HPSG-Based Syntactic And Semantic Parsing

Syntactic and semantic parsing has become one of the hot research topics of modern linguistics. Syntactic parsing is aimed at obtaining the syntactic expressions of input sentences. Semantic parsing is aimed at obtaining the semantic expressions of input sentences. The main approaches of parsing include parsing as derivation with rewriting rules and the more recent new approach of parsing as deduction with logic. HPSG (Head-Driven Phrase Structure Grammar) was forwarded by Polland and Sag in 1994. HPSG adopts unification as the basic operation for syntactic and semantic parsing, which is ideal. Due to the distance between linguistic theories and actual application of natural languages as well as the fact that the specific recognition processes are not included in the grammars overtly, parsing should comprise grammars, algorithms and other representation methods. However, for parsing, there are problems for some logic deductive systems like Categorial Minimalist Grammars (CMG), the glue semantics for HPSG, as well as some HPSG parsing algorithms. For example, sometimes the syntax-semantics interface for sentences involving multi-quantifier could not be established; some kind of sentence attribute passing was missing; information in the parsing processes was not all recorded; and the normalization of semantic representation of UDC sentences still needs improving.Aimed at the above problem, our work from the following five perspectives is carried out in this dissertation:Chapter Three forwards HDS. HDS is a HPSG deductive system as a fragment of pCLL (partially commutative linear logic). The features of HDS are: 1. We keep the type psoa (parameterized state of affairs) of HPSG that the glue semantics approach removes. 2. HDS handles the Subcategorization Principle, the Head-Feature Principle, the Trace Principle, and the Semantic Principle of HPSG. HDS is composed of general lexical entries and ten deductive rules. Lexical entries are actually axioms in HDS. They are composed of the PHONOLOGY attribute, type, andλ?term. Lexical entries also represent the Curry-Howard correspondence between types and PHONOLOGY attributes. Rules represent the deduction from antecedents to conclusions. The HDS rules include the introduction rule and elimination rule of the logic implication symbol, the composition law of products, and an entropy rule which may relax the order of the discharging of hypotheses. The function of HDS is that through parsing as logic deduction, we are able to establish syntactic objects and semantic objects, as well as the syntax-semantics interface.Application example illustrates that HDS is able to solve the following problem that CMG and the glue semantics for HPSG sometimes cannot establish the syntax-semantics interface for sentences involving multi-quantifier.Chapter Four forwards an HPSG-deductive-system-based automatic parsing algorithm and solves the parsability problem for HDS. Firstly, we raise"the simultaneous substitution of variables by the product"to the mechanism of"extension through projection of the product."Based on this mechanism, Chapter Four forwards the elimination rule [?E] of HDS for n-component product ? (n≥2), and thus enlarge the range of application of HDS. Then we forward the automatic parsing algorithm for HDS. HDS itself is a logical deductive system and it does not overtly include concrete syntax-semantics parsing deduction processes. The automatic parsing algorithm for HDS can realize the concrete deductive process for HDS. Our algorithm has the features that it is type-driven, deduces inductively, and can carry out local deductions without unnecessary repetitions. The advantages are that it can reduce the number of deductions since once local deductions are built by the algorithm, they don't have to be built repeatedly. That is to say a local deduction established will become a component of many later local deductions. Application example shows that it is applicable.Chapter Five proposes a new HPSG attribute parsing algorithm, which is based on the inductive mechanism for the attribute dependencies in the attributed structure trees.1. Based on the attribute grammars, we define the attribute production rules for context-sensitive grammars. We define the attribute dependencies which suit various situations. And we give the inductive definition of attribute dependencies induced in the syntactic attributed tree. The feature of the parsing algorithm is that it has inductive mechanism T-IDP (induced attribute dependencies of structure tree), which induces the general conclusion of the whole attributed structure tree out of separate instances of attribute dependencies.2. We forward the T-IDP-based HPSG attribute parsing algorithm. An application example shows that the algorithm is applicable. The time complexity of the T-IDP-based HPSG Parsing Algorithm is O(n3), and thus the algorithm is efficient.Our work can solve the following problem. Some HPSG parsing algorithm do not consider the cross-schemata attribute passing paths or bounded attribute-passing paths which exist in certain natural language structures.Chapter Six presents a Predicate/Transition Petri-Net (Pr/T-net) which combines the function of HPSG attribute parsing and the function of recording the parsing information.The Pr/T-net consists of two parts. The first part analyzes the structure of the input string. The second part processes the attribute relations among the components of the input string, establishes the corresponding attributed structure tree, and records the information during the parsing processes. The Pr/T-Net is composed of various tokens, places, and transitions. It is featured by rather complicated data structure carried by individualized tokens, the information stored in the places of dynamic predicates, as well as the processing and analysis conducted by transitions directed by conditions. And thus it is able to carry out HPSG syntactic/attribute parsing, record the rich context-sensitive information in the parsing process, and answer the questions concerning syntactic/attribute structures. We give an application example and the result shows that the Pr/T-Net model can do HPSG parsing and record parsing information successfully.The Pr/T-Net formalism can overcome the limitations of some other currently available formalisms such as pushdown automata and BFSA, which were not powerful enough to process the complex attribute relations of HPSG, or could not record all the information during the parsing process.Chapter Seven forwards the fusion approach of formal semantics (FAFS). This approach combinesλ?abstraction,λ?calculus, and HPSG (Head-Driven Phrase Structure Grammar).FAFS is aimed at solving the problem of obtaining the normalized forms which represent the UDC sentence meaning. Firstly, employing the context-sensitive attribute grammar of HPSG, FAFS establishes the S-structures of UDCs, distinguishes between strong and weak UDCs based on HPSG syntactic attributed structure trees, and extracts the corresponding fillers for the traces. Then FAFS obtains the WFF (well-formed formulas) of UDCs according to the HPSG parse trees. With the result of HPSG parsing and filler extraction, as well as the WFF of UDCs, FAFSλ?abstracts the PHON labels of the sentence, carries outλ?application to the fillers, and doesλ?calculus to derive the normal forms of semantic representation of UDC sentences. Here in the identical intuitionistic explanation of sentences, a normal form is a simplest form. The overall computational efficiency of FAFS is in the cubic time, and thus FAFS is efficient. Finally, application examples show that FAFS is feasible.FAFS can overcome the limitation of rewriting rules which cannot fill the corresponding fillers into the trace positions in the UDC sentence, and cannot fill the corresponding co-indexed words, which do not appear in the sentence, into the trace positions.The innovative points of this dissertation are summarized as follows:(1) Extending HPSG towards HDS as a fragment of partially commutative linear logic as well as forwarding the automatic parsing algorithm for HDS.(2) A new HPSG attribute-parsing algorithm with the inductive mechanism for attribute dependencies.(3) HPSG-predicate/transition Petri Net (HPSG-Pr/T-Net).(4) A new approach combiningλ-abstraction,λ-calculus and HPSG.