Initial commit

Author: mmxgn

README.md

@@ -0,0 +1,66 @@
+# SPRL-Spacy
+
+This repository implements an easy to use Spatial Role Labeling module trained on
+three entities (`TRAJECTOR`, `SPATIAL_INDICATOR`, `LANDMARK`) and the relations appearing
+on the SpRL 2013 IAPR TC-12 dataset. 
+
+## Requirements
+
+- [git lfs](https://git-lfs.github.com/) for cloning the large files in this repository
+- `spacy >=2.0.0a18` and the necessary requirements
+- `sklearn`
+- `scipy`
+- `pickle` for python 3.7.0
+- `problog` for use with ProbLog.
+
+## Usage
+
+1. Clone this repository where you want to use it. 
+2. Download the two models from the Releases page
+3. Import `spacy` and `sprl` and use them like the following example:
+
+
+```
+import spacy
+from sprl import *
+
+nlp = spacy.load('en_core_web_lg-sprl')
+
+sentence = "An angry big dog is behind us."
+
+rel = sprl(sentence, nlp)
+```
+
+If everything went fine you should get something like:
+
+```
+[(An angry big dog, behind, us, 'direction')]
+```
+
+If you happen to have problog installed, you can see `example.pl` on how to use it from
+within problog.
+
+## Credits
+
+While the model has been trained by me, the relation extraction part uses features from
+the paper for Sprl-CWW (see below), and the dataset from SemEval 2013 Task 3: Spatial Role Labeling.
+
+The features for relation extraction:
+
+```
+Nichols, Eric, and Fadi Botros. 
+"SpRL-CWW: Spatial relation classification with independent multi-class models." 
+Proceedings of the 9th International Workshop on Semantic Evaluation.
+```
+
+Semeval 2013 task 3: Spatial Role Labeling
+
+```
+Kolomiyets, Oleksandr, et al. 
+"Semeval-2013 task 3: Spatial role labeling." 
+Second Joint Conference on Lexical and Computational Semantics
+```
+
+So please cite the papers above, as well as spacy and ProbLog (if you use it) in your work :)
+
+

example.pl

@@ -0,0 +1,4 @@
+:-use_module('sprl_pl.py').
+
+query(sprl('A book and a ball on the table', Tr, Sp, Lm, Type)).
+%query(in_range(1,10,X1, X2)).

sprl/__init__.py

@@ -0,0 +1 @@
+from .sprl import *

sprl/__pycache__/__init__.cpython-37.pyc

@@ -0,0 +1,162 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+Created on Mon Sep  3 17:14:47 2018
+
+@author: Emmanouil Theofanis Chourdakis <e.t.chourdakis@qmul.ac.uk>
+
+Functions that do spatial role labeling. Relation extraction is done
+using sklearn with features extracted from the sentence based on the following paper:
+cd
+Nichols, Eric, and Fadi Botros. 
+"SpRL-CWW: Spatial relation classification with independent multi-class models." 
+Proceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015)
+"""
+
+from sklearn.externals import joblib
+import spacy
+
+
+def get_dep_path(span1, span2):
+    assert span1.sent == span2.sent, "sent1: {}, span1: {}, sent2: {}, span2: {}".format(span1.sent, span1, span2.sent, span2)
+    
+    up = []
+    down = [] 
+    
+    head = span1[0]
+    while head.dep_ != 'ROOT':
+        up.append(head)
+        head = head.head
+    up.append(head)
+
+    head = span2[0]
+    while head.dep_ != 'ROOT':
+        down.append(head)
+        head = head.head
+    down.append(head)
+    down.reverse()
+    
+    for n1, t1 in enumerate(up):
+        for n2, t2 in enumerate(down):
+            if t1 == t2:
+                return ["{}::{}".format(u.dep_, 'up') for u in up[1:n1]] + ["{}::{}".format(d.dep_, 'down') for d in down[n2:]]
+
+def extract_relation_features(relation):
+    F = {} # Feature dict
+
+    trigger = relation[1]
+    args = [relation[0], relation[2]]
+
+    # Extract features relating to trigger
+    #trigger_head = get_head(trigger)
+
+    for n, token in enumerate(trigger):
+        F['TF1T{}'.format(n)] = token.text
+        F['TF2T{}'.format(n)] = token.lemma_
+        F['TF3T{}'.format(n)] = token.pos_
+        F['TF4T{}'.format(n)] = "::".join([token.lemma_, token.pos_]) # RF.2 concat RF.1
+
+    # Extract features relating to the two arguments
+    for a, arg in enumerate(args):
+        if arg is not None:
+            for n, token in enumerate(arg):
+                F['A{}F5T{}'.format(a, n)] = token.text
+                F['A{}F6T{}'.format(a, n)] = token.lemma_
+                F['A{}F7T{}'.format(a, n)] = token.pos_
+                F['A{}F8T{}'.format(a, n)] = "::".join([token.lemma_, token.pos_])
+
+
+            if arg[-1].i < trigger[0].i:
+                F['A{}F12'.format(a)] = 'LEFT'
+                F['A{}F22'.format(a)] = trigger[0].i - arg[-1].i
+            elif arg[0].i > trigger[-1].i:
+                F['A{}F12'.format(a)] = 'RIGHT'  
+                F['A{}F22'.format(a)] = arg[0].i - trigger[-1].i
+
+
+            path = get_dep_path(arg, trigger)
+            for np, p in enumerate(path):
+                F['A{}F17E{}'.format(a, np)] = p
+            F['A{}F20'.format(a)] = len(path)
+            F['A{}F24'.format(a)] = False
+        else:
+            F['A{}F24'.format(a)] = True
+
+    # Joint features
+    if 'A0F12' in F and 'A1F12' in F:
+        F['F13'] = "::".join([F['A0F12'],F['A1F12']])
+        if F['A0F12'] == F['A1F12']:
+            F['14'] = True
+        else:
+            F['14'] = False
+
+    if 'F13' in F:
+        for n, token in enumerate(trigger):
+            F['14T{}'.format(n)] = '::'.join([F['F13'], token.lemma_])
+
+    if 'A0F22' in F and 'A1F22' in F:
+        F['F23'] = F['A0F22'] + F['A1F22']
+
+    return F        
+
+def extract_candidate_relations_from_sents(sents, gold_relations):
+    candidate_relations = []
+    candidate_labels = []
+
+    for sent in sents:
+
+        triggers = [t for t in sent.ents if t.label_ == 'SPATIAL_INDICATOR']
+        trajectors = [t for t in sent.ents if t.label_ == 'TRAJECTOR']
+        landmarks = [t for t in sent.ents if t.label_ == 'LANDMARK']
+
+      #  print(trajectors, triggers, landmarks)
+
+        for trigger in triggers:
+            for trajector in trajectors:
+                for landmark in landmarks:
+                    if not (trajector is None and landmark is None):
+                        assert trajector.sent == trigger.sent == landmark.sent, "{}: {}".format(sent, sent.ents)
+                        crel = (trajector, trigger, landmark)
+                        if crel not in gold_relations:
+                            candidate_relations.append(crel)
+                            candidate_labels.append('NONE')
+                        else:
+                            #print("In gold relations already", crel)
+                            pass
+    return candidate_relations, candidate_labels
+
+def sprl(sentence, 
+         nlp,
+         model_relext_filename='model_svm_relations.pkl'):
+    output = []
+    doc = nlp(sentence)
+    sents = [nlp(s.text) for s in doc.sents]
+    candidate_relations, _ = extract_candidate_relations_from_sents(sents, [])
+    clf, dv = joblib.load(model_relext_filename)
+    for relation in candidate_relations:
+        F = extract_relation_features(relation)
+        feat_vec = dv.transform(F)
+        general_type = clf.predict(feat_vec)[0]
+        if general_type != 'NONE':
+            output.append((relation[0], relation[1], relation[2], general_type))
+        
+    return output
+        
+
+def sprl_str(sentence,
+             nlp,
+             model_relext_filename='model_svm_relations.pkl'):
+    """ Returns triples where every element is string """
+    output = []
+    doc = nlp(sentence)
+    sents = [nlp(s.text) for s in doc.sents]
+    candidate_relations, _ = extract_candidate_relations_from_sents(sents, [])
+    clf, dv = joblib.load(model_relext_filename)
+    for relation in candidate_relations:
+        F = extract_relation_features(relation)
+        feat_vec = dv.transform(F)
+        general_type = clf.predict(feat_vec)[0]
+        if general_type != 'NONE':
+            output.append((str(relation[0]), str(relation[1]), str(relation[2]), general_type))
+        
+    return output    

sprl/__pycache__/sprl.cpython-37.pyc

@@ -0,0 +1,53 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+Created on Thu Sep 20 13:51:33 2018
+
+@author: Emmanouil Theofanis Chourdakis
+
+Problog bindings for spacy-sprl 
+
+"""
+
+from problog.extern import problog_export, problog_export_nondet, problog_export_raw
+from problog.logic import Constant as c_
+
+import spacy
+import sprl as mod_sprl
+
+global lut_sprl # For memoization
+lut_sprl = {}
+
+global nlp_sprl # Initialize the model the first time it is neede
+nlp_sprl = None
+
+@problog_export_nondet('+str', '-str', '-str', '-str', '-str')
+def sprl(sentence):
+    global nlp_sprl, lut_sprl
+    
+    # Load NLP model if it is not loaded already
+    if not nlp_sprl:
+        print("[II] Loading NLP model...")
+        nlp_sprl = spacy.load('en_core_web_lg-sprl')
+        print("[II] done.")
+    
+    # Remove "'"S
+    sent = sentence
+    if sent[0] == "'":
+        sent = sent[1:]
+    if sent[-1] == "'":
+        sent = sent[:-1]
+    
+    if sentence in lut_sprl:
+        return lut_sprl[sent]
+    else:
+        triples = mod_sprl.sprl(sent, nlp_sprl)
+        lut_sprl[sent] = [tuple([str(t) for t in l]) for l in triples]
+        L = lut_sprl[sent] 
+        return L
+
+
+@problog_export_nondet('+int','+int','-int', '-int')
+def in_range(a, b):
+     L= list([(aa+bb, aa*bb) for aa in range(a, b) for bb in range(a, b)])
+     return L