S. R. K. Branavan’s research while affiliated with Massachusetts Institute of Technology and other places

... For the non-NLP (multi-modal) tasks, most focused on environment-grounded language learning, i.e., driving the agent to associate natural language instructions with the environments and make corresponding reactions, such as selecting mentioned objects from an image/video (Matuszek et al. 2012;Krishnamurthy and Kollar 2013;Puig et al. 2018), following navigational instructions to move the agent (Tellex et al. 2011;Kim and Mooney 2012;Chen 2012;Artzi and Zettlemoyer 2013;Bisk, Yuret, and Marcu 2016), plotting corresponding traces on a map (Vogel and Jurafsky 2010; Chen and Mooney 2011), playing soccer/card games based on given rules (Kuhlmann et al. 2004;Eisenstein et al. 2009;Branavan, Silver, and Barzilay 2011;Babeş-Vroman et al. 2012;Goldwasser and Roth 2011), generating real-time sports broadcast (Chen and Mooney 2008;Liang, Jordan, and Klein 2009), controlling software (Branavan, Zettlemoyer, and Barzilay 2010), and querying external databases (Clarke et al. 2010), among others. Meanwhile, instructions are also widely adapted to help communicate with the system in solving NLP tasks, for example, following instructions to manipulate strings (Gaddy and Klein 2019), classifying e-mails based on the given explanations Mitchell 2017, 2018), and text-to-code generation (Acquaviva et al. 2022). ...

Reference:

Large Language Model Instruction Following: A Survey of Progresses and Challenges

... Understanding action preconditions and effects in text is a crucial yet challenging task. Branavan et al. (2012) pioneer work in this area using reinforcement learning to extract high-level planning knowledge from text with the guidance of action preconditions and effects. Dalvi et al. (2018) develop a dataset and models for paragraph comprehension, and highlight the importance of tracking state changes in procedural text. ...

Reference:

Making Large Language Models into World Models with Precondition and Effect Knowledge

... Another straightforward approach to using natural language for reward shaping can be seen in [26], which describes rules using natural language to decide whether to choose an action. [27][28] also perform language-to-reward mapping by designing rewards based on if the agent arrives at the right location or not by following the instruction provided. These approaches still face the problem of efficiently defining reward functions since it requires an expert programmer who can make decisions about how language instructions are mapped to the environment. ...

Reference:

Temporal Video-Language Alignment Network for Reward Shaping in Reinforcement Learning

... A few sentiment analysis systems take as input the sample speech of the user and perform prosodic processing of the speech signals. [16] takes the speech signal as input, converts it into text and then, only considers the text content for the affect analysis. The feature extraction is performed on the text review and adjective-noun pairs are identified as potential sentiment phrases. ...

Reference:

Issues and Dependencies on Affect Data Models For Assessment of Disorders-A Survey

... One of the most well studied class of models on the space of rankings/permutations are the celebrated Mallows models, first introduced in [24]. Since then, Mallows models (and their variants) have received significant attention in statistics ( [9,10,13,14,15,25,28]), probability ( [3,6,12,17,19,29,30]), and machine learning ( [2,8,21,22,23,26,27]). In [28], the authors introduce a class of exponential family models on the space of permutations, which includes some of the commonly studied Mallows models. ...

Reference:

Inference on a class of exponential families on permutations

... Before the prevalence of Large Language Model, traditional autonomous agents primarily implement through reinforcement learning (Branavan et al., 2009;Shvo et al., 2021;Gur et al., 2022), semantic parsing (Li et al., 2020) and imitation learning (Humphreys et al., 2022) that clones human's keyboard and mouse actions . The recent trend is to use Large Language Model to generate GUI instructions and actions. ...

Reference:

AutoEval: A Practical Framework for Autonomous Evaluation of Mobile Agents

... Fine-tuning Virtual Agents traditionally relies on humanannotated datasets, which are labor-intensive and timeconsuming . Methods such as imitation learning (Humphreys et al., 2022) and reinforcement learning (Branavan et al., 2009;2010) have been employed to fine-tune agents based on curated expert trajectories or outcome rewards, but these approaches often suffer from compounding errors and limited exploration (Christianos et al., 2023;Xi et al., 2024;Song et al., 2024). Recent advancements, such as reject sampling fine-tuning (RFT) (Yuan et al., 2023) and direct policy optimization (DPO) (Rafailov et al., 2023), have sought to reduce reliance on human annotations by leveraging both successful and failure trajectories (Lai et al., 2024;Zhang et al., 2024c). ...

Reference:

Boosting Virtual Agent Learning and Reasoning: A Step-wise, Multi-dimensional, and Generalist Reward Model with Benchmark

... The originality of the work stems from using the n-gram concept to model nodes in a tree in addition to modeling sequential entities. Branavan, Deshpande and Barzilay (2007) propose a method that automatically generates a table-of-contents structure for long documents such as books. They first segment a document hierarchically and then generate an informative title for each segment. ...

Reference:

A tree-based learning approach for document structure analysis and its application to web search

... Mais ces modèles ne sont pas adéquats dans le cas où le domaine etudié possède une multitude d'aspects dont certains apparaissent trés rarement dans le corpus. Ainsi, certains auteurs [6,10,12,18] ont plutôt adopté des méthodes basées sur la modélisation des sujets. ...

Reference:

Extraction of lexico-grammatic features and coupling of CRF (Conditional Random Field) units to the deep neural network for aspect extraction

... Text semantic segmentation necessitates that sentences within the same paragraph cohesively revolve around a central topic, while maintaining minimal semantic overlap between distinct paragraphs. Early unsupervised methodologies identified segmentation boundaries through Bayesian models (Chen et al., 2009;Riedl and Biemann, 2012) or graph-based methods, where sentences were treated as nodes (Glavaš et al., 2016). On the other hand, supervised methods have leveraged pretrained language models (PLMs) derived from extensive corpora, subsequently fine-tuning them on annotated text semantic segmentation datasets. ...

Reference:

CrossFormer: Cross-Segment Semantic Fusion for Document Segmentation