It seem like the negative seeds aren't updated.

for t in list(self.candidate_tuples.keys()):
                    if t.confidence >= self.config.instance_confidence:
                        seed = Seed(t.e1, t.e2)
                        self.config.positive_seed_tuples.add(seed)

Only the positive is updated.

Line 18: if tuple is not None:
Line 19: self.tuples.add(t)

tuple is a Python keyword for the tuple class, and the statement in Line 18 will always return True.

Should it be:

if t is not None:
self.tuples.add(t)

? Thank you.

Hi again !

The errors mentionned in the title are due to the use of Sentence.py (from BREDS.Sentence import Sentence at line 15 of large-scale-evaluation-freebase.py)
At the begining I assumed that using the module Sentence.py located in /BREDS/ would work. But in fact, I found clues that you may have used a different version of this module when developping large-scale-evaluation-freebase.py.

The clues are :

• In the /BREDS/Sentence.py file, you don't handle the case where the pos_tagger and config arguments of __init__ are None (which happen to be the case in the call at line 141 of large-scale-evaluation-freebase.py)
• In large-scale-evaluation-freebase.py, in process_corpus(), line 144 you wrote tokens = word_tokenize(r.between) even if r.between is already a pos-tagged string (see line 199 of /BREDS/Sentence.py)

In the function pattern.update_selectivity, if you have the situation where you hit s.e1 == t.e1 but not s.e2 == t.e2 you will have matched_both = False and matched_e1 = True. Meaning that you will increment the negative length of the pattern. BUT you will also hit the
if matched_both is False: self.unknown += 1.

This will increment the negative and unknown length.

In your paper "Semi-Supervised Bootstrapping of Relationship Extractors with Distributional Semantics", you said "BREDS also tries to identify the passive voice using part-of-speech (PoS) tags ,..., it will swap the order of the entities when producing a relational triple".

But I can't find this process in your code. If BREDS detect the passive voice and swap the entities, about the words before the first entity and words after the second entity? Can you give
me some advice? Thank you!

According to the paper: Semi-Supervised Bootstrapping of Relationship Extractors with Distributional Semantics, it looks like only positive seeds are used in algorithm. Would you mind explaining more on the usage of negative seeds here?

The README states that there is a requirements.txt file, but it's missing from the repo. A requirements.txt with the right versions of the appropriate libraries would be helpful!

Hi !
I can't find the requirements.txt file that you evoke in your readme.md.
This causes me having an error at the call of the method jellyfish.jaro_winkler at lines 787-789 of the script large-scale-evaluation-freebase.py located in /automatic-evaluation/. I think that my version of jellyfish is too recent as compared to yours.
If you still have it, could you please add it to the repo ?

By the way I fixed this script for windows users (there is a difference in the behavior of the multiprocessing package between windows and linux, look at http://rhodesmill.org/brandon/2010/python-multiprocessing-linux-windows/ if you are interrested). Tell me if you would like to see what I have changed.

Thank you for your work, I really appreciate it !

The process of feature extraction can be parallelized by first loading all the files into a Queue and then launch separate processes/threads do generate Tuples from each of sentence.

Hi, I'd like to check on the variable self.config.min_pattern_support.

In line 293 (breds_parallel.py), the check for len(p.tuples) > self.config.min_pattern_support occurs under the if condition of len(self.patterns) == 0

However, under the else case, there is no such checks to remove patterns with tuples less than the min support. Is this deliberate by design? Otherwise, should the check also be done for len(self.patterns) != 0 too?

Thank you.

Hi there,

Can you please include the exact commands with parameters (including running pre_processing_KB scripts: easy_freebase_clean.py, select_sentences.py and index_whoosh.py) to replicate evaluation numbers of BREDS reported in paper?
Thanks for adding some notes for the ReadMe for automatic-evaluation, but it would really help us if you can also include the exact commands.
[I would really appreciate a timely reply]

Thank you!

Hi,

Thank you for sharing your (clean) code. In the parallel version of BREDS when computing the 'tuples confidence' why you are taking a weighted combination of confidence score from last & current iteration. I understand self.config.wUpdt is a hyperparameter but still why is it needed only in the parallel version ?

Hi,

Sorry for bothering again regarding the automatic-evaluation but this is important for me to know.
In large-scale-evaluation-freebase.py while loading freebase the whole KB is loaded irrespective of the relation type, now consider string_matching_parallel() it is only checked if e1 and e2 occur in the KB but not for which relation? which means if my model predicts a Founder-Of relation between Marissa Mayer and Google than it will be correct becasuse of this line:
if len(database_1[(r.ent1.decode("utf8"), r.ent2.decode("utf8"))]) > 0:
But KB actually contained Marissa Mayer Employment history Google.
Is my understanding correct that it is only checked if two entites occur in a database but not for which relation? (please indicate yes/no for the Marissa Mayer Founder-Of Google example).

Many Thanks!

Thank you for your appliment, my google account is invalid , could you release another way？Thanks

Hello:

When I want to extract a relation between two entities 'ORG' and 'Date' the system doesn't extract any relations. for example
<ORG>NOKIA</ORG> founded in <Date>1865</Date>
However, when I change the 'Date' tag to 'LOC' or 'PER' it works perfectly for example
<ORG>NOKIA</ORG> founded in <LOC>1865</LOC>

NOTE: I changed the negative and positive file ( e1 to ORG and e2 to Date) but still the system can not extract any relation between ORG and Date

Hi, I'm really sorry to bother you with some simple questions. How do I place the downloaded Word2VEc model?

FileNotFoundError: [Errno 2] No such file or directory: 'afp_apw_xin_embeddings.bin'

In tuple.py, lines 72, 78 and 79 passes bet_filtered, bef_no_tags and and aft_no_tags to function pattern2vector_sum. Note that these variables passed are only the literal words (without POS tags).

However, at line 87, the word vector retrieval appears to be retrieving the zero-th position character of the token passed into this function instead of the expected vector retrieval of the entire word.

i.e.
line 87 :: vector = config.word2vec[t[0].strip()]

Changing line 87 to
vector = config.word2vec[t.strip()]

works for me.