Exploratory Collocation Extraction

نویسنده

  • Stefan Evert
چکیده

Lexical collocations are a fuzzy phenomenon that has not yet been satisfactorily explained by linguistic theory. At the same time, they are important both for understanding the structure of language and for many applications such as lexicography and natural language processing. Corpus-based studies of collocations as well as collocation extraction tools have been influenced by two basic views: (a) An empirical notion of lexical collocations as recurrent combinations of words, which developed from the ideas of Firth (1957). Proponents of this view are typically interested in studying sets of collocations extracted from a corpus. Since Firth was mostly concerned with semantically motivated cooccurrences (such as “dark”/“night” and “milk”/“cow”) that provide information about the objects and concepts in the world and their properties, collocation extraction is based on spans of a few tokens (or a full sentence) around the instances of a given keyword. (b) A phraseological notion of collocations as pre-constructed syntactic units (Grossmann & Tutin 2003) or lexically determined elements of grammatical structures (e.g. Choueka 1988), which is prevalent in the lexicographic treatment of word combinations and in computational linguistics. In this view, collocations are characterised by semantic, syntactic and distributional irregularity (cf. Manning & Schütze 1999: 184), i.e. by intrinsic properties of the word combinations rather than their actual occurrences in corpora. The goal of such approaches is to extract a specific type of collocation – according to an intensional definition – with high precision and recall. In order to improve accuracy, it is common to consider only words that cooccur in a specific syntactic relation (e.g. verb+object), based on a (partial) syntactic analysis of the corpus text. Views (a) and (b) approach the phenomenon of lexical collocations from opposite directions. Approach (a) starts from recurrent word combinations, which are defined by their empirical distribution in corpora, and aims to describe and understand their observed linguistic properties. Approach (b), on the other hand, starts from a theoretical analysis of lexical collocations (often resulting in a taxonomy of subtypes). Its goal is to develop methods to extract the desired type of collocation with high accuracy. This situation has led to much controversy (if not open hostilities) between adherents of the two views, which culminated in the recent publication of Hausmann (2004). However, a closer look reveals that both approaches face essentially the same problem: the difficulty of giving their object of study a precise definition. For (a), it is necessary to operationalise the concept of “recurrence”. Most researchers use a statistical criterion, viz. significant association, which may seem to be an objective and indisputably definition at first sight. However, statistical association can be quantified in many different ways, neither of which is obviously right or wrong (cf. the long-standing debate in mathematical statistics reported by Yates (1984)). In addition, methods for establishing the significance of an observed association face various mathematical problems that can often be traced back to characteristic properties of language data such as Zipf’s law and the untenability of independence assumptions (cf. Evert 2004). As a result, a wide range of equally plausible association measures will extract entirely different sets of “recurrent word combinations” from a given corpus. Approach (b) seems to have a clearer goal to guide the choice of a suitable association measure. Here, the problem lies in the theoretical analysis, namely the lack of a precise definition of lexical collocations and a clear delineation of relevant subtypes. The classifications that have been developed up to now – figurative expressions, support verb constructions, idioms, proverbs, etc. – are problematic for various reasons. While they often function well for a core set of instances, they invariably leave open a grey area of word combinations that exhibit properties of different classes of collocations. An example is the distinction between support verb constructions and figurative expressions in German, which can be operationalised fairly well (cf. Krenn 2000). Nevertheless, a considerable number of instances are difficult to assign unanimously to one class or the other. We have thus identified three key problems for corpus-based studies of lexical collocations: (i) to develop suitable (statistical) definitions of recurrent word combinations; (ii) to achieve a better theoretical understanding of the linguistic phenomenon of collocations; and (iii) to investigate the relation between (different definitions of) recurrence and (different types of) collocativity. The “traditional” approaches concentrate on (i) and (iii), respectively, to the extent that they have all but forgotten their common ground (ii). It is now obvious, though, that both sides must address all three issues in order to achieve their goals. Combining approaches (a) and (b), we suggest an incremental exploratory strategy that works in the following way: 1. sketch a provisional classification of lexical collocations with clear definitions for core instances 2. perform evaluation experiments to find a suitable association measure for each class of collocations 3. extract recurrent word combinations from large corpora, using the measures identified in step 2 4. make a detailed linguistic analysis of the extracted data, paying special attention to the grey areas where candidates cannot be clearly assigned to one class by the association measures 5. refine the theoretical definition and classification of collocations, then repeat from step 2 An essential component of this exploratory approach is the large number of evaluation experiments carried out in step 2, which require manual and conscientious annotation of candidate data according to the provisional classification. Such time-consuming tasks are only practicable when the amount of manual work can be reduced. Fortunately, this is indeed possible by carrying out evaluation experiments on a random sample and extrapolating the results to the full data set (Evert and Krenn, to appear).

برای دسترسی به متن کامل این مقاله و 10 میلیون مقاله دیگر ابتدا ثبت نام کنید

ثبت نام کنید

اگر عضو سایت هستید لطفا وارد حساب کاربری خود شوید

منابع مشابه

Exploratory Search on the Mobile Web

We present a mobile touchable application for online topic graph extraction and exploration of web content. The system has been implemented for operation on a tablet computer, i.e. an Apple iPad, and on a mobile device, i.e. Apple iPhone or iPod touch. The topics are extracted from web snippets which are determined by a standard search engine. We consider the extraction of topics as a specific ...

متن کامل

MobEx: A System for Exploratory Search on the Mobile Web

We present MobEx, a mobile touchable application for exploratory search on the mobile web. The system has been implemented for operation on a tablet computer, i.e. an Apple iPad, and on a mobile device, i.e. Apple iPhone or iPod touch. Starting from a topic issued by the user the system collects web snippets that have been determined by a standard search engine in a first step and extracts asso...

متن کامل

Collocation Extraction for Machine Translation

This paper reports on the development of a collocation extraction system that is designed within a commercial machine translation system in order to take advantage of the robust syntactic analysis that the system offers and to use this analysis to refine collocation extraction. Embedding the extraction system also addresses the need to provide information about the source language collocations ...

متن کامل

Collocation Extraction: Needs, Feeds And Results Of An Extraction System For German

This paper provides a specification of requirements for collocation extraction systems, taking as an example the extraction of noun + verb collocations from German texts. A hybrid approach to the extraction of habitual collocations and idioms is presented, aiming at a detailed description of collocations and their morphosyntax for natural language generation systems as well as to support learne...

متن کامل

Normalized (Pointwise) Mutual Information in Collocation Extraction

In this paper, we discuss the related information theoretical association measures of mutual information and pointwise mutual information, in the context of collocation extraction. We introduce normalized variants of these measures in order to make them more easily interpretable and at the same time less sensitive to occurrence frequency. We also provide a small empirical study to give more ins...

متن کامل

افزودن به منابع من


  با ذخیره ی این منبع در منابع من، دسترسی به آن را برای استفاده های بعدی راحت تر خواهید کرد

برای دسترسی به متن کامل این مقاله و 10 میلیون مقاله دیگر ابتدا ثبت نام کنید

ثبت نام کنید

اگر عضو سایت هستید لطفا وارد حساب کاربری خود شوید

عنوان ژورنال:

دوره   شماره 

صفحات  -

تاریخ انتشار 2005