This paper presents a solution for mining the biographical information from commentaries on Leo Tolstoy's letters. It is implemented as a part of Tolstoy Digital Project - a semantically marked-up web publication of the 90-volume complete collection of Leo Tolstoy's works. Extraction of relevant biographical information will be used to create an open database for all the persons who were somehow connected with Tolstoy or Tolstoy's works. The paper also accounts for various subtleties of the commentary apparatus and pays special attention to specific difficulties of biographical information extraction, such as the problem of defining the boundaries of expressions denoting profession, or the problem of non-standardized syntactic constructions for kinship relations.
The Tolstoy Digital project1 aims to prepare a
webpublished semantically marked-up version of the
90volume complete collection of Leo Tolstoy’s works2. The
digital version of the 90-volume edition has become easy
to access thanks to the mass crowdsourcing campaign
“All of Tolstoy in one click”3 . The next step of Tolstoy
digitization is devoted to the semantic tagging of
Tolstoy’s text and the creation of a comprehensive
database of all of the additional reference information that
goes along with the Tolstoy oeuvres and private archive.
The 90-volume edition
The main aim of our current experiment was to estimate the effectiveness of a rule-based approach that could be used for data analysis and fact extraction for our type of texts. We started with the most basic concepts and facts: names, dates, professions and family relations. The 1 See the project description here: http://tolstoy.ru/projects/tolstoy-digital/ 2 http://tolstoy.ru/creativity/90-volume-collection-of-theworks/ 3 See, for example, review in the Guardian: http://www.theguardian.com/books/2013/oct/16/all-leotolstoy-one-click-project-digitisation problem of automatic named entity recognition and fact extraction seems to be very well elaborated, see (Grishman 2003, and Jurafsky Martin 2009) for a comprehensive outline and reference list. Still, information extraction in academic philological texts pursues slightly different aims than mining the news flow: there are graphical, linguistic and conceptual problems. Specific graphics (such as abbreviations, inner references, font highlighting) may have non-trivial semiotic functions. Some patterns that are uncommon in general language may be used to render the latent recurring logics of the commentary structure. Finally, the notions that are relevant for characterizing historical events or persons may have little in common with the conceptual space of data mining in the web (the domain which is primarily regarded in all the works on information extraction). This is the reason why the task of biographical information extraction from academic editions cannot be solved with the help of already existing solutions, but requires some specific process modifications. This paper aims to report on the first steps that we have made in this direction. We used a rule-based toolbox Tomita to write and apply some basic grammar rules that are used to extract relevant ontology concepts. In part (2) we will present a short outline of the biographical ontology we are going to elaborate and we will briefly describe preliminary data preparation. Part (3) will be dedicated to the description of our grammar and the rules for biographical fact extraction. Evaluation and analysis are presented in part (4).
The commentary apparatus of epistolary volumes of the complete 90-volume edition consists of letter commentaries, which constitute about 40% of entire text. Usually they are organized in a non-structured way as a sequence of factual comments. Some of them are hard to classify, and many of them seem not to fit to any database, making them redundant. This lack of explicit structure can be explained by the fact that the commentaries have been created by different authors, and each of them had used his own text template. As a result, commentaries are represented as an accompanying text, but not as an enumeration of properties and parameters of a common database. An example of a commentary is provided in Table 1.
English Анатолий Иванович Anatoly Ivanovich Фаресов (1852—1928), Faresov (1852-1928), публицист-народник. writer-populist. [He was [Судился по «делу 193»; tried for “Case 193”; He был амнистирован в was pardoned in 1880 and 1880 г. и перешел в joined the camp of лагерь умеренных moderate liberals.] He либералов.] worked in the “News,” in Сотрудничал в the “Week,” and other «Новостях», в «Неделе» publications. He had met и других изданиях. with Tolstoy February 1, Познакомился с 1898 [and left unpublished Толстым 1 февраля 1898 memories of him]. г. [и оставил Faresov’s article titled By неопубликованные yesterday incident edition воспоминания о нем]. of “Week”" was published Статья Фаресова под in the "New Times" , № заглавием «Ко 7208 of 23 March вчерашнему происшествию редакции в «Недели» была напечатана в «Новом времени», № 7208 от 23 марта
This example illustrates the complexity of the information in the comments. We have marked with square brackets all information that cannot be structurally formalized. That means that though we have at our disposable a considerable corpus of short biographical texts, but it is still a collection of unstructured texts with high degree of lexical variation and a mixture of relevant and excessive factual statements. We aim to transfer an unstructured commentary text into a structured database with given semantic relationships between the elements. For the first stage of our project, we limit ourselves to a small list of relationships. Table 2 shows the relevant facts, which have been filtered out of the text, given in table 1.
Ivanovich Farezov (18521928) writer-populist He was tried for “Case 193”; He was pardoned in 1880 and joined the camp of moderate liberals.
“News,” in the profession fact “Week,” and other publications.
Tolstoy February 1, 1898 and left unpublished memories of him.
publicistic fact Faresov’s Article Faresov’s titled "By publicistic fact yesterday incident edition of “Week”" was published in the "New Times" , № 7208 of 23 March
YES NO NO YES NO YES
The corpus of letters, used for rule development and
testing, was based on the 63rd volume of the Complete
edition, which contains 281 letters by Tolstoy of the
period of 1880-1886. We have created a manual markup
of 50 letters using the annotation module from GATE
framework
We used Tomita parser to create rules for fact extraction
Yandex.ru4 and is available for free download. Tomita provides modules of morphological analysis, as well as ready-made rules for extracting names and numbers. Tomita grammars consist of rules. The user may create his or her own grammars and dictionaries for a certain language. To construct a grammar, the user should write a number of transducer rules, for which one can use regular expressions, words lists, and other rules, built-in or composed. Each rule has a left and a right side; the transducer operation is denoted by arrow, separating left and right sides. The left side can be only represented by a terminal, while the right side can be both terminals or nonterminals. An example of a rule sequence can be found in Figures 2-3.
This rule says that a date is a number, consisting of four characters from 0 to 9. Then, we can define date chains (for example dates of birth and death) using previously set rules. The operator interp in Figure 3 assigns the extracted text span to a specific fact. Here Dt1 stands for the birth date, while Dt2 corresponds to the date of death :
Using Tomita parser we have written a grammar with specific rules and dictionaries aimed to extract the following biographic facts from letter commentaries: ProfessionFact (person’s name, dates of life and professions) and FamilyFact (kinship). We quickly realized that the built-in date-extraction and nameextraction rules are not well-sutied to philological commentaries, so we also made some specific modifications. For description of kinship relations of a person, we have created several dictionaries containing different kinship types. The kinship dictionary is applied after the pattern is recognized. If no kinship term is detected in the pattern, then the rules that extract profession are applied. Overall there have been created 31 rules and 3 dictionaries. The processing scheme is presented by Figure 4.
The performance rank of these rules has been measured with the help of testing commentary corpus of 560,000 tokens. The results are considerably acceptable for precision, but are much lower for recall, which is quite typical for rule-based approaches. The overall score for both facts is demonstrated in Table 3
ProfessionFact FamilyFact
Precision 1 0,78
Recall 0,82 0,3
F-measure 0,9 0,43 The analysis of results reveals that some important problems lie in the conceptual part. First of all, it is problematic to define a strict classification of what may (or may not) be considered as a profession. Thus in example 1 the profession extracted is a peasant. But in context of biographical commentaries this person’s origin (village Shevelevo) becomes even more important than the profession of peasant itself. It’s obvious that Tolstoy had met a lot of peasants during his life, so any special attributes are of much value in the context. There remains a problem of distinguishing those cases, which are meaningless without additional information. RUS: Александр Андреевич Иванов (Шешин) (1806 — 1858) — художник .
ENG: Alexandr Andreevich Ivanov (Sheshin) (1806 — 1858) — а painter.
► PATTERN RECOGNITION ► KINSHIP DICTIONARY CHECK ► CHOICE OF RULE SET ► FACT EXTRACTION OUTPUT <has name:first> RUS: Александр ENG: Aleksandr <has name:middle> RUS: Андреевич ENG: Andreevich <has name:last> RUS: Иванов ENG: Ivanov <has name:last> RUS: Shenshin ENG: Шеньшин <date:birth> 1806 <date:death> 1858 <profession> RUS: художник
ENG: painter Another problem is the boundaries of the text span which refers to the profession itself. Thus, in example 2, we see a description of Shamil’s social activities. So the question is what part of the complicated NP should be extracted as Shamil’s profession. Should it be the leader and consolidator or the full NP (leader and consolidator of hillmen of Daghestan and Chechnya)? 1) Василий Кириллович Сютаев (1819—1892) — ... крестьянин дер. Шевелино.
Vasiliy Kirillovich Syutaev (1819—1892) — ... peasant of village Shevelino. 2) Шамиль (1797—1874) — знаменитый вождь и объединитель горцев Дагестана и Чечни в их борьбе с русскими.
Shamil (1797-1874) – a popular leader and consolidator of hillmen of Daghestan and Chechnya in their struggle with Russians.
To resume, if such types of professions as writer, musician, and philosopher are good categorizers, that allow to define a group of people in one way or another connected with Tolstoy or his work, such status as a peasant or consolidator are meaningless without their attributes (genitive groups), in this case – “village Shevelino” and “hillmen of Daghestan and Chechnya in their struggle with Russians.” Accordingly, the significant problem is with the boundary of the nominal group that determines the professional status. The second problem is concerned with intricate chains of kinships that have a very specific syntax, as shown in Example 3. In this phrase there is an inversion (see Figure 5), probably made especially for publication, so that Sophia Tolstaya, wife of Leo Tolstoy, would take a more prominent position.
the husband of the sister of SA, TA
Another problem is the discrepancy between singular and plural forms when it comes to the descriptions of relations between a person and a family group, as shown in Example 4.
Nikolay Mihaylovich Nagornov (1845—1896) — the son of Mihail Mihaylovich and Nadejda Ivanovna Nagornovs.
Unlike profession extraction which shows good quality by rule-based approach, kinship relation patterns seem to be much less regular, so perhaps it is worth trying to extract them using algorithms of machine learning. In general, during the next stage of the research we intend to develop the ontology (i.e. add important locations and relations between person and location), to process all the 31 volumes of letters, to make an open database with all the persons connected to Tolstoy, and to provide every input referring to a person with a short semantically marked-up biography. The database is to be used as an interlinked reference base to Leo Tolstoy’s 90 volume edition, and also as an aggregator of exterior information from other sources. The specific syntactic constructions, that are intrinsic for the commentary genre (such as abundance of names and titles in a sentence, special means of reference inside text, etc.) will be certainly taken into account.