Structured Output Learning with Abstention: Application to Accurate Opinion Prediction

Authors: Alexandre Garcia, Chloé Clavel, Slim Essid, Florence d’Alche-Buc

ICML 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Instantiated on a hierarchical abstention-aware loss, SOLA is shown to be relevant for fine-grained opinion mining and gives state-of-the-art results on this task. Moreover, the abstention-aware representations can be used to competitively predict user-review ratings based on a sentence-level opinion predictor. Section 5 presents the numerical experiments and Section 6 draws a conclusion.
Researcher Affiliation Academia 1LTCI, Telecom Paris Tech, Paris, France.
Pseudocode No The paper describes algorithms and steps but does not include a formal pseudocode block or an algorithm section labeled as such.
Open Source Code No The paper does not provide any explicit statement about making its source code available, nor does it provide a link to a code repository.
Open Datasets Yes We test our model on the problem of aspect-based opinion mining on a subset of the Trip Advisor dataset released in (Marcheggiani et al., 2014).
Dataset Splits No The paper mentions 'predefined train and test sets' for the Trip Advisor dataset, but it does not specify a validation set split or methodology for creating one. It also does not provide specific percentages or counts for these splits.
Hardware Specification No The paper does not explicitly describe the hardware (e.g., specific CPU, GPU, or memory details) used to run its experiments.
Software Dependencies No The paper mentions using 'Infer Sent representation' and states that it built a 'vector-valued regressor', but it does not specify any software versions for libraries, frameworks (like PyTorch, TensorFlow), or specific solvers. It also mentions 'ridge regression' but without specific software versions.
Experiment Setup Yes In all our experiments, we rely on the expression of the Haloss presented in 4. The linear programming formulation of the pre-image problem used in the branch-and-bound solver is derived in the supplementary material and involves a decomposition similar to the one described in Section 2 for the H-loss. Implementing the Ha-loss requires choosing the weights ci, c Ai and c Aci. We first fix the ci weights in the following way : ci = cp(i) |siblings(i)| i {1, . . . , d}. Here, 0 is assumed to be the index of the root node. As far as the abstention weights c Ai and c Aci are concerned, making an exhaustive analysis of all the possible choices is impossible due to the number of parameters involved. Therefore, our experiments focus on weighting schemes built in the following way: c Ai = KAci c Aci = KAcci The effect of the choices of KA and KAc will be illustrated below on the opinion prediction task.