- https://www.google.com/search?q=Same+overcast+sky%2C+wet+roads+from+overnight+..
- https://www.google.com/search?q=which+is+more+appropriate+for+academic+writing%3A+%C2%A7+or+%22Section%22
✦ AI Overview
In general academic and scientific writing, spelling out the word "Section" (or "section") is much more appropriate and readable. The section symbol (§) is heavily reserved for legal writing and highly abbreviated in-text citations.
- https://chatgpt.com/share/6a239139-aec4-83ab-9897-4a2033a7a84e (This a figure at the start of "Section 3. Method". Is it mainly a TOC of the METHOD or pseudocode of the overall METHOD or both? What might you call it?)
- https://www.google.com/search?q=Is+it+more+common+to+use+figure%2Bnumber+to+refer+to+a+visual+part+with+or+without+use+the+word+%22above%22+or+%22below%22%3F
✦ AI Overview
It is much more common and preferred in academic and technical writing to refer to a visual part using only the figure number without the words "above" or "below".
Major style guides—such as APA, Chicago, and IEEE—explicitly instruct writers not to use "the figure above" or "the figure below". (SFU Library: www.lib.sfu.ca/help/cite-write/citation-style-guides/apa/tables-figures)
- Ask Google: When do an organization paragraph in writing the method section with subsection 3.1, 3.2, 3,3 and subsubsection 3.2.1, 3.2.2 3.2.3, one should cover the subsections only and ignore the subsubsection. Is that right?
✦ AI Overview
When an organization paragraph outlines the Method section, it summarizes the main subsections only (e.g., 3.1, 3.2, 3.3). You should ignore the subsubsections (e.g., 3.2.1, 3.2.2) to maintain a macro-level view of the section's flow without getting bogged down in microscopic details.Why cover only the main subsections?An organization paragraph is designed to act as a "road map" for the reader, providing a high-level preview of how the overarching topic is tackled. Because subsubsections merely break down a subsection into granular, isolated points, including them makes the overview clunky, overly dense, and difficult to follow. [1]How it looks in practiceYour organization paragraph (usually placed at the end of your introductory text) should flow like this:
- "This section is organized into three main subsections. Section 3.1 details the participant demographics, Section 3.2 describes the experimental materials, and Section 3.3 outlines the step-by-step data collection and analysis procedures."
Fig. 2. Outline of the process used to train ECP
3. Method
Submitting base words(e.g., “earn”) to collocation tools often does not provide enough information for language learning. Existing collocation tools typically organize collocations according to frequency counts or grammatical relations. Unfortunately, the collocations with the best pedagogical value may require information beyond syntactic and statistical patterns, such as proficiency levels and semantic tags. Such useful collocations may then be difficult for learners to use appropriately. To support collocation learning, a promising approach is to automatically transform corpus-derived collocations into structured collocation entries that contain CEFR proficiency levels and semantic tags expected to be useful for learners.
3.1 Problem Statement
We focus on the collocation extraction and annotation process: extracting a set of collocations likely to be useful for a given base word and transforming them into structured entries. These collocations are then returned as the output of the system. The returned collocations can be examined by a language learner directly, or passed on to educational applications such as reading comprehension test generation or essay scoring systems. Thus, it is crucial that the collocations of interest be present in this initially retrieved set. At the same time, the set of retrieved collocations cannot be so large that it overwhelms the learner or the subsequent educational components. Therefore, our goal is to return a reasonable-sized set of collocations that, at the same time, contains proficiency levels and semantic tags for the given base word. We now formally state the problem that we are addressing.
Problem Statement: We are given a web-scale corpus C (e.g., web1t5), an English database P of words with proficiency levels (e.g., English Vocabulary Profile and CEFR), a collocations dictionary D (e.g., Macmillan Collocations Dictionary), a pre-trained text-to-text model M (e.g., T5), and a base word w. Our goal is to generate a set of collocations W = {(w, coll1, lev1, tag1), ..., (w, colln, levn, tagn),}, where (w, colli) is a collocation extracted from C, levi is its CEFR level, and tagi is its semantic tag. For this, we extract collocations from C and create a collection T of triples, <collocation, level, semantic tag> from P and D, in order to fine-tune M into M’ such that M’ maximizes the probability of T. At run-time, the system starts with the given base word w and retrieves collocations of w with proficiency levels and semantic tags.
In the rest of this section, we describe our solution to this problem. First, we define a strategy for extracting candidate collocations from a web-scale corpus (Section 3.2.1). The annotation strategy relies on a collection of collocation, proficiency level, and semantic tag triples for training, derived from lexicographic sources (Section 3.2.2). In this section, we also describe our method for fine-tuning the pre-trained text-to-text model using the generated training triples (Section 3.2.3). Finally, we show how ECP processes a base word at run-time by applying the fine-tuned model to retrieve collocations with proficiency levels and semantic tags (Section 3.3).
3.2 Learning Structured Collocation Entries
We attempt to transform corpus-derived collocations into structured collocation entries that contain information expected to be useful for language learning. Our training process is shown in Figure 2.
3.2.1 Extracting Collocations. In the first stage of the learning process (Step (1) in Figure 2), we generate a set of collocations that can later be transformed into structured collocation entries. For example, the base word “earn” may have collocations such as “earn income,” “earn respect,” and “earn degree.” The goal of this stage is to identify word combinations that are likely to represent fluent and natural word combinations which is a hallmark of native speakers.
The input to this stage is a web-scale corpus and corpus-derived statistics associated with word combinations in the corpus. Although syntactic parsers can be used to extract collocations from running text, we use a simpler corpus-statistical procedure based on n-gram and dependency statistics. This keeps the extraction process scalable and transparent, while still allowing the system to capture both adjacent and gappy collocations. Figure 3 shows the procedure for extracting candidate collocations for a base word.
In Step (1), we collect adjacent n-gram candidates containing the base word. These candidates capture linear collocations whose component words occur next to each other in the corpus. In Step (2), we collect dependency-based candidates containing the base word. These candidates capture syntactic or gappy collocations whose component words may not be adjacent but are connected by a grammatical relation.
Fig. 3. Extract Collocations from the Web-scale Corpus.
In Step (3), we merge the n-gram and dependency-based candidates into a single candidate set. For each candidate in this set, we record its component words, grammatical relation when available, frequency information, and association scores (Step (4)).
This method for extracting candidate collocations can produce many candidates, potentially including a significant number of word combinations that are noisy without strong association to be widely used. Therefore, in Step (5), we apply MCD-calibrated dual-track filtering: one track retains candidates with strong bigram association, and the other retains candidates with strong dependency association. In Step (6), we apply structural and part-of-speech validation, focusing on collocation types that are well represented in collocation dictionaries, including verb-noun, adjective-noun, noun-noun, and adverbial collocations. The exact thresholds and filtering parameters are reported in the experimental setting (Section 4).
The output of this stage is a filtered set of corpus-derived collocations that can be used in the following annotation stages.
3.2.2 Constructing Training Triples. In the second stage of the learning process (Step (2) in Figure 2), we construct training triples that can be used for fine-tuning the text-to-text model. Each training triple consists of a collocation, a CEFR proficiency level, and a semantic tag. The procedure for constructing these triples consists of a number of steps, namely, assigning CEFR levels from proficiency-labeled lexical resources, deriving semantic tags from a collocations dictionary, and filtering the resulting triples for consistency and reliability. Each step is described below in detail.
Fig. 4. Construct Training Triples from lexicographic sources.
The input to this stage is the filtered set of candidate collocations from the
previous stage, a proficiency database P, and a collocations dictionary D. The
proficiency database provides CEFR-labeled lexical, sense-level, and example
evidence, while the collocations dictionary provides collocation groups,
definitions, examples, and semantic evidence. The output of this stage is a
collection T of <collocation, level, semantic tag> triples. Figure 4 shows the procedure for constructing training triples.
In Step (1), we consider each candidate collocation generated in the previous stage. In Steps (2) and (3), we assign a CEFR proficiency level by referring to the proficiency database. When a collocation is directly attested in an EVP example, the CEFR level of the corresponding EVP sense is used as direct evidence. If a complete collocation is not directly attested in an EVP example, the system assigns its level by aligning the collocation with CEFR-labeled sense evidence from the component words.
In Steps (4) and (5), we assign a semantic tag using evidence from the collocations dictionary. In the current semantic-tagging stage, we focus on verb-noun collocations, where dictionary definitions and collocation groups provide relatively clear evidence for representative collocation labels. To derive these tags, the system uses dictionary-based evidence such as local collocation groups, definitions, and selects representative labels using weighted TF-IDF evidence.
In Step (6), the assigned level and semantic tag are combined with the collocation to form a structured training triple. In Step (7), triples are retained only when the required annotations are supported by the corresponding lexicographic resources. This filtering step is necessary because the extraction stage may produce collocations for which proficiency or semantic evidence is incomplete or unreliable. The exact filtering criteria and resource coverage are reported in the experimental setting.
The result of this stage is a collection of training triples T, automatically
constructed from corpus-derived collocations and lexicographic evidence. These
triples are used in the next stage to fine-tune the pre-trained text-to-text model.
3.2.3 Fine-tuning the Text-to-text Model. In the third and final stage of the learning process (Step (3) in Figure 2), we fine-tune the pre-trained text-to-text model using the constructed training triples. The goal of this stage is to learn a mapping from a collocation input to a structured output that contains its CEFR level and semantic tag.
The input to this stage is the training triple collection T constructed in the previous stage and a pre-trained text-to-text model M. Each triple t in T is converted into an input-output training example. The model input is the collocation, and the model output is a structured sequence containing the collocation, its CEFR level, and its semantic tag. The output of this stage is a fine-tuned model M' that can generate structured collocation entries. Figure 5 shows the procedure for fine-tuning the text-to-text model.
In Step (1), we consider each training triple constructed from corpus-derived collocations and lexicographic evidence. In Steps (2), (3), and (4), we convert the triple into the text-to-text format required by the model. For example, a collocation such as “earn admiration” can be used as the input sequence, while the corresponding structured entry, including its CEFR level and semantic tag, is used as the target output sequence.
In Step (5), we fine-tune the pre-trained model M using the constructed input-output examples. The fine-tuning objective is to maximize the probability of generating the target structured entry given the input collocation. More formally, for each training example, the model learns to estimate:
P(<lev(c), tag(c)> | c)
where c is the input collocation, lev(c) is the CEFR level assigned from the proficiency resource, and tag(c) is the semantic tag derived from the collocations dictionary.
In Step (6), we select the fine-tuned checkpoint based on validation performance. The detailed training parameters, validation setting, and model selection criteria are reported in the experimental setting(Section 4). The result of this stage is M', a fine-tuned text-to-text model that can be applied offline to produce CEFR-labeled and semantically tagged collocation entries for the ECP database. At run-time, the system retrieves these precomputed entries for a given base word.
BEFORE
Submitting base words(e.g., “earn”) to collocation tools often does not provide enough information for language learning. Existing collocation tools typically organize collocations according to frequency counts or grammatical relations. Unfortunately, the collocations with the best pedagogical value may require information beyond syntactic and statistical patterns, such as proficiency levels and semantic tags. Such useful collocations may then be difficult for learners to use appropriately. To support collocation learning, a promising approach is to automatically transform corpus-derived collocations into structured collocation entries that contain CEFR proficiency levels and semantic tags expected to be useful for learners.
3.1 Problem Statement
We focus on the collocation extraction and annotation process: extracting a set of collocations likely to be useful for a given base word and transforming them into structured entries. These collocations are then returned as the output of the system. The returned collocations can be examined by a language learner directly, or passed on to educational applications such as reading comprehension test generation or essay scoring systems. Thus, it is crucial that the collocations of interest be present in this initially retrieved set. At the same time, the set of retrieved collocations cannot be so large that it overwhelms the learner or the subsequent educational components. Therefore, our goal is to return a reasonable-sized set of collocations that, at the same time, contains proficiency levels and semantic tags for the given base word. We now formally state the problem that we are addressing.
Problem Statement: We are given a web-scale corpus C (e.g., web1t5), an English database P of words with proficiency levels (e.g., English Vocabulary Profile and CEFR), a collocations dictionary D (e.g., Macmillan Collocations Dictionary), a pre-trained text-to-text model M (e.g., T5), and a base word w. Our goal is to generate a set of collocations W = {(w, coll1, lev1, tag1), ..., (w, colln, levn, tagn),}, where (w, colli) is a collocation extracted from C, levi is its CEFR level, and tagi is its semantic tag. For this, we extract collocations from C and create a collection T of triples, <collocation, level, semantic tag> from P and D, in order to fine-tune M into M’ such that M’ maximizes the probability of T. At run-time, the system starts with the given base word w and retrieves collocations of w with proficiency levels and semantic tags.
In the rest of this section, we describe our solution to this problem. First, we define a strategy for extracting candidate collocations from a web-scale corpus (Section 3.2.1). The annotation strategy relies on a collection of collocation, proficiency level, and semantic tag triples for training, derived from lexicographic sources (Section 3.2.2). In this section, we also describe our method for fine-tuning the pre-trained text-to-text model using the generated training triples (Section 3.2.3). Finally, we show how ECP processes a base word at run-time by applying the fine-tuned model to retrieve collocations with proficiency levels and semantic tags (Section 3.3).
3.2 Constructing CEFR-Labeled and Semantically Tagged Collocation Entries
We attempt to transform corpus-derived collocations into structured collocation entries that contain information expected to be useful for language learning. Our training process is shown in Figure 2.
Use the form of
In Step (1), we …
We … (in Step (2))
For A in B ({A1, A2, })
with or without pseudo code (academic writing recipe)
3.2.1 Extracting Collocations. In the first stage of the learning process (Step (1) in Figure 2), we generate a set of collocations that can later be transformed into structured collocation entries. For example, the base word “earn” may have collocations such as “earn income,” “earn respect,” and “earn degree.” The goal of this stage is to identify word combinations that are likely to represent fluent and natural word combinations which is a hallmark of native speakers.
The input to this stage is a web-scale corpus and corpus-derived statistics associated with word combinations in the corpus. We use this information to collect potential word combinations associated with each base word.
The output of this stage is a filtered set of corpus-derived candidate collocations that can be used in the following annotation stages.
This method for extracting candidate collocations can produce many candidates, potentially including a significant number of word combinations that are noisy without strong association to be widely used. Because the following stages of the process are intended to construct proficiency-labeled and semantically tagged collocation entries, we apply a filtering procedure to reduce noise and retain candidates that are more likely to be useful for language learning. Specifically, we use MCD-calibrated dual-track filtering: a bigram MI track for candidates with strong linear association, and a dependency PMI track for candidates with strong syntactic association. We also apply structural and part-of-speech validation, focusing on collocation types that are well represented in collocation dictionaries, including verb-noun, adjective-noun, noun-noun, and adverbial collocations.(預告會在實驗的時候說明數據，雖然parser也可以做，但我決定用ngram/gappy bigram簡單的方法，用三步驟說明In step 1/psuedocode)
3.2.2 Constructing Training Triples. In the second stage of the learning process (Step (2) in Figure 2), we construct training triples that can be used for fine-tuning the text-to-text model. Each training triple consists of a collocation, a CEFR proficiency level, and a semantic tag. The procedure for constructing these triples consists of a number of steps, namely, assigning proficiency levels from proficiency-labeled lexical resources, deriving semantic tags from a collocations dictionary, and filtering the resulting triples for consistency and reliability. Each step is described below in detail.
For this stage of the process, we use the collocations extracted in the previous stage, together with two types of lexicographic resources. The first resource is a proficiency database, such as the English Vocabulary Profile, which provides CEFR-labeled lexical and sense-level information. The second resource is a collocations dictionary, such as the Macmillan Collocations Dictionary, which provides collocation evidence, semantic groupings, definitions, and examples. These resources allow us to construct resource-grounded annotations for corpus-derived collocations.(input標出來)
For each candidate collocation, we first assign a CEFR proficiency level by aligning the component words of the collocation with proficiency-labeled lexical entries. Since the difficulty of a collocation may depend on the meanings of both component words, the resulting collocation level is determined from the aligned proficiency information rather than from the surface word forms alone. In our implementation, the final level of a collocation is assigned based on the more advanced of the aligned component levels.（不要去提SBERT，對照EVP裡面有等級的collocation，越簡單越好，實驗再詳實報告，一句話帶過）
We then assign a semantic tag using evidence from the collocations dictionary. In the current semantic-tagging stage, we focus on verb-noun collocations, where dictionary evidence provides relatively clear semantic organization. To derive these tags, we use dictionary-based evidence such as local collocation groups, definitions, and examples, and select representative tags using weighted TF-IDF evidence.
This method for constructing training triples can produce annotations that vary in reliability, potentially including triples whose level or semantic tag is weakly supported by the available resources. Because the following stage fine-tunes a text-to-text model using these triples, we apply filtering to reduce noisy supervision. We retain triples when the collocation is supported by the candidate inventory and the assigned level and semantic tag are grounded in the corresponding lexicographic resources. The output of this stage is a collection T of <collocation, level, semantic tag> triples that will be used to fine-tune the pre-trained text-to-text model in the next stage.
Outline of the Process:
https://chatgpt.com/share/6a2246d6-9be0-83aa-8adc-cbf79095a571
Method:
https://chatgpt.com/share/6a2296de-5354-83a5-8a14-950830a1ddf6
Teacher’ version:
https://chatgpt.com/share/6a22a792-3878-83ab-a248-dd28f5bbdcc2
留言
張貼留言