Grammar

Revision as of 21:10, 21 September 2012

In the UNL framework, a grammar is a set of rules that are used to generate UNL out of natural language, and natural language out of UNL. Along with the UNL-NL dictionaries, they constitute the basic resource for UNLization and NLization.

Networks, Trees and Lists

Natural language sentences and UNL graphs are supposed to convey the same amount of information in different structures: whereas the former arranges data as an ordered list of words, the latter organizes it as a network. In that sense, going from natural language into UNL and from UNL into natural language is ultimately a matter of transforming lists into networks and vice-versa.

The UNL framework assumes that such transformation can be carried out progressively, i.e., through a transitional data structure: the tree, which could be used as an interface between lists and networks. Accordingly, there are seven different types of rules (LL, TT, NN, LT, TL, TN, NT), as indicated below:

• ANALYSIS (NL-UNL)
  • LL - List Processing (list-to-list)
  • LT - Surface-Structure Formation (list-to-tree)
  • TT - Syntactic Processing (tree-to-tree)
  • TN - Deep-Structure Formation (tree-to-network)
  • NN - Semantic Processing (network-to-network)

• GENERATION (UNL-NL)
  • NN - Semantic Processing (network-to-network)
  • NT - Deep-Structure Formation (network-to-tree)
  • TT - Syntactic Processing (tree-to-tree)
  • TL - Surface-Structure Formation (tree-to-list)
  • LL - List Processing (list-to-list)

The NL original sentence is supposed to be preprocessed, by the LL rules, in order to become an ordered list. Next, the resulting list structure is parsed with the LT rules, so as to unveil its surface syntactic structure, which is already a tree. The tree structure is further processed by the TT rules in order to expose its inner organization, the deep syntactic structure, which is supposed to be more suitable to the semantic interpretation. Then, this deep syntactic structure is projected into a semantic network by the TN rules. The resultant semantic network is then post-edited by the NN rules in order to comply with UNL standards and generate the UNL Graph.

The reverse process is carried out during natural language generation. The UNL graph is preprocessed by the NN rules in order to become a more easily tractable semantic network. The resulting network structure is converted, by the NT rules, into a syntactic structure, which is still distant from the surface structure, as it is directly derived from the semantic arrangement. This deep syntactic structure is subsequently transformed into a surface syntactic structure by the TT rules. The surface syntactic structure undergoes many other changes according to the TL rules, which generate a NL-like list structure. This list structure is finally realized as a natural language sentence by the LL rules.

As sentences are complex structures that may contain nested or embedded phrases, both the analysis and the generation processes may be interleaved rather than pipelined. This means that the natural flow described above is only "normal" and not "necessary". During natural language generation, a LL rule may apply prior to a TT rule, or a NN rule may be applied after a TL rule. Rules are recursive and must be applied in the order defined in the grammar as long as their conditions are true, regardless of the state.

Types of rules

Main article: Grammar Specs

In the UNL framework there are two basic types of rules:

• Transformation rules, or T-rules, are used to manipulate data structures, i.e., to transform lists into trees, trees into lists, trees into networks, networks into trees, etc. They follow the very general formalism

α:=β;

where the left side α is a condition statement, and the right side β is an action to be performed over α.

• Disambiguation rules, or D-rules, are used to improve the performance of transformation rules by constraining or forcing their applicability. The Disambiguation Rules follows the formalism:

α=P;

where the left side α is a statement and the right side P is an integer from 0 to 255 that indicates the probability of occurrence of α.

Direction

In the UNL framework, grammars are not bidirectional, although they share the same syntax:

• UNL-NL T-Grammar: used for natural language generation
• UNL-NL D-Grammar: used for improving the results of the UNL-NL T-Grammar
• NL-UNL T-Grammar: used for natural language analysis
• NL-UNL D-Grammar: used for tokenization and for improving the results of the NL-UNL T-Grammar

Units

In the UNL framework, grammars may have different processing units:

• Text-driven grammars operate over texts and process the source document as a single unit
• Sentence-driven grammars operate over sentences and process each sentence or graph separately
• Word-driven grammars operate over words and process each word or node separately

Text-driven grammars are normally used in summarization and simplification, when the rhetorical structure of the source document is important. Sentence-driven grammars are used mostly in translation, when the source document can be treated as a list of non-semantically related units, to be processed one at a time. Word-driven grammars are used in information retrieval and opinion mining, when each word or node can be treated in isolation.
All these grammars share the same type of rule.

Recall

Grammars may target the whole source document or only parts of it (e.g. main clauses):

• Chunk grammars target only a part of the source document
• Full grammars target the whole source document

Precision

Grammars may target the deep or the surface structure of the source document:

• Deep grammars focus on the deep dependency relations of the source document
• Shallow grammars focus only on the surface dependency relations of the source document