Zlecenia dla freelancerów

Zlecenia dla freelancerów

Topic modeling

Python — podano nieprawidłowe kategorie?

205 PLN

Use data from https://github.com/thedenaas/hse_seminars/tree/master/2018/seminar_13/data.zip
Implement model in pytorch from "An Unsupervised Neural Attention Model for Aspect Extraction, He et al, 2017", also desribed in https://github.com/thedenaas/hse_seminars/blob/master/2019/seminar_13/topic.ipynb .

You can use sentence embeddings with attention :
$z_s = \sum_{i}^n \alpha_i e_{w_i}, z_s \in R^d$ sentence embedding $\alpha_i = softmax(d_i)$ attention weight for i-th token $d_i = e_{w_i}^T M y_s$ attention with trainable matrix $M \in R^{dxd}$ $y_s = \frac 1 n \sum_{i=1}^n e_{w_i}, y_s \in R^d$ sentence context $e_{w_i} \in R^d$ , token embedding of size d - number of tokens in a sentence

Or just use sentence embedding as an average over word embeddings :
$z_s = \frac 1 n \sum_{i=1}^n e_{w_i}, z_s \in R^d$ sentence embedding $e_{w_i} \in R^d$ , token embedding of size d - number of tokens in a sentence

$p_t = softmax(W z_s + b), p_t \in R^K$ topic weights for sentence , with trainable matrix $W \in R^{dxK}$ and bias vector $b \in R^K$ $r_s = T^T p_t, r_s \in R^d$ reconstructed sentence embedding as a weighted sum of topic embeddings $T \in R^{Kxd}$ trainable matrix of topic embeddings, K=number of topics

Training objective: $J = \sum_{s \in D} \sum_{i=1}^n max(0, 1-r_s^T z_s + r_s^T n_i) + \lambda ||T^T T - I ||^2_F$ where random sentences are sampled as negative examples from dataset for each sentence $n_i = \frac 1 n \sum_{i=j}^n e_{w_j}$ average of word embeddings in the i-th sentence regularizer, that enforces matrix to be orthogonal $||A||^2_F = \sum_{i=1}^N\sum_{j=1}^M a_{ij}^2, A \in R^{NxM}$ Frobenius norm

Compute topic coherence for at least for 3 different number of topics. Use 10 nearest words for each topic. It means you have to train one model for each number of topics. You can use code from seminar notes with word2vec similarity scores.

Oferty 2 Odrzucone 1

Sergey Galitovsky

0 0

Zlecenia -
Ocena -
Ranking 60

Budżet: 5500 RUB Termin: 2 dni

Могу выполнить Ваше задание. Есть опыт работы с нейронными сетями и имплементации их с научных статей на pytorch/tensorflow.

О сроках и цене можем договориться.

Vladislav Bandurin

8 0

Zlecenia 8
Ocena -
Ranking 356

Budżet: 8000 UAH Termin: 7 dni

Добрый день!
У меня имеется большой опыт в DataScience. Есть экспертиза в NLP и разработке моделей на pytorch. Пишите, обсудим детали:)

Fdmklhi Kjlkn
Moskwa, Rosja

Zleceń 4
Ocena -
Ranking 112

Topic modeling

Sergey Galitovsky

Vladislav Bandurin

Aktualnie brak ofert

Oferty ukryte

Aktualne zlecenia dla freelancerów w kategorii Python

Opracowanie automatyzacji dla Google Sheets + cotygodniowe raportowanie w Telegramie

Program do automatycznego montażu wideo (lokalny, do użytku osobistego)

Real-Time Trener

Skrypt Pythona do automatycznego wypełniania Google Sheets danymi z wewnętrznej usługi (API / Pla

Opracowanie mikrousługi dynamicznego zarządzania sekretami i rotacji kluczy API