Lecture Notes-Finite State Automata for NLP.pdf

Lecture Notes-Finite State Automata for NLP.pdf

• 1.
  Prof. Deptii Chaudhari,Department of Computer Engineering, I2IT Lecture Notes - Finite State Automata for NLP Finite state automata (deterministic and nondeterministic finite automata) provide decision regarding acceptance and rejection of a string while transducers provide some output for a given input. Thus, the two machines are quite useful in language processing tasks. Finite state automata are useful in deciding whether a given word belongs to a particular language or not. Similarly, transducers are useful in parsing and generation of words from their lexical form. Areas where finite-state methods have been proved to be particularly efficient are • phonological • morphological • syntactic analyses • tokenization • shallow parsing • word sense disambiguation • spell checking/correcting Each path in a finite-state network encodes a single string or a set of strings. All paths in such a network encode a finite-state language or a finite-state relation. Consider the following figure encoding the language of the type {clear, clever, ear, ever, fat, fatter}. FSA for Morphological Analysis Morphology is a branch of linguistics that studies the inner structure of words. It deals with the ways the words are formed (word formation rules) using the smallest units bearing some lexical or grammatical meaning. In modeling natural language morphology, it is necessary to distinguish between surface forms of words and their lemmas. A lemma for a surface form is typically its dictionary entry form together with some terms that shows the morphological properties of that form. So, the lemma for the word “bigger” is represented as big+Adj+Comp which shows that the word is the comparative form of the adjective big.
• 2.
  Prof. Deptii Chaudhari,Department of Computer Engineering, I2IT There are two processes by which the morphemes can be combined to form words: inflection and derivation. Inflection is the process of adding a grammatical affix to a word stem, forming a word of the same part of speech as stem. Adding plural s to a noun stem is an example in this respect. Derivation, on the other hand, is the process of adding an affix to a word stem resulting in a word having a part of speech different from that of the stem. Adding ness to the adjective good, making a noun as goodness is an example in this respect. Most of the inflectional endings in English are regular; however, there are some cases in which there are irregular inflectional morphology at work. Examples in this regard are foot vs. feet, and take vs. took. Finite-state automaton for inflectional morphology of English nouns For some other kinds of complexities in English inflectional morphology, the finite-state technology can also find rules to represent them. Fox vs. foxes and mouse vs. mice are examples in this respect. The transducer mapping mice to mouse can be demonstrated as follows: A finite-state transducer mapping mice to mouse
• 3.
  Prof. Deptii Chaudhari,Department of Computer Engineering, I2IT An FSA for another fragment of English derivational morphology This FSA models several derivational facts, such as the well-known generalization that any verb ending in -ize can be followed by the nominalizing suffix -ation. Thus, since there is a word fossilize, we can predict the word fossilization by following states q0, q1, and q2. Similarly, adjectives ending in -al or -able at q5 (equal, formal, realizable) can take the suffix -ity, or sometimes the suffix -ness to state q6 (naturalness, casualness) Exercise: Design a Finite State Automata for baa+! Q: Finite set of states = q0, q1, q2, q3, q4 ∑ : Set of input alphabets = {a,b,!} q0: Start state F: Set of final states {q4} (q, i ) defined by the transition table
• 4.
  Prof. Deptii Chaudhari,Department of Computer Engineering, I2IT a b ! q0 ϕ q1 ϕ q1 q2 ϕ ϕ q2 q3 ϕ ϕ q3 q3 ϕ q4 q4 ϕ ϕ ϕ Finite State Transducer for NLP A finite-state transducer is a non-simple finite-state automaton in which at least one arc is labeled by a symbol pair such as a:b where a designates the symbol on the upper side of the arc and b the symbol on the lower side. There are two types of transducers: • a functional transducer in which there is at most one output for any input, and • a sequential transducer in which no state has more than one arc having the same symbol on the input side. FST as recognizer: Takes a pair of strings and accepts or rejects them FST as generator: Outputs a pair of strings for a language FST as translator: Reads a string and outputs another string and Morphological parsing: letters (input); morphemes (output) FST as relater: Computes relations between sets Below Figure demonstrates a lexical transducer through which the relations {<leaf+NN, leaf>,<leaf+NNS, leaves>, <left+JJ, left>,<leave+NN, leave>,<leave+NNS, leaves>,<leave+VB. Leave>,<leave+VBZ, leaves>,<leave+VBD, left>} are encoded. As figure shows each word has a distinct path. 0 is the epsilon symbol standing for the empty string. Such lexical transducers are used both for analysis and generation. To analyze the word leaves, for instance, one
• 5.
  Prof. Deptii Chaudhari,Department of Computer Engineering, I2IT needs to follow the path containing the symbols l, e, a, v, e and s on the lower side of the arc label. There are three such paths in the above network as follows: a) 0 – l – 1 – e – 2 – a – 3 – v – 4 – e – 5 - +NNS:s – 6 b) 0 – l – 1 – e – 2 – a – 3 – v – 4 – e – 5 - +VBZ:s – 6 c) 0 – l – 1 – e – 2 – a – 3 – f:v –8 – +NNS:e – 9 – 0:s – 6