Our WSD algorithm has three components, as depicted in Figure 1. Taking as input the

* This work was supported by the EPSRC while the author was at the University of Sheffield.

test data parsed using the Briscoe and Car­roll (1993) parser (which uses the grammar de­scribed in Carroll and Briscoe (1996)), the first step is to identify and discard the pleonastic pronouns. Our pleonastic component is de­scribed in section 2.2.

In the next phase (section 2.3), third person pronouns are resolved to a noun antecedent and replaced in the text by the noun antecedent. The purpose of this is to increase the number of topic words in the text, to aid the disam­biguation of other nouns. This approach as­sumes firstly that pronouns refer mainly to topic words, and secondly that repeating topic words in the text helps overall disambiguation.

The final phase of the algorithm is the WSD component, described in section 2.4. Using sim­ulated annealing, it attempts to find a sense assignment for every noun that minimizes an overall 'distance' function using the WordNet hierarchy. In addition, for the repeated nouns added in the previous phase, the senses are tied together. This means that if the sense of one word in a tie is changed during simulated an­nealing, the sense of all words in the tie are si­multaneously changed.

The advantage of this approach can be shown on the following discourse: The parrot, like the chicken, is kept by people as a domesticated bird. It can speak. Suppose firstly that there is no anaphora resolution phase. The words par­rot, chicken, person, bird are given to the word sense disambiguation algorithm, and the system chooses senses which are related to people (par­rot in the sense of mimicking people, chicken a wimp and so on). This is clearly incorrect. Now suppose we resolve the pronoun it to par­rot, and repeat the word parrot in the text. Now the words parrot, chicken, person, bird, parrot are passed to the WSD system (where the two parrots are sense-tied together), and the system now chooses the correct bird-related senses.

It can be a pleonastic pronoun (pronoun with no antecedent), for example in the sentence: It is raining. We label the pronoun it as pleonastic if it is a subject of a raising verb (these were ex­tracted from the ANLT lexicon (Boguraev and Briscoe, 1987)) or if it was used in conjunctions with the verb to be and one of a particular set of adjectives (for example It is possible to go to town.).

The component was evaluated on a manually anaphorically resolved portion of the bnc (the initial 2000 sentences of wOl). It has a preci­sion (proportion of pronouns deemed pleonastic which really are pleonastic) of 94% and recall (proportion of pleonastic pronouns recognized as pleonastic) of 61%.

The pronominal anaphora resolution is carried out by our re-implementation of the Kennedy and Boguraev (Kennedy and Boguraev, 1996) anaphora resolution algorithm. This algorithm is based on that of Lappin and Leass (Lappin and Leass, 1994), but does not require a full parse. It treats the cases of third person pro­nouns and lexical anaphors. Its cited accuracy is 75% on general corpora (Kennedy and Bogu­raev, 1996), but note that their published algo­rithm uses the lingsoft morphosyntactic tag­ger.

The algorithm creates coreference classes which join together words which are believed by the algorithm to be referring to the same ob­ject. These classes are assigned a salience value based on the presence of the features in Table 1. The salience value of a class is the sum of the feature weights of its members, scaled down by the number of sentences ago that the feature last occurred. The correct antecedent is chosen to be the closest word from the coreference class with the highest salience.

We define a notion of distance between any two WordNet noun senses which is based on the

WordNet hierarchy. As pointed out by Resnik (1999), it is naive to assume that the distance between any two nodes in the hierarchy is equal. We therefore assign a weight w to every noun sense x:

weight (x) =

number of children below x in hierarchy total nodes in hierarchy

This is used to define the distance between two distinct noun senses x and y:

where h(s) denotes the hypernym chain of noun sense s. If the hypernym chains of x and y do not intersect, the distance is set to the max­imum value of 1. In Preiss (2001), we investi­gated scaling the distance function such that for noun senses x and y at positions in the corpus n and m respectively:

Note that we do not explicitly use a window of surrounding nouns, but the |n — m\ denom­inator means that contributions from far away nouns are usually negligible. We showed that it was not possible to guess the optimal value of a

Pleonastic component

Anaphora resolution

in advance for any set of texts covered in sem-cor. However, averaged over all words there is a slight peak around a = 1, so this is the value we take.

The distance between two adjacent nodes in the hierarchy may now not be equal. To il­lustrate this, consider the following example adapted from a paper of Resnik (1999). In WordNet 1.7 (prerelease), valve is the parent node of safety valve, and machine is the parent of information processing system. However, the intuitive distance between the first pair of nodes seems to be less than the distance between the second pair. Using our distance function outlined above, the distance between safety valve and valve is 0.000121, while the distance between information process­ing system and machine is 0.00229. This is depicted in Figure 2.

We want to assign precisely one sense to each noun in the text; we call this a path. We find the 'optimal' path by simulated annealing (Bertsi-mas and Tsitsiklis, 1992). Simulated annealing is a probabilistic method for finding the global optimum of a function which may have a num­ber of local optima. We define the function to be minimized, the energy function, to be the sum of all the pairwise scaled distances.

Our version of simulated annealing starts with a randomly chosen path which it attempts to improve. It performs a number of iterations in which it randomly chooses a word and then chooses a new sense for this word. If this change is an improvement in terms of the en­ergy function, it is kept. Otherwise, it may or may not be accepted depending on the current value of the temperature. Over time the tem­perature decreases, making it less likely to keep changes that increase the energy. The algorithm terminates when no changes were made in the last 1000 iterations.

When simulated annealing terminates, it out­puts what it deems the optimal sense assign­ment for all the nouns in the text. For a more detailed description of the WSD algo­rithm, please refer to Preiss (2001).

This algorithm was implemented in C and ex­ecuted on a Pentium III 500MHz. Each text took 1 hour to initialize, and 2 hours to perform 20 runs of simulated annealing. A majority vote then decided the sense assignment.

The WSD component enhanced with the anaphora resolution algorithm was submitted for the English all-words task in Senseval-2. The test data for this task consisted of three ar­ticles, and information gathered from each ar­ticle is displayed in Table 2. The words col­umn shows the number of words marked as nouns by the part of speech tagger in the parser. The senses column contains the total number of senses for all of these words. The ties column shows the number of ties in the text, where each tie contains a noun and some pronouns that re­fer to it. The system achieved 44% precision and 20% recall fine-grained, and 45.2% preci­sion and 20.5% recall coarse-grained.