C. Lyon. The Representation of Natural Language to Enable Neural Networks to Detect Syntactic Structures. Phd. Thesis, Computer Science Department, University of Hertfordshire, UK, 1994.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....algorithm, but also include some examples of recurrent networks and ensembles of several single neural networks. Other examples addressing more complex problems, sometimes in combination with symbolic approaches, are: Identification of clause boundaries [95] parsing and sentence analysis [115, 43, 128, 129], grammatical inference [111] PP attachment disambiguation [218, 125] WSD [224] text categorization [233] and detecting spelling errors [116] In [125] there is a present day survey of neural networks with application to NLP 5 . 2.3.2 Genetic Algorithms Genetic Algorithms [87, 219] have ....

....semantic, etc. of language acquisition. 12 DTs HMMs ME IBL Clause Boudaries [181] Shallow Parsing [45, 1, 16, 212] 211] 7, 227, 33, 58] Parsing [12, 132, 92] 178] 210, 37, 36, 38] PP attachment disambiguation [180] 246] TBL NB NNs LSM EC Clause Boudaries [95] Shallow Parsing [18] [128, 129] [130] 223] Parsing [115, 43] 93] PP attachment disambiguation [20] 52] 125, 218] 107] 2] Table 2: References corresponding to syntactic analysis and structural ambiguity NLP problems DLs DTs NB TBL EM WSD [240, 150] 26, 150] 86, 150, 112] 67] 203, 166] Text categorization and ....

C. Lyon. The Representation of Natural Language to Enable Neural Networks to Detect Syntactic Structures. Phd. Thesis, Computer Science Department, University of Hertfordshire, UK, 1994.

....syntactic constituents The advantages of using neural methods Data driven, neural methods bear comparison with well used stochastic approaches. Neural nets, however, can capture more of the implicit information in the training data since they can model negative as well as positive relationships [2, 4]. Stochastic methods have no obvious way of representing negative relationships. There are negative correlations in natural language that are an important source of information [5, page 80] We need to distinguish between examples that are never going to be correct and those that just have not ....

....the natural language field, where data is often found to have a Zipfian distribution (see Section 3) Neural methods also have the advantage that training is done in advance, so the run time computational load is comparatively low. 2 Data Representation A full description of this process is in [4]. Further details are also found in [2] The innards of the system can be seen in the experimental prototype accessible via telnet (details from the authors) Linguistic data is converted to a set of binary input vectors for each sentence, one of which will represent the desired parse, in the ....

[Article contains additional citation context not shown here]

C Lyon. The representation of natural language to enable neural networks to detect syntactic structures. PhD thesis, University of Hertfordshire, 1994.

....the string with the correct placement. This paper gives an overview of how natural language is converted to a representation that the neural nets can handle, and how the problem is reduced to a manageable size. It then outlines the neural net selection process. A comprehensive account is given in Lyon (1994); descriptions of the neural net process are also in Lyon (1993) and Lyon and Frank (1992) This is a hybrid system. The core process is data driven, as the parameters of the neural networks are derived from training text. The neural net is trained in supervised mode on examples that have been ....

....of this number: if the classifier marked all strings incorrect the percentage wrongly classified would only be 0:03 , yet it would be quite useless. In order to find the correct string most of the outside candidates must be dropped, The skeletal grammatic framework A minimal grammar, set out in Lyon (1994) in EBNF form, is composed of 9 rules. For instance, the subject must contain a noun type word. Applying this particular rule to sentence (3) above would eliminate candidate strings (3.1) and (3.2) We also have the 2 arbitrary limits on length of pre subject and subject. There is a small set of 4 ....

[Article contains additional citation context not shown here]

C Lyon. 1994. The representation of natural language to enable neural networks to detect syntactic features. PhD thesis.

....for a formal language [8] We suggest that they can also be used to evaluate representations for natural language. 4 Automatic subject location In this section we give an overview of the ALPINE parser, designed to locate the subject of a declarative sentence automatically. For more details see [1, 9, 10, 11]. The principle on which this hybrid system is based is to generate all possible parses, prune them, then pick the right one out of the remainder with a neural network. Currently there are arbitrary limits on the length of the subject (12 words) and the length of the pre subject (15 words) The ....

....generation of strings is limited by (i) the grammatical framework and (ii) local and semi local constraints. The grammatical framework first asserts that the sentence can be decomposed into consecutive segments. It then expands each segment in a very shallow analysis, based on an EBNF formalism [11]. The local and semi local constraints are derived from the ideas of Barton et al. 12] Applying these constraints the generation of any string is zapped if a prohibited feature is produced. An example of a local prohibition is that the adjacent pair (verb, verb) is not allowed. Of course ....

C Lyon. The representation of natural language to enable neural networks to detect syntactic structures. PhD thesis, University of Hertfordshire, 1994.

....a limited number of part of speech tag classes. This also make syntactic patterns more pronounced. Devising optimal tagsets is a significant task, on which further work remains to be done. For the purpose of this paper we take as given the tagsets used in the demonstration prototype, described in [12]. At the stage of processing described in this paper 19 tags are used. 3.3 Grammatical structure There is an underlying hierarchical structure to all natural languages, a phenomenon that has been extensively explored. Sentences will usually conform to certain structural patterns, as is shown in ....

....a start symbol and 2 hypertags we have 22 tags in all. Thus, there are potentially 22 2 22 3 = 11132 pairs and triples. In practice only a small proportion of tuples are actually realised see Tables 3 and 4 . At other stages of the parsing process larger tagsets are required (see [12]) 5.4 Rule based pruning: the Prohibition Table Strings can potentially be generated with the hypertags in all possible positions, in all possible sequences of ambiguous tags. However, this process would produce an unmanageable amount of data, so it is pruned by rule based methods integrated ....

[Article contains additional citation context not shown here]

C Lyon. The representation of natural language to enable neural networks to detect syntactic structures. PhD thesis, University of Hertfordshire, 1994.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC