Natural Language Learning

This thesis is a small step towards automated learning of natural languages. With the use of a parser that incorporates machine-learning algorithms, our algorithm is able to learn mean-ings of words representing relations in simple sentences, that describe relative positions of two points on a 2D plane. Our SentenceLearner program can create simple sentences describing rela-tions between two points on another 2D plane using data, collected by a statistical parser from sentences given for training, based on n-grams of five words. In this thesis I show that association of simple relations expressed in training sentences with the positional relations of a corresponding pair of points on a 2D plane is possible without the use of any machine-learning algorithm in some circumstances.

Natural Language Learning

This thesis is a small step towards automated learning of natural languages. With the use of a parser that incorporates machine-learning algorithms, our algorithm is able to learn mean-ings of words representing relations in simple sentences, that describe relative positions of two points on a 2D plane. Our SentenceLearner program can create simple sentences describing rela-tions between two points on another 2D plane using data, collected by a statistical parser from sentences given for training, based on n-grams of five words. In this thesis I show that association of simple relations expressed in training sentences with the positional relations of a corresponding pair of points on a 2D plane is possible without the use of any machine-learning algorithm in some circumstances.

CustNER

This article describes CustNER: a system for named-entity recognition (NER) of person, location, and organization. Realizing the incorrect annotations of existing NER, four categories of false negatives have been identified. The NEs not annotated contain nationalities, have corresponding resource in DBpedia, are acronyms of other NEs. A rule-based system, CustNER, has been proposed that utilizes existing NERs and DBpedia knowledge base. CustNER has been trained on the open knowledge extraction (OKE) challenge 2017 dataset and evaluated on OKE and CoNLL03 (Conference on Natural Language Learning) datasets. The OKE dataset has also been annotated with the three types. Evaluation results show that CustNER outperforms existing NERs with F score 12.4% better than Stanford NER and 3.1% better than Illinois NER. On another standard evaluation dataset for which the system is not trained, the CoNLL03 dataset, CustNER gives results comparable to existing systems with F score 3.9% better than Stanford NER, though Illinois NER F score is 1.3% better than CustNER.

Is regularisation uniform across linguistic levels? Comparing learning and production of unconditioned probabilistic variation in morphology and word order.

Languages exhibit variation at all linguistic levels, from phonology, to the lexicon, to syntax. Importantly, that variation tends to be (at least partially) conditioned on some aspect of the social or linguistic context. When variation is unconditioned, language learners regularise it—removing some or all variants, or conditioning variant use on context. Previous studies using artificial language learning experiments have documented regularising behaviour in learning of lexical, morphological, and syntactic variation. These studies implicitly assume that regularisation reflects uniform mechanisms and processes across linguistic levels. However, studies on natural language learning and pidginisation suggest that morphological and syntactic variation may be treated differently. In particular, there is evidence that morphological variation may be more susceptible to regularisation (Good 2015;Siegel 2006; Slobin 1986). Here we provide the first systematic comparison of the strength of regularisation across these two linguistic levels. In line with previous studies, we find that the presence of a favoured variant can induce different degrees of regularisation. However, when input languages are carefully matched—with comparable initial variability, and no variant-specific biases—regularisation can be comparable across morphology and word order. This is the case regard-less of whether the task is explicitly communicative. Overall, our findings suggest an overarching regularising mechanism at work, with apparent differences among levels likely due to differences in inherent complexity or variant-specific biases. Differences between production and encoding in our tasks further suggests this overarching mechanism is driven by production