Semeval (Semantic Evaluation)

Semeval is a Python package for modelling various aspects of meaning in languages.

Installation

The package can be installed via PIP: pip install -U semeval

Download models

The following command downloads word vector embeddings for English and Finnish

python3 -m semeval.download -l eng fin -m embeddings

We currently suppor eng, fin, kpv, myv, mdf, rus, liv and sms.

Relatedness models can be downloaded using the command:

python3 -m semeval.download -l eng fin -m relatedness

Currently, only English and Finnish are supported in the relatedness model.

Usage

Word embeddings

Load embeddings of a language by running:

from semeval import Embeddings
e = Embeddings("eng")

After this, you can find related words

e.theme(['shoe', 'clothes']) #outputs a word describing all input words
>> ('clothing', 0.8700831532478333)

e.neighbours('hi') #outputs the nearest neighbors
>> [('hey', 0.6976003050804138), ('jeez', 0.6230848431587219), ('ya', 0.6213312149047852), ('hello', 0.6144036650657654) ...]

e.analogy('man', 'king', 'woman') #analogous words
>> [('monarch', 0.6457855105400085), ('regnant', 0.6354122161865234)...]

e.most_similar(positive=['hello', 'world'], negative=['king'], topn=10)
>> [('Tamana', 1.1119599342346191), ('Ibaraki', 1.087041974067688), ('Makuhari', 1.0793628692626953),... ]

Get vector representations

e.centroid(['hi', 'hello']) #the centroid vector of the input words
>> [0. 0.10486 -0.23701501 0.1293 0.052645 -0.027155...]

e.to_vector('this is a great api !'.split(' ')) # a vector representing the sentence
>> [1.1630000e-02 -2.2303334e-02  8.4316663e-02 -2.1063333e-02...]

e.vector("king")
>> [0.05574 -0.16716 0.10282 -0.10851 0.08783 -0.09499 0.16031...]

Get similar words in another language

e2 = Embeddings("fin")
e.project('king', e2)
>>[('Ahasveros', 0.5551087856292725), ('kuningas', 0.5522325038909912), ('kuninkas', 0.5242758393287659)..]

Word similarity

e.similarity('hi', 'bye')
>> 0.28805155

Vocabulary

e.vocabulary()
>> {'astrologically': <Vocab object>, 'spinto': <Vocab object>, 'NortelNet': <Vocab object>...}

Server-mode

Server-mode is optimal for some cases such debugging or not wanting to wait for multiple models to load. To start word embeddings server, run the below command in the terminal: python -m semeval.server --service embeddings

Once the server is loaded, the service is accessible through EmbeddingsAPI class. Note that the language/s must be passed every call, otherwise the server cannot know which model to use. Here is an example of accessing the service from Python.

from semeval import EmbeddingsAPI

api = EmbeddingsAPI()
api.theme(words=['shoe', 'clothes'], lang='eng')
api.neighbours(word='hi', threshold=0.4, lang='eng')
api.analogy('man', 'king', 'woman', topn=10, lang='fin')
api.centroid(words=['hi', 'hello'], lang='eng')
api.to_vector(tokens='this is a great api !'.split(' '), lang='eng')
api.align(word='king', lang1='eng', lang2='fin')
api.similarity(w1='hi', w2='bye', lang='eng')
api.most_similar(positive=['hello', 'world'], negative=['king'], topn=10, lang='eng')
api.vector(word='king', lang='eng')
api.vocabulary(lang='eng')

Relatedness

Load the relatedness model:

from semeval import Relatedness
m = Relatedness(lang='eng')

Get most 5 related word to the word car:

m.get_sorted_rel('car')[:5]
>> [('park', 0.20570222), ('insurance', 0.085593514), ('parking', 0.06783264), ('hire', 0.036158927), ('cheap', 0.028782822)]

Get relatedness score between two words:

m.get_rel('car')['insurance']
>> 0.085593514

Top 10 interpretations for the metaphor "Alcohol is a Crutch". NOTE: unlike the original paper, there is no filtering (e.g., POS filtering) applied in this function. Read the paper for further details and post-processing steps to improve the results.

m.interpret('alcohol', 'crutch')[:10]
>> [('use', 0), ('cane', 0), ('smoking', 1), ('psychological', 2), ('dependence', 2), ('emotional', 3), ('cigarette', 3), ('drug', 4), ('bandage', 4), ('week', 5)]

Metaphoricity scores for the tenor computer, vehicle creative and expression 'The algorithm for painting.'

m.metaphoricity('computer', 'creative', ['the', 'algorithm', 'for', 'painting'], 300) # 0 to select all
>> (3.2977940826411467e-07, 0.000510800164192915, 0.00025556497180058955) # (magnitude score, difference score, avg if both positive)

Business solutions

When your NLP needs grow out of what UralicNLP can provide, we have your back! Rootroo offers consulting related to a variety of NLP tasks. We have a strong academic background in the state-of-the-art AI solutions for every NLP need. Just contact us, we won't bite.