Marco Ernandes

A Web-Based Agent Challenges Human Experts on Crosswords

Crosswords are very popular and represent a useful domain of investigation for modern artificial intelligence. In contrast to solving other celebrated games (such as chess), cracking crosswords requires a paradigm shift towards the ability to handle tasks for which humans require extensive semantic knowledge. This article introduces WebCrow, an automatic crossword solver in which the needed knowledge is mined from the web: clues are solved primarily by accessing the web through search engines and applying natural language processing techniques. In competitions at the European Conference on Artificial Intelligence (ECAI) in 2006 and other conferences this web-based approach enabled WebCrow to outperform its human challengers. Just as chess was once called “the Drosophila of artificial intelligence,” we believe that crossword systems can be useful Drosophila of web-based agents.

Solving italian crosswords using the web

We designed and implemented a software system, called WebCrow, that represents the first solver for Italian crosswords and the first system that tackles a language game using the Web as knowledge base. Its core feature is the Web Search Module that produces a special form of web-based question answering that we call clue-answering. This paper will focus its attention on this task. The web-search approach has proved itself to be very consistent: using a limited set of documents the clue-answering process is able to retrieve over two thirds of the correct answers. In many cases the targeted word is given in output among the very first most probable candidates (15% of correct answers in first position). To complete the crosswords solving problem the system has to fill the grid with the best set of word answers. Currently, WebCrows performances are interesting: crosswords that are “easy” for expert humans (i.e. crosswords from the cover pages of La Settimana EnigmisticaTM) are solved, in a 15 minutes time limit, with 80% of correct words and over 90% of correct letters. With crosswords that are designed for experts, WebCrow places correctly two thirds of the words and around 80% of the letters.

Semantic Labeling of Data by Using the Web

This paper proposes a system for automatically categorizing terms or lexical entities into a predefined set of semantic domains. We present an approach that exploits the knowledge available in the Web to create a model of each term or entity (entity context lexicons - ECLs). Each profile is simply a list of terms (similar to the bag-of-words representation in text categorization) and it is composed primarily by the words often appearing in the same contexts of the entity. These profiles model the contexts in which the entity usually appears and they can be subsequently processed by an automatic classifier. Moreover, we propose and validate a profile-based categorization model developed for this particular task which uses the ECLs of the training entities to build a profile for each class (class context lexicon - CCL). Finally, we propose a technique for dealing with multi-label classification based on a decision module that exploits a neural network. We show the effectiveness of the proposed approach on a term categorization task using a standard benchmark composed of a set of domain-specific lexicons (WordNetDomains)

Cracking crosswords: the computer challenge

Crosswords is over 90 years old, yet it is still one of the most popular puzzles around the world. It is in fact a linguistic game which requires a wide knowledge in different domains and the ability to crack enigmatic clues, that are often regarded as inherent human capabilities. Unlike chess, crossword solving does not require strong skills for the actuation of strategic plans, but the linguistic specifications is in itself a source of enormous difficulty for machines. This paper discusses the problem of automatic crossword solving with special emphasis to the WebCrow project carried out at the University of Siena. After a brief historical description of the evolution of crosswords, the paper gives a formalization of the main problems to be faced and provides a number of relevant architectural issues behind cracking crosswords. In particular, it is claimed that the Web is likely to be the most important source for the development of challenging programs based on clue answering, a sort of question answering mechanism in which the machine is expected to return candidate word solutions.

Webcrow: a web-based crosswords solver

Webcrow is a software system whose aim is to solve crosswords. Problems of like solving crosswords have been informally defined as AI-Complete and are extremely challenging for machines. Webcrow represents the first solver for Italian crosswords and the first system that tackles this language game using the Web as knowledge base. Currently, Webcrow is competitive against an average crossword player. Crosswords that are “easy” for expert humans (i.e. crosswords from the cover pages of La Settimana Enigmistica) are solved, in a 15 minutes time limit, with 80% of correct words and over 90% of correct letters. Webcrow well performs on crosswords that are designed for experts and, in general, it outperforms an average undergraduate student on this kind of crosswords.