineReader v. 7 or v. 8, and an ABBYY FineReader v. 11 re-OCRed version for comparison with Tesseract’s OCR. Using this sample and its page image originals we have developed a re-OCRing procedure using the open source software package Tesseract v. 3.04.01. Our method achieved initially 27.48% improvement vs. ABBYY FineReader 7 or 8 and 9.16% improvement vs. ABBYY FineReader 11 on document level. On word level our method achieved 36.25% improvement vs. ABBYY FineReader 7 or 8 and 20.14% improvement vs. ABBYY FineReader 11. Our final precision and recall results on word level show clear improvement in the quality: recall is 76.0 and precision 92.0 in comparison to GT OCR. Other measures, such as recognizability of words with a morphological analyzer and character accuracy rate, show also steady improvement after re-OCRing.

This paper introduces a work in progress for implementing a free full text semantic tagger for Finnish, FiST. The tagger is based on a 46 226 lexeme semantic lexicon of Finnish that was published in 2016. The basis of the semantic lexicon was developed in the early 2000s in an EU funded project Benedict (Löfberg et al., 2005). Löfberg (2017) describes compilation of the lexicon and evaluates a proprietary version of the Finnish Semantic Tagger, the FST2. The FST and its lexicon were developed using the English Semantic Tagger (The EST) of University of Lancaster as a model. This semantic tagger was developed at the University Centre for Corpus Research on Language (UCREL) at Lancaster University as part of the UCREL Semantic Analysis System (USAS3 ) framework. The semantic lexicon of the USAS framework is based on the modified and enriched categories of the Longman Lexicon of Contemporary English (McArthur, 1981). We have implemented a basic working version of a new full text semantic tagger for Finnish based on freely available components. The implementation uses Omorfi and FinnPos for morphological analysis of Finnish words. After the morphological recognition phase words from the 46K semantic lexicon are matched against the morphologically unambiguous base forms. In our comprehensive tests the lexical tagging coverage of the current implementation is around 82–90% with different text types. The present version needs still some enhancements, at least processing of semantic ambiguity of words and analysis of compounds, and perhaps also treatment of multiword expressions. Also a semantically marked ground truth evaluation collection should be established for evaluation of the tagger.

The aim of the study was to test whether query expansion by approximate string matching methods is beneficial in retrieval from historical newspaper collections in a language rich with compounds and inflectional forms Finnish. First, approximate string matching methods were used to generate lists of index words most similar to contemporary query terms in a digitized newspaper collection from the 1800s. Top index word variants were categorized to estimate the appropriate query expansion ranges in the retrieval test. Second, the effectiveness of approximate string matching methods, automatically generated inflectional forms, and their combinations were measured in a Cranfield-style test. Finally, a detailed topic-level analysis of test results was conducted. In the index of historical newspaper collection the occurrences of a word typically spread to many linguistic and historical variants along with optical character recognition OCR errors. All query expansion methods improved the baseline results. Extensive expansion of around 30 variants for each query word was required to achieve the highest performance improvement. Query expansion based on approximate string matching was superior to using the inflectional forms of the query words, showing that coverage of the different types of variation is more important than precision in handling one type of variation.

Named Entity Recognition (NER), search, classification and tagging of names and name like frequent informational elements in texts, has become a standard information extraction procedure for textual data. NER has been applied to many types of texts and different types of entities: newspapers, fiction, historical records, persons, locations, chemical compounds, protein families, animals etc. In general a NER system's performance is genre and domain dependent and also used entity categories vary [16]. The most general set of named entities is usually some version of three partite categorization of locations, persons and organizations. In this paper we report evaluation result of NER with data out of a digitized Finnish historical newspaper collection Digi. Experiments, results and discussion of this research serve development of the Web collection of historical Finnish newspapers.Digi collection contains 1,960,921 pages of newspaper material from years 1771-1910 both in Finnish and Swedish. We use only material of Finnish documents in our evaluation. The OCRed newspaper collection has lots of OCR errors; its estimated word level correctness is about 70-75% [7]. Our baseline NER tagger is a rule-based tagger of Finnish, FiNER, provided by the FIN-CLARIN consortium. Three other available tools are also evaluated: a Finnish Semantic Tagger (FST), Connexor's NER tool and Polyglot's NER.

Many operational IR indexes are non-normalized, i.e. no lemmatization or stemming techniques, etc. have been employed in indexing. This poses a challenge for dictionary-based cross-language retrieval (CLIR), because translations are mostly lemmas. In this study, we face the challenge of dictionary-based CLIR in a non-normalized index. We test two optional approaches: FCG (Frequent Case Generation) and s-gramming. The idea of FCG is to automatically generate the most frequent inflected forms for a given lemma. FCG has been tested in monolingual retrieval and has been shown to be a good method for inflected retrieval, especially for highly inflected languages. S-gramming is an approximate string matching technique (an extension of n-gramming). The language pairs in our tests were English-Finnish, English-Swedish, Swedish-Finnish and Finnish-Swedish. Both our approaches performed quite well, but the results varied depending on the language pair. S-gramming and FCG performed quite equally in all the other language pairs except Finnish-Swedish, where s-gramming outperformed FCG.

PurposeThe purpose of this article is to discuss advantages and disadvantages of various means to manage morphological variation of keywords in monolingual information retrieval.Design/methodology/approachThe authors present a compilation of query results from 11 mostly European languages and a new general classification of the language dependent techniques for management of morphological variation. Variants of the different techniques are compared in some detail in terms of retrieval effectiveness and other criteria. The paper consists mainly of an overview of different management methods for keyword variation in information retrieval. Typical IR retrieval results of 11 languages and a new classification for keyword management methods are also presented.FindingsThe main results of the paper are an overall comparison of reductive and generative keyword management methods in terms of retrieval effectiveness and other broader criteria.Originality/valueThe paper is of value to anyone who wants to get an overall picture of keyword management techniques used in IR.

In this article, we are studying the differences between the European Union languages using statistical and unsupervised methods. The analysis is conducted in the different levels of language: the lexical, morphological and syntactic. Our premise is that the difficulty of the translation could be perceived as differences or similarities in different levels of language. The results are compared to linguistic groupings. Two approaches are selected for the analysis. A Kolmogorov complexity-based approach is used to compare the language structure in syntactic and morphological levels. A morpheme-level comparison is conducted based on an automated segmentation of the languages into morpheme-like units. The way the languages convey information in these levels is taken as a measure of similarity or dissimilarity between languages and the results are compared to classical linguistic classifications. The results have a significant impact on the design of (statistical) machine translation systems. If the s...

There has been a huge interest in digitization of both hand-written and printed historical material in the last 10–15 years and most probably this interest will only increase in the ongoing Digital Humanities era. As a result of the interest we have lots of digital historical document collections available and will have more of them in the future.

Digital collections of the National Library of Finland (NLF) contain over 10 million pages of historical newspapers, journals and some technical ephemera. The material ranges from the early Finnish newspapers from 1771 until the present day. The material up to 1910 can be viewed in the public web service, where as anything later is available at the six legal deposit libraries in Finland. A recent user study noticed that a different type of researcher use is one of the key uses of the collection. National Library of Finland has gotten several requests to provide the content of the digital collections as one offline bundle, where all the needed content is included. For this purpose we introduced a new format, which contains three different information sets: the full metadata of a publication page, the actual page content as ALTO XML, and the raw text content. We consider these formats most useful to be provided as raw data for the researchers. In this paper we will describe how the export format was created, how other parties have packaged the same data and what the benefits are of the current approach. We shall also briefly discuss word level quality of the content and show a real research scenario for the data.

Word form normalization through lemmatization or stemming is a standard procedure in information retrieval because morphological variation needs to be accounted for and several languages are morphologically non-trivial. Lemmatization is effective but often requires expensive resources. Stemming is also effective in most contexts, generally almost as good as lemmatization and typically much less expensive; besides it also has a query expansion effect. However, in both approaches the idea is to turn many inflectional word forms to a single lemma or stem both in the database index and in queries. This means extra effort in creating database indexes. In this paper we take an opposite approach: we leave the database index un-normalized and enrich the queries to cover for surface form variation of keywords. A potential penalty of the approach would be long queries and slow processing. However, we show that it only matters to cover a negligible number of possible surface forms even in morphologically complex languages to arrive at a performance that is almost as good as that delivered by stemming or lemmatization. Moreover, we show that, at least for typical test collections, it only matters to cover nouns and adjectives in queries. Furthermore, we show that our findings are particularly good for short queries that resemble normal searches of web users. Our approach is called FCG (for Frequent Case (form) Generation). It can be relatively easily implemented for Latin/Greek/Cyrillic alphabet languages by examining their (typically very skewed) nominal form statistics in a small text sample and by creating surface form generators for the 3---9 most frequent forms. We demonstrate the potential of our FCG approach for several languages of varying morphological complexity: Swedish, German, Russian, and Finnish in well-known test collections. Applications include in particular Web IR in languages poor in morphological resources.

The paper presents a new method for handling of morphological variation of query terms in best-match IR. The method is based on enhanced inflectional stems. Use of inflectional stems has earlier been shown to be a good retrieval method in inflected indexes in a best-match environment for a highly inflected and compound-rich language, Finnish. In this paper the earlier stem method is elaborated upon by enhancing the stems with regular expressions. Contrary to our expectations the results show that the enhanced stem queries do not outperform basic inflectional stems, but neither are they considerably worse with long queries. With short web-like queries they perform relatively better than with long queries and outperform clearly stemming (Finnish stemmer of Snowball) and plain, unprocessed query words. The main benefits of the proposed method, besides fairly good precision and recall (P-R) performance, are shorter and more manageable queries, which is of practical importance, e.g. with large web indexes.

This paper presents results of a generative method for the management of morphological variation of query keywords in Bengali, Gujarati and Marathi. The method is called Frequent Case Generation (FCG). It is based on the skewed distributions of word forms in natural languages and is suitable for languages that have either fair amount of morphological variation or are morphologically very rich. We participated in the ad hoc task at FIRE 2011 and applied the FCG method on monolingual Bengali, Gujarati and Marathi test collections. Our evaluation was carried out with title and description fields of test topics, and the Lemur search engine. We used plain unprocessed word index as the baseline, and n-gramming and stemming as competing methods. The evaluation results show 30%, 16% and 70% relative mean average precision improvements for Bengali, Gujarati and Marathi respectively when comparing the FCG method to plain words. The method shows competitive performance in comparison to n-gramming and stemming.

This paper discusses information retrieval of Finnish and keyword variation management by generating inflected variant keyword forms. Finnish is a highly inflectional language, and thus keyword variation management of queries and query indexes is of utter importance for successful Finnish full-text retrieval. In the paper we show that generation of a quite small number of variant keyword forms leads to good retrieval performance using a probabilistic best-match retrieval system (Lemur). Generation of almost the full paradigm of inflected nominal forms improves the results slightly. We have also interesting results with regards to different index types: our evaluation shows that generated inflected queries behave extremely well in a lemmatized index, which is supposedly not suitable for this query type. We also show that in a research environment even inexact generation that produces lots of incorrect inflected forms achieves high precision-recall performance without considerable loss in query throughput effectiveness. We use two different word form generators and their variants and compare the results to commonly used reductive word form variation management methods, stemming and lemmatization. The paper includes also a short discussion about usage of the variant keyword method with Web search engines.

This study describes first usage of a particular implementation of Normalized Compression Distance (NCD) as a machine translation quality evaluation tool. NCD has been introduced and tested for clustering and classification of different types of data and found a reliable and general tool. As far as we know NCD in its Complearn implementation has not been evaluated as a MT quality tool yet, and we wish to show that it can also be used for this purpose. We show that NCD scores given for MT outputs in different languages correlate highly with scores of a state-of-the-art MT evaluation metrics, METEOR 0.6. Our experiments are based on translations between one source and three target languages with a smallish sample that has available reference translations, UN's Universal Declaration of Human Rights. Secondly we shall also briefly describe and discuss results of a larger scale evaluation of NCD as an MT metric with WMT08 Shared Task Evaluation Data. These evaluations confirm further that NCD is a noteworthy MT metric both in itself and also enriched with basic language tools, stemming and Wordnet.

This paper describes usage of MT metrics in choosing the best candidates for MT-based query translation resources. Our main metrics is METEOR, but we also use NIST and BLEU. Language pair of our evaluation is English *** German, because MT metrics still do not offer very many language pairs for comparison. We evaluated translations of CLEF 2003 topics of four different MT programs with MT metrics and compare the metrics evaluation results to results of CLIR runs. Our results show, that for long topics the correlations between achieved MAPs and MT metrics is high (0.85-0.94), and for short topics lower but still clear (0.63-0.72). Overall it seems that MT metrics can easily distinguish the worst MT programs from the best ones, but smaller differences are not so clearly shown. Some of the intrinsic properties of MT metrics do not also suit for CLIR resource evaluation purposes, because some properties of translation metrics, especially evaluation of word order, are not significant in CLIR.

This paper presents implementations of generative management method for morphological variation of query keywords. The method is called FCG, Frequent Case Generation. It is based on the skewed distributions of word forms in natural languages and is suitable for languages that either have fair amount of morphological variation or are morphologically very rich. The paper reports implementation and evaluation of automatic procedures of variant query keyword form generation with short and long queries of CLEF collections for English, Finnish, German and Swedish. The evaluated languages show varying degrees of morphological complexity.

This paper presents a new management method for morphological variation of keywords. The method is called FCG, Frequent Case Generation. It is based on the skewed distributions of word forms in natural languages and is suitable for languages that have either fair amount of morphological variation or are morphologically very rich. The proposed method has been evaluated so far with four languages, Finnish, Swedish, German and Russian, which show varying degrees of morphological complexity.

In this article, we are studying the differences between the European languages using statistical and unsupervised methods. The analysis is conducted in different levels of language, lexical, morphological and syntactic. Our premise is that the difficulty of the translation could be perceived as differences or similarities in different levels of language. The results are compared to linguistic groupings. The analyses of this paper are based on the concept of Kolmogorov complexity, which is used to compare the language structure in syntactic and morphological levels. The way the languages convey information in these levels is taken as a measure of similarity or dissimilarity between languages and the results are compared to classical linguistic classification. The results will serve as a tool in developing machine translation system(s), e.g., in the following way: if source language conveys more information in the morphological level and the target language more in the syntactic level, it is clear that the (machine) translator must be able to transfer the information from one level to another.

PurposeTo show that stem generation compares well with lemmatization as a morphological tool for a highly inflectional language for IR purposes in a best‐match retrieval system.Design/methodology/approachEffects of three different morphological methods – lemmatization, stemming and stem production – for Finnish are compared in a probabilistic IR environment (INQUERY). Evaluation is done using a four‐point relevance scale which is partitioned differently in different test settings.FindingsResults show that stem production, a lighter method than morphological lemmatization, compares well with lemmatization in a best‐match IR environment. Differences in performance between stem production and lemmatization are small and they are not statistically significant in most of the tested settings. It is also shown that hitherto a rather neglected method of morphological processing for Finnish, stemming, performs reasonably well although the stemmer used – a Porter stemmer implementation – is far from optimal for a morphologically complex language like Finnish. In another series of tests, the effects of compound splitting and derivational expansion of queries are tested.Practical implicationsUsefulness of morphological lemmatization and stem generation for IR purposes can be estimated with many factors. On the average P‐R level they seem to behave very close to each other in a probabilistic IR system. Thus, the choice of the used method with highly inflectional languages needs to be estimated along other dimensions too.Originality/valueResults are achieved using Finnish as an example of a highly inflectional language. The results are of interest for anyone who is interested in processing of morphological variation of a highly inflected language for IR purposes.