Proceeding of the earlier proposed MediExtract distil-tuning framework, which has been reforged to support decoder-based models (LLMs) with role-based prompt formation.
The key contribution of the novel framework is its scalable potentials to other modalities.
RuOpinionNE-2024 proceeds the past year RuSentNE-2023
that go further with:
โ๏ธ annotation of other sources of opinion causes: entities, out-of-context object (None), and
๐ evaluation of factual statements that support the extracted sentiment.
In this talk we cover LLM evaluation concepts through the detailed overview of manual reporting directions in liver cancer imaging.
We cover radiological imaging aspects that doctors tend to mention โ๏ธ and see how these aspects could be retrieved by LLM ๐ค by means of NLP and IR techniques.
We provide recommendations on model scale choice and findings out of result analysis. This contributes to the similar studies presented at CASA-2025.
This milestone grants me skills for evaluating Autonomous Intelligent Systems (AIS).
Generative AI and Large Language Models as the most recent advances in AI inevitably takes part in products of various fields:
transportation, manufacturing, retail / customer service, healthcare, finance, education (AI learning platforms).
The completed training makes me a right person for ensuring quality of the AIS-powered products and services, especially with the focus on personal expertise in:
AI, NLP, and Information Retrieval.
MediExtract Distillation Framework (MEDF) is a
a hybrid teacher-student distillation process that leverages
the power of LLMs in information capturing to enhance the performance
of a smaller student model.
Model training involves two feedforward branches per iteration:
one using ground truth as labels and another using generated structured medical key information as an auxiliary supervision.
A tiny Python no-string package for performing translation of a massive stream of texts with native support of pre-annotated fixed-spans that are invariant for translator.
Powered by AREkit-pipelines.
In this talk, by saying reading in between the lines, we refer to performing such and
Implicit IR that involves extraction of such information that is related to, i.e. โ ๏ธauthor /
๐ฉโโ๏ธ patient / ๐งโ๐ฆฐ character etc.
Presented an update on NLP advances in MMI-NLP project in end-to-end system for training novice practitioners.
We showcase the importance on NLP application for processing medical narratives of liver-related MRI/CT scan series,
such as one mentioned in "Series Descriptions" of the DICOM metadata.
The end-to-end concept advances were demonstrated as well.
A lightweight, no-strings-attached Chain-of-Thought framework for your LLM, ensuring reliable results for bulk input requests stored in CSV / JSONL / sqlite.
It allows applying series of prompts formed into schema
A no-strings framework for Named Entity Recognition (NER) in large textual collection
with third-party AI models via very accessible API.
Powered by AREkit-pipelines.
The talk is devoted to application of Large Language Models (LLMs) for retrieving implicit information from non-structured texts via reasoning the result sentiment label.
To enhance model reasoning capabilities ๐ง , we adopt Chain-of-Thought technique and explore its proper adaptation in Sentiment Analysis task.
This paper proposes the workflow of automatic profiling fictional character from literature novel books.
The workflow is aimed at character personalities construction by solely rely on their comments in book: dialogue utterances and
surrounding text, paragraphs.
We propose a loss function which combines: (1) structured contrastive loss and (2) Pearson loss.
For exploiting the related function in BERT optimization process, we propose "Adversarial Training with Fast Gradient Method (FGM)".
To improve model generalization ability we exploit "mix-up" as a data augmentation technique to mix inputs with the labels in specific range.
We adopt two-stage method SFT (๐ฅ) + inference (๐ฉ) as contrastive reasoning calibration (CRC).
For the training, we enrich input samples: standard, role-play, contrastive.
We concatenate the article and task content togerther in input to train model predict all track results (1,2,3).
We fine-tune Flan-T5-base (250M) with the reforged THoR-ISA framework for Emotion Cause Extraction.
Our final submission is 3'rd place by F1-proportional and 4-5'th by F1-strict
which counts the emotion cause span borders. Our THOR-ECAC is publicly available.
We propose a two-phase SFT training strategy: (i) data-processing and (ii) model training. In (i) we combine CoT + knowledge-distillation concept (using GPT-3.5) + CoT steps for CoT-NumHG; this resource is then adopted in (ii) for the full-param SFT
This is the Russian version of the talk devoted to application of Large Language Models (LLMs) for retrieving implicit information from non-structured texts via reasoning the result sentiment label.
To enhance model reasoning capabilities ๐ง , we adopt Chain-of-Thought technique and explore its proper adaptation in Sentiment Analysis task.
Presented a multi-disciplinary research to transform the UK/global healthcare sector and allied
training using digital technologies derived from the creative industries sector.
In particular, the personal findings on Multimodal LLM-based system development
and research to be conducted has been demonstrated in a form of the demo setups.
We explore LLMs reasoning capabilities in Targeted Sentiment Analysis task.
In particular we assess LLM models in two modes: (i) zero-shot-learning (ZSL) (ii) fine-tuning with prompt and CoT THoR
proposed at ACL-2023.
The fine-tuned Flan-T5-xl outperforms
the prior top submission at
RuSentNE-2023.
AREkit-based application for a granular view onto sentiments
between entities for large document collections, including books, mass-media, Twitter/X, and more.
See our online demo
Unlike to the similar talks in past,
in this one we overview the capabilities the most-recent instructive Large Language Models (LLM).
This list includes: Mistral, Mixtral-7B, Flan-T5, Microsoft-Phi2, LLama2, etc.
We also cover advances of LLM-fine tuning by exploiting Chain-of-Thought techniques
to get the most out of the LLM capabilities.
We propose an automated system that uses character dialogues from literary works.
We used the Chinese classic, Dream of the Red Chamber.
Our system efficiently extracts dialogues and personality traits from the book, creates a personality map for each character, generates responses that reflect these traits.
Participants are offered the task of extracting sentiments of three classes
(neg, pos, neutral) from news texts in relation to pre-marked entities such as
PERSON, ORG, PROFESSION, COUNTRY, NATIONALITY within a separate sentence.
A first LongT5-based model pre-trained on a large amount of unlabeled Vietnamese
texts and fine-tuned within the manually summarized texts from ViMS and
VMDS and VLSP2022 collections
In this talk we cover the advances of machine-learning approaches in sentiment
analysis of large mass-media documents. Complements the past talk at Wolfson College (Oxford)
with the detailed description of long-input transformers, RLHF training overview.
In this talk we examine the limitations of BERT model from the input size perspective.
To address the shortcommings, authors propose a related position embeddings in
order to implement local-window based sparse attention. To attend the distant tokens,
authors propose a Global-Attention mechanism in addition to the sparsed one for the main input.
Another main contribution is input structuring in attention mechanism.
In this talk we cover the advances of machine-learning approaches in sentiment
analysis of large mass-media documents. We provide both evolution of the task over time
including a survey of task-oriented models starting from the conventional linear
classification approaches to the applications findings of the recently announced
ChatGPT model.
About Tensorflow-based framework which lists attentive implementation of the conventional
neural network models (CNN, RNN-based), applicable for Relation Extraction classification
tasks as well as API for custom model implementation
For Biomedical domain. We investigate:
(1) which search interface elements searchers are look at when searching for docs to answer complex questions,
and
(2) relationship between individual differences and the interface elements which users are looked at.
This is a curated list of works devoted to sentiment attitude extraction domain.
The latter considered as an area of studies, in which sentiment analysis
and relation extraction are inextricably linked
This repository is an official results benchmark for automatic sentiment attitude
extraction task within RuSentRel-1.1
collection, for the following models:
conventional approaches (SVM, RandomForest, kNN, NB, Gradient Boosting),
CNN-based networks,
RNN-based networks,
Attentive models,
BERT-based language models.
Application of Distant Supervision in model training process results in a weight distribution
biasing: frames in between subject and object of attitude got more weight values;
the latter reflects the pattern of frame-based approach, utilized in RuAttitudes
collection development.
Pages: 169-179. The paper describes the task of automatic extraction of attitudes between
the subjects mentioned in the text, as well as their connection with the implicit
expression of the author's attitude to these subjects. A RuSentiFrames vocabulary is presented, in which the basic attitudes associated of
Russian predicate words are described.
Utilizing a set of sentence level attitudes with related
metrics to perform a sentiment prediction. As for
input of neural network based models, in prior works all the
models deals with an attitudes limited by a single sentence.
An application of CNN based architecture (adapted for relation
extraction) towards sentiment attitudes extraction task.
Presented at Third Workshop "Computational linguistics and
language science" (CCLS-2018), HSE, Moscow.
Provides details on lexicon development using twitter messages (related works
[pmi],
[dev],
[sota]).
Sentiment classification of user reviews using SVM.
Master degree paper.