Muralidaran, Vigneshwaran
2022.
Automatic grammar induction from free text using insights from cognitive grammar.
PhD Thesis,
Cardiff University.
![]() Item availability restricted. |
Preview |
PDF
- Accepted Post-Print Version
Download (5MB) | Preview |
![]() |
PDF (Cardiff University Electronic Publication Form)
- Supplemental Material
Restricted to Repository staff only Download (100kB) |
Abstract
Automatic identification of the grammatical structure of a sentence is useful in many Natural Language Processing (NLP) applications such as Document Summarisation, Question Answering systems and Machine Translation. With the availability of syntactic treebanks, supervised parsers have been developed successfully for many major languages. However, for low-resourced minority languages with fewer digital resources, this poses more of a challenge. Moreover, there are a number of syntactic annotation schemes motivated by different linguistic theories and formalisms which are sometimes language specific and they cannot always be adapted for developing syntactic parsers across different language families. This project aims to develop a linguistically motivated approach to the automatic induction of grammatical structures from raw sentences. Such an approach can be readily adapted to different languages including low-resourced minority languages. We draw the basic approach to linguistic analysis from usage-based, functional theories of grammar such as Cognitive Grammar, Computational Paninian Grammar and insights from psycholinguistic studies. Our approach identifies grammatical structure of a sentence by recognising domain-independent, general, cognitive patterns of conceptual organisation that occur in natural language. It also reflects some of the general psycholinguistic properties of parsing by humans - such as incrementality, connectedness and expectation. Our implementation has three components: Schema Definition, Schema Assembly and Schema Prediction. Schema Definition and Schema Assembly components were implemented algorithmically as a dictionary and rules. An Artificial Neural Network was trained for Schema Prediction. By using Parts of Speech tags to bootstrap the simplest case of token level schema definitions, a sentence is passed through all the three components incrementally until all the words are exhausted and the entire sentence is analysed as an instance of one final construction schema. The order in which all intermediate schemas are assembled to form the final schema can be viewed as the parse of the sentence. Parsers for English and Welsh (a low-resource minority language) were developed using the same approach with some changes to the Schema Definition component. We evaluated the parser performance by (a) Quantitative evaluation by comparing the parsed chunks against the constituents in a phrase structure tree (b) Manual evaluation by listing the range of linguistic constructions covered by the parser and by performing error analysis on the parser outputs (c) Evaluation by identifying the number of edits required for a correct assembly (d) Qualitative evaluation based on Likert scales in online surveys.
Item Type: | Thesis (PhD) |
---|---|
Date Type: | Completion |
Status: | Unpublished |
Schools: | English, Communication and Philosophy |
Subjects: | P Language and Literature > P Philology. Linguistics |
Date of First Compliant Deposit: | 17 August 2022 |
Last Modified: | 11 Mar 2023 02:22 |
URI: | https://orca.cardiff.ac.uk/id/eprint/151977 |
Actions (repository staff only)
![]() |
Edit Item |