# Natural Language Toolkit: TnT Tagger # # Copyright (C) 2001-2024 NLTK Project # Author: Sam Huston # # URL: # For license information, see LICENSE.TXT """ Implementation of 'TnT - A Statisical Part of Speech Tagger' by Thorsten Brants https://aclanthology.org/A00-1031.pdf """ from math import log from operator import itemgetter from nltk.probability import ConditionalFreqDist, FreqDist from nltk.tag.api import TaggerI class TnT(TaggerI): """ TnT - Statistical POS tagger IMPORTANT NOTES: * DOES NOT AUTOMATICALLY DEAL WITH UNSEEN WORDS - It is possible to provide an untrained POS tagger to create tags for unknown words, see __init__ function * SHOULD BE USED WITH SENTENCE-DELIMITED INPUT - Due to the nature of this tagger, it works best when trained over sentence delimited input. - However it still produces good results if the training data and testing data are separated on all punctuation eg: [,.?!] - Input for training is expected to be a list of sentences where each sentence is a list of (word, tag) tuples - Input for tag function is a single sentence Input for tagdata function is a list of sentences Output is of a similar form * Function provided to process text that is unsegmented - Please see basic_sent_chop() TnT uses a second order Markov model to produce tags for a sequence of input, specifically: argmax [Proj(P(t_i|t_i-1,t_i-2)P(w_i|t_i))] P(t_T+1 | t_T) IE: the maximum projection of a set of probabilities The set of possible tags for a given word is derived from the training data. It is the set of all tags that exact word has been assigned. To speed up and get more precision, we can use log addition to instead multiplication, specifically: argmax [Sigma(log(P(t_i|t_i-1,t_i-2))+log(P(w_i|t_i)))] + log(P(t_T+1|t_T)) The probability of a tag for a given word is the linear interpolation of 3 markov models; a zero-order, first-order, and a second order model. P(t_i| t_i-1, t_i-2) = l1*P(t_i) + l2*P(t_i| t_i-1) + l3*P(t_i| t_i-1, t_i-2) A beam search is used to limit the memory usage of the algorithm. The degree of the beam can be changed using N in the initialization. N represents the maximum number of possible solutions to maintain while tagging. It is possible to differentiate the tags which are assigned to capitalized words. However this does not result in a significant gain in the accuracy of the results. """ def __init__(self, unk=None, Trained=False, N=1000, C=False): """ Construct a TnT statistical tagger. Tagger must be trained before being used to tag input. :param unk: instance of a POS tagger, conforms to TaggerI :type unk: TaggerI :param Trained: Indication that the POS tagger is trained or not :type Trained: bool :param N: Beam search degree (see above) :type N: int :param C: Capitalization flag :type C: bool Initializer, creates frequency distributions to be used for tagging _lx values represent the portion of the tri/bi/uni taggers to be used to calculate the probability N value is the number of possible solutions to maintain while tagging. A good value for this is 1000 C is a boolean value which specifies to use or not use the Capitalization of the word as additional information for tagging. NOTE: using capitalization may not increase the accuracy of the tagger """ self._uni = FreqDist() self._bi = ConditionalFreqDist() self._tri = ConditionalFreqDist() self._wd = ConditionalFreqDist() self._eos = ConditionalFreqDist() self._l1 = 0.0 self._l2 = 0.0 self._l3 = 0.0 self._N = N self._C = C self._T = Trained self._unk = unk # statistical tools (ignore or delete me) self.unknown = 0 self.known = 0 def train(self, data): """ Uses a set of tagged data to train the tagger. If an unknown word tagger is specified, it is trained on the same data. :param data: List of lists of (word, tag) tuples :type data: tuple(str) """ # Ensure that local C flag is initialized before use C = False if self._unk is not None and self._T == False: self._unk.train(data) for sent in data: history = [("BOS", False), ("BOS", False)] for w, t in sent: # if capitalization is requested, # and the word begins with a capital # set local flag C to True if self._C and w[0].isupper(): C = True self._wd[w][t] += 1 self._uni[(t, C)] += 1 self._bi[history[1]][(t, C)] += 1 self._tri[tuple(history)][(t, C)] += 1 history.append((t, C)) history.pop(0) # set local flag C to false for the next word C = False self._eos[t]["EOS"] += 1 # compute lambda values from the trained frequency distributions self._compute_lambda() def _compute_lambda(self): """ creates lambda values based upon training data NOTE: no need to explicitly reference C, it is contained within the tag variable :: tag == (tag,C) for each tag trigram (t1, t2, t3) depending on the maximum value of - f(t1,t2,t3)-1 / f(t1,t2)-1 - f(t2,t3)-1 / f(t2)-1 - f(t3)-1 / N-1 increment l3,l2, or l1 by f(t1,t2,t3) ISSUES -- Resolutions: if 2 values are equal, increment both lambda values by (f(t1,t2,t3) / 2) """ # temporary lambda variables tl1 = 0.0 tl2 = 0.0 tl3 = 0.0 # for each t1,t2 in system for history in self._tri.conditions(): (h1, h2) = history # for each t3 given t1,t2 in system # (NOTE: tag actually represents (tag,C)) # However no effect within this function for tag in self._tri[history].keys(): # if there has only been 1 occurrence of this tag in the data # then ignore this trigram. if self._uni[tag] == 1: continue # safe_div provides a safe floating point division # it returns -1 if the denominator is 0 c3 = self._safe_div( (self._tri[history][tag] - 1), (self._tri[history].N() - 1) ) c2 = self._safe_div((self._bi[h2][tag] - 1), (self._bi[h2].N() - 1)) c1 = self._safe_div((self._uni[tag] - 1), (self._uni.N() - 1)) # if c1 is the maximum value: if (c1 > c3) and (c1 > c2): tl1 += self._tri[history][tag] # if c2 is the maximum value elif (c2 > c3) and (c2 > c1): tl2 += self._tri[history][tag] # if c3 is the maximum value elif (c3 > c2) and (c3 > c1): tl3 += self._tri[history][tag] # if c3, and c2 are equal and larger than c1 elif (c3 == c2) and (c3 > c1): tl2 += self._tri[history][tag] / 2.0 tl3 += self._tri[history][tag] / 2.0 # if c1, and c2 are equal and larger than c3 # this might be a dumb thing to do....(not sure yet) elif (c2 == c1) and (c1 > c3): tl1 += self._tri[history][tag] / 2.0 tl2 += self._tri[history][tag] / 2.0 # otherwise there might be a problem # eg: all values = 0 else: pass # Lambda normalisation: # ensures that l1+l2+l3 = 1 self._l1 = tl1 / (tl1 + tl2 + tl3) self._l2 = tl2 / (tl1 + tl2 + tl3) self._l3 = tl3 / (tl1 + tl2 + tl3) def _safe_div(self, v1, v2): """ Safe floating point division function, does not allow division by 0 returns -1 if the denominator is 0 """ if v2 == 0: return -1 else: return v1 / v2 def tagdata(self, data): """ Tags each sentence in a list of sentences :param data:list of list of words :type data: [[string,],] :return: list of list of (word, tag) tuples Invokes tag(sent) function for each sentence compiles the results into a list of tagged sentences each tagged sentence is a list of (word, tag) tuples """ res = [] for sent in data: res1 = self.tag(sent) res.append(res1) return res def tag(self, data): """ Tags a single sentence :param data: list of words :type data: [string,] :return: [(word, tag),] Calls recursive function '_tagword' to produce a list of tags Associates the sequence of returned tags with the correct words in the input sequence returns a list of (word, tag) tuples """ current_state = [(["BOS", "BOS"], 0.0)] sent = list(data) tags = self._tagword(sent, current_state) res = [] for i in range(len(sent)): # unpack and discard the C flags (t, C) = tags[i + 2] res.append((sent[i], t)) return res def _tagword(self, sent, current_states): """ :param sent : List of words remaining in the sentence :type sent : [word,] :param current_states : List of possible tag combinations for the sentence so far, and the log probability associated with each tag combination :type current_states : [([tag, ], logprob), ] Tags the first word in the sentence and recursively tags the reminder of sentence Uses formula specified above to calculate the probability of a particular tag """ # if this word marks the end of the sentence, # return the most probable tag if sent == []: (h, logp) = current_states[0] return h # otherwise there are more words to be tagged word = sent[0] sent = sent[1:] new_states = [] # if the Capitalisation is requested, # initialise the flag for this word C = False if self._C and word[0].isupper(): C = True # if word is known # compute the set of possible tags # and their associated log probabilities if word in self._wd: self.known += 1 for history, curr_sent_logprob in current_states: logprobs = [] for t in self._wd[word].keys(): tC = (t, C) p_uni = self._uni.freq(tC) p_bi = self._bi[history[-1]].freq(tC) p_tri = self._tri[tuple(history[-2:])].freq(tC) p_wd = self._wd[word][t] / self._uni[tC] p = self._l1 * p_uni + self._l2 * p_bi + self._l3 * p_tri p2 = log(p, 2) + log(p_wd, 2) # compute the result of appending each tag to this history new_states.append((history + [tC], curr_sent_logprob + p2)) # otherwise a new word, set of possible tags is unknown else: self.unknown += 1 # since a set of possible tags, # and the probability of each specific tag # can not be returned from most classifiers: # specify that any unknown words are tagged with certainty p = 1 # if no unknown word tagger has been specified # then use the tag 'Unk' if self._unk is None: tag = ("Unk", C) # otherwise apply the unknown word tagger else: [(_w, t)] = list(self._unk.tag([word])) tag = (t, C) for history, logprob in current_states: history.append(tag) new_states = current_states # now have computed a set of possible new_states # sort states by log prob # set is now ordered greatest to least log probability new_states.sort(reverse=True, key=itemgetter(1)) # del everything after N (threshold) # this is the beam search cut if len(new_states) > self._N: new_states = new_states[: self._N] # compute the tags for the rest of the sentence # return the best list of tags for the sentence return self._tagword(sent, new_states) ######################################## # helper function -- basic sentence tokenizer ######################################## def basic_sent_chop(data, raw=True): """ Basic method for tokenizing input into sentences for this tagger: :param data: list of tokens (words or (word, tag) tuples) :type data: str or tuple(str, str) :param raw: boolean flag marking the input data as a list of words or a list of tagged words :type raw: bool :return: list of sentences sentences are a list of tokens tokens are the same as the input Function takes a list of tokens and separates the tokens into lists where each list represents a sentence fragment This function can separate both tagged and raw sequences into basic sentences. Sentence markers are the set of [,.!?] This is a simple method which enhances the performance of the TnT tagger. Better sentence tokenization will further enhance the results. """ new_data = [] curr_sent = [] sent_mark = [",", ".", "?", "!"] if raw: for word in data: if word in sent_mark: curr_sent.append(word) new_data.append(curr_sent) curr_sent = [] else: curr_sent.append(word) else: for word, tag in data: if word in sent_mark: curr_sent.append((word, tag)) new_data.append(curr_sent) curr_sent = [] else: curr_sent.append((word, tag)) return new_data def demo(): from nltk.corpus import brown sents = list(brown.tagged_sents()) test = list(brown.sents()) tagger = TnT() tagger.train(sents[200:1000]) tagged_data = tagger.tagdata(test[100:120]) for j in range(len(tagged_data)): s = tagged_data[j] t = sents[j + 100] for i in range(len(s)): print(s[i], "--", t[i]) print() def demo2(): from nltk.corpus import treebank d = list(treebank.tagged_sents()) t = TnT(N=1000, C=False) s = TnT(N=1000, C=True) t.train(d[(11) * 100 :]) s.train(d[(11) * 100 :]) for i in range(10): tacc = t.accuracy(d[i * 100 : ((i + 1) * 100)]) tp_un = t.unknown / (t.known + t.unknown) tp_kn = t.known / (t.known + t.unknown) t.unknown = 0 t.known = 0 print("Capitalization off:") print("Accuracy:", tacc) print("Percentage known:", tp_kn) print("Percentage unknown:", tp_un) print("Accuracy over known words:", (tacc / tp_kn)) sacc = s.accuracy(d[i * 100 : ((i + 1) * 100)]) sp_un = s.unknown / (s.known + s.unknown) sp_kn = s.known / (s.known + s.unknown) s.unknown = 0 s.known = 0 print("Capitalization on:") print("Accuracy:", sacc) print("Percentage known:", sp_kn) print("Percentage unknown:", sp_un) print("Accuracy over known words:", (sacc / sp_kn)) def demo3(): from nltk.corpus import brown, treebank d = list(treebank.tagged_sents()) e = list(brown.tagged_sents()) d = d[:1000] e = e[:1000] d10 = int(len(d) * 0.1) e10 = int(len(e) * 0.1) tknacc = 0 sknacc = 0 tallacc = 0 sallacc = 0 tknown = 0 sknown = 0 for i in range(10): t = TnT(N=1000, C=False) s = TnT(N=1000, C=False) dtest = d[(i * d10) : ((i + 1) * d10)] etest = e[(i * e10) : ((i + 1) * e10)] dtrain = d[: (i * d10)] + d[((i + 1) * d10) :] etrain = e[: (i * e10)] + e[((i + 1) * e10) :] t.train(dtrain) s.train(etrain) tacc = t.accuracy(dtest) tp_un = t.unknown / (t.known + t.unknown) tp_kn = t.known / (t.known + t.unknown) tknown += tp_kn t.unknown = 0 t.known = 0 sacc = s.accuracy(etest) sp_un = s.unknown / (s.known + s.unknown) sp_kn = s.known / (s.known + s.unknown) sknown += sp_kn s.unknown = 0 s.known = 0 tknacc += tacc / tp_kn sknacc += sacc / tp_kn tallacc += tacc sallacc += sacc # print(i+1, (tacc / tp_kn), i+1, (sacc / tp_kn), i+1, tacc, i+1, sacc) print("brown: acc over words known:", 10 * tknacc) print(" : overall accuracy:", 10 * tallacc) print(" : words known:", 10 * tknown) print("treebank: acc over words known:", 10 * sknacc) print(" : overall accuracy:", 10 * sallacc) print(" : words known:", 10 * sknown)