Learning logic rules from text using statistical methods for natural language processing

Official URL: http://risc01.sabanciuniv.edu/record=b1655287 (Table of Contents)

The field of Natural Language Processing (NLP) examines how computers can be made to do beneficial tasks by understanding the natural language. The foundations of NLP are diverse and include scientific fields such as electrical and electronic engineering, linguistics, and artificial intelligence. Some popular NLP applications are information extraction, machine translation, text summarization, and question answering. This dissertation proposes a new methodology using Answer Set programming (ASP) as our main formalism to predict Interpretable Semantic Textual Similarity (iSTS) with a rule-based approach focusing on hard-coded rules for our system, Inspire. We next propose an intelligent rule learning methodology using Inductive Logic Programming (ILP) and modify the ILP-tool eXtended Hyrbid Abductive Inductive Learning (XHAIL) in order to test if we are able to learn the ASP-based rules that were hard-coded earlier on the chunking subtask of the Inspire system. Chunking is the identification of short phrases such as noun phrases which mainly rely on Part-of-Speech (POS) tags. We next evaluate our results using real data sets obtained from the SemEval2016 Task-2 iSTS competition to work with a real application which could be evaluated objectively using the test-sets provided by experts. The Inspire system participated at the SemEval2016 Task-2 iSTS competition in the subtasks of predicting chunk similarity alignments for gold chunks and system generated chunks for three different Datasets. The Inspire system extended the basic ideas from SemEval2015 iSTS Task participant NeRoSim, by realising the rules in logic programming and obtaining the result with an Answer Set Solver. To prepare the input for the logic program, the PunktTokenizer, Word2Vec, and WordNet APIs of NLTK, and the Part-of-Speech (POS) and Named-Entity-Recognition (NER) taggers from Stanford CoreNLP were used. For the chunking subtask, a joint POS-tagger and dependency parser were used based on which an Answer Set program determined chunks. The Inspire system ranked third place overall and first place in one of the competition datasets in the gold chunk subtask. For the above mentioned system, we decided to automate the sentence chunking process by learning the ASP rules using a statistical logical method which combines rule-based and statistical artificial intelligence methods, namely ILP. ILP has been applied to a variety of NLP problems some of which include parsing, information extraction, and question answering. XHAIL, is the ILP-tool we used that aims at generating a hypothesis, which is a logic program, from given background knowledge and examples of structured knowledge based on information provided by the POS-tags One of the main challenges was to extend the XHAIL algorithm for ILP which is based on ASP. With respect to processing natural language, ILP can cater for the constant change in how language is used on a daily basis. At the same time, ILP does not require huge amounts of training examples such as other statistical methods and produces interpretable results, that means a set of rules, which can be analysed and tweaked if necessary. As contributions XHAIL was extended with (i) a pruning mechanism within the hypothesis generalisation algorithm which enables learning from larger datasets, (ii) a better usage of modern solver technology using recently developed optimisation methods, and (iii) a time budget that permits the usage of suboptimal results. These improvements were evaluated on the subtask of sentence chunking using the same three datasets obtained from the SemEval2016 Task-2 competition. Results show that these improvements allow for learning on bigger datasets with results that are of similar quality to state-of-the-art systems on the same task. Moreover, the hypotheses obtained from individual datasets were compared to each other to gain insights on the structure of each dataset. Using ILP to extend our Inspire system not only automates the process of chunking the sentences but also provides us with interpretable models that are useful for providing a deeper understanding of the data being used and how it can be manipulated, which is a feature that is absent in popular Machine Learning methods.