Natural language processing (NLP) is a field of computer science concerned with the interactions between computers and human (natural) languages. Natural language generation systems convert information from computer databases into readable human language. Natural language understanding systems convert samples of human language into more formal representations that are easier for computer programs to manipulate. Many problems within NLP apply to both generation and understanding; for example, a computer must be able to model morphology (the structure of words) in order to understand an English sentence, but a model of morphology is also needed for producing a grammatically correct English sentence.
NLP has significant overlap with the field of computational linguistics, and is often considered a sub-field of artificial intelligence. The term natural language is used to distinguish human languages (such as Spanish, Swahili or Swedish) from formal or computer languages (such as C++, Java or LISP). Although NLP may encompass both text and speech, work on speech processing has evolved into a separate field.
Contents |
In theory, natural-language processing is a very attractive method of human-computer interaction. Early systems such as SHRDLU, working in restricted "blocks worlds" with restricted vocabularies, worked extremely well, leading researchers to excessive optimism, which was soon lost when the systems were extended to more realistic situations with real-world ambiguity and complexity.
Natural-language understanding is sometimes referred to as an AI-complete problem, because natural-language recognition seems to require extensive knowledge about the outside world and the ability to manipulate it. The definition of "understanding" is one of the major problems in natural-language processing.
Statistical natural-language processing uses stochastic, probabilistic and statistical methods to resolve some of the difficulties discussed above, especially those which arise because longer sentences are highly ambiguous when processed with realistic grammars, yielding thousands or millions of possible analyses. Methods for disambiguation often involve the use of corpora and Markov models. Statistical NLP comprises all quantitative approaches to automated language processing, including probabilistic modeling, information theory, and linear algebra[1]. The technology for statistical NLP comes mainly from machine learning and data mining, both of which are fields of artificial intelligence that involve learning from data.
Some examples of the problems faced by natural-language-understanding systems:
English is particularly challenging in this regard because it has little inflectional morphology to distinguish between parts of speech.
The goal of NLP evaluation is to measure one or more qualities of an algorithm or a system, in order to determine whether (or to what extent) the system answers the goals of its designers, or meets the needs of its users. Research in NLP evaluation has received considerable attention, because the definition of proper evaluation criteria is one way to specify precisely an NLP problem, going thus beyond the vagueness of tasks defined only as language understanding or language generation. A precise set of evaluation criteria, which includes mainly evaluation data and evaluation metrics, enables several teams to compare their solutions to a given NLP problem.
The first evaluation campaign on written texts seems to be a campaign dedicated to message understanding in 1987 (Pallet 1998). Then, the Parseval/GEIG project compared phrase-structure grammars (Black 1991). A series of campaigns within Tipster project were realized on tasks like summarization, translation and searching (Hirshman 1998). In 1994, in Germany, the Morpholympics compared German taggers. Then, the Senseval and Romanseval campaigns were conducted with the objectives of semantic disambiguation. In 1996, the Sparkle campaign compared syntactic parsers in four different languages (English, French, German and Italian). In France, the Grace project compared a set of 21 taggers for French in 1997 (Adda 1999). In 2004, during the Technolangue/Easy project, 13 parsers for French were compared. Large-scale evaluation of dependency parsers were performed in the context of the CoNLL shared tasks in 2006 and 2007. In Italy, the evalita campaign was conducted in 2007 to compare various tools for Italian evalita web site. In France, within the ANR-Passage project (end of 2007), 10 parsers for French were compared passage web site.
Adda G., Mariani J., Paroubek P., Rajman M. 1999 L'action GRACE d'évaluation de l'assignation des parties du discours pour le français. Langues vol-2
Black E., Abney S., Flickinger D., Gdaniec C., Grishman R., Harrison P., Hindle D., Ingria R., Jelinek F., Klavans J., Liberman M., Marcus M., Reukos S., Santoni B., Strzalkowski T. 1991 A procedure for quantitatively comparing the syntactic coverage of English grammars. DARPA Speech and Natural Language Workshop
Hirshman L. 1998 Language understanding evaluation: lessons learned from MUC and ATIS. LREC Granada
Pallet D.S. 1998 The NIST role in automatic speech recognition benchmark tests. LREC Granada
Depending on the evaluation procedures, a number of distinctions are traditionally made in NLP evaluation.
Intrinsic evaluation considers an isolated NLP system and characterizes its performance mainly with respect to a gold standard result, pre-defined by the evaluators. Extrinsic evaluation, also called evaluation in use considers the NLP system in a more complex setting, either as an embedded system or serving a precise function for a human user. The extrinsic performance of the system is then characterized in terms of its utility with respect to the overall task of the complex system or the human user. For example, consider a syntactic parser that is based on the output of some new part of speech (POS) tagger. An intrinsic evaluation would run the POS tagger on some labelled data, and compare the system output of the POS tagger to the gold standard (correct) output. An extrinsic evaluation would run the parser with some other POS tagger, and then with the new POS tagger, and compare the parsing accuracy.
Black-box evaluation requires one to run an NLP system on a given data set and to measure a number of parameters related to the quality of the process (speed, reliability, resource consumption) and, most importantly, to the quality of the result (e.g. the accuracy of data annotation or the fidelity of a translation). Glass-box evaluation looks at the design of the system, the algorithms that are implemented, the linguistic resources it uses (e.g. vocabulary size), etc. Given the complexity of NLP problems, it is often difficult to predict performance only on the basis of glass-box evaluation, but this type of evaluation is more informative with respect to error analysis or future developments of a system.
In many cases, automatic procedures can be defined to evaluate an NLP system by comparing its output with the gold standard (or desired) one. Although the cost of producing the gold standard can be quite high, automatic evaluation can be repeated as often as needed without much additional costs (on the same input data). However, for many NLP problems, the definition of a gold standard is a complex task, and can prove impossible when inter-annotator agreement is insufficient. Manual evaluation is performed by human judges, which are instructed to estimate the quality of a system, or most often of a sample of its output, based on a number of criteria. Although, thanks to their linguistic competence, human judges can be considered as the reference for a number of language processing tasks, there is also considerable variation across their ratings. This is why automatic evaluation is sometimes referred to as objective evaluation, while the human kind appears to be more subjective.
An ISO sub-committee is working in order to ease interoperability between Lexical resources and NLP programs. The sub-committee is part of ISO/TC37 and is called ISO/TC37/SC4. Some ISO standards are already published but most of them are under construction, mainly on lexicon representation (see LMF), annotation and data category registry.