Tagging with Hidden Markov Models Using Ambiguous Tags

Tagging with Hidden Markov Models Using Ambiguous Tags Alexis Nasr LaTTice - Universit´e Paris 7 anasr@linguist.jussieu.fr Fr´ed´eric B´echet Laboratoire d’Informatique d’Avignon frederic.bechet@lia.univ-avignon.fr Alexandra Volanschi LaTTice - Universit´e Paris 7 avolansk@linguist.jussieu.fr Abstract Part of speech taggers based on Hidden Markov Models rely on a series of hypothe- ses which make certain errors inevitable. The idea developed in this paper consists in allowing a limited, controlled ambiguity in the output of the tagger in order to avoid a number of errors. The ambiguity takes the form of ambiguous tags which denote subsets of the tagset. These tags are used when the tagger hesitates between the dif- ferent components of the ambiguous tags. They are introduced in an existing lexicon and 3-gram database. Their lexical and syntactic counts are computed on the basis of the lexical and syntactic counts of their constituents, using impurity functions. The tagging process itself, based on the Viterbi algorithm, is unchanged. Experiments con- ducted on the Brown corpus show a recall of 0.982, for an ambiguity rate of 1.233 which is to be compared with a baseline recall of 0.978 for an ambiguity rate of 1.414 using the same ambiguous tags and with a recall of 0.955 corresponding to the one best solu- tion of standard tagging (without ambigu- ous tags). 1 Introduction Taggers are commonly used as pre-processors for more sophisticated treatments like full syn- tactic parsing or chunking. Although taggers achieve high accuracy, they still make some mistakes that quite often impede the following stages. There are at least two solutions to this problem. The first consists in devising more so- phisticated taggers either by providing the tag- ger with more linguistic knowldge or by refining the tagging process, through better probability estimation, for example. The second strategy consists in allowing some ambiguity in the out- put of the tagger. It is the second solution that was chosen in this paper. We believe that this is an instance of a more general problem in se- quential natural language processing chains, in which a module takes as input the output of the preceding module. Since we cannot, in most cases, expect a module to produce only correct solutions, modules should be able to deal with ambiguous input and ambiguous output. In our case, the input is non ambiguous while the out- put is ambiguous. From this perspective, the quality of the tagger is evaluated by the trade- off it achieves between accuracy and ambiguity. The introduction of ambiguous tags in the tagger output raises the question of the process- ing of these ambiguous tags in the post-tagging stages of the application. Leaving some ambigu- ity in the output of the tagger only makes sense if these other processes can handle it. In the case of a chunker, ambiguous tags can be taken into account through the use of weighted finite state machines, as proposed in (Nasr and Volan- schi, 2004). In the case of a syntactic parser, such a device can usually deal with some ambi- guity …

Content truncated to 500 words