In this paper, we describe a process of extracting information from meteorological texts in Serbian. The text corpus consists of almost 46000 sentences. Having in mind the specifics of Serbian and characteristics of meteorological sublanguage, we develop a classification schema for structuring extracted information and transducers for annotating pieces of information in the text corpus. We describe the transducer for extracting information about daily temperatures and give some evaluation parameters for all other transducers used in the information extraction process.
Weather forecast reports are interesting for natural language
processing because of their properties and the possibility of
various uses of extracted data. These texts have been studied over
the years in the areas such as information extraction, text mining
or text understanding, and the obtained data were used for
machine translation from one language to another
(TAUMMETEO system developed in Canada for machine translation
from English to French and vice versa [
In this paper we present the process of extracting information about weather conditions from meteorological texts in Serbian, which can be used for different purposes (for example, for automatic creation of lexicon or annotation of texts). The main goal of this research was to provide foundations for developing electronic resources in Serbian, construction of sublanguages, ontologies, machine translation system from Serbian to English, and vice versa, and different kinds of linguistic researches in the domain of weather forecast. Some specifics of Serbian that are important for this research are presented in Section 2. The corpus of meteorological texts in Serbian, collected during 2010, 2011, and 2012 years from several sources is described in Section 3. The main goal of the extraction process was to annotate information contained in a text description. Three types of information were of interest: location, time, and meteorological phenomena. Semantic classes of information used to structure the data are described in Section 4.
The process of information extraction is presented in Section 5.
The extraction rules are defined by finite state transducers (FST)
([
Finally, we evaluate the information extraction process and give the directions for the future research.
The specific features of Serbian determine, to a great extent, approach and method that will be used for the information extraction from texts written in Serbian.
Serbian is a language with rich morphology. For example, the
most adjectives in Serbian may take more than 40 different forms.
There are algorithms for different Natural Language Processing
(NLP) tasks that have excellent results when applied to texts in
English, but very bad when it comes to texts in a language with
rich morphology, such as Serbian. The rich morphological system
of Serbian requires the use of additional linguistic resources, such
as electronic dictionaries and grammars, for text processing. Thus,
it is possible to develop systems for the information extraction
that would be efficient when applied to texts in Serbian.
This paper describes a process of extracting information from
texts in Serbian, in which the electronic dictionary for Serbian
([
Meteorological texts have been collected during 2010, 2011, and 2012 years from several sources (Republic Hydrometeorological Service of Serbia,1 the Meteos agency,2 the Politika daily news,3 B92,4 SMedia5 and Internet portal Krstarica6). The created text corpus contains 13705 text descriptions, which consist of a total of 45862 sentences.
The language used for describing weather conditions in textual reports is very specific and easily recognizable. A limited set of words from natural language, which is used to describe the meteorological phenomenon, can be treated as a sublanguage, along with its characteristics: – – – limited vocabulary – the same words are used to describe a meteorological phenomenon in almost every weather report; irregular syntax – sentences in meteorological reports typically do not contain auxiliary verb, and often do not have a predicate (“Vetar slab, jugoistočni.” – “Wind weak, southeast.”) or adverbs; text structure – it is not possible to distinguish different statements based only on punctuation, since a sentence often contains multiple statements, and a few sentences sometimes merges into one separated with commas.
On the one hand, the existence of such sublanguage facilitates the text processing, since many syntactic rules are simplified in comparison to natural language. On the other hand, it is contempt of natural language syntax rules that prevents the use of existing electronic grammars, developed and available for a given natural language.
The descriptions of weather conditions consist of smaller fragments (sentences and parts of sentences), which carry three types of information (meteorological phenomenon, location and time), combined together in a statement. Therefore, every semantic unit of the text structure (particular statement) can be treated as a triple <location, time, phenomenon>. The ideal information extraction process from the following description in Serbian “Ujutru i pre podne u nižim delovima grada magla ili sumaglica.“ (“In the morning and before the noon in the lower parts of the city fog or haze.”) would extracts the following triples: <“niži delovi grada”, “ujutru”, “magla ili sumaglica”> <“niži delovi grada”, “pre podne”, “magla ili sumaglica”> (<“the lower parts of the city”, “In the morning”, “ fog or haze”> <“the lower parts of the city”, “before the noon”, “ fog or haze”>) The statements mutually overlap in the textual descriptions, usually with no clear boundary between two different statements. This semantic structure requires a special approach, semantically oriented, in order to resolve coreferences between different parts. However, the first steps in this process are the detection and isolation of the values of individual features. This paper describes exactly this process, while merging isolated pieces of information and their values into the statements will be the subject of a future research. 1 http://www.hidmet.gov.rs 2 http://www.meteos.rs 3 http://www.politika.rs 4 http://www.b92.net 5 http://www.smedia.rs 6 http://www.krstarica.com
The information contained in the textual descriptions of weather conditions, which were of interest in the research, are grouped into semantic classes of different levels. A semantic class, together with possible additional classification, should be assigned to each separate fragment of the text. Hierarchical classes are shown in Table 1. (TTeemmppeerraattuurrea) (CKTaaettmeTgpeoemrryap)tuerrea-ture (nmaajvxiišmau,mju,tmaorrnnjiang..….) (DOTepesmicsrpTiepertmaiotupnr)eer-ature (hcloalddn,wo,atrompelirj,er,ispiongra…s)t ...
Pojava TipPojave magla, oluja ... (Phenomenon) (PhenomenonType) (fog, storm...)
ImeTeritorije Srbija, Evropa, Beograd n Teritorija (TeritoryName) ... ito (Teritory) DeoTeritorije severoistok, južni delovi aco (TeritoryPart) (northeast, southern parts) L Lokalitet Lokalitet na planinama, lokalno... (Locality) (Locality) (in the mountains, localy) Datum 15. januar (Date) (January 15th) Dan ImeDana ponedeljak, utorak .. (Day) (DayName) (Monday, Tuesday...) ieTm (DDeaoyDPaarnta) (auifjntuettrhrneuo,mopnoo)rnsliengp, oindnthee sledeće nedelje, tokom Period Period februara (Period) (Period) (next week, during February…) The names of the features, given in Table 1, are used for annotating pieces of information in the text.
The annotations had the following syntax: <Feature>text segment</Feature>
Hence, the example sentence “U većem delu zemlje promenljivo oblačno, mestimično slaba kiša, pljuskovi i grmljavina.” (“In most of the country variable cloudiness, with areas of light rain, showers, and thunder.”), should be annotated as follows: <lokalitet>U većem delu zemlje</ lokalitet> <obimOblacnosti>promenljivo</obimOblacnosti> <prisustvoOblaka>oblačno</prisustvoOblaka>, <lokalitet>mestimično</lokalitet> <obimPadavina>slaba</obimPadavina> <tipPadavina>kiša </tipPadavina>, <tipPojave>pljuskovi</tipPojave> i <tipPojave>grmljavina</tipPojave>. (<Locality>In most of the country </Locality> <CloudAmount>variable</CloudAmount> <CloudPresence>cloudiness</CloudPresence>, <Locality>with areas</Locality> of <PrecipitationAmount>light </PrecipitationAmount> <PrecipitationType>rain</PrecipitationType>, <PhenomenonType>showers</PhenomenonType> and <PhenomenonType>thunder</PhenomenonType>.)
We used transducers (FST and RTN) as extraction rules. The
transducer that describes the rule for extracting particular piece of
information was created for each feature given in Table 1. The
rules were applied through the software system UNITEX, where
the structuring of data was done by annotating text segments that
carry information. The application of transducers was performed
sequentially, one by one. The application order was not important
for the majority of created transducers, although it is possible to
organize the information extraction process so that the successive
application of transducers improves the efficiency of the process
(a cascade of transducers, one operating after the other using the
results of previously applied transducers [
Temperature data have been presented in the texts as values (12 stepeni – 12 degrees, 12°C, 12 C, dva stepena – two degrees, ispod nule – below zero, minus 5 ...) or descriptive (hladno - cold, hladnije - colder, toplo - warm, toplije – warmer, pad temperature – the temperature drop, temperatura u porastu - the temperature rising ...). For each way of representing temperature, a special extraction rule has been created. Figure 1 shows the main transducer (temperatura.grf) in the RTN for extracting information related to the temperature.
Subgraph calls are marked with gray colour. Subgraph vrednost.grf recognizes different expressions for the specific value (number of degrees) of the temperature. This subgraph is shown in Figure 2.
The lexical mask <NB> recognizes successive digits. The lexical mask <NUM> recognizes all the words in the dictionary that are marked with a code NUM (jedan, dva, tri - one, two, three, ...). Thus, this subgraph recognizes, among others, the following expressions: 10, minus dva – minus two, +5 ili jedanaest – eleven. The main transducer temperatura.grf (Figure 1) contains a subgraph call stepen.grf. This graph is intended to recognize expressions that describe the degrees on the Celsius scale, as the common unit of temperature measure, in the texts in Serbian language. Subgraph stepen.grf is shown in Figure 3.
The lexical mask which refers to a dictionary word (<stepen>) recognizes any form of the word stepen – degree (stepena, stepeni, stepenima etc.). Graph temperatura.grf recognizes the following phrases: oko +8 °C (approximately +8 ° C), - 1C, - 30 ° C,- 4 stepena (- 4 degrees), od -1 C do 1 C (from -1 C to 1 C), -1 do +3 stepena (-1 to +3 degrees), -12 do -8 (-12 to -8), od 11 do 15 stepeni (from 11 to 15 degrees), 11 stepeni (11 degrees), od pet do devet stepeni (from five to nine degrees), oko četiri (about four), ispod 0 (below 0) etc.
Similarly, for each feature in the Table 1 a rule extraction is created for annotation of the text segments that carry specific information.
The process of information extraction from the meteorological texts is in the initial phase. During this phase, the analysis of the texts from the described corpus was performed and the transducers for extracting simple features were created. Since the extraction rules are still evolving, and the text corpus over which the extraction is carried out is fairly large (45862 sentences with more than one million tokens), a comprehensive evaluation of the system’s efficiency, which would accurately assess the precision and recall, is not currently possible. However, an initial analysis of the created transducers, which would determine the directions for further development, is possible.
Table 2 lists the transducers which were used to extract information, in order of their implementation. The number of extracted text segments is shown in the third column of the table, while the evaluation of precision is presented in the fourth.
opisTemp OpisTemperature (Temperature
Description) temperature Temperatura
(Temperature) katTemp KatTemperature (Temperature-Category) vetarPre JacinaVetra (WindAmount) and PravacVetra (WindDirection) vetarPost JacinaVetra (WindAmount) and PravacVetra (WindDirection) padavine TipPadavina (PrecipitationType) and ObimPadavina (Precipitation-Amount) oblacnost ObimOblacnosti (CloudAmount) and PrisustvoOblaka (CloudPresence) deoTeritorije DeoTeritorije (TeritoryPart) imeTeritorije ImeTeritorije
(TeritoryName) lokalitet Lokalitet (Locality) pojava Pojava (Phenomenon) Number of Evaluation extracted of text precision segments 11518 100% 25618 14817 7720 99.6% 100% 100% 1559
100% 18878
100% 18875
98% 4918 6036 7623 3737 99.8% 95% 98% 100%
The high precision of the transducers is expected, given that this is an early stage of the system design and the extraction rules creation process. Further development of the process, in order to extract a larger number of individual pieces of information (i.e. to increase recall), will surely reduce the precision. However, it is expected the transducers will still maintain high efficiency. We would like to emphasize that the next step in the process, after the extraction of simple features, is merging the extracted data into classes of higher semantic level. During that process, it will be possible to further improve efficiency, by resolving ambiguities or correcting wrongly interpreted text segments.
This research was conducted through the projects 178006 and III 47003, financed by the Serbian Ministry of Science.