In the information society large amounts of information are being generated and transmitted constantly, especially in the most natural way for humans, i.e., natural language. Social networks, blogs, forums, and Q&A sites are a dynamic Large Knowledge Repository. So, Web 2.0 contains structured data but still the largest amount of information is expressed in natural language. Linguistic structures for text recognition enable the extraction of structured information from texts. However, the expressiveness of the current structures is limited as they have been designed with a strict order in their phrases, limiting their applicability to other languages and making them more sensible to grammatical errors. To overcome these limitations, in this paper we present a linguistic structure named “linguistic schema”, with a richer expressiveness that introduces less implicit constraints over annotations.
Text understanding covers a series of tasks such as document classification [
Although recognition structures are widely used, and many examples with different interpretations can be found, it is not so easy to find a specification of the language in which these linguistic structures are expressed, nor the formalization used to express the restrictions involved. Furthermore, these representations of structures have been focused more on human legibility than on machine interpretation, although computational systems need a formal form of representations to work. In fact, these systems use a formal representation, but this is implicit and has not been fully explained. For that reason, sharing the structures, defined following a specific representation, is not a trivial issue.
In this paper we present a well defined proposal of formal representation to express linguistic structures of recognition. For the purpose of this work, we have named them “linguistic schemas”, in which the meaning of all the elements appearing in the structures is made explicit. Moreover, a formal representation of these linguistic schemas, which is also interpretable by a computational application, is specified. The main aim is to provide these recognition structures with the capability of being reused and shared by different tools and systems, and to allow this formal representation to be explicit, well defined and computationally interpretable. This proposal aims at solving the complex problem of expressiveness in linguistic structures for NLP.
Thus, section 2 presents the representation specifications of linguistic structures. Section 3 offers a view of the linguistic scenario in which we can find the need for these new linguistic structures. The representation of the linguistic schemas is presented in section 4 and they are exemplified. Section 5 analyzes the expressiveness of the existing recognition structures comparing them with the new one developed and presents the results and future work. Finally references are also included.
Linguistic patterns are used in Computational Linguistics to understand natural
language texts. Among the most outstanding projects it is worth noting the program
PHRAN (PHRasal Analysis) [
In the field of information acquisition from machine readable dictionaries (MRDs),
Hearst [
More recently, the development of systems for automatic knowledge extraction has
generated a substantial amount of works focused both on representations and systems.
A detailed analysis can be found in the compilatory study by [
All in all, the lexical-syntactic patterns are generally expressed by means of
operators in the Backus-Naus Form (BNF) in order to compose regular expressions in
context-free grammars. Jacobs et al. [
root (ej. “shoot”), lexical category (ej. “adj.”) Variable assignment from pattern components: ?X =
Wild cards: $ - 0 or 1 token, * - 0 or more tokens, + - 1 or more tokens Grouping operators: <> for grouping, [] for disjunctive grouping Repetition: * - 0 or more, + - 1 or more Range: *N - 0 to N, +N - 1 to N
Linguistic patterns, be they lexical-syntactic, semantic or, as in the case of
PALKA, structures of phrase frames, are always ordered sets of components that
express characteristics or constraints on the phrase elements. In every case, the phrase
element order and the pattern component order will be the same, even if not explicitly
indicated, as all of them are patterns for English, a language with a strict phrase order
[
This problem is partially solved by Hazez [
Linguistic patterns based on annotations can be found in other cases, as in Specia
and Motta’s work [
The purpose of this work is the understanding of Spanish texts annotated electronically by software tools. In order to enable the automatic application of these patterns to large scale corpora, we have established some constraints over the phrases in several levels, specifically in orthographical, morpho-syntactic, and syntactic levels. However, the possibility of including annotations of any other level (such as semantic, pragmatic or discursive) remains open.
As for the works about annotation and creation of linguistic patterns to extract
information from texts in Spanish, the initiatives grow in number and importance as
the multilinguality significance increases in the Internet. In the framework of the
European project SEKT1, one of the use cases was focused on the Spanish legal
terminology for the creation of ontologies in the legal domain. For this task Hearst’s
taxonomic relation patterns [
Related with knowledge extraction for ontology enrichment and population in
Spanish we can find another classification attempt in Álvarez de Mon y Rego y
Aguado de Cea [
Nica’s et al. [
We decided to represent linguistic structures in XML format to work computationally with these structures in an easier way as XML is the language most widely used for knowledge representation and many tools can process it. However, files in XML cannot be easily read by humans because of the verbosity of its syntax. 1 http://www.sekt-project.com/ A linguistic schema is a set of constraints over the tokens of a phrase (token contraints) and over the relations between these tokens (phrase constraints). Token constraints are expressed as a set of values of characteristics of annotations of a token. Phrase constraints are expressed using operators (optimality, grouping, etc.) over token constraints or other phrase constraints.
As previously stated, the complete representation of the schemas is stored in XML files for an easier computational processing. Although these files can be read by a person, this task is rather tedious and can be untractable if the number or size of the schemas grows significantly.
For this reason a shortened and user-friendly annotation is defined. This annotation may serve as a mnemonic of the schemas that appear in a file. It does not comply with the XML conventions and may not contain all the information available in the schema. However, the annotation is much easier to read, and, if used correctly, it may identify the schema that is referred to without any short of ambiguity.
Furthermore, this notation has been extended with additional operators, which are not present in the XML notation, to increase the expressiveness and improve the shortness. These operators are replaced by combinations of the operators available in the XML notation. As an example, the optionality operator (see section 4.2) applied to a token would be replaced with a disjunction between this token and the negation of the same token.
A brief summary of the notation proposed (for a friendly representation) is: Token constraints (Elements): constant (ej. “shirt”), identifier (ej. “ANIMAL”)
Order operators: A ⊕ B – A appears before B, A + B – A appears immediately before B
Disjunction operators: A | B – A and B can appears, A / B – A or B can appear
For the purposes of this work, a term in a linguistic schema is the set of constraints, in other words, the set of elements applied to one single token. In the user-friendly syntax, these terms may be displayed with two different types of symbols: constants and identifiers. • Constants are words written as they appear in the text, for example clasifica. • Identifiers are used to retrieve values instead of restricting them, and they appear as strings in uppercase, for example “ACTOR”.
In those cases in which an identifier ("ACTOR") or a lemma (clasificar) is specified this will be shown in the set of constraints of a token.. For example, when a token has as a constraint the lemma clasificar, and its morpho-syntactic value is “main verb”, only the lemma will be shown. If both data about the same term are used in the information, then the identifier will be shown. For example, when a token has as a constraint the lemma “clasificar” and as text the identifier “CONJUGATED_FORM”, then only the identifier “CONJUGATED_FORM” will be shown. If a set of identifiers is specified for a token, then the identifier whose value has previously appeared will be used, according to the annotation standard used. For instance, if two identifiers are assigned to a token, such as the values of gender (“GENDER”) and syntactic function (“FUNCTION”), only the former will be shown, i.e., “GENDER”.
It is possible to use the identifier to refer to any of the non constant terms. For instance, the next schema can be written using the identifiers A, B, C and Z: A + come + B + y + C + en + Z
This schema would match a phrase such as “Pepe come pan y chocolate en el patio de la escuela”, and in this matching the identifiers will take the values corresponding to this specific phrase: A=Pepe, B=pan, C=chocolate, Z=patio.
For those terms for which no identifier or lemma are specified, but the value of, at least, a category in an annotation level is defined, the name of that category will be shown. Taking as reference the previous example, the values “verb” or “direct object” will appear instead of “come” and “B”.
If an abbreviated form is specified for any category in the standard used2 and possibly with information about additional attributes (for example “Fused_Prep-At” for “Fused Preposition-Article”), then the most specific abbreviated form will be shown for each annotation level, being the most specific form the one that includes more information about the additional attributes. Thus, when we want to identify a token that is an ordinal pronoun, but of which we do not want to obtain any other information, its lemma or value for any other category (as in the phrase “el primero es el grande”) is described as “Ordinal_pronoun” as we only want to restrict this word to this type of pronoun.
As previously mentioned, operators define the relations among the different parts of a schema. It is necessary to point out that the order in which the parts of the phrase must appear is not specified by the element appearance order in the schema; therefore, if it is necessary to set this order, then it must be specified explicitly. This can be done with two symbols: With the symbol ‘+’: the expression “symbol1 + symbol2” means that “symbol2” must appear immediately after “symbol1”. 2 http://pln.oeg-upm.net/annotation/ontotag With the symbol ‘⊕’: the expression “symbol1 ⊕ symbol2” means that “symbol2” must appear after “symbol1”, immediately or not.
There are other symbols besides the previous ones which express different relations. These symbols are the following: ‘*’: expresses repetition.
For example, “symbol*” means that “symbol” may appear more than once. ‘(’ and ‘)’: groups several symbols.
For example, “(symbol1 + symbol2)*” means that “symbol1” may appear several times, all of them followed by “symbol2”. ‘[’ and ‘]’: means that whatever is between both square brackets is optional. For example, “[symbol]” means that “symbol” may appear or not in the phrase. ‘|’: means that either what is in the left side or what is in the right side must appear.
For example “a|an” means that “a” or “an” must appear. ‘/’: means that either what is on the left side or what is on the right side must appear, but not both of them.
For example “a/an” means that “a” or “an” must appear, but not “a” and “an” at the same time. ‘¬’: means that the next element must not appear in the phrase. When combined with the symbols + and ⊕, it may indicate that the said symbol must not appear in some specific positions of the phrase.
For example, “¬symbol” means that “symbol” may not appear in the phrase.
To show the versatility and possibilities of the linguistic schemas we include some examples, expressed in the user-friendly notation.
We want to identify who buys things to María, and which those things are. Hence, we express these constraints in a linguistic schema setting the main verb (“compra”) and the indirect object (“a María”). The rest of the phrase and the order of appearance are not restricted. These constraints may be expressed with the next linguistic schema: X compra Y a + María
This schema would match phrases like “Pepe compra a María flores”, “Pepe a María flores compra”, “Pepe compra flores a María en domingo”, “A María Pepe le compra flores” and “Juan a María compra bombones de licor en Santander”.
It is worth mentioning that the previous schema would be equivalent to the next one, since linguistic schemas have no implicit order, as it happens in the case of lexical-syntactic patterns. Also, the name assigned to the identifier does not change the recognition capabilities of a linguistic schema:
SOMEONE a + María SOMETHING compra
This schema would match with exactly the same phrases as the previous one. However, it would take 24 lexical-syntactic patterns (P(4,4) = 4! = 24) to match the same phrases using patterns, as there are four pattern components in the previous example, (1) SOMEONE, (2) a+María, (3) SOMETHING and (4) compra. Moreover, these patterns could also have additional elements in the phrase, resulting in a larger list of lexical-syntactic patterns.
Because of this combinatorial explosion and the open world assumption, processing a schema requires more computational power than a pattern. However, our proposal for a schema represents a set of patterns in a more compact way, enabling a further optimization and more efficient algorithms.
The application of these linguistic schemas to Spanish does not mean that they cannot be used for other languages. For the lexical-syntactic pattern
X buys Y for María the equivalent linguistic schema would be: X + buys + Y + for + María
An example can be seen in pln.oeg-upm.net/process/linguisticschemas.
Once we have described and exemplified the notation proposed, we will compare the
expressiveness of our notation with the lexical-syntactic pattern notation, accepted by
Jacobs et al.[
The first point is that the notation we propose assumes the open world assumption. This assumption means that everything that is not described in the schema is not restricted, thus, it can appear or not.
As future work, the implementation of an assistant (already designed) for editing schemas will make linguists work easier and will contribute to a greater automatization.
The assistant should allow the definition of many schemas comfortably. Presumably, this combination of quality and quantity should allow a greater automation of NLP tasks, improving the results when processing large scale corpora.
This work has been supported by the project BabeLData (TIN-2010-17550), funded by INIA, under Plan Nacional de Investigación Científica, Desarrollo e Innovación Tecnológica (I+D+i) of Ministerio de Ciencia e Innovación.