Elizabeth Jasmi George received her MS-Dual Degree in Computer Science and Engineering (CSE). Her thesis presents three linguistically motivated approaches that leverage the dependency trees to (i) compute inferences, (ii) summarize, and (iii) comprehend documents.Elizabeth Jasmi George wrote the thesis as the culmination of her MS by research in Computer science and Engineering, based on her research in NLP at Language Technologies Research Centre (LTRC) from December 2016 – July 2020 under the supervision of Dr. Radhika Mamidi. Her thesis reviewed by Prof. Rajeev Sangal and Dr. Kavita Vemuri.
Elizabeth’s M.S thesis, Leveraging Dependency Structure for Inference Computation Summarization , and comprehension presents an introductory dataset to aid the generation of inferences from a response in a context. The research paper associated with this data collection work is published by Elsevier, in Procedia Computer Science Journal, Volume 171, 2020, Pages 2316-2323, https://doi.org/10.1016/j.procs.2020.04.251. The dataset can be accessed from https://figshare.com/articles/Implicature%20dataset/10315505 The three computational approaches of inference computation, summarization and machine reading comprehension are accepted for publication by Springer. The inference computation work is published as Chapter 36 in the book “Computational Linguistics” https://doi.org/10.1007/978- 981-15-6168-9_36, and the other two conferences are yet to happen as virtual oral presentations in September 2020. Inferences help find the indirect meanings of human utterances. The inference computation is becoming a popular task in conversational AI. Her research started with computing the pragmatic inferences such as (i) presuppositions (ii) conventional implicatures and (iii) conversational implicatures and later spanned to the tasks of summarization and machine reading comprehension with a similar approach of utilizing dependency structures. Computing inferences from an utterance will reduce the conversational failures (Refer to Figure 1) happening while humans try to interact with the virtual assistant as if it is another human.
Figure 1: Illustration of conversational failures in a chat with a virtual assistant
Inferences such as presuppositions and conventional implicatures depend on the syntactic structure of utterances. To compute these inferences, she formulated a method of generating inferences from the written representation of utterances based on the utterance’s syntax. This method works by iterating through the dependency tuples obtained from the Stanford dependency parser. Elizabeth introduced an extractive summarization method, achieved by attaching the dependency trees of the sentences in the document. The advantage of this low-resource summarization method is the smooth construction of the graph representation for any document from the tuples obtained from the parser. She also attained the right answers from passages, for free form natural language questions using a syntax-based machine reader focusing on the verbs in the questions and the context passage. This method works by comparing the lemmatized and synonymous form of verbs in the question and the passage. This method is useful in processing action-oriented genres like news where many statements contain a main verb.
The thesis can be accessed from the following link. http://web2py.iiit.ac.in/research_centres/publications/view_publication/mastersthesis/875