Jean Beney

Multi-classification of Patent Applications with Winnow

The Winnow family of learning algorithms can cope well with large numbers of features and is tolerant to variations in document length, which makes it suitable for classifying large collections of large documents, like patent applications.

Phrase-based document categorization revisited

This paper takes a fresh look at an old idea in Information Retrieval: the use of linguistically extracted phrases as terms in the automatic categorization (aka classification) of documents. Until now, there was found little or no evidence that document categorization benefits from the application of linguistics techniques. Classification algorithms using the most cleverly designed linguistical representations typically do no better than those using simply the bag-of-words representation. Shallow linguistical techniques are used routinely, but their positive effect on the accuracy is small at best. We have investigated the use of dependency triples as terms in document categorization, which are derived according to a dependency model based on the notion of aboutness. The documents are syntactically analyzed by a parser and transduced to dependency trees, which in turn are unnested into dependency triples following the aboutness-based model. In the process, various normalizing transformations are applied to enhance recall. We describe a sequence of large-scale experiments with different document representations, test collections and even languages, presenting evidence that adding such triples to the words in a bag-of-terms document representation may lead to a significant increase in the accuracy of document categorization.

On the importance of parameter tuning in text categorization

Text Categorization algorithms have a large number of parameters that determine their behaviour, whose effect is not easily predicted objectively or intuitively and may very well depend on the corpus or on the document representation. Their values are usually taken over from previously published results, which may lead to less than optimal accuracy in experimenting on particular corpora. In this paper we investigate the effect of parameter tuning on the accuracy of two Text Categorization algorithms: the well-known Rocchio algorithm and the lesser-known Winnow. We show that the optimal parameter values for a specific corpus are sometimes very different from those found in literature. We show that the effect of individual parameters is corpus-dependent, and that parameter tuning can greatly improve the accuracy of both Winnow and Rocchio. We argue that the dependence of the categorization algorithms on experimentally established parameter values makes it hard to compare the outcomes of different experiments and propose the automatic determination of optimal parameters on the train set as a solution.

Phrase-Based Document Categorization

This chapter takes a fresh look at an old idea in Information Retrieval: the use of linguistically extracted phrases as terms in the automatic categorization of documents, and in particular the pre-classification of patent applications. In Information Retrieval, until now there was found little or no evidence that document categorization benefits from the application of linguistic techniques. Classification algorithms using the most cleverly designed linguistic representations typically did not perform better than those using simply the bag-of-words representation. We have investigated the use of dependency triples as terms in document categorization, according to a dependency model based on the notion of aboutness and using normalizing transformations to enhance recall. We describe a number of large-scale experiments with different document representations, test collections and even languages, presenting evidence that adding such triples to the words in a bag-of-terms document representation may lead to a statistically significant increase in the accuracy of document categorization.

STARLET: an affix-based compiler compiler designed as a logic programming system

We present STARLET, a new compiler compiler which compiles Extended Affix Grammars defining a translation into an executable program : the translator. We look at its operational semantics and we focus on the points which are close to or different from Prolog procedural semantics. We discuss the two interwoven issues which are Program Reliability (due to many static checks) and Program Efficiency (optimizations at compile time). Both are reached through a systematic use of grammatical properties.

On the borderline between grammars and programs

We describe some of the engineering considerations and trade-offs in the design of a new Compiler Description Language, CDL3. The language is based on Extended Affix Grammars, where the affix rules are used to define tree types. The execution model is deterministic and depth-first, except that part of the work can be delayed until a second pass over the implicit parse tree. It is checked statically whether the program can indeed be executed in two passes without backtracking. A simple module structure allows separate compilation in a safe way.